# Gigabit (Gigabit Agency)

> AI-native software factory. We own the outcome. Forward-deployed senior engineers and AI agents that build, ship, and operate the production systems your business runs on — brand to backend — at published, transparent prices. Model: Audit → Build → Operate. Offices: Wyoming (US anchor) and Dhaka (delivery). Production deployments since 2016; 13 published case studies.

Key facts:
- AI Citation Audit: free — see whether AI assistants recommend you today, across major assistants for your category's buying queries. The front door to Ascent.
- Ascent: a 12-month growth program — brand, website, content, SEO, and GEO by one team — from $4,000 a month with no upfront cost. The exact monthly fee depends on scope and is put in writing after the free audit. Measured on citation rate.
- Gigabit Agents: production AI agents (support, chat, voice, booking, lead-qualification, and more), each built, integrated, and deployed in your stack for one flat fee — from $8,000 per agent; advanced builds (voice, multi-system, high-volume) scope up to about $150,000.
- AI Transformation Sprint: $25,000 fixed, 2 weeks — a diagnostic that ends in a deployed pilot, not a slide deck; credits toward a Gigabit Agents build.
- Managed AI Operations: Light $3,000/mo (1 system), Standard $8,000/mo (3 systems), Heavy $20,000/mo (5+ systems or production agent platforms).
- Embedded AI Teams: $3,000–$7,500 per engineer per month, assembled in 2 weeks, first PRs by Day 5, no lock-in.
- AI Quick Win Sprint: $8,000 fixed, 3 weeks — one workflow scoped, built, and deployed.
- Prototype Production Review: $2,500 fixed, 5 business days — A senior review of an app built with Lovable, Bolt, Replit, Cursor or Claude Code: security (Supabase row-level security, exposed keys, auth), data model, performance and deploy, with a graded report and a fix plan. Credited in full to a build within 60 days.
- AI Security Assessment: from $12,000 fixed, 2–4 weeks — threat-model and red-team the agents, LLMs and RAG pipelines you ship, with a graded report and a remediation plan.
- Website Build: from $6,000 fixed, from 4 weeks. Store Build: from $9,000 fixed, from 6 weeks. Platform Migration: from $7,000 fixed. CRM Setup: from $6,000 fixed. Odoo Implementation: from $15,000 fixed, phased; vendor licence fees are separate.
- Custom builds (software, SaaS, mobile, integrations, data, cloud): a fixed price in writing after scoping.
- Senior engineers plus AI agents on a productized, fixed-price line — quality-gated with evals and observability. Output, not hours; not a body shop, not a coding-agent tool.
- Tool-agnostic and independent: not affiliated with any model vendor's deployment arm. 4.6★ from 18 Google reviews.

Full text of the case studies and articles: https://gigabit.agency/llms-full.txt

## Pricing
- [Pricing — every published price, one page](https://gigabit.agency/pricing/): Every fixed-scope offer with its dollar amount; builds larger than a base scope are quoted in writing

## Services
- [Services hub](https://gigabit.agency/services/): The full ladder, Audit → Build → Operate, and what the factory builds
- [Ascent — the growth program](https://gigabit.agency/services/geo/): 12-month growth program, from $4,000/mo with no upfront; measured on citation rate
- [Gigabit Agents](https://gigabit.agency/services/ai-agents/): Production AI agents built and deployed for one flat fee, from $8,000 per agent
- [Managed AI Operations](https://gigabit.agency/services/managed-ai-operations/): Eval runs, drift detection, incident response, $3,000–$20,000/mo
- [AI Transformation Sprint](https://gigabit.agency/services/ai-transformation/): $25,000, 2 weeks diagnostic that ends in a deployed pilot
- [Embedded AI Teams](https://gigabit.agency/services/embedded-ai-teams/): A dedicated development team of senior AI engineers inside your team
- [AI Security](https://gigabit.agency/services/ai-security/): AI security assessment for agents, LLMs and RAG pipelines, from $12,000 fixed
- [Prototype Production Review](https://gigabit.agency/services/prototype-to-production/): $2,500 fixed, 5 business days, for apps built with Lovable, Bolt, Replit, Cursor or Claude Code

## What the factory builds
The same senior team builds the software around the AI. Each build gets a fixed price in writing after scoping.
- [Custom software development](https://gigabit.agency/services/custom-software-development/): Custom software and web app development by senior engineers working with AI agents. Past builds: a $48K MVP in 11 weeks, a $118K payments platform.
- [SaaS development](https://gigabit.agency/services/saas-development/): SaaS development from first release to scale: multi-tenant architecture, billing, analytics and AI features. A $48K first build reached $19K MRR and a Series A.
- [MVP development](https://gigabit.agency/services/mvp-development/): MVP development for founders: scope cut to what users will pay for, built in weeks by senior engineers. Published MVPs: $48K in 11 weeks, $72K in 14 weeks.
- [Cloud migration](https://gigabit.agency/services/cloud-migration/): Cloud migration and DevOps consulting: Heroku, legacy hosting or on-premise to AWS, Azure or GCP. Published: zero downtime at $14M/month, 35% lower cost.
- [Data engineering](https://gigabit.agency/services/data-engineering/): Data engineering services: pipelines, a Snowflake or BigQuery warehouse, dbt models and dashboards. Published build: reporting cut from 56 hours a month to 4.
- [Legacy modernization](https://gigabit.agency/services/legacy-modernization/): Legacy application modernization without a big-bang rewrite: services extracted one at a time, CI/CD, cloud and data clean-up. Published: deploys 38 min to 2.5.
- [CRM implementation](https://gigabit.agency/services/crm-implementation/): CRM implementation from $6,000 fixed on HubSpot, Zoho or Pipedrive, plus Salesforce and Odoo projects: clean data, pipelines, automation and AI agents.
- [ERP and API integration](https://gigabit.agency/services/erp-integration/): ERP and API integration: NetSuite, Odoo, QuickBooks and Xero connected to your store, CRM and apps, with retries and audit logs. One integration $8,000 fixed.
- [Mobile app development](https://gigabit.agency/services/mobile-app-development/): Mobile app development in React Native for iOS and Android, with the back end, integrations and AI features. Published: a $124K telehealth app in 5 months.
- [Offshore development](https://gigabit.agency/services/offshore-development/): Senior-led offshore software development from Dhaka: an embedded team at $3,000–$7,500 per engineer a month, or a fixed price. US and European overlap hours.
- [AI integration](https://gigabit.agency/services/ai-integration/): AI integration services: models and agents built into the software and workflows you already run, with evaluations and monitoring. One workflow from $8,000.
- [SaaS website design](https://gigabit.agency/services/saas-website-design/): SaaS and B2B website design on Webflow, Framer or WordPress from $6,000 fixed: positioning, pages that convert, fast builds and content AI assistants can cite.
- [Business process automation](https://gigabit.agency/services/business-process-automation/): Business process automation services: workflows across your CRM, ERP, store and inbox on n8n, Make or code, with AI where rules fail. One workflow $8,000 fixed.

## Advice and oversight
- [Fractional CTO](https://gigabit.agency/services/fractional-cto/): Fractional CTO support and technical due diligence from senior engineers: architecture, AI strategy, codebase reviews, hiring and vendor decisions.
- [AI governance](https://gigabit.agency/services/ai-governance/): AI governance consulting for teams that build or buy AI: AI inventory, risk tiers mapped to the EU AI Act, NIST AI RMF and ISO 42001, oversight and audit logs.

## Industries
- [AI for SaaS](https://gigabit.agency/industries/saas/): AI support & CX automation
- [AI for Healthcare](https://gigabit.agency/industries/healthcare/): Patient intake + clinical documentation, HIPAA-aware
- [AI for Commerce](https://gigabit.agency/industries/commerce/): Support, inventory, catalog and returns agents
- [AI for Fintech](https://gigabit.agency/industries/fintech/): Compliance-aware agents with audit trails
- [AI for KYC and KYB](https://gigabit.agency/industries/fintech/kyc-kyb-ai-agents/): Where agents prepare onboarding cases and where a person decides; our fintech case is payment-compliance verification, not KYC
- [AI for Professional Services](https://gigabit.agency/industries/professional-services/): Document handling, time capture, knowledge agents

## Free tools
- [AI Citation Audit](https://gigabit.agency/tools/geo-citation-audit/): Whether AI assistants recommend you for your category's buying queries; the report is sent by email
- [AI Readiness Assessment](https://gigabit.agency/tools/ai-readiness-assessment/): Six questions across six dimensions, with a readiness score and a recommended next step
- [AI ROI Calculator](https://gigabit.agency/tools/ai-roi-calculator/): Annual saving, payback and 3-year net of automating one workflow, with the method on the page

## Platforms we build on
Engineering (builds, migrations, integrations, custom apps) and AI agents on the platform a client already runs, priced from the same published list.
- [All platforms](https://gigabit.agency/platforms/): Commerce, websites and CMS, CRM and business systems, automation
- [Shopify](https://gigabit.agency/platforms/shopify/): Builds, migrations, custom apps and Functions, plus support and returns agents grounded in live order data.
- [WooCommerce](https://gigabit.agency/platforms/woocommerce/): Custom extensions, checkout and payments, performance at scale, and an honest view on when to stay and when to move.
- [BigCommerce](https://gigabit.agency/platforms/bigcommerce/): B2B and multi-storefront builds, headless storefronts, and the ERP integrations that make or break them.
- [Adobe Commerce (Magento)](https://gigabit.agency/platforms/adobe-commerce/): Upgrades, performance, custom modules and ERP integrations, and a straight answer on whether to stay.
- [WordPress](https://gigabit.agency/platforms/wordpress/): Custom themes, blocks and plugins, headless front ends, performance and security, and AI through the 7.0 Abilities API.
- [Webflow](https://gigabit.agency/platforms/webflow/): Marketing sites, CMS architecture, custom code components and migrations, built so your team can edit without us.
- [Framer](https://gigabit.agency/platforms/framer/): Design-led sites with custom code components, CMS structure, working analytics and CRM hand-off.
- [Wix](https://gigabit.agency/platforms/wix/): Custom apps, integrations and automation on Wix, and a clean migration when you have outgrown it.
- [HubSpot](https://gigabit.agency/platforms/hubspot/): Workflow automation, custom objects, integrations with billing and product data, and agents that span the CRM and beyond.
- [Salesforce](https://gigabit.agency/platforms/salesforce/): Integrations, Flow migration, Apex and Agentforce actions at a fixed price, for teams that already run Salesforce.
- [Zoho](https://gigabit.agency/platforms/zoho/): CRM customisation, Deluge functions, integrations across Zoho One and outside it, and agents on Zia or beyond.
- [Pipedrive](https://gigabit.agency/platforms/pipedrive/): Pipeline design, automation, integrations and agents that qualify leads and chase follow-ups.
- [Odoo](https://gigabit.agency/platforms/odoo/): Implementation, custom modules, upgrades and store integrations from an Odoo Partner, plus agents on Odoo 20.
- [NetSuite](https://gigabit.agency/platforms/netsuite/): Integrations with Shopify, CRMs and billing, SuiteScript and RESTlets, SOAP-to-REST migration, and agents on NetSuite data.
- [QuickBooks](https://gigabit.agency/platforms/quickbooks/): Stripe, Shopify, CRM and app integrations with QuickBooks Online, built to reconcile, plus finance agents and reporting.
- [n8n](https://gigabit.agency/platforms/n8n/): Self-hosted n8n, workflows and AI agents with error handling, tests and monitoring. In production for a $28M seller.
- [Zapier & Make](https://gigabit.agency/platforms/zapier-make/): An inventory of what is running, the fragile parts rebuilt, and a migration when the task bill stops making sense.

## Where we work
One contracting entity (Gigaverse Global LLC, Wyoming, USA), one price list in US dollars, delivery from Dhaka (UTC+6). No other offices. Delivery is in English; the site is also published in Japanese at /ja/, German at /de/, French at /fr/, Spanish at /es/ and Arabic at /ar/ (same paths, prefixed with the language code).
- [Regions](https://gigabit.agency/regions/): Working-hours overlap and data rules by region
- [United States](https://gigabit.agency/regions/united-states/): A Wyoming-registered company with engineering in Dhaka. US contract, US dollars, published prices, BAAs signed, and a team whose working day is your night.
- [Australia](https://gigabit.agency/regions/australia/): Senior engineers four hours behind Sydney, so every afternoon is shared. Fixed prices in USD, and builds documented for the December 2026 Privacy Act change.
- [United Kingdom](https://gigabit.agency/regions/united-kingdom/): Senior engineers who share your morning, fixed prices, and builds designed so personal data need not leave the UK. Read how transfers and UK GDPR are handled.
- [Europe](https://gigabit.agency/regions/europe/): Senior engineers sharing your morning, fixed prices, EU-region deployment by default, and agents classified against the EU AI Act before they are built.
- [Middle East](https://gigabit.agency/regions/middle-east/): Two hours from Dubai with seven shared working hours a day. Fixed prices in USD, in-region deployment, and a straight account of what we cannot do in the Gulf.
- [Japan](https://gigabit.agency/regions/japan/): Three hours behind Tokyo with six shared hours a day. Enquiries answered in Japanese, a bridge engineer where a project needs one, and fixed prices published.
- [Asia-Pacific](https://gigabit.agency/regions/asia-pacific/): Two hours behind Singapore and Hong Kong with seven shared hours a day. Agents built to Singapore's agentic AI governance framework, at published prices.

## 日本語 (Japanese)
英語サイトと同じ構成の日本語ページです。URLは英語ページのパスの前に /ja を付けたものです。料金はすべて米ドル建て。日本国内に拠点・法人はありません。
- [/ja/](https://gigabit.agency/ja/)
- [AI導入支援・AI統合](https://gigabit.agency/ja/services/ai-integration/): 今お使いのソフトウェアと業務フローに、AIモデルとAIエージェントを組み込みます。
- [CTO代行・技術顧問](https://gigabit.agency/ja/services/fractional-cto/): シニアエンジニアによる技術リーダーシップと技術デューデリジェンス。
- [業務プロセス自動化](https://gigabit.agency/ja/services/business-process-automation/): CRM、ERP、ECストア、メールをまたぐワークフローを構築し、固定のルールでは扱えない部分にAIを使います。
- [AIガバナンス](https://gigabit.agency/ja/services/ai-governance/): AIの棚卸し、リスク分類、ポリシー、そしてそれを実行する統制。
- [Shopify](https://gigabit.agency/ja/platforms/shopify/): 構築、移行、独自アプリ、Shopify Functionsの開発に加え、実際の注文データに基づいて動くサポート・返品対応エージェントを開発します。
- [WooCommerce](https://gigabit.agency/ja/platforms/woocommerce/): 独自の拡張機能、チェックアウトと決済、大規模ストアのパフォーマンス改善。使い続けるべきか移行すべきかも率直にお伝えします。
- [BigCommerce](https://gigabit.agency/ja/platforms/bigcommerce/): B2Bとマルチストアフロントの構築、ヘッドレス構成のストアフロント、そして成否を分けるERP連携。
- [Adobe Commerce（Magento）](https://gigabit.agency/ja/platforms/adobe-commerce/): アップグレード、パフォーマンス改善、独自モジュール、ERP連携。使い続けるべきかどうかも率直にお答えします。
- [WordPress](https://gigabit.agency/ja/platforms/wordpress/): 独自のテーマ・ブロック・プラグイン開発、ヘッドレス構成、高速化とセキュリティ対策、7.0のAbilities APIによるAI連携。
- [Webflow](https://gigabit.agency/ja/platforms/webflow/): マーケティングサイト、CMS設計、カスタムコードコンポーネント、移行。納品後は貴社チームだけで更新できるように構築します。
- [Framer](https://gigabit.agency/ja/platforms/framer/): デザインを重視したサイトに、カスタムコードコンポーネント、CMSの構造、正しく動くアクセス解析、CRMへの受け渡しを加えます。
- [Wix](https://gigabit.agency/ja/platforms/wix/): Wix上の独自アプリ開発、システム連携、業務自動化。Wixでは足りなくなった場合は、確実に移行します。
- [HubSpot](https://gigabit.agency/ja/platforms/hubspot/): ワークフローの自動化、カスタムオブジェクト、請求・製品データとの連携、CRMの外まで扱えるエージェント。
- [Salesforce](https://gigabit.agency/ja/platforms/salesforce/): すでにSalesforceを運用しているチーム向けに、連携、Flowへの移行、Apex、Agentforceのアクションを固定価格で開発します。
- [Zoho](https://gigabit.agency/ja/platforms/zoho/): CRMのカスタマイズ、Deluge関数、Zoho One内外の連携、Ziaの上でも外でも動くエージェント。
- [Pipedrive](https://gigabit.agency/ja/platforms/pipedrive/): パイプラインの設計、自動化、連携、リードの選別とフォローアップを担うエージェント。
- [Odoo](https://gigabit.agency/ja/platforms/odoo/): Odoo Partnerによる導入、カスタムモジュール、バージョンアップ、ECストアとの連携。Odoo 20で動くエージェントにも対応します。
- [n8n](https://gigabit.agency/ja/platforms/n8n/): セルフホストのn8nで、エラー処理、テスト、監視を備えたワークフローとAIエージェントを構築します。年商2,800万ドルのEC事業者で本番稼働中です。
- [Zapier & Make](https://gigabit.agency/ja/platforms/zapier-make/): 稼働中の自動化の棚卸し、壊れやすい部分の再構築、タスクの請求額が見合わなくなったときの移行に対応します。
- [物流SaaSのMVP開発事例：企画段階から11週間で有料顧客を獲得](https://gigabit.agency/ja/proof/saas-mvp-logistics-platform/)
- [HerokuからAWSへの無停止移行：月間1,400万ドルの決済を処理しながら](https://gigabit.agency/ja/proof/cloud-migration-fintech-zero-downtime/)
- [AIサポートエージェントで年間$217,200を削減、顧客満足度も改善したD2Cブランド](https://gigabit.agency/ja/proof/ai-automation-ecommerce-support/)
- [遠隔医療プラットフォーム開発：FHIR連携とHIPAA準拠で90日間に患者1,200人](https://gigabit.agency/ja/proof/healthtech-telehealth-platform/)
- [ラボ型開発事例：エンジニア3名でHRテックSaaSのスプリント速度が2.3倍に](https://gigabit.agency/ja/proof/team-augmentation-hrtech/)
- [週32時間の手作業を4.5時間に：マルチチャネルEC運営のAI業務自動化](https://gigabit.agency/ja/proof/ai-workflow-automation-ecommerce-ops/)
- [マーケットプレイス開発事例：構想から初年度の流通総額240万ドルへ](https://gigabit.agency/ja/proof/marketplace-platform-build/)
- [AI導入事例：経営コンサルティング会社が四半期1,200時間を取り戻す](https://gigabit.agency/ja/proof/ai-transformation-professional-services/)
- [B2B決済プラットフォーム開発事例：構想から初年度4,100万ドルの決済処理へ](https://gigabit.agency/ja/proof/fintech-payment-platform/)
- [SaaSのデータ基盤構築事例：手作業のスプレッドシートからセルフサービス分析へ](https://gigabit.agency/ja/proof/data-engineering-saas-analytics/)
- [患者受付・EHR連携・分析基盤の統合事例：受付事務の時間を41%削減](https://gigabit.agency/ja/proof/healthcare-unified-intake-ehr-analytics/)
- [専門クリニックの生成AI患者アシスタント：受付への定型問い合わせを62%削減](https://gigabit.agency/ja/proof/clinic-llm-assistant-front-desk/)
- [複数拠点の歯科グループがオンライン予約を導入：予約どおりの来院が27%増加](https://gigabit.agency/ja/proof/dental-online-booking-kept-appointments/)
- [AIエージェント開発の費用相場（2026年版）：構築費と運用費の内訳](https://gigabit.agency/ja/insights/ai-agent-cost/)
- [AIエージェントとチャットボットの違い：自社に必要なのはどちらか](https://gigabit.agency/ja/insights/ai-agent-vs-chatbot/)
- [AIのPoCが本番稼働に進まない理由と、本番に届けるための5つの関門](https://gigabit.agency/ja/insights/why-your-ai-pilot-never-reached-production/)
- [RPA（ルールベースの自動化）とAIエージェントの違い：切り替えの目安](https://gigabit.agency/ja/insights/rules-based-automation-vs-ai-agents/)
- [AI開発会社の選び方：発注前に聞くべき9つの質問](https://gigabit.agency/ja/insights/questions-to-ask-an-ai-agency/)
- [GEO・LLMO対策の実践ガイド：ChatGPTやPerplexityに引用される方法](https://gigabit.agency/ja/insights/generative-engine-optimization-field-guide/)
- [llms.txtとは何か：書き方と効果、robots.txtとの違い](https://gigabit.agency/ja/insights/llms-txt-explained/)
- [ChatGPTで上位表示する方法：LLMO実践ガイド](https://gigabit.agency/ja/insights/how-to-rank-in-chatgpt/)
- [n8nテンプレートの選び方と本番前に直す8項目](https://gigabit.agency/ja/insights/n8n-templates-for-business/)
- [/ja/offshore-development/](https://gigabit.agency/ja/offshore-development/)

## Deutsch (German)
Deutsche Fassung der englischen Website. Die URL ist der englische Pfad mit vorangestelltem /de. Alle Preise in US-Dollar. Keine Niederlassung in Deutschland, Österreich, der Schweiz oder der EU. Anfragen auf Deutsch oder Englisch, Antworten und Projektarbeit auf Englisch.
- [/de/](https://gigabit.agency/de/)
- [KI-Integration](https://gigabit.agency/de/services/ai-integration/): Modelle und Agenten, eingebaut in die Software und Abläufe, die Sie bereits nutzen.
- [Fractional CTO](https://gigabit.agency/de/services/fractional-cto/): Technische Führung und Due Diligence durch Senior-Engineers.
- [Prozessautomatisierung](https://gigabit.agency/de/services/business-process-automation/): Workflows über CRM, ERP, Shop und Postfach hinweg, mit KI dort, wo feste Regeln versagen.
- [KI-Governance](https://gigabit.agency/de/services/ai-governance/): Ein KI-Inventar, Risikostufen, Richtlinien und die Kontrollen, die sie durchsetzen.
- [Shopify](https://gigabit.agency/de/platforms/shopify/): Shops, Migrationen, eigene Apps und Functions, dazu Support- und Retouren-Agenten, die mit echten Bestelldaten arbeiten.
- [WooCommerce](https://gigabit.agency/de/platforms/woocommerce/): Eigene Erweiterungen, Checkout und Zahlung, Performance bei großen Shops und eine ehrliche Einschätzung, wann Sie bleiben und wann Sie wechseln sollten.
- [BigCommerce](https://gigabit.agency/de/platforms/bigcommerce/): B2B und Multi-Storefront, Headless-Storefronts und die ERP-Integrationen, an denen solche Projekte gelingen oder scheitern.
- [Adobe Commerce (Magento)](https://gigabit.agency/de/platforms/adobe-commerce/): Upgrades, Performance, eigene Module und ERP-Integrationen, dazu eine klare Antwort auf die Frage, ob Sie bleiben sollten.
- [WordPress](https://gigabit.agency/de/platforms/wordpress/): Eigene Themes, Blöcke und Plugins, Headless-Frontends, Performance und Sicherheit sowie KI über die Abilities API aus Version 7.0.
- [Webflow](https://gigabit.agency/de/platforms/webflow/): Marketing-Websites, CMS-Architektur, Code Components und Migrationen. So gebaut, dass Ihr Team ohne uns weiterarbeiten kann.
- [Framer](https://gigabit.agency/de/platforms/framer/): Designgetriebene Websites mit Code Components, CMS-Struktur, funktionierender Webanalyse und Übergabe ans CRM.
- [Wix](https://gigabit.agency/de/platforms/wix/): Eigene Apps, Integrationen und Automatisierung auf Wix und eine saubere Migration, wenn Sie der Plattform entwachsen sind.
- [HubSpot](https://gigabit.agency/de/platforms/hubspot/): Workflow-Automatisierung, benutzerdefinierte Objekte, Integrationen mit Abrechnungs- und Produktdaten und Agenten, die über das CRM hinaus arbeiten.
- [Salesforce](https://gigabit.agency/de/platforms/salesforce/): Integrationen, Flow-Migration, Apex und Agentforce-Aktionen zum Festpreis, für Teams, die Salesforce bereits einsetzen.
- [Zoho](https://gigabit.agency/de/platforms/zoho/): CRM-Anpassung, Deluge-Funktionen, Integrationen innerhalb von Zoho One und nach außen, dazu Agenten auf Zia oder außerhalb davon.
- [Pipedrive](https://gigabit.agency/de/platforms/pipedrive/): Pipeline-Design, Automatisierung, Integrationen und Agenten, die Leads qualifizieren und nachfassen.
- [Odoo](https://gigabit.agency/de/platforms/odoo/): Einführung, eigene Module, Upgrades und Shop-Integrationen vom Odoo Partner, dazu Agenten auf Odoo 20.
- [n8n](https://gigabit.agency/de/platforms/n8n/): Selbst gehostetes n8n, Workflows und KI-Agenten mit Fehlerbehandlung, Tests und Monitoring. Im Produktivbetrieb bei einem Händler mit 28 Mio. $ Jahresumsatz.
- [Zapier & Make](https://gigabit.agency/de/platforms/zapier-make/): Eine Bestandsaufnahme dessen, was läuft, die anfälligen Teile neu gebaut und eine Migration, wenn die Rechnung für Tasks nicht mehr aufgeht.
- [MVP für Logistik-SaaS: in 11 Wochen zu zahlenden Kunden](https://gigabit.agency/de/proof/saas-mvp-logistics-platform/)
- [Von Heroku zu AWS ohne Ausfall, bei 14 Mio. $ Volumen pro Monat](https://gigabit.agency/de/proof/cloud-migration-fintech-zero-downtime/)
- [KI-Support-Agent spart einer D2C-Marke 217.200 $ pro Jahr](https://gigabit.agency/de/proof/ai-automation-ecommerce-support/)
- [Telemedizin mit FHIR und HIPAA: 1.200 Patienten in 90 Tagen](https://gigabit.agency/de/proof/healthtech-telehealth-platform/)
- [HR-Tech-SaaS: 3 eingebettete Engineers, 2,3-fache Sprint-Velocity](https://gigabit.agency/de/proof/team-augmentation-hrtech/)
- [KI-Workflow-Automatisierung: von 32 auf 4,5 Stunden pro Woche](https://gigabit.agency/de/proof/ai-workflow-automation-ecommerce-ops/)
- [Marktplatz-Entwicklung: vom Konzept zu 2,4 Mio. $ GMV in Jahr 1](https://gigabit.agency/de/proof/marketplace-platform-build/)
- [KI-Transformation: 1.200 Stunden pro Quartal zurückgewonnen](https://gigabit.agency/de/proof/ai-transformation-professional-services/)
- [B2B-Zahlungsplattform: vom Konzept zu 41 Mio. $ im ersten Jahr](https://gigabit.agency/de/proof/fintech-payment-platform/)
- [Data Engineering für SaaS: von Tabellen zu Self-Service-Analytics](https://gigabit.agency/de/proof/data-engineering-saas-analytics/)
- [Patientenaufnahme mit EHR-Anbindung: 41 % weniger Zeitaufwand](https://gigabit.agency/de/proof/healthcare-unified-intake-ehr-analytics/)
- [LLM-Patientenassistent: 62 % weniger Routineanfragen am Empfang](https://gigabit.agency/de/proof/clinic-llm-assistant-front-desk/)
- [Online-Terminbuchung: 27 % mehr wahrgenommene Zahnarzttermine](https://gigabit.agency/de/proof/dental-online-booking-kept-appointments/)
- [Was kostet ein KI-Agent? Entwicklung und Betrieb](https://gigabit.agency/de/insights/ai-agent-cost/)
- [KI-Agent oder Chatbot: der Unterschied](https://gigabit.agency/de/insights/ai-agent-vs-chatbot/)
- [Warum KI-Piloten nicht in Produktion gehen](https://gigabit.agency/de/insights/why-your-ai-pilot-never-reached-production/)
- [RPA oder KI-Agent: wann Regeln reichen](https://gigabit.agency/de/insights/rules-based-automation-vs-ai-agents/)
- [KI-Agentur: 9 Fragen vor der Beauftragung](https://gigabit.agency/de/insights/questions-to-ask-an-ai-agency/)
- [GEO: von ChatGPT und Perplexity zitiert werden](https://gigabit.agency/de/insights/generative-engine-optimization-field-guide/)
- [Was ist llms.txt? Unterschied zu robots.txt](https://gigabit.agency/de/insights/llms-txt-explained/)
- [In ChatGPT ranken: Praxisleitfaden zu ChatGPT SEO](https://gigabit.agency/de/insights/how-to-rank-in-chatgpt/)
- [n8n Templates für Unternehmen: die Checkliste](https://gigabit.agency/de/insights/n8n-templates-for-business/)
- [/de/offshore-entwicklung/](https://gigabit.agency/de/offshore-entwicklung/)
- [/de/impressum/](https://gigabit.agency/de/impressum/)

## Français (French)
Version française du site anglais. L’URL est le chemin anglais précédé de /fr. Tous les prix sont en dollars US. Aucun établissement en France, en Belgique, en Suisse ni dans l’Union européenne. Demandes en français ou en anglais, réponses et projets en anglais.
- [/fr/](https://gigabit.agency/fr/)
- [Intégration de l’IA](https://gigabit.agency/fr/services/ai-integration/): Des modèles et des agents intégrés aux logiciels et aux workflows que vous utilisez déjà.
- [CTO à temps partagé](https://gigabit.agency/fr/services/fractional-cto/): Leadership technique et due diligence par des ingénieurs seniors.
- [Automatisation des processus métier](https://gigabit.agency/fr/services/business-process-automation/): Des workflows entre votre CRM, votre ERP, votre boutique et votre messagerie, avec l’IA là où les règles fixes ne suffisent plus.
- [Gouvernance de l’IA](https://gigabit.agency/fr/services/ai-governance/): Un inventaire de l’IA, des niveaux de risque, des politiques et les contrôles qui les font respecter.
- [Shopify](https://gigabit.agency/fr/platforms/shopify/): Créations, migrations, applications sur mesure et Functions, plus des agents de support et de retours appuyés sur les données de commande réelles.
- [WooCommerce](https://gigabit.agency/fr/platforms/woocommerce/): Extensions sur mesure, checkout et paiements, performance à grande échelle, et un avis franc sur le moment de rester et celui de partir.
- [BigCommerce](https://gigabit.agency/fr/platforms/bigcommerce/): Projets B2B et multi-boutiques, fronts headless, et les intégrations ERP dont dépend leur réussite.
- [Adobe Commerce (Magento)](https://gigabit.agency/fr/platforms/adobe-commerce/): Montées de version, performance, modules sur mesure et intégrations ERP, et une réponse franche à la question : faut-il rester ?
- [WordPress](https://gigabit.agency/fr/platforms/wordpress/): Thèmes, blocs et extensions sur mesure, front-ends headless, performance et sécurité, et l’IA via l’Abilities API de la version 7.0.
- [Webflow](https://gigabit.agency/fr/platforms/webflow/): Sites marketing, architecture CMS, composants de code et migrations, construits pour que votre équipe modifie le site sans nous.
- [Framer](https://gigabit.agency/fr/platforms/framer/): Des sites portés par le design, avec composants de code sur mesure, structure CMS, mesure d’audience qui fonctionne et transmission au CRM.
- [Wix](https://gigabit.agency/fr/platforms/wix/): Applications sur mesure, intégrations et automatisation sur Wix, et une migration propre quand la plateforme ne vous suffit plus.
- [HubSpot](https://gigabit.agency/fr/platforms/hubspot/): Automatisation des workflows, objets personnalisés, intégrations avec la facturation et les données produit, et agents qui travaillent au-delà du CRM.
- [Salesforce](https://gigabit.agency/fr/platforms/salesforce/): Intégrations, migration vers Flow, Apex et actions Agentforce à prix fixe, pour les équipes qui utilisent déjà Salesforce.
- [Zoho](https://gigabit.agency/fr/platforms/zoho/): Personnalisation du CRM, fonctions Deluge, intégrations dans Zoho One et vers l’extérieur, et agents sur Zia ou en dehors.
- [Pipedrive](https://gigabit.agency/fr/platforms/pipedrive/): Conception du pipeline, automatisation, intégrations et agents qui qualifient les leads et assurent les relances.
- [Odoo](https://gigabit.agency/fr/platforms/odoo/): Déploiement, modules spécifiques, montées de version et intégrations e-commerce par un Odoo Partner, et agents sur Odoo 20.
- [n8n](https://gigabit.agency/fr/platforms/n8n/): n8n auto-hébergé, workflows et agents IA avec gestion des erreurs, tests et supervision. En production chez un vendeur à 28 M$ de chiffre d’affaires.
- [Zapier & Make](https://gigabit.agency/fr/platforms/zapier-make/): Un inventaire de ce qui tourne, les parties fragiles reconstruites, et une migration quand la facture à la tâche ne se justifie plus.
- [MVP SaaS logistique : des clients payants en 11 semaines](https://gigabit.agency/fr/proof/saas-mvp-logistics-platform/)
- [De Heroku à AWS sans interruption, avec 14 M$ traités par mois](https://gigabit.agency/fr/proof/cloud-migration-fintech-zero-downtime/)
- [Agent IA de support client : 217 200 $ économisés par an](https://gigabit.agency/fr/proof/ai-automation-ecommerce-support/)
- [Télémédecine, FHIR et HIPAA : 1 200 patients en 90 jours](https://gigabit.agency/fr/proof/healthtech-telehealth-platform/)
- [SaaS RH : 3 ingénieurs intégrés, 2,3 fois plus de vélocité](https://gigabit.agency/fr/proof/team-augmentation-hrtech/)
- [Workflows IA en e-commerce : de 32 à 4,5 heures par semaine](https://gigabit.agency/fr/proof/ai-workflow-automation-ecommerce-ops/)
- [Place de marché B2B : du concept à 2,4 M$ de GMV en un an](https://gigabit.agency/fr/proof/marketplace-platform-build/)
- [Transformation IA : 1 200 heures récupérées par trimestre](https://gigabit.agency/fr/proof/ai-transformation-professional-services/)
- [Plateforme de paiement B2B : du concept à 41 M$ en un an](https://gigabit.agency/fr/proof/fintech-payment-platform/)
- [Data engineering SaaS : du tableur à la BI en libre-service](https://gigabit.agency/fr/proof/data-engineering-saas-analytics/)
- [Admission des patients et EHR : 41 % de temps en moins](https://gigabit.agency/fr/proof/healthcare-unified-intake-ehr-analytics/)
- [Assistant patient LLM : 62 % de demandes répétitives en moins](https://gigabit.agency/fr/proof/clinic-llm-assistant-front-desk/)
- [Prise de rendez-vous en ligne : +27 % de rendez-vous honorés](https://gigabit.agency/fr/proof/dental-online-booking-kept-appointments/)
- [Coût d'un agent IA : build et exploitation](https://gigabit.agency/fr/insights/ai-agent-cost/)
- [Agent IA ou chatbot : la différence](https://gigabit.agency/fr/insights/ai-agent-vs-chatbot/)
- [Pourquoi les pilotes IA ratent la production](https://gigabit.agency/fr/insights/why-your-ai-pilot-never-reached-production/)
- [RPA ou agent IA : quand les règles suffisent](https://gigabit.agency/fr/insights/rules-based-automation-vs-ai-agents/)
- [Agence IA : 9 questions avant de signer](https://gigabit.agency/fr/insights/questions-to-ask-an-ai-agency/)
- [GEO : comment être cité par ChatGPT et Perplexity](https://gigabit.agency/fr/insights/generative-engine-optimization-field-guide/)
- [llms.txt : à quoi ça sert, et robots.txt](https://gigabit.agency/fr/insights/llms-txt-explained/)
- [Référencement ChatGPT : comment y être cité](https://gigabit.agency/fr/insights/how-to-rank-in-chatgpt/)
- [Modèles n8n : du template à la production](https://gigabit.agency/fr/insights/n8n-templates-for-business/)
- [/fr/externalisation-developpement/](https://gigabit.agency/fr/externalisation-developpement/)
- [/fr/mentions-legales/](https://gigabit.agency/fr/mentions-legales/)

## Español (Spanish)
Versión en español del sitio en inglés. La URL es la ruta en inglés precedida de /es. Todos los precios están en dólares estadounidenses (USD). Sin oficinas ni filiales en España, América Latina o la Unión Europea. Consultas en español o en inglés, respuestas y proyectos en inglés.
- [/es/](https://gigabit.agency/es/)
- [Integración de IA](https://gigabit.agency/es/services/ai-integration/): Modelos y agentes integrados en el software y los flujos de trabajo que ya utiliza.
- [CTO fraccional](https://gigabit.agency/es/services/fractional-cto/): Liderazgo técnico y due diligence a cargo de ingenieros sénior.
- [Automatización de procesos de negocio](https://gigabit.agency/es/services/business-process-automation/): Flujos de trabajo entre su CRM, su ERP, su tienda y su correo, con IA donde las reglas fijas no bastan.
- [Gobernanza de IA](https://gigabit.agency/es/services/ai-governance/): Un inventario de IA, niveles de riesgo, políticas y los controles que las hacen cumplir.
- [Shopify](https://gigabit.agency/es/platforms/shopify/): Tiendas nuevas, migraciones, aplicaciones a medida y Functions, más agentes de soporte y de devoluciones basados en los datos reales de los pedidos.
- [WooCommerce](https://gigabit.agency/es/platforms/woocommerce/): Extensiones a medida, checkout y pagos, rendimiento a gran escala y una opinión franca sobre cuándo quedarse y cuándo migrar.
- [BigCommerce](https://gigabit.agency/es/platforms/bigcommerce/): Proyectos B2B y multitienda, tiendas headless y las integraciones con el ERP de las que depende que funcionen.
- [Adobe Commerce (Magento)](https://gigabit.agency/es/platforms/adobe-commerce/): Actualizaciones de versión, rendimiento, módulos a medida e integraciones con el ERP, y una respuesta directa sobre si conviene quedarse.
- [WordPress](https://gigabit.agency/es/platforms/wordpress/): Temas, bloques y plugins a medida, front-ends headless, rendimiento y seguridad, e IA mediante la Abilities API de la versión 7.0.
- [Webflow](https://gigabit.agency/es/platforms/webflow/): Sitios de marketing, arquitectura del CMS, componentes de código a medida y migraciones, construidos para que su equipo edite sin nosotros.
- [Framer](https://gigabit.agency/es/platforms/framer/): Sitios guiados por el diseño, con componentes de código a medida, estructura del CMS, analítica que funciona y traspaso al CRM.
- [Wix](https://gigabit.agency/es/platforms/wix/): Aplicaciones a medida, integraciones y automatización en Wix, y una migración ordenada cuando la plataforma ya no le basta.
- [HubSpot](https://gigabit.agency/es/platforms/hubspot/): Automatización de workflows, objetos personalizados, integraciones con la facturación y los datos de producto, y agentes que trabajan más allá del CRM.
- [Salesforce](https://gigabit.agency/es/platforms/salesforce/): Integraciones, migración a Flow, Apex y acciones de Agentforce a precio fijo, para equipos que ya usan Salesforce.
- [Zoho](https://gigabit.agency/es/platforms/zoho/): Personalización del CRM, funciones Deluge, integraciones dentro y fuera de Zoho One, y agentes sobre Zia o fuera de Zia.
- [Pipedrive](https://gigabit.agency/es/platforms/pipedrive/): Diseño del pipeline, automatización, integraciones y agentes que califican leads y hacen el seguimiento.
- [Odoo](https://gigabit.agency/es/platforms/odoo/): Implementación, módulos a medida, actualizaciones de versión e integraciones con su tienda, a cargo de un Odoo Partner, además de agentes en Odoo 20.
- [n8n](https://gigabit.agency/es/platforms/n8n/): n8n autoalojado, flujos de trabajo y agentes de IA con gestión de errores, pruebas y supervisión. En producción en un vendedor que factura 28 millones de USD al año.
- [Zapier & Make](https://gigabit.agency/es/platforms/zapier-make/): Un inventario de lo que está funcionando, las partes frágiles reconstruidas y una migración cuando la factura por tareas deja de tener sentido.
- [MVP SaaS de logística: clientes de pago en 11 semanas](https://gigabit.agency/es/proof/saas-mvp-logistics-platform/)
- [De Heroku a AWS sin caídas, con 14 millones de USD al mes](https://gigabit.agency/es/proof/cloud-migration-fintech-zero-downtime/)
- [Agente de IA de soporte: 217 200 USD de ahorro al año](https://gigabit.agency/es/proof/ai-automation-ecommerce-support/)
- [Telemedicina con FHIR y HIPAA: 1200 pacientes en 90 días](https://gigabit.agency/es/proof/healthtech-telehealth-platform/)
- [Equipo dedicado de 3 ingenieros: 2,3 veces más velocidad](https://gigabit.agency/es/proof/team-augmentation-hrtech/)
- [Automatización con IA: de 32 a 4,5 horas por semana](https://gigabit.agency/es/proof/ai-workflow-automation-ecommerce-ops/)
- [Marketplace B2B: 2,4 millones de USD de GMV en el primer año](https://gigabit.agency/es/proof/marketplace-platform-build/)
- [Transformación con IA: 1200 horas recuperadas por trimestre](https://gigabit.agency/es/proof/ai-transformation-professional-services/)
- [Plataforma de pagos B2B: 41 millones de USD en el primer año](https://gigabit.agency/es/proof/fintech-payment-platform/)
- [Ingeniería de datos SaaS: de hojas de cálculo al autoservicio](https://gigabit.agency/es/proof/data-engineering-saas-analytics/)
- [Admisión de pacientes integrada con el EHR: 41 % menos tiempo](https://gigabit.agency/es/proof/healthcare-unified-intake-ehr-analytics/)
- [Asistente LLM para pacientes: 62 % menos preguntas repetitivas](https://gigabit.agency/es/proof/clinic-llm-assistant-front-desk/)
- [Reserva de citas en línea: 27 % más de citas cumplidas](https://gigabit.agency/es/proof/dental-online-booking-kept-appointments/)
- [¿Cuánto cuesta un agente de IA?](https://gigabit.agency/es/insights/ai-agent-cost/)
- [Agente de IA o chatbot: diferencias y cuál elegir](https://gigabit.agency/es/insights/ai-agent-vs-chatbot/)
- [Por qué los pilotos de IA no llegan a producción](https://gigabit.agency/es/insights/why-your-ai-pilot-never-reached-production/)
- [RPA o agente de IA: cuándo bastan las reglas](https://gigabit.agency/es/insights/rules-based-automation-vs-ai-agents/)
- [Agencia de IA: 9 preguntas antes de firmar](https://gigabit.agency/es/insights/questions-to-ask-an-ai-agency/)
- [GEO: que ChatGPT y Perplexity citen su empresa](https://gigabit.agency/es/insights/generative-engine-optimization-field-guide/)
- [llms.txt: qué es y diferencia con robots.txt](https://gigabit.agency/es/insights/llms-txt-explained/)
- [Cómo posicionar en ChatGPT: guía práctica de SEO](https://gigabit.agency/es/insights/how-to-rank-in-chatgpt/)
- [Plantillas n8n para empresas: qué corregir primero](https://gigabit.agency/es/insights/n8n-templates-for-business/)
- [/es/outsourcing-desarrollo-software/](https://gigabit.agency/es/outsourcing-desarrollo-software/)
- [/es/aviso-legal/](https://gigabit.agency/es/aviso-legal/)

## العربية (Arabic)
النسخة العربية من الموقع الإنجليزي. عنوان كل صفحة هو المسار الإنجليزي مسبوقًا بـ /ar. جميع الأسعار بالدولار الأمريكي. لا مكاتب ولا كيانات قانونية في أي بلد عربي. تُستقبل الرسائل بالعربية أو بالإنجليزية، والرد وتنفيذ المشاريع بالإنجليزية. لم نسلّم بعد وكيل ذكاء اصطناعي يعمل باللغة العربية.
- [/ar/](https://gigabit.agency/ar/)
- [دمج الذكاء الاصطناعي](https://gigabit.agency/ar/services/ai-integration/): نماذج ووكلاء ذكاء اصطناعي داخل البرمجيات وسير العمل التي تعملون بها اليوم.
- [مدير تقني بدوام جزئي](https://gigabit.agency/ar/services/fractional-cto/): قيادة تقنية وفحص نافٍ للجهالة يقدّمهما مهندسون ذوو خبرة.
- [أتمتة العمليات التجارية](https://gigabit.agency/ar/services/business-process-automation/): سير عمل يربط نظام CRM ونظام ERP والمتجر والبريد، مع الذكاء الاصطناعي حيث تتعطل القواعد الثابتة.
- [حوكمة الذكاء الاصطناعي](https://gigabit.agency/ar/services/ai-governance/): جرد لأنظمة الذكاء الاصطناعي، ومستويات مخاطر، وسياسات، والضوابط التي تطبّقها.
- [Shopify](https://gigabit.agency/ar/platforms/shopify/): بناء المتاجر وترحيلها، وتطبيقات مخصصة مع Shopify Functions، ووكلاء للدعم والمرتجعات يعتمدون على بيانات الطلبات الفعلية.
- [WooCommerce](https://gigabit.agency/ar/platforms/woocommerce/): إضافات مخصصة، وصفحة الدفع وبوابات الدفع، وأداء يتحمّل الحجم الكبير، ورأي صريح في متى تبقون ومتى تنتقلون.
- [BigCommerce](https://gigabit.agency/ar/platforms/bigcommerce/): مشاريع B2B ومتاجر متعددة الواجهات، وواجهات headless، وتكاملات ERP التي يتوقف عليها نجاح ذلك كله.
- [Adobe Commerce (Magento)](https://gigabit.agency/ar/platforms/adobe-commerce/): ترقية الإصدارات، وتحسين الأداء، ووحدات مخصصة، وتكاملات ERP، وجواب صريح عن سؤال البقاء أو المغادرة.
- [WordPress](https://gigabit.agency/ar/platforms/wordpress/): قوالب ومكوّنات وإضافات مخصصة، واجهات headless، الأداء والأمان، والذكاء الاصطناعي عبر Abilities API في الإصدار 7.0.
- [Webflow](https://gigabit.agency/ar/platforms/webflow/): مواقع تسويقية، هيكلة CMS، مكوّنات برمجية مخصصة وترحيل، تُبنى ليحرّرها فريقكم دون الرجوع إلينا.
- [Framer](https://gigabit.agency/ar/platforms/framer/): مواقع يقودها التصميم، مع مكوّنات برمجية مخصصة وبنية CMS وتحليلات تعمل وربط بنظام CRM.
- [Wix](https://gigabit.agency/ar/platforms/wix/): تطبيقات مخصصة وتكاملات وأتمتة على Wix، وترحيل منظَّم عندما تتجاوز أعمالكم حدود المنصة.
- [HubSpot](https://gigabit.agency/ar/platforms/hubspot/): أتمتة سير العمل، وكائنات مخصصة، وتكاملات مع بيانات الفوترة والمنتج، ووكلاء يتجاوز عملهم حدود CRM.
- [Salesforce](https://gigabit.agency/ar/platforms/salesforce/): تكاملات، وترحيل إلى Flow، وتطوير Apex، وإجراءات Agentforce بسعر ثابت، للفرق التي تستخدم Salesforce أصلًا.
- [Zoho](https://gigabit.agency/ar/platforms/zoho/): تخصيص CRM، ودوال Deluge، وتكاملات داخل Zoho One وخارجه، ووكلاء على Zia أو خارجه.
- [Pipedrive](https://gigabit.agency/ar/platforms/pipedrive/): تصميم مسار المبيعات، والأتمتة، والتكاملات، ووكلاء يؤهّلون العملاء المحتملين ويتولّون المتابعة.
- [Odoo](https://gigabit.agency/ar/platforms/odoo/): تطبيق، ووحدات مخصصة، وترقيات، وربط للمتاجر من Odoo Partner، إضافة إلى وكلاء على Odoo 20.
- [n8n](https://gigabit.agency/ar/platforms/n8n/): استضافة ذاتية لمنصة n8n، وسير عمل ووكلاء ذكاء اصطناعي مع معالجة الأخطاء والاختبارات والمراقبة. يعمل في الإنتاج لدى بائع إيراداته السنوية 28 مليون دولار.
- [Zapier & Make](https://gigabit.agency/ar/platforms/zapier-make/): جرد لما يعمل، وإعادة بناء للأجزاء الهشّة، وترحيل عندما تفقد فاتورة المهام جدواها.
- [منصة SaaS لوجستية: من الفكرة إلى عملاء يدفعون خلال 11 أسبوعًا](https://gigabit.agency/ar/proof/saas-mvp-logistics-platform/)
- [الترحيل من Heroku إلى AWS بلا توقف: 14 مليون دولار شهريًا](https://gigabit.agency/ar/proof/cloud-migration-fintech-zero-downtime/)
- [وكيل ذكاء اصطناعي لدعم العملاء يوفّر 217,200 دولار سنويًا](https://gigabit.agency/ar/proof/ai-automation-ecommerce-support/)
- [منصة رعاية صحية عن بُعد: تكامل FHIR و1,200 مريض خلال 90 يومًا](https://gigabit.agency/ar/proof/healthtech-telehealth-platform/)
- [فريق تطوير مخصص من 3 مهندسين يرفع سرعة الإنجاز 2.3 مرة](https://gigabit.agency/ar/proof/team-augmentation-hrtech/)
- [أتمتة عمليات متجر إلكتروني: من 32 ساعة إلى 4.5 ساعة أسبوعيًا](https://gigabit.agency/ar/proof/ai-workflow-automation-ecommerce-ops/)
- [سوق إلكترونية: معاملات بقيمة 2.4 مليون دولار في أول سنة](https://gigabit.agency/ar/proof/marketplace-platform-build/)
- [التحول بالذكاء الاصطناعي: استعادة 1,200 ساعة كل ربع سنة](https://gigabit.agency/ar/proof/ai-transformation-professional-services/)
- [منصة مدفوعات B2B: معالجة 41 مليون دولار في السنة الأولى](https://gigabit.agency/ar/proof/fintech-payment-platform/)
- [هندسة بيانات SaaS: من الجداول اليدوية إلى تحليلات ذاتية](https://gigabit.agency/ar/proof/data-engineering-saas-analytics/)
- [استقبال المرضى مع ربط EHR: وقت أقل بنسبة 41% في 14 فرعًا](https://gigabit.agency/ar/proof/healthcare-unified-intake-ehr-analytics/)
- [مساعد مرضى بنموذج لغوي كبير: استفسارات متكررة أقل 62%](https://gigabit.agency/ar/proof/clinic-llm-assistant-front-desk/)
- [حجز مواعيد الأسنان عبر الإنترنت: حضور المواعيد يرتفع 27%](https://gigabit.agency/ar/proof/dental-online-booking-kept-appointments/)
- [كم تكلفة وكيل الذكاء الاصطناعي؟ البناء والتشغيل](https://gigabit.agency/ar/insights/ai-agent-cost/)
- [وكيل ذكاء اصطناعي أم روبوت محادثة: أيهما؟](https://gigabit.agency/ar/insights/ai-agent-vs-chatbot/)
- [لماذا لا تصل تجارب الذكاء الاصطناعي إلى الإنتاج؟](https://gigabit.agency/ar/insights/why-your-ai-pilot-never-reached-production/)
- [أتمتة بالقواعد أم وكيل ذكاء: متى تكفي؟](https://gigabit.agency/ar/insights/rules-based-automation-vs-ai-agents/)
- [كيف تختارون شركة ذكاء اصطناعي: 9 أسئلة قبل التوقيع](https://gigabit.agency/ar/insights/questions-to-ask-an-ai-agency/)
- [GEO: كيف يجعل ChatGPT و Perplexity يستشهد بكم](https://gigabit.agency/ar/insights/generative-engine-optimization-field-guide/)
- [llms.txt: ما هو والفرق عن robots.txt](https://gigabit.agency/ar/insights/llms-txt-explained/)
- [الظهور في ChatGPT: دليل عملي لبحث ChatGPT](https://gigabit.agency/ar/insights/how-to-rank-in-chatgpt/)
- [قوالب n8n للشركات: ما تصلحونه قبل الإنتاج](https://gigabit.agency/ar/insights/n8n-templates-for-business/)
- [/ar/software-outsourcing/](https://gigabit.agency/ar/software-outsourcing/)

## Insights
- [All insights](https://gigabit.agency/insights/)

### AI Visibility & GEO
- [AI Visibility & GEO: topic hub](https://gigabit.agency/insights/topics/geo-visibility/): When a buyer asks ChatGPT, Perplexity, or AI Overviews, you are named or invisible. How to measure and improve that citation rate.
- [How to rank in ChatGPT: a practical guide to ChatGPT search](https://gigabit.agency/insights/how-to-rank-in-chatgpt/): How to rank in ChatGPT: let OAI-SearchBot crawl you, appear in the results ChatGPT search reads, publish answer-first pages, and measure your citation rate.
- [When ChatGPT and Perplexity disagree](https://gigabit.agency/insights/when-chatgpt-and-perplexity-disagree/): When ChatGPT and Perplexity disagree, a blended citation rate hides a zero on the engine buyers use. Score each assistant. Run the free AI Citation Audit.
- [What gets a company excluded from AI answers](https://gigabit.agency/insights/what-gets-a-company-excluded-from-ai-answers/): Companies get excluded from AI answers when crawlers, published facts, or the brand name conflict. The page can still exist. Run the free AI Citation Audit.
- [How to check your own AI visibility in 30 minutes](https://gigabit.agency/insights/check-your-ai-visibility-in-30-minutes/): Check AI visibility free this afternoon: a spreadsheet, four tabs, what to ask, what to record, and three mistakes that ruin a DIY audit.
- [Your buyers are asking ChatGPT for a shortlist. Are you on it?](https://gigabit.agency/insights/are-you-on-the-ai-shortlist/): B2B AI search visibility: shortlists are drafted by an assistant before anyone calls. You can check in five minutes whether you're on them.
- [llms.txt explained: the new file most sites don't have yet](https://gigabit.agency/insights/llms-txt-explained/): A one-page machine-readable summary of your site, in Markdown, at /llms.txt. What it is, what it isn't, and an honest answer on whether it does anything.
- [Answer-Market Fit: the metric that replaces rankings in the AI era](https://gigabit.agency/insights/answer-market-fit/): Product-market fit means the market wants what you built. Answer-Market Fit means its AI assistants recommend you by default. How to measure and reach it.
- [AI citation rate: what counts as a good one?](https://gigabit.agency/insights/what-is-a-good-ai-citation-rate/): We ran 18 buying questions through Google AI Mode: the most-named firm appeared in 20% of answers. What a good AI citation rate is.
- [How ChatGPT decides which companies to recommend](https://gigabit.agency/insights/how-ai-assistants-choose-recommendations/): When a buyer asks an AI assistant for a vendor, the answer isn't ranked, it's assembled. What the engines retrieve, what they trust, and how to be cited.
- [Generative Engine Optimization: how to get cited by ChatGPT, Claude, and Perplexity](https://gigabit.agency/insights/generative-engine-optimization-field-guide/): Generative engine optimization is how you get cited when search splits into blue links and answer engines. Be the source an LLM quotes, not the one it skips.

### AI Agents & Automation
- [AI Agents & Automation: topic hub](https://gigabit.agency/insights/topics/ai-agents/): A chatbot answers; an agent acts. What a production agent is made of, and which workflow to start with. Agents from $8,000.
- [AI in logistics: where it pays off, and what it needs first](https://gigabit.agency/insights/ai-in-logistics/): AI in logistics pays off on email quotes, order entry, documents, tracking updates and exceptions. Sourced figures, what it needs first, and a decision table.
- [n8n templates for business: which to start from, and what to fix before production](https://gigabit.agency/insights/n8n-templates-for-business/): n8n lists more than 13,000 workflow templates. Which kinds suit a business, how to import one safely, and the eight fixes each one needs for production.
- [What are agentic workflows? A plain definition, six patterns and when to use one](https://gigabit.agency/insights/what-are-agentic-workflows/): Agentic workflows explained: workflow vs agent, six patterns from prompt chaining to tool-using agents, with business examples and when to skip them.
- [AI agent use cases: 20 jobs agents can do, rated for how well they work](https://gigabit.agency/insights/ai-agent-use-cases/): AI agent use cases by function: 20 jobs from support to finance, each with the data it touches, the human checkpoint, the metric and a realism rating.
- [AI agent security framework: the 3Cs and beyond](https://gigabit.agency/insights/ai-agent-security-framework/): Docker's 3Cs (contain, curate, control) checked against OWASP, Google, Meta, NIST and real agent incidents, plus the gaps you still have to close yourself.
- [When a multi-step AI agent fails in production](https://gigabit.agency/insights/when-a-multi-step-ai-agent-fails-in-production/): A multi-step AI agent at 95% per step is about 66% end to end after eight steps. Split or gate it. Book the $25,000, two-week AI Transformation Sprint.
- [Patient intake agents + EHR: the pattern that works](https://gigabit.agency/insights/ai-patient-intake-ehr-integration/): AI patient intake EHR integration that cuts admin time when eligibility, consent, and chart write are one path. Book the $25,000 AI Transformation Sprint.
- [Where AI agents pay off first in ecommerce ops](https://gigabit.agency/insights/where-ai-agents-pay-off-first-in-ecommerce-ops/): AI agents for ecommerce operations pay back first on returns, inventory, and exceptions — with published ROI. Book the $25,000 AI Transformation Sprint.
- [Rules-based automation vs. AI agents: when you need the upgrade](https://gigabit.agency/insights/rules-based-automation-vs-ai-agents/): Your Zapier flows and RPA bots aren't obsolete: they're cheaper and more predictable than an agent for the work they fit. The line is your exception rate.
- [When should an AI agent stop and ask a human?](https://gigabit.agency/insights/ai-agent-human-in-the-loop/): An AI agent should stop when an action is hard to undo, a mistake would go unnoticed, or it isn't confident. Four supervision patterns, and how to choose.
- [Five signs your business is ready for an AI agent](https://gigabit.agency/insights/signs-your-business-is-ready-for-an-ai-agent/): Readiness isn't enthusiasm, it's five specific conditions. The checklist we run before we agree to build, and what to do if you only meet three of them.
- [AI agent vs. chatbot: what's the difference (and which do you need)?](https://gigabit.agency/insights/ai-agent-vs-chatbot/): A chatbot answers; an agent acts. The real distinction — autonomy, tools, and multi-step work — and how to tell which one your workflow actually needs.
- [HIPAA compliant AI: what it takes, which vendors sign a BAA, and what to check](https://gigabit.agency/insights/hipaa-compliant-ai-agents/): HIPAA compliant AI needs a BAA at every hop touching PHI, minimum necessary access, audit logs and human review. Which AI vendors sign BAAs, and a checklist.
- [AI support automation for SaaS: the playbook behind 60%+ autonomous resolution](https://gigabit.agency/insights/ai-support-automation-saas/): How AI support automation SaaS hits majority autonomous resolution without torching CSAT: tier design, escalation, and evals, with real numbers.
- [The production AI agent stack: what we actually deploy](https://gigabit.agency/insights/production-ai-agent-stack/): The six layers a production AI agent stack needs: model, orchestration, retrieval, evals, observability, guardrails, and where teams cut corners.

### Vendor Evaluation
- [Vendor Evaluation: topic hub](https://gigabit.agency/insights/topics/vendor-evaluation/): Pricing traps, diligence questions, and failure patterns before you sign an AI vendor. Compare the options.
- [What is a forward deployed engineer?](https://gigabit.agency/insights/what-is-a-forward-deployed-engineer/): A forward deployed engineer works inside one customer's business and ships software against its real problem. Where the role came from and when to hire one.
- [Composable commerce: what it is, what it costs, and when it pays](https://gigabit.agency/insights/composable-commerce/): Composable commerce builds a store from separate API-connected services. How it differs from headless and MACH, what it costs, and when to stay on one platform.
- [When to hire vs embed AI engineers](https://gigabit.agency/insights/hire-vs-embed-ai-engineers/): When to hire vs embed AI engineers: two senior seats open four months is the signal to stop recruiting. Embedded AI Teams start at $3,000–$7,500/engineer/mo.
- [Vendor Roulette: why switching AI vendors keeps you at square one](https://gigabit.agency/insights/vendor-roulette-switching-ai-vendors/): Third vendor, same line. Why switching AI vendors resets you to zero, what the last engagement should have left, and how to make this the last.
- [Solution looking for a problem: how to spot a doomed AI project](https://gigabit.agency/insights/solution-looking-for-a-problem/): Five tells in the first meeting, before any code. Each is a question you can ask, and the answers say whether an AI project fails or should be funded.
- [In-house, freelancer, or agency: who should build your AI agent](https://gigabit.agency/insights/in-house-freelancer-or-agency-ai/): You've decided to build custom. Hire AI in-house vs agency depends on whether AI is core to your product, and who operates it in year two.
- [Bait-and-Bill: the AI pricing trap to watch for](https://gigabit.agency/insights/bait-and-bill-ai-pricing-trap/): The headline number isn't the price, it's the entry fee for an open-ended meter. The five contract clauses that give the trap away, and what to ask for.
- [Nine questions to ask any AI agency before you sign](https://gigabit.agency/insights/questions-to-ask-an-ai-agency/): The questions that separate a firm that ships from one that demos — what each is really testing, and what a bad answer sounds like. Ask ours the same ones.
- [Software factory, not software agency: how the mid-market ships in the AI era](https://gigabit.agency/insights/software-factory-vs-software-agency/): A software factory is a quality-gated line of senior engineers and AI agents, not bespoke projects, and why the mid-market should hire one.

### Cost & ROI
- [Cost & ROI: topic hub](https://gigabit.agency/insights/topics/cost-roi/): What an AI build costs, what running it costs, and how to model the return before you commit. We publish our prices.
- [n8n pricing in 2026: every plan, and what self-hosting really costs](https://gigabit.agency/insights/n8n-pricing/): n8n pricing in 2026: Cloud from €20 a month for 2,500 executions, Business €667, a free self-hosted edition, and the cost at 5,000 and 50,000 runs.
- [Webflow pricing in 2026: plans, seats, add-ons and what a site really costs](https://gigabit.agency/insights/webflow-pricing/): Webflow pricing in 2026: Premium is $25 a month billed yearly, plus $15 to $39 per editor seat. Every plan, add-on and fee, with three worked examples.
- [Odoo pricing in 2026: plans, hosting, implementation and what a rollout really costs](https://gigabit.agency/insights/odoo-pricing/): Odoo pricing in the US: Standard $31.10 and Custom $61 per user a month billed yearly, 20% off year one. Plus hosting, implementation and three worked budgets.
- [Claude Code pricing in 2026: every plan, API costs and three worked examples](https://gigabit.agency/insights/claude-code-pricing/): Claude Code pricing in 2026: Pro from $17 a month, Max at $100 or $200, Team seats from $20, or pay per token on the API. What a session and a team cost.
- [WordPress to Webflow migration: a step-by-step guide with a redirect map](https://gigabit.agency/insights/wordpress-to-webflow-migration/): WordPress to Webflow migration in 9 steps: inventory, CMS design, CSV import, a copyable 301 redirect map, SEO carry-over and a cutover checklist.
- [What does an ERP implementation cost in 2026? A real budget breakdown](https://gigabit.agency/insights/erp-implementation-cost/): What a mid-market ERP implementation costs, why the licence is the cheapest line, and where budgets blow up. From a firm that publishes its own prices.
- [AI for consulting firms: start with the archive](https://gigabit.agency/insights/ai-for-consulting-firms/): AI for consulting firms should start with the archive, not a proposal bot. One firm got 1,200 hours a quarter back. Book the $25,000 AI Transformation Sprint.
- [What does an AI MVP actually cost? A build budget breakdown](https://gigabit.agency/insights/ai-mvp-cost/): Four cost centres, and the model is the smallest. What an AI MVP cost covers to reach real users, and the line most budgets forget.
- [How to calculate ROI on an AI project before you build it](https://gigabit.agency/insights/calculate-ai-project-roi/): An AI project ROI calculator is simple arithmetic. The hard part is four inputs, and three are inflated by whoever builds the case.
- [What should a mid-market company budget for AI in 2026?](https://gigabit.agency/insights/mid-market-ai-budget/): Not what one agent costs — the whole line item. How to size a mid-market AI budget across audit, build, and operate, and the split that survives year two.
- [The hidden costs of AI projects nobody puts in the proposal](https://gigabit.agency/insights/hidden-costs-of-ai-projects/): Not the vendor's tricks, your own. Seven internal line items that never appear in a quote, why the biggest is your team's time, and how to price them.
- [How much does generative engine optimization (GEO) cost?](https://gigabit.agency/insights/how-much-does-geo-cost/): GEO has no standard sticker price. The three ways it is priced, what drives the number, what agencies charge, and our own: Ascent from $4,000 a month.
- [How much does an AI agent cost in 2026? A real budget breakdown](https://gigabit.agency/insights/ai-agent-cost/): What an AI agent cost covers to build and run, why the token bill is the cheapest line, and where teams overspend. From a firm that publishes its prices.

### Strategy & Readiness
- [Strategy & Readiness: topic hub](https://gigabit.agency/insights/topics/strategy-readiness/): Most AI projects are decided before engineering starts. How to tell if you are ready, and what to fix first.
- [AI policy template: a company AI acceptable use policy you can copy](https://gigabit.agency/insights/ai-policy-template/): A free AI policy template: a 12-section company AI acceptable use policy with data rules and human review, mapped to NIST AI RMF, ISO 42001 and the EU AI Act.
- [AI security review vs a pentest: what each one covers](https://gigabit.agency/insights/what-an-ai-security-assessment-covers/): A network pentest skips prompt injection, RAG poisoning and agent tool misuse. What an AI red team tests instead, mapped to OWASP LLM Top 10 questionnaire rows.
- [Calibrate an LLM-as-judge for AI agents](https://gigabit.agency/insights/calibrate-llm-as-judge-for-ai-agents/): LLM-as-judge for AI agents is usable only after a human-labeled holdout calibration. Trust auto-scores then. Book the $25,000 AI Transformation Sprint.
- [Build an AI agent eval set from real tickets](https://gigabit.agency/insights/build-ai-agent-eval-set-from-tickets/): Build an AI agent eval set from real tickets. Stratify volume and failures, label correct answers, gate deploys. Book the $25,000 AI Transformation Sprint.
- [What to measure in an AI agent eval suite](https://gigabit.agency/insights/ai-agent-eval-metrics/): AI agent eval metrics that should gate every deploy: golden-set size, accuracy thresholds, confidence gates. Book the $25,000 AI Transformation Sprint.
- [What is an AI readiness assessment, and do you actually need one?](https://gigabit.agency/insights/what-is-an-ai-readiness-assessment/): An AI readiness assessment scores your organisation: whether you can ship, what to fix first, what a good one contains, and when to skip it.
- [Why your AI pilot never reached production — and the five gates that get it there](https://gigabit.agency/insights/why-your-ai-pilot-never-reached-production/): Pilot purgatory is an engineering problem, not an ambition problem. Here are the eval, ownership, and rollback gates that separate a demo from a deployed agent.

### Operate
- [Operate: topic hub](https://gigabit.agency/insights/topics/operate/): AI systems decay rather than crash. What breaks after go-live, and what operating one properly costs. From $3,000/mo.
- [The security checklist to run before real users reach your AI-built app](https://gigabit.agency/insights/ai-built-app-security-checklist/): Vibe coding security checklist for Lovable, Bolt, Replit, Cursor and Claude Code apps: 13 checks from Supabase RLS to AI spend caps, each with a 10-minute test.
- [AI agent observability tools: what to capture, and which tool fits your team](https://gigabit.agency/insights/ai-agent-observability-tools/): AI agent observability tools compared: Langfuse, LangSmith, Phoenix, Helicone, Braintrust, Weave, Datadog and OpenTelemetry, priced on 23 Oct 2026.
- [Why AI systems break after launch — and what "operating" one means](https://gigabit.agency/insights/why-ai-systems-break-after-launch/): AI systems don't crash. They decay while dashboards stay green. Five ways a live system degrades, and what AI system maintenance actually is.
- [What an AI operations runbook actually contains](https://gigabit.agency/insights/ai-operations-runbook/): Seven entries, each a question someone asks at an inconvenient hour. The runbook that turns a live AI system into something a person can be on call for.
- [How to hand over an AI system — to a new vendor or in-house](https://gigabit.agency/insights/ai-system-handover/): AI system handover: the inventory to collect, the test that proves it worked, and the four things that always go missing.

## Decision guides
- [All comparisons](https://gigabit.agency/compare/)
- [Gigabit vs consultancies](https://gigabit.agency/compare/gigabit-vs-consultancies/): Cost, speed, and lock-in compared with real numbers
- [Gigabit vs automation agencies](https://gigabit.agency/compare/gigabit-vs-automation-agencies/): Where an AI-native software factory differs from an automation agency
- [Build vs buy vs embed](https://gigabit.agency/compare/build-vs-buy-vs-embed/): Which path fits your team
- [Embedded team vs staff augmentation vs in-house hire](https://gigabit.agency/compare/embedded-vs-staff-augmentation/): Four ways to add AI engineers on cost, time to first shipped change and who owns the system in 18 months
- [Odoo vs Zoho vs ERPNext](https://gigabit.agency/compare/odoo-vs-zoho-vs-erpnext/): Odoo, Zoho One and ERPNext on licence, published price, hosting, manufacturing, customisation and built-in AI, from each vendor's own pages. Checked 10 October 2026
- [Fixed-price AI pilots compared](https://gigabit.agency/compare/fixed-price-ai-pilots/): Published AI pilot, proof-of-concept and discovery-sprint offers from 13 firms side by side: price, duration, what runs at the end, what happens to the fee. Checked 10 October 2026
- [Custom support agent vs Intercom Fin](https://gigabit.agency/compare/custom-support-agent-vs-intercom-fin/): Intercom Fin ($0.99 per outcome) and Zendesk AI agents ($1.50–$2.00 per automated resolution) against a custom agent (from $8,000 build, from $3,000/mo to run): the breakeven formula and a five-volume cost table. Checked 22 October 2026
- [AI agencies that publish their prices](https://gigabit.agency/compare/ai-agencies-with-published-prices/): 16 AI development agencies whose own sites show a price for a defined scope (Gigabit among them, listed alphabetically), plus 9 that publish only ranges, minimums or cost guides. Read on each firm's own site. Checked 22 October 2026
- [Webflow vs WordPress](https://gigabit.agency/compare/webflow-vs-wordpress/): Webflow and WordPress on licence, published price (Webflow Basic $15 and Premium $25 a month billed yearly; WordPress free software plus hosting), editor seats, hosting, code export, plugins, languages and built-in AI, from each vendor's own pages. Gigabit builds on both, from $6,000 fixed. Checked 22 October 2026
- [Shopify vs WooCommerce](https://gigabit.agency/compare/shopify-vs-woocommerce/): Shopify (US: Basic $19, Grow $49, Advanced $299 a month billed yearly; Plus from $2,300) and WooCommerce (free plugin, paid hosting and extensions) on payment fees, hosting, PCI, checkout, catalog limits, B2B and built-in AI, from each vendor's own pages. Gigabit builds stores on both, from $9,000 fixed. Checked 22 October 2026
- [HubSpot vs Salesforce](https://gigabit.agency/compare/hubspot-vs-salesforce/): HubSpot vs Salesforce on seat prices, onboarding fees, contracts, customisation, AI agents and API limits, with 10- and 50-seat costs. Checked 23 Oct 2026. Checked 23 October 2026
- [n8n vs Zapier vs Make](https://gigabit.agency/compare/n8n-vs-zapier-vs-make/): n8n bills per workflow run, Zapier per task, Make per credit. Prices at 5,000 and 50,000 runs, self-hosting, licence, AI and errors, checked 23 Oct 2026. Checked 23 October 2026
- [Framer vs Webflow](https://gigabit.agency/compare/framer-vs-webflow/): Framer vs Webflow on price per site, editor seats, CMS limits, languages, export, SEO, AI and ecommerce, read from each vendor's own pages in October 2026. Checked 23 October 2026
- [Shopify vs Shopify Plus](https://gigabit.agency/compare/shopify-vs-shopify-plus/): Shopify Plus costs $2,300/mo on a 3-year term ($2,500 on 1 year) vs $299 for Advanced. Fees, checkout, B2B, POS and when Plus pays. Checked 23 October 2026
- [Odoo vs NetSuite](https://gigabit.agency/compare/odoo-vs-netsuite/): Odoo vs NetSuite: Odoo publishes $31.10 and $61 per user; NetSuite quotes only. Hosting, modules, customisation, multi-company and exit compared. Checked 23 October 2026
- [Claude Code vs Cursor](https://gigabit.agency/compare/claude-code-vs-cursor/): Claude Code vs Cursor on pricing, usage limits, models, agent autonomy, privacy and team admin, with what one developer and a team of five pay. Oct 2026. Checked 23 October 2026
- [Lovable vs Bolt vs Replit](https://gigabit.agency/compare/lovable-vs-bolt-vs-replit/): Replit vs Lovable vs Bolt on pricing, credits, stacks, Supabase, auth, hosting, GitHub sync and security scans, plus what breaks first. Checked 23 Oct 2026. Checked 23 October 2026

## Research
- [Research](https://gigabit.agency/research/): Original studies with method and CSV data
- [Platform price tracker](https://gigabit.agency/research/platform-price-tracker/): List prices for 45 plans across n8n, Zapier, Make, Webflow, Framer, HubSpot, Salesforce, Odoo, NetSuite, Shopify, Claude Code, Cursor, Lovable, Bolt and Replit, read on vendor pages on 23 October 2026, CSV under CC BY 4.0, updated monthly
- [AI services price index, October 2026](https://gigabit.agency/research/ai-services-price-index/): What 122 AI service firms publish about price on their own sites, read 22 October 2026. 50% publish no price, 27% only ranges or rates, 23% a fixed or "from" price for a defined offer; none of 13 large IT and consulting firms publishes one. Median entry prices: discovery $4,500, pilot $12,000, agent build $20,000, retainer $2,500 a month, hourly $30. Gigabit is the author and is not counted. Method, sources and CSV data under CC BY 4.0.

## Evidence
- [Proof](https://gigabit.agency/proof/): 13 case studies with investment, return, and timeline on each
- [From Pitch Deck to Paying Customers in 11 Weeks: A Logistics SaaS MVP](https://gigabit.agency/proof/saas-mvp-logistics-platform/): Pre-seed logistics technology startup, US-based solo founder. Investment $48,000; Return $2.1M Series A closed four months after launch; Timeline 11 weeks to production. A $48K, 11-week logistics SaaS MVP that reached 14 paying customers and unlocked a $2.1M Series A.
- [Zero-Downtime Migration: Heroku to AWS While Processing $14M/Month](https://gigabit.agency/proof/cloud-migration-fintech-zero-downtime/): Series B payments platform, US-based, $14M monthly transaction volume. Investment $74,000; Return $60,000+ annual infrastructure savings, growing with volume; Timeline 14 weeks (2-week assessment + 12-week migration). A 14-week strangler-fig migration off Heroku with zero downtime: -35% infra cost, -93% deploy time, +800% deploy frequency.
- [How an AI Support Agent Saved a D2C Brand $218K Per Year — and Made Their Customers Happier](https://gigabit.agency/proof/ai-automation-ecommerce-support/): Direct-to-consumer home goods brand on Shopify Plus, $12M annual revenue. Investment $92,000 (Year 1 total); Return $217,200 Year 1 savings — 136% net ROI; Timeline 10 weeks (2-week assessment + 8-week build). A confidence-gated AI support agent now resolves 64% of tickets autonomously, cutting support costs 62% — $217,200 saved in year one.
- [Telehealth Platform: FHIR Integration, HIPAA Compliance, 1,200 Patients in 90 Days](https://gigabit.agency/proof/healthtech-telehealth-platform/): Physician-founded chronic disease management company (Type 2 diabetes and hypertension). Investment $124,000; Return $4.5M Series A raised 6 months post-launch; 3.6x coordinator capacity; Timeline 5 months — launched 1 week ahead of deadline. HIPAA-compliant telehealth platform with 3 EHR integrations enrolled 1,200 patients in 90 days and backed a $4.5M Series A.
- [How 3 Embedded Engineers Doubled Sprint Velocity for an HR Tech SaaS](https://gigabit.agency/proof/team-augmentation-hrtech/): Series B HR technology SaaS, US-based, $8M ARR. Investment $14,400/month (3 engineers); Return $392,400 annual savings vs US hiring (65%); Timeline First PRs in 5 days; full sprint velocity in 3 weeks. Three embedded engineers lifted sprint velocity 2.3x and shipped 4 stalled features in 3 months — saving $392,400/year vs US hiring.
- [32 Hours Per Week Recovered: AI Workflow Automation for a Multi-Channel E-Commerce Operation](https://gigabit.agency/proof/ai-workflow-automation-ecommerce-ops/): Multi-channel e-commerce brand, $28M annual revenue, 65 employees. Investment $74,000 (Year 1 total; $44,000 build); Return ~$87,000 Year 1 impact; Timeline 8 weeks (1-week assessment + 7-week build). Five n8n + GPT-4o-mini automations cut a 4-channel ops team's manual workload from 32 to 4.5 hours/week — ~$87K in year-one impact.
- [Two-Sided Marketplace: From Concept to $2.4M GMV in Year 1](https://gigabit.agency/proof/marketplace-platform-build/): B2B services marketplace startup, US-based, pre-seed ($500K). Investment $72,000; Return $2.4M GMV and $360K commission revenue in Year 1; Timeline 14 weeks to launch. A $72K, 14-week two-sided marketplace build processed $2.4M GMV in year one and powered a $3.8M seed round.
- [AI Transformation: 1,200 Hours Per Quarter Recovered for a Management Consulting Firm](https://gigabit.agency/proof/ai-transformation-professional-services/): Management consulting firm, US-based, 140 employees, $32M annual revenue. Investment $210,000 (Year 1 total; $156,000 build); Return $480,000/year in recovered value — 129% Year 1 ROI; Timeline 6 months across 3 phases. A phased AI transformation recovered 1,200 consultant-hours per quarter and $480K/year for a 140-person consulting firm — 129% Year 1 ROI.
- [B2B Payment Platform: From Concept to $41M Processed in Year 1](https://gigabit.agency/proof/fintech-payment-platform/): B2B construction payments startup, US-based, seed-funded ($2.1M). Investment $118,000; Return $41.2M processed in Year 1; $6.2M Series A raised; Timeline 4 months to MVP launch. A construction payments platform built in 4 months for $118K processed $41.2M in year one with a 0.02% error rate.
- [From Manual Spreadsheets to Self-Serve Analytics: A SaaS Data Engineering Build](https://gigabit.agency/proof/data-engineering-saas-analytics/): B2B project management SaaS, $6M ARR, 85 employees. Investment $78,200 (Year 1 total; $44,000 build); Return $270,000+ Year 1 value; Timeline 7 weeks (1-week assessment + 6-week build). A 7-week modern data stack eliminated 56 hours/month of manual reporting, cut month-end close from 5 days to 2, and preserved $186K in ARR.
- [Regional Provider Network — Unified Intake, EHR Integration & Analytics](https://gigabit.agency/proof/healthcare-unified-intake-ehr-analytics/): Multi-site medical group (primary + specialty care), 14 locations. Timeline 12 weeks. Unified HIPAA-aware intake with EHR integration cut intake/admin time 41% and lifted clinician throughput 16% across 14 locations.
- [Specialty Clinic — LLM Patient Assistant Reduces Front-Desk Load](https://gigabit.agency/proof/clinic-llm-assistant-front-desk/): Single-specialty dermatology clinic — ~30 staff, 8 providers, 2 locations. Timeline 8 weeks. A RAG patient assistant with clinical guardrails cut repetitive front-desk inquiries 62% and time-to-answer 68% in 8 weeks.
- [Multi-Location Dental Group — Self-Serve Booking = More Kept Appointments](https://gigabit.agency/proof/dental-online-booking-kept-appointments/): General + cosmetic dentistry group — 6 locations, 45 staff, 12 providers. Timeline 9 weeks. HIPAA-aware self-serve booking with two-way reminders lifted kept appointments 27% and cut no-shows 29% across 6 dental locations.
- [What client projects cost](https://gigabit.agency/proof/project-costs/): One table of client projects with one-off cost, monthly cost after launch, timeline, result and payback
- [Approach](https://gigabit.agency/approach/): The Audit → Build → Operate method
- [About](https://gigabit.agency/about/): Who runs Gigabit, where delivery happens, and the team

# Case studies

## From Pitch Deck to Paying Customers in 11 Weeks: A Logistics SaaS MVP
URL: https://gigabit.agency/proof/saas-mvp-logistics-platform/
Client: Pre-seed logistics technology startup, US-based solo founder
Investment: $48,000 · Return: $2.1M Series A closed four months after launch · Timeline: 11 weeks to production

- Total investment: $48,000
- Time to production: 11 weeks
- Paying customers (month 6): 14
- MRR at month 6: $19,000
- Series A raised: $2.1M at $6.2M pre-money

### Challenge
The founder — a 15-year freight logistics veteran — had spotted a gap. Mid-size freight brokers handling 50–500 shipments a month were running operations on 2000s-era TMS software, Excel, email chains, and phone calls. The alternatives were enterprise platforms with 6-figure annual contracts and 6-month implementations, or basic tools that couldn't handle multi-modal freight coordination.
He had $180K in pre-seed funding, LOIs from three pilot customers, and a Series A investor whose commitment was contingent on a working product with live traction within 4 months. The constraints were hard: $50K maximum for MVP development, 12 weeks to a deployable product, and pilot customers committed to a 60-day evaluation starting no later than week 14.
The feature wishlist was 47 items long. It had to be cut to the 12 features that would make a broker switch off their spreadsheet. Three US development shops had quoted $120K–$220K with 16–24 week timelines — numbers that would have consumed the entire raise.

### Approach
Sprint Zero locked scope and architecture in one week for $4,500. The build ran as four 2-week sprints: foundation (auth, tenant management, roles, CI/CD), core workflow (shipment creation, multi-stop routing with Mapbox, carrier assignment, document upload), intelligence (real-time tracking on a map, automated notifications, KPI dashboard), then monetization and polish (Stripe per-shipment billing, customer portal, reporting, CSV import).
The cadence was strict: Monday sprint planning with the founder on video, daily async standups across the Dhaka–US time gap, and a Friday demo on staging where the founder and a pilot customer tested live. Feedback fed directly into the next sprint.
Weeks 10–11 were testing and launch. A QA engineer ran 180 test cases covering multi-tenant isolation, payment failure handling, and concurrent editing with optimistic locking. Load testing validated 50 concurrent users with 500 active shipments at under 200ms response times, with headroom for 200+ concurrent users. Production deployment used a zero-downtime ECS rolling update — the founder onboarded his first pilot customer the same afternoon.

### Results
The platform shipped in 11 weeks — one week ahead of the 12-week target — for $48,000, under the $50K budget. All 12 committed MVP features delivered, zero critical bugs at launch, 78% automated test coverage, and $180/month infrastructure costs at launch.
Traction compounded: 3 pilot customers in month 1 grew to 7 by month 3 and 14 by month 6. MRR went from $1,700 to $19,000 over the same period, with NPS holding between +58 and +64 and uptime at 99.97%+.
Four months after launch the founder closed a $2.1M Series A at a $6.2M pre-money valuation. The lead investor cited product quality, development speed, and capital efficiency — and audited the codebase. After the raise, the founder retained 3 Gigabit engineers as his dedicated development team at $14,500/month.

## Zero-Downtime Migration: Heroku to AWS While Processing $14M/Month
URL: https://gigabit.agency/proof/cloud-migration-fintech-zero-downtime/
Client: Series B payments platform, US-based, $14M monthly transaction volume
Investment: $74,000 · Return: $60,000+ annual infrastructure savings, growing with volume · Timeline: 14 weeks (2-week assessment + 12-week migration)

- Downtime during migration: Zero
- Deployment time: 38 min → 2.5 min (-93%)
- Deployment frequency: 2×/week → 18×/week (+800%)
- Infrastructure cost: $14,200 → $9,230/month (-35%)
- Payment p99 latency: 340ms → 145ms (-57%)

### Challenge
The client processes roughly $14M in monthly transaction volume across 180,000+ transactions for approximately 900 SMB merchants in the US and Canada. The entire platform ran as a monolithic Ruby on Rails application on Heroku — and the architecture had become the primary constraint on the business.
The pain was quantifiable. Every deploy was a 38-minute full restart, allowed only twice a week inside a maintenance window; one failed deployment had caused a 47-minute outage during peak hours and roughly $62,000 in failed transactions. Heroku couldn't scale components independently, so month-end reporting spikes forced the whole application to scale — costs had hit $14,200/month and were climbing 8–12% monthly. And a single bug anywhere could take down everything: a formatting error in reporting once crashed the payment queue for 23 minutes, costing two enterprise prospects.
The CTO's constraint was absolute: no migration window, no merchant-visible disruption. The migration had to happen underneath live payment traffic.

### Approach
A $6,000 infrastructure assessment came first: review of the ~85,000-line Rails codebase, mapping of the 47-table schema (12 tables with >1M rows), and request tracing that showed reporting queries consuming 60% of database CPU at month-end while the payment path itself ran at p99 340ms. We recommended the strangler fig pattern — extract services one at a time, run old and new in parallel, shift traffic gradually — with extraction order ranked by risk: notifications, reporting, onboarding, webhooks, and the payment core last.
Weeks 3–4 built the foundation before any extraction: every AWS resource in Terraform, GitHub Actions CI/CD with security scanning and blue-green deploys to EKS, and a Datadog observability stack deployed up front to catch discrepancies between old and new systems within seconds.
Low-risk extractions ran weeks 5–8. The notification service ran in parallel for 5 days with a 0.00% discrepancy rate before cutover. Moving reporting to a read replica dropped primary database CPU from 78% to 31% and improved payment p99 latency from 340ms to 180ms as a side effect. The payment core migrated last with maximum caution: 48 hours of shadow traffic across 12,000 transactions at 0.00% discrepancy, then a progressive shift from 1% to 100% of traffic over 10 days with a tested rollback plan at every step. Week 14 closed with runbooks, knowledge transfer, and a 3-hour failure-injection game day.

### Results
Measured over the first 90 days post-migration versus the 90 days prior: deployment time fell from 38 minutes to 2.5 minutes per service (-93%), deployment frequency rose from 2×/week to 18×/week (+800%), monthly infrastructure cost dropped from $14,200 to $9,230 (-35%), payment p99 latency improved from 340ms to 145ms (-57%), and month-end reporting queries went from 12–45 seconds to 2–8 seconds (-82%). Unplanned downtime: 70 minutes before, zero after. Downtime during the migration itself: zero.
At the company's 15% quarterly volume growth, Heroku costs would have reached $22,000/month by year end; the AWS architecture projects to $11,500/month at the same volume — a growing annual savings of $120,000+. Payback on the $74,000 migration is 14.8 months on infrastructure savings alone, before counting the eliminated $62K-per-incident downtime risk.
The client retained a Gigabit DevOps engineer on an ongoing SRE retainer at $5,500/month, with multi-region deployment, a real-time fraud detection pipeline, and SOC 2 Type II infrastructure on the current roadmap.

## How an AI Support Agent Saved a D2C Brand $218K Per Year — and Made Their Customers Happier
URL: https://gigabit.agency/proof/ai-automation-ecommerce-support/
Client: Direct-to-consumer home goods brand on Shopify Plus, $12M annual revenue
Investment: $92,000 (Year 1 total) · Return: $217,200 Year 1 savings — 136% net ROI · Timeline: 10 weeks (2-week assessment + 8-week build)

- Annual savings: $217,200
- Tickets resolved by AI: 3,149/month (64%)
- First response time: 4.2 hours → 6 minutes
- CSAT: 3.9 → 4.5
- Payback period: 5.1 months

### Challenge
At $12M in annual revenue and 40% year-over-year growth, the brand was winning on product and marketing while support broke underneath. Six agents handled roughly 4,800 tickets a month through Gorgias. Average first response time had ballooned to 4.2 hours — spiking past 12 hours during launches — and CSAT had dropped from 4.6 to 3.9 over six months.
Ticket analysis made the root cause obvious. About 70% of volume fell into predictable categories: order status (23%), sizing and compatibility (18%), returns and exchanges (16%), discount codes (8%), and shipping policy (5%). Each rules-based inquiry still consumed 6–8 minutes of agent time, while the 1,500 monthly tickets that genuinely needed a human sat in queue.
The founder put it bluntly: six people at $58K each were answering the same 15 questions 3,000 times a month. A rule-based Gorgias chatbot had been tried — it handled about 8% of tickets before customers bypassed it.

### Approach
An $8,000 AI readiness assessment came first: classification of 28,400 tickets from 6 months of Gorgias history, an audit of every knowledge source (420-SKU Shopify catalog, policies, sizing guides, ShipStation tracking data), and an architecture evaluation. Fine-tuning was rejected — it would bake in the inconsistencies in historical agent answers. A RAG pipeline with GPT-4o and tool use won: the agent grounds every answer in current product, policy, and order data. The assessment projected 60–70% of tickets automatable and $180K–$240K in annual savings.
The build put product, policy, and order knowledge behind semantic retrieval (Pinecone, with nightly Shopify sync) and wired secure API access to Shopify and ShipStation for real-time order lookup. The agent is confidence-gated: it replies autonomously only when intent classification confidence exceeds 0.85 and retrieval relevance exceeds 0.80 — below either threshold the ticket routes to a human with the AI's draft and context pre-loaded. It can also act: initiate returns, generate labels, apply discount codes, and update addresses, each with customer confirmation.
Testing against 2,000 historical tickets with known resolutions produced 94.2% intent classification accuracy, 91.7% response accuracy, 99.1% policy compliance, and a 4.3/5.0 tone match against the team's best agent. Rollout was staged over two weeks — 10% of tickets with full human review, then 30% with flagged-only review, then 100% of eligible tickets autonomous.

### Results
Over the first 90 days of full deployment versus the prior 90 days: the AI resolved 3,149 tickets a month (64%) with no human involvement, tickets requiring a human fell 63% to 1,771/month, and first response time dropped from 4.2 hours to 6 minutes for AI replies and 1.8 hours for human ones. CSAT recovered from 3.9 to 4.5.
Monthly support operating cost fell from $29,200 to $11,100 — a 62% reduction worth $217,200 annualized. Against $92,000 in Year 1 investment ($8,000 assessment, $42,000 build, $3,500/month operation), net Year 1 ROI was $125,200 (136%) with a 5.1-month payback.
Three agents were reassigned to proactive outreach, VIP management, and post-purchase experience — contributing to a 14% lift in repeat purchase rate the following quarter. During a sale event at 2x normal ticket volume, the AI held a 4-minute average response time; the all-human team had averaged 14 hours in comparable events.

## Telehealth Platform: FHIR Integration, HIPAA Compliance, 1,200 Patients in 90 Days
URL: https://gigabit.agency/proof/healthtech-telehealth-platform/
Client: Physician-founded chronic disease management company (Type 2 diabetes and hypertension)
Investment: $124,000 · Return: $4.5M Series A raised 6 months post-launch; 3.6x coordinator capacity · Timeline: 5 months — launched 1 week ahead of deadline

- Patients enrolled (90 days): 1,200
- Vitals adherence: 34% → 78%
- Patients per coordinator: 50 → 180 (3.6x)
- EHR systems integrated: 3 (Epic, Cerner, athenahealth)
- Series A raised: $4.5M

### Challenge
The founder — a physician with 20 years of clinical experience — was running a chronic-care pilot on Zoom, spreadsheets, and manual phone calls. Clinical outcomes were strong: pilot patients showed a 1.8-point average A1C reduction over 6 months. But the manual workflow capped each care coordinator at 50 patients.
The product requirements were clinically specific: a patient mobile app for daily vitals logging, secure messaging, and video visits; a provider dashboard with risk stratification and configurable vitals alerts (e.g., blood glucose >300 mg/dL triggers immediate notification); bidirectional EHR integration with Epic, Cerner, and athenahealth via SMART on FHIR; and HIPAA compliance throughout — encryption, access controls, audit logging, BAA-covered infrastructure.
The budget was $130K maximum, because the seed round also had to fund 12 months of operations, clinical staff, and regulatory work. The timeline was 5 months, fixed by a health system contract that required a live platform by a specific date.

### Approach
Month 1 built the foundation. Two weeks of clinical workflow mapping — watching coordinators work with pilot patients — surfaced unstated requirements like 6-month vitals trends at a glance and regimen-specific medication reminders. The HIPAA-compliant AWS infrastructure went up before application code: private-subnet VPC, encrypted RDS PostgreSQL, encrypted S3, CloudTrail audit logging, WAF, field-level PHI encryption, RBAC, 15-minute session timeouts, and MFA for all provider accounts.
Months 2–3 delivered the core platform. The React Native patient app used a large-button, high-contrast interface for an older population — 18px minimum fonts, 48px touch targets, vitals entry optimized to complete in under 60 seconds. The provider dashboard added color-coded risk panels, 30/90/180-day vitals trends, and per-patient alert thresholds with audited acknowledgment.
Month 4 was the FHIR work — the most technically complex phase. OAuth2 SMART on FHIR authorization per EHR, patient matching across systems, pulls of demographics, problem lists, medications, and labs, and pushes of coordinator notes and vitals back to the EHR. Because Epic, Cerner, and athenahealth each implement FHIR R4 differently, a normalization layer mapped every EHR's output to one internal data model. Month 5 covered testing (alert thresholds verified against 500 synthetic patient datasets, every violation alerting within 30 seconds), a third-party security assessment with two medium findings remediated pre-launch, and full compliance documentation.

### Results
The platform launched 1 week ahead of the 5-month deadline, fulfilling the health system contract on time. Within 90 days, enrollment grew from 20 migrated pilot patients to 1,200. Daily vitals adherence rose from 34% in the manual pilot to 78%, coordinator response time fell from 4.2 hours to 22 minutes, and critical vitals alerts were acknowledged in 4.2 minutes on average. Uptime: 99.96%. Security incidents: zero.
Patient capacity per coordinator increased 3.6x — from 50 to 180 — fundamentally changing the unit economics of the care model. The founder used the traction data (1,200 patients, 78% adherence, 1.8-point A1C reduction) to raise a $4.5M Series A 6 months post-launch, and two additional health systems signed contracts on the strength of the clinical outcomes.
Total build investment was $124,000 across four phases, with ongoing maintenance and support continuing at $5,500/month.

## How 3 Embedded Engineers Doubled Sprint Velocity for an HR Tech SaaS
URL: https://gigabit.agency/proof/team-augmentation-hrtech/
Client: Series B HR technology SaaS, US-based, $8M ARR
Investment: $14,400/month (3 engineers) · Return: $392,400 annual savings vs US hiring (65%) · Timeline: First PRs in 5 days; full sprint velocity in 3 weeks

- Sprint velocity: 2.3x (34 → 78 points)
- Annual savings vs US hiring: $392,400 (65%)
- Test coverage: 34% → 71%
- Time to first PR: 5 days
- Enterprise deals influenced: 12

### Challenge
The client's workforce management platform serves 600+ mid-market companies. At $8M ARR and 45% year-over-year growth, the 7-engineer internal team was stretched across too many priorities — 14 months of P0/P1 roadmap against a velocity that would ship only 8 months of it.
Hiring had failed. Two senior positions sat open for 4 months: 200+ applications, 40 phone screens, 3 offers extended — all 3 lost to counter-offers or competitors. Fully-loaded cost per hire in their market: $195K–$220K/year. They needed capacity immediately, not in 6 months.

### Approach
Week 1 was team assembly. Against the VP of Engineering's spec — 2 senior full-stack engineers (React, TypeScript, NestJS, PostgreSQL) plus 1 QA automation engineer (Cypress) — we presented 5 candidate profiles within 4 business days. He interviewed all 5 in 30-minute architecture conversations and selected 3.
Week 2 was Sprint Zero: access provisioned on days 1–2, then guided codebase exploration where each engineer traced a core workflow and wrote a technical summary — forcing deep understanding in 48 hours and producing documentation the internal team had never written. First PRs were submitted by end of day Friday.
From week 3 the engineers ran inside the client's own cadence: biweekly sprints, the same Linear board, repos, and Slack channels. Async daily updates posted at 5:30 PM Dhaka landed at 7:30 AM Eastern; sync standups ran twice weekly in the 2-hour overlap window; PRs were reviewed within 24 hours across time zones.

### Results
Over the first 3 months versus the prior 3-month average: story points per sprint rose from 34 to 78 (+129%, 2.3x), tickets completed from 18 to 41, PRs merged from 22 to 52, and sprint commitment hit rate from 72% to 91%. The QA engineer wrote 240+ Cypress tests in 2 months, lifting coverage from 34% to 71% and cutting regression bugs 60%.
Four features stalled on the roadmap for 6+ months shipped in the first quarter: a compensation benchmarking module, a drag-and-drop custom workflow builder, an advanced reporting engine (replacing static reports that drove 30% of support tickets), and SAML SSO with directory sync — required by 12 enterprise prospects in active sales cycles.
The financials: $207,600 total annual cost for the Gigabit team versus $600,000 for the US-equivalent hires — $392,400 (65%) saved annually, with productivity in 2 weeks instead of a 4–6 month hiring cycle. The engagement has been extended indefinitely; the client expanded from 3 to 5 engineers and plans 2 more in Q3 2026 ahead of a Series C.

## 32 Hours Per Week Recovered: AI Workflow Automation for a Multi-Channel E-Commerce Operation
URL: https://gigabit.agency/proof/ai-workflow-automation-ecommerce-ops/
Client: Multi-channel e-commerce brand, $28M annual revenue, 65 employees
Investment: $74,000 (Year 1 total; $44,000 build) · Return: ~$87,000 Year 1 impact · Timeline: 8 weeks (1-week assessment + 7-week build)

- Weekly manual hours: 32 → 4.5 (-86%)
- Returns processed autonomously: 85%
- Inventory accuracy: 94.2% → 99.1%
- Exception resolution time: 25 min → 6 min
- Year 1 impact: ~$87,000

### Challenge
The client sells consumer electronics through four channels — Shopify Plus (55% of revenue), Amazon (25%), Walmart Marketplace (12%), and B2B wholesale (8%) — with an 8-person operations team handling fulfillment, returns, inventory, vendor communication, and reporting across all of them. Nothing was catastrophically broken; everything was 10 minutes more manual than it should be, multiplied by 300 orders a day.
The assessment quantified it: returns processing consumed 9 hours/week at 12–18 minutes per return; daily inventory reconciliation across 4 systems took 7 hours/week; vendor follow-up another 5; manual reporting 6 (and the reports were 2 hours stale on delivery); order exception handling 5 more. In total, 32+ hours per week of automatable work costing roughly $78,000 a year in direct labor.

### Approach
A 1-week, $6,000 AI readiness assessment mapped every workflow across the 4 channels, interviewed 5 team members, and ranked automations by ROI: returns first (highest volume, most standardized), then reporting, inventory reconciliation, vendor follow-up, and order exception routing.
The system was built on self-hosted n8n as the orchestration engine, custom Python for complex logic, and GPT-4o-mini for classification and communication tasks. Returns automation (live week 3) validates eligibility, creates NetSuite records, generates ShipStation labels, emails confirmations via Klaviyo, and auto-processes refunds on delivery — with GPT-4o-mini classifying free-text return reasons and flagging potential defects by SKU. The reporting suite (week 4) delivers four automated reports with AI-generated narrative analysis before anyone reaches the office.
Nightly inventory reconciliation (week 5) auto-corrects known discrepancy causes and surfaces only true exceptions. Vendor follow-up automation (week 6) sends POs, has the AI parse wildly inconsistent vendor reply emails into structured ship dates and tracking numbers, and chases delays automatically. The exception router (week 7) classifies order exceptions and preps one-click resolution actions — deliberately keeping a human approval in the loop because the stakes are too high for fully autonomous action.

### Results
Over the first 60 days of full deployment: weekly hours on the automated processes fell from 32 to 4.5 (-86%), per-return handling from 12–18 minutes to zero for the 85% of returns that process autonomously, daily reports moved from 10:30 AM to 7:00 AM delivery, inventory accuracy rose from 94.2% to 99.1%, vendor PO follow-up dropped from 5 hours/week to 45 minutes, and exception resolution time fell from 25 to 6 minutes.
The financial impact totals roughly $87,000/year: $40,040 in direct labor savings (27.5 hours/week recovered), an estimated $35,000 from inventory accuracy gains, and an estimated $12,000 from faster exception resolution. Against $74,000 in Year 1 investment ($6,000 assessment, $38,000 build, $2,500/month operation), payback lands at 10.2 months with ROI accelerating in Year 2.
The remaining 4.5 weekly hours are high-judgment work — exception review, allocation decisions during shortages, vendor negotiation. The team's role shifted from data processing to operational decision-making.

## Two-Sided Marketplace: From Concept to $2.4M GMV in Year 1
URL: https://gigabit.agency/proof/marketplace-platform-build/
Client: B2B services marketplace startup, US-based, pre-seed ($500K)
Investment: $72,000 · Return: $2.4M GMV and $360K commission revenue in Year 1 · Timeline: 14 weeks to launch

- Year 1 GMV: $2.4M
- Commission revenue: $360K (15% avg rate)
- Vendors on platform: 35 → 840
- Uptime: 99.94%
- Seed round raised: $3.8M at $16M valuation

### Challenge
The founder targeted an underserved niche in the $50B US corporate events industry: connecting event planners with independent vendors — photographers, caterers, florists, AV technicians, designers. The incumbent process ran on personal referrals, email, spreadsheets, manual quote collection, and check payments.
The technical complexity was real. Two distinct user types with different UIs, onboarding, and incentives — plus the classic cold-start problem. Single events requiring 5–8 vendors meant multi-vendor carts, coordinated scheduling, and Stripe Connect split payments across multiple vendors per transaction. Search needed faceting by service type, location, availability, budget, and rating. And the budget was $75K maximum, because the $500K pre-seed also had to fund 12 months of operations and marketplace bootstrapping.

### Approach
Weeks 1–2 nailed the data model — the critical foundation for marketplace many-to-many relationships across users, vendor profiles, events, bookings, payments, and bidirectional reviews. Stripe Connect ran in destination-charges mode: each booking is an independent charge with an automatic transfer to the vendor's connected account, the platform retaining its commission. Multi-vendor events check out as a batch of individual charges — simpler and more reliable than splitting one charge.
Weeks 3–10 built the platform: vendor onboarding completable in under 15 minutes, planner event creation with recommended vendor types, Algolia-powered search with geo-ranking (2 days of integration versus 2+ weeks building comparable quality on PostgreSQL), booking workflow with in-platform Socket.io messaging, 48-hour payment holds with dispute handling, and an anti-gaming review system — double-blind until both parties submit, 30-day windows, moderation queue.
Weeks 11–14 covered quality and launch. Load testing at 500 concurrent users exposed two bottlenecks — an unindexed search-ranking query (fixed: 800ms to 120ms) and a WebSocket connection leak — both resolved pre-launch. Server-rendered vendor profiles with Schema.org structured data and city-level landing pages laid the SEO foundation. Soft launch shipped with 35 pre-onboarded New York vendors; the first booking came 3 days later.

### Results
Year 1 totals: $2.4M in gross merchandise value and $360K in platform commission revenue at a 15% average rate. Vendors grew from 35 at launch to 840 by month 12, planners from 12 to 620, monthly bookings from 8 to 310, and monthly GMV from $18,000 to $340,000. Average booking value declined from $2,250 to $1,097 as the platform broadened beyond large events — a healthy sign.
Technical performance held: 99.94% uptime with a single 45-minute incident in month 7, 140ms average search response, 99.7% payment processing success, and 72% vendor onboarding completion. Organic search drove vendor sign-ups from month 3; a planner referral program drove 40% of new planner sign-ups from month 4.
On the back of Year 1 traction — $2.4M GMV, 840 vendors, 620 planners, 15% month-over-month Q4 growth — the founder raised a $3.8M seed round at a $16M valuation, 18 months after the initial $500K pre-seed. The platform is expanding to 8 additional US cities, with ongoing development continuing at $7,500/month.

## AI Transformation: 1,200 Hours Per Quarter Recovered for a Management Consulting Firm
URL: https://gigabit.agency/proof/ai-transformation-professional-services/
Client: Management consulting firm, US-based, 140 employees, $32M annual revenue
Investment: $210,000 (Year 1 total; $156,000 build) · Return: $480,000/year in recovered value — 129% Year 1 ROI · Timeline: 6 months across 3 phases

- Hours recovered: 1,200/quarter
- Year 1 value: $480,000
- Proposal creation time: 4–8 hours → 45 minutes
- Time capture rate: 80–85% → 94%
- Payback period: 5.3 months

### Challenge
The firm — 95 consultants and 45 operations staff serving manufacturing and logistics clients — was growing 20% year-over-year with operational overhead scaling linearly alongside. The managing partner's math: consultants billing $250–$400/hour were spending 35% of their time on work that didn't require a consultant — roughly $4.8 million a year in wasted billable capacity.
The assessment quantified the leaks. Proposal generation consumed 320 hours/quarter at 4–8 hours per proposal across 40+ proposals. Knowledge retrieval cost 280 hours/quarter hunting through SharePoint, Outlook, Teams, and senior partners' heads. Deliverable formatting ate 240 hours/quarter. Memory-based Friday time entry left 15–20% of billable time unrecorded — on $32M of revenue, roughly $4.8M in potential unbilled work. Internal operations consumed another 200 hours/quarter.
Underneath it all sat one structural problem: 8 years of deliverables, proposals, and methodologies spread across 47,000 SharePoint files with no naming convention, no tagging, and no search beyond filename matching.

### Approach
A 2-week, $12,000 readiness assessment — 15 interviews across partners, consultants, and operations, plus shadowing a full proposal and delivery cycle — set the sequence: knowledge base first (the foundation), then proposals, time entry, deliverables, and operational automation.
Phase 1 (weeks 1–8) ingested 12,000 of the most relevant documents into a RAG knowledge base: structure-aware chunking, OpenAI embeddings in pgvector, access control by practice area and client authorization, delivered as a Slack bot and web interface. A consultant asks a question in Slack and gets a sourced answer in 4 seconds with links to the underlying deliverables. The proposal assistant built on top generates full draft proposals from a structured brief — cutting creation time from 4–8 hours to 45 minutes of review.
Phase 2 (weeks 9–14) shipped the time entry assistant — every morning it suggests entries from calendar, email, document, and channel activity for one-tap approval — plus report-drafting and presentation-builder tools that handle structure, formatting, and data presentation while consultants supply the analysis. Phase 3 (weeks 15–24) automated client intake, monthly partnership reporting with AI-generated narrative, and compliance documentation.

### Results
Measured over the first full quarter with all three phases live: the knowledge base recovered 280 hours ($84,000 at the $300/hour average), the proposal assistant 260 hours ($78,000), deliverable tools 240 hours ($72,000), operational automation 200 hours ($36,000 at the $45/hour ops rate), and monthly reporting 40 hours ($12,000). The time entry assistant lifted capture from 80–85% to 94%, worth $95,000 in the quarter — the managing partner estimates $380K/year in additional billings. Quarterly total: 1,020+ hours and $377,000.
Annualized with continued optimization, that's $480,000/year in recovered value. Against $210,000 in Year 1 investment ($12,000 assessment, $156,000 build, $4,500/month operation), Year 1 ROI was 129% with a 5.3-month payback.
The qualitative shift mattered as much: consultants now open engagements by asking the knowledge base what the firm has done before, and junior consultants — previously slowed for months by tribal knowledge — reported the largest productivity gains.

## B2B Payment Platform: From Concept to $41M Processed in Year 1
URL: https://gigabit.agency/proof/fintech-payment-platform/
Client: B2B construction payments startup, US-based, seed-funded ($2.1M)
Investment: $118,000 · Return: $41.2M processed in Year 1; $6.2M Series A raised · Timeline: 4 months to MVP launch

- Year 1 payment volume: $41.2M
- Processing error rate: 0.02%
- Subcontractor DSO: 67 days → 12 days
- Uptime: 99.97%
- Series A raised: $6.2M

### Challenge
Subcontractors on construction projects wait 60–90 days for payment. General contractors juggle payments across dozens of subcontractors per project — each with different payment terms, retainage requirements, and lien waiver workflows — on spreadsheets, manual check cutting, and phone calls.
The founder's vision: one platform where general contractors manage subcontractor payments from contract to release, with automated compliance documentation and flexible payment methods. The technical requirements were demanding — Stripe Connect multi-method payments with custom splits, a compliance engine that blocks payment release until lien waivers, insurance certificates, contract terms, and retainage calculations all verify, clean many-to-many data modeling across contractors, projects, and subcontractors, and a full audit trail with SOC 2 readiness from launch.
The budget ceiling was $150K, because the $2.1M seed round had to fund 18 months of operations including sales — not just engineering.

### Approach
Months 2–3 built the core platform: a general contractor dashboard with real-time views of outstanding payments, upcoming draws, expiring compliance documents, and audited payment history; and a subcontractor portal for payment schedules, document upload, payment status tracking, and bank account management.
The compliance engine is rules-based and explicit. Before any payment releases it verifies the lien waiver (signed, correct amount and period), the insurance certificate (current, meets contract minimums), contract terms (draw matches schedule, retainage correct), and prior payment settlement. Every rule returns pass/fail with a specific failure reason, so contractors see exactly what blocks each payment. Retainage — typically 5–10% withheld per payment — is tracked per subcontractor per project and released as its own compliance-verified payment event.
Month 4 hardened the system: full payment-flow testing with real Stripe test transactions including 5-day ACH settlement, failures, refunds, and multi-subcontractor splits; edge-case testing like documents expiring mid-settlement and partial draws; SOC 2-ready infrastructure (audit logging, RBAC, encrypted storage, documented change management) without premature certification; and a third-party penetration test whose two medium findings were remediated before launch.

### Results
Year 1 performance: $41.2M in total payment volume across 3,847 transactions at a $10,710 average size, serving 67 general contractors and 412 subcontractors. The payment processing error rate was 0.02% — one failed payment, caused by a bank return rather than the platform. Compliance documents auto-verified at 94.6% accuracy with the remaining 5.4% flagged for manual review. Uptime: 99.97%.
For contractors, payment processing dropped from 3–5 business days of check cutting to same-day or next-day ACH. For subcontractors, Days Sales Outstanding fell from a 67-day average to 12 days.
The Year 1 volume at the platform's fee structure demonstrated product-market fit, and the founder raised a $6.2M Series A 14 months after the seed round. The MVP came in at $118,000 over 4 months, with ongoing development and support continuing at $8,500/month.

## From Manual Spreadsheets to Self-Serve Analytics: A SaaS Data Engineering Build
URL: https://gigabit.agency/proof/data-engineering-saas-analytics/
Client: B2B project management SaaS, $6M ARR, 85 employees
Investment: $78,200 (Year 1 total; $44,000 build) · Return: $270,000+ Year 1 value · Timeline: 7 weeks (1-week assessment + 6-week build)

- Manual reporting hours: 56+ → 4/month (-93%)
- Month-end close: 5 days → 2 days
- Time to answer a business question: 4–8 hours → 30 seconds
- ARR preserved via churn saves: ~$186,000
- Year 1 value: $270,000+

### Challenge
The company — a project management SaaS used by 1,200+ companies, growing 40% year-over-year — generated data across 7 systems: HubSpot, Stripe, a PostgreSQL production database, Intercom, Google Analytics, QuickBooks, and Gusto. Answering basic business questions took 2 analysts days of spreadsheet work.
Computing Net Revenue Retention meant exporting from Stripe, HubSpot, and the product database, reconciling mismatched customer IDs with VLOOKUPs, and producing a number nobody fully trusted. The CEO wanted a Monday-morning dashboard showing ARR, NRR, churn, expansion, and usage trends with no analyst intervention. The VP of Finance was burning 5 days on month-end close because financial reporting required manual aggregation from 4 systems.

### Approach
A 1-week, $4,000 data assessment audited all 7 sources and prioritized the reporting needs — the ARR/MRR dashboard alone cost 4 manual hours weekly, NRR 8 hours monthly, and the financial close pack 20 hours monthly. The plan: ingest everything into Snowflake, model metrics in dbt, serve self-serve dashboards in Metabase.
Weeks 2–3 stood up the infrastructure. Fivetran connected all 7 sources with sync schedules matched to need — hourly for product usage, every 6 hours for Stripe and HubSpot, daily for GA and QuickBooks. Snowflake used a three-layer architecture: an immutable raw layer as audit trail, a staging layer with customer IDs unified across systems, and business-ready marts for finance, customer, product, and support domains.
Weeks 4–5 built 34 dbt models — MRR per customer with proration logic, an ARR waterfall, cohort-based NRR, a composite customer health score, and feature adoption models — each guarded by data quality tests so corrupt source data fails the pipeline loudly instead of producing wrong metrics. Weeks 6–7 shipped 6 Metabase dashboards (executive, financial close, customer health, product analytics, support, marketing) and trained the teams in a 2-hour session.

### Results
Manual reporting fell from 56+ hours a month to 4 hours of review (-93%). Month-end close dropped from 5 days to 2 (-60%). Time to answer a business question went from 4–8 hours of manual analysis to 30 seconds on a dashboard. All 7 data sources now live in Snowflake as a single source of truth, with real-time health scoring across all 1,200 customers.
The health score dashboard drove proactive saves: the customer success team identified 47 accounts trending toward churn and retained 31 (66% save rate), preserving approximately $186,000 in ARR. Product analytics revealed that time tracking — an underused feature — correlated 2.4x more strongly with retention than any other feature, redirecting the next release cycle.
Year 1 investment totaled $78,200 ($44,000 build, $850/month platform costs, $2,000/month support retainer) against $270,000+ in Year 1 value from analyst time saved and churn prevention.

## Regional Provider Network — Unified Intake, EHR Integration & Analytics
URL: https://gigabit.agency/proof/healthcare-unified-intake-ehr-analytics/
Client: Multi-site medical group (primary + specialty care), 14 locations
Investment: TBD · Return: TBD · Timeline: 12 weeks

- Intake/admin time per patient: -41% (median)
- Throughput (visits per clinician session): +16%
- Form error/rework rate: -32%
- Lighthouse scores: 96 desktop / 89 mobile

### Challenge
The 14-location medical group ran on disjointed web forms and double data entry into EHR and practice management systems. Front-desk staff re-keyed patient information, producing frequent errors and rework, and intake throughput bottlenecked at the desk.
Leadership also had blind spots: no single view of bookings, show rate, marketing source mix, or provider utilization, and an inconsistent web experience across locations. The quarterly targets were explicit — 30–50% less intake/admin time, 10–20% more throughput, and unified dashboards by location, provider, and service line.

### Approach
We built a unified HIPAA-aware intake experience for web, tablet, and kiosk: a Next.js app stepping through demographics, insurance, consent, and prep, with 270/271 insurance eligibility checks, e-signature, document upload, and validation guardrails to cut incomplete submissions.
An HL7/FHIR bridge (via vendor APIs/Redox) eliminated double entry with event-driven sync from intake through appointment to encounter — field mapping, least-privilege credentials, retries with a dead-letter queue, audit logs on every write, and encryption in transit and at rest.
The analytics layer combined GA4/GTM funnel events with a BigQuery warehouse, dbt models, and Looker Studio/Metabase dashboards — KPIs sliced by location, provider, service, and source, with cohort views for show rate and lead-to-visit time. A centralized content system standardized service and location pages with FAQ schema, and a performance and accessibility pass enforced Core Web Vitals budgets and WCAG 2.1 AA.

### Results
Over 12 weeks: intake/admin time per patient fell 41% (median), throughput rose 16% in visits per clinician session, and the form error/rework rate dropped 32%.
The group gained a single source of truth for bookings, show rate, provider utilization, and marketing source mix — visibility from campaign to clinic that did not exist before. Site performance landed at Lighthouse 96 desktop / 89 mobile.

## Specialty Clinic — LLM Patient Assistant Reduces Front-Desk Load
URL: https://gigabit.agency/proof/clinic-llm-assistant-front-desk/
Client: Single-specialty dermatology clinic — ~30 staff, 8 providers, 2 locations
Investment: TBD · Return: TBD · Timeline: 8 weeks

- Repetitive inquiries: -62%
- Median time-to-answer: 4:10 → 1:20 (-68%)
- CSAT: +18%
- After-hours interactions self-served: 40%

### Challenge
The clinic ran phone-first support. Repetitive questions — prep instructions, insurance, medications — overwhelmed staff and produced long hold times. SOPs were scattered across documents and email, so answers were inconsistent and new-hire training was slow.
There was no after-hours coverage, so a backlog spiked every morning, and no reporting on deflection or time-to-answer. The 8-week targets: at least a 50% reduction in repetitive inquiries, under 90 seconds median time-to-answer for common questions, and higher CSAT with fewer call-backs.

### Approach
We curated a source-of-truth corpus — SOPs, prep guides, insurance FAQs — versioned in a secure repo with a clinician-review editorial workflow, then built a RAG pipeline (LlamaIndex/LangChain) serving cited answers with PHI minimization and encryption at rest and in transit.
Conversation design was deliberately narrow: administrative and prep topics only — hours, directions, insurance, forms, pre/post-op. Guardrails included explicit not-medical-advice framing, refusal rules, keyword and intent escalation triggers, PHI redaction, rate limits, and human-in-the-loop handling for edge cases. Handoffs to staff carry a conversation summary.
Rollout was omni-channel: a React web widget site-wide, SMS/WhatsApp via Twilio, and IVR deflection to self-serve, with after-hours auto-response. An intake assist collects visit type, insurance carrier, and allergy/medication checklists (no diagnoses) and sends procedure-specific prep instructions. PostHog dashboards tracked deflection, time-to-answer, escalations, and CSAT, with weekly transcript reviews driving coverage expansion and tone calibration.

### Results
After 8 weeks: repetitive inquiries handled by the front desk fell 62%, median time-to-answer on common questions dropped 68% (4:10 to 1:20), CSAT rose 18% on the post-interaction micro-survey, and 40% of after-hours interactions resolved self-serve.
Documenting tribal knowledge into approved SOPs also cut staff onboarding time for FAQs — and by week three the morning backlog was gone.

## Multi-Location Dental Group — Self-Serve Booking = More Kept Appointments
URL: https://gigabit.agency/proof/dental-online-booking-kept-appointments/
Client: General + cosmetic dentistry group — 6 locations, 45 staff, 12 providers
Investment: TBD · Return: TBD · Timeline: 9 weeks

- Kept appointments: +27%
- No-shows: -29%
- Review velocity: +41% (rating 4.3 → 4.6)
- High-value case inquiries: +32%

### Challenge
The group was leaking leads: missed calls, long hold times, and web forms routed to a shared inbox. One-way reminders with no easy rescheduling drove high no-show rates.
Local presence was thin — few procedure pages, weak internal linking, inconsistent NAP data — and there was no source-to-appointment reporting by location or service. The 60–90 day targets: +20% kept appointments, -20% no-shows, and +25% high-value case inquiries (implants, veneers).

### Approach
We built a HIPAA-aware Next.js booking and intake flow with insurance capture, consent, e-signature, and PHI minimized and encrypted — synced in real time with Dentrix through a secure bridge with conflict handling and rules for high-value services.
Two-way SMS/WhatsApp reminders via Twilio carried smart links to confirm, reschedule, or request a callback, plus procedure-specific prep instructions and automated post-op follow-ups. A reputation engine requested reviews at the right moments after completed visits, and NAP cleanup standardized listings across locations.
A content hub added procedure explainers (implants, veneers, Invisalign) with plain-language FAQs, FAQPage/Organization/Service schema, and internal links from treatments to locations to booking. GA4/GTM events fed a Looker Studio dashboard tracking kept appointments, no-show rate, and source mix by location and service line.

### Results
Within 60 days: kept appointments rose 27% versus baseline and no-shows fell 29% across all locations.
Review velocity increased 41% with the average rating climbing from 4.3 to 4.6, and high-value case inquiries (implants, veneers) grew 32%. Site performance reached Lighthouse 95 desktop / 88 mobile.

# Articles

## Generative Engine Optimization: how to get cited by ChatGPT, Claude, and Perplexity
URL: https://gigabit.agency/insights/generative-engine-optimization-field-guide/
Published 2026-06-15, updated 2026-10-22

For twenty years, being found meant ranking on a results page. That motion isn't going away, but a second one has opened beside it: the answer engine. When a buyer asks ChatGPT, Claude, or Perplexity for a recommendation, there's no page of ten blue links to climb. There's one answer — and either you're cited inside it or you're invisible.
This is no longer a fringe channel. ChatGPT reached **800 million weekly active users** in October 2025 ([TechCrunch](https://techcrunch.com/2025/10/06/sam-altman-says-chatgpt-has-hit-800m-weekly-active-users/)), Google's AI Overviews serve **2 billion monthly users** ([TechCrunch](https://techcrunch.com/2025/07/23/googles-ai-overviews-have-2b-monthly-users-ai-mode-100m-in-the-us-and-india/)), and Perplexity handled **780 million queries in a single month** ([TechCrunch](https://techcrunch.com/2025/06/05/perplexity-received-780-million-queries-last-month-ceo-says/)). Your buyers are already asking machines who to hire. Generative Engine Optimization (GEO) is the discipline of being the company those machines name.
### Why GEO matters now, not later
Because the click is disappearing, and the citation is replacing it. When Google shows an AI summary, only **8% of users click any traditional link — versus 15% without one — and just 1% click a source inside the summary itself** ([Pew Research, 2025](https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/)). The old game of ranking to earn a click is being quietly zeroed out. The new game is being the source the answer is built from.
And the traffic that remains is worth more. Retail visits from generative-AI sources grew **1,200%** between mid-2024 and early 2025 ([Adobe Analytics](https://blog.adobe.com/en/publish/2025/03/17/adobe-analytics-traffic-to-us-retail-websites-from-generative-ai-sources-jumps-1200-percent)), and those visitors convert better: Semrush found an AI-search visitor is roughly **4.4× more valuable** than a traditional organic one ([Semrush, 2025](https://www.semrush.com/blog/ai-search-seo-traffic-study/)), and one 94-site study measured **31% higher conversion** from ChatGPT referrals than non-branded organic ([via Search Engine Land, 2026](https://searchengineland.com/chatgpt-vs-non-branded-organic-search-conversions-470321)). Fewer clicks, higher intent — which is exactly why the cited brands win and the rest fade.
### GEO vs SEO: what's actually different
SEO competes for a rank on a page of ten links. GEO competes to be the one source quoted when there's no list to climb. They're related — an AI answer often draws on well-ranked pages — but the link is loosening: **38% of AI Overview citations now come from the top-10 organic results, down from about 76% a year earlier** ([Ahrefs, 2026](https://ahrefs.com/blog/ai-overview-citations-top-10/)). Ranking still helps; it's no longer sufficient. GEO leans on machine-legible structure — clear claims, named entities, real numbers, schema, and unambiguous phrasing a model will reach for when it has to be correct. It starts with understanding [how AI assistants choose what to recommend](https://gigabit.agency/insights/how-ai-assistants-choose-recommendations/), then building for it on purpose.
| | SEO | GEO |
| --- | --- | --- |
| Competes for | A position on a page of ten links | Being the source quoted inside one answer |
| Unit of success | A click | A citation |
| You optimize | Keywords, backlinks, rankings | Extractable claims, entities, schema |
| You measure | Rank and organic sessions | Citation rate on a fixed query panel |
| You win when | You out-rank the competitor | The model can quote you *and* verify you |
So is SEO dead? No — and anyone selling you that is selling something. The retrieval layer underneath every answer engine is still a search index, so crawlable pages and real topical coverage remain the entry fee. What's changed is that they're now the entry fee rather than the win condition. The practical move isn't to defund SEO; it's to stop treating rank as the finish line and start treating the citation as the thing you're actually buying. Budget accordingly — we break down [what a GEO program costs](https://gigabit.agency/insights/how-much-does-geo-cost/) against what it replaces.
### What actually moves your citation rate
The levers are measurable, not mystical. The foundational study on this — Princeton's GEO paper — found that adding **citations, quotations, and statistics to your content boosted visibility in generative-engine answers by up to 40%** ([Aggarwal et al., KDD 2024](https://arxiv.org/abs/2311.09735)). In practice, that means:
- **Pages that answer the exact question** a buyer asks — not adjacent ones — with the answer stated up front, in a form a model can lift verbatim.
- **Quotable, verifiable facts**: published prices, named numbers, dated claims. Vague marketing copy doesn't get quoted; a specific figure with a source does.
- **Structure the machine can parse**: FAQ and schema markup, clean headings, entity consistency an engine can cross-check.
- **Third-party corroboration**: what others say about you across the web, because the engines weight it heavily.
### The five technical checks underneath
None of the levers above works if the engine cannot read you. Being cited takes three things in sequence: the engine has to **retrieve** a page about you, **extract** a claim it can repeat, and **verify** that you are who the page says you are. Five site-level conditions decide whether that sequence can run:
1. **The content is in the HTML.** GPTBot, ClaudeBot and PerplexityBot fetch your HTML and largely do not run JavaScript. If prices and claims render in the browser, the crawler sees an empty shell. View source on your most important page; if the substance is not in the raw response, server-render it.
2. **The AI crawlers are allowed.** Read `robots.txt` for `GPTBot`, `OAI-SearchBot`, `ChatGPT-User`, `ClaudeBot`, `Claude-User`, `PerplexityBot`, `Perplexity-User` and `Google-Extended`. Blocking them on purpose is a real choice; inheriting the block from a CDN preset is an accident.
3. **One page per question, answered in the first two sentences.** A heading phrased the way a buyer asks it, and an answer under it that still makes sense when lifted with no surrounding context.
4. **Facts a model can defend.** Named prices, dated figures, cited sources. A model avoids adjectives, because quoting *industry-leading* commits it to a claim it cannot support.
5. **One consistent entity across the web.** The same name, address and description on your site, your `Organization` schema, LinkedIn and directories. Schema states plainly who you are; it helps an engine resolve you as one company, but on its own it is not a citation lever.
| Check | How to test it | Fails if |
| --- | --- | --- |
| HTML-first content | View source on a key page | Your copy isn't in the raw response |
| Crawler access | Read robots.txt for named agents | A blanket disallow you inherited |
| Answer-shaped pages | Read the first two sentences under each heading | They need the preceding section to make sense |
| Quotable facts | Count the numbers on the page | Adjectives outnumber figures |
| Entity consistency | Compare name and address across five sources | Any two disagree |
Do them in that order: there is no point writing quotable answers into pages a crawler cannot read. And be clear about what the checklist does. It makes you *citable*. It does not, by itself, make you *cited*: retrieval still depends on topical relevance and on third-party corroboration you do not control, which is why a company can pass all five and still lose to a competitor with a thinner site and a thicker footprint across the web. That gap is diagnosed in [what gets a company excluded from AI answers](https://gigabit.agency/insights/what-gets-a-company-excluded-from-ai-answers/).
### How to measure it
By **citation rate** — the share of your category's real buying queries where an assistant names you. It's the closest thing GEO has to rank tracking, and it's the number a serious program is held to. We cover exactly [what a good citation rate looks like](https://gigabit.agency/insights/what-is-a-good-ai-citation-rate/), measured on a fixed query panel across ChatGPT, Perplexity, and Google AI Overviews, run monthly so the trend — not any single stochastic answer — is the signal.
### Where to start
Start with a baseline. A [citation audit](https://gigabit.agency/tools/geo-citation-audit/) runs your category's buying questions through the major assistants and measures where you appear, where a competitor appears instead, and where the answer is simply wrong. That score is the starting line; everything after it is closing the cheapest, highest-impact gaps first — the work behind [our GEO program](https://gigabit.agency/services/geo/). The companies that move now win a category with almost no serious players in it yet — what we call reaching [Answer-Market Fit](https://gigabit.agency/insights/answer-market-fit/). The ones that wait will be optimizing to be cited in answers their competitors already own.

### FAQ
**What is generative engine optimization (GEO)?**
Generative engine optimization (GEO) is the practice of structuring your content and entity data so AI answer engines — ChatGPT, Claude, Perplexity, and Google AI Overviews — cite your company when buyers ask for a recommendation. Unlike traditional SEO, which competes for blue-link rankings, GEO competes to be the source quoted inside a single generated answer.
**How is GEO different from SEO?**
SEO optimizes for ranking on a results page of ten links; GEO optimizes for being the one source an answer engine quotes when there is no list to climb. GEO leans on machine-legible structure — clear claims, named entities, real numbers, schema, and unambiguous phrasing a model will reach for when it has to be correct.
**How do you measure GEO results?**
By citation rate: you run your category's real buying questions through the major assistants and measure how often you are cited, how often a competitor is cited instead, and where the answer is simply wrong. That baseline score is what every later change is measured against.
**Is SEO dead now that AI answers exist?**
No. Every answer engine still retrieves from a search index, so crawlable pages and real topical coverage remain the entry fee for being considered at all. What changed is that ranking is no longer the win condition — only 38% of AI Overview citations now come from the top-10 organic results, down from about 76% a year earlier. Keep the SEO foundation; stop treating rank as the finish line.
**Should I move budget from SEO to GEO?**
Not as a wholesale swap. The retrieval layer GEO depends on is the one SEO builds, so defunding SEO undercuts GEO. The useful reallocation is within the content budget: fewer pages written to rank for a keyword, more pages written to answer a specific buying question with quotable, verifiable facts a model can lift.
**What does a website need before an AI assistant will cite it?**
Five things: content present in the raw HTML rather than assembled by JavaScript, a robots.txt that allows the named AI crawlers, one page per buying question with a self-contained answer in the first two sentences, quotable facts like named prices and dated figures rather than adjectives, and one consistent entity across your site, schema, LinkedIn, and directories. They make you citable; being cited still depends on topical relevance and third-party corroboration.
**Do AI crawlers execute JavaScript?**
Largely no. GPTBot, ClaudeBot, and PerplexityBot fetch your HTML and read what is in it, so prices, claims, and copy rendered client-side are invisible to them. View source on your most important page: if the substance is not in the raw response, no amount of content work will help until it is.

## Why your AI pilot never reached production — and the five gates that get it there
URL: https://gigabit.agency/insights/why-your-ai-pilot-never-reached-production/
Published 2026-06-04, updated 2026-10-10

Most AI pilots do not fail because the idea was wrong. They fail because nobody built the machinery that lets a probabilistic system run unattended in a workflow that matters. A demo proves the model can do the task once. Production proves it can do the task ten thousand times without a human catching every miss.
We treat the gap between the two as **five gates**. Each one answers a question someone will eventually ask before letting the system run unattended, and each one is cheap to build early and expensive to retrofit late. Skip any single gate and the pilot stalls — usually not in engineering, but in the review meeting where nobody can answer the question that gate exists to answer.
### The five gates
| Gate | The question it answers | What it looks like when it's missing |
| --- | --- | --- |
| Evals | Is it good enough, and did that change make it better? | Every release is an argument about vibes |
| Observability | What did it actually do on this case? | You find out from an angry customer |
| Guardrails | What's the worst thing it can do? | It never clears security or legal review |
| Rollback | What do we do when the model changes under us? | A vendor update silently breaks it |
| Ownership | Whose number is this? | A committee approves it and nobody ships it |
### Gate 1 — An eval suite that measures the thing you care about
An eval suite is a fixed set of real cases with known-good outcomes, scored automatically every time anything changes. Not a spot-check, and not *does this output look right* — a number you can compare against last week's number.
Without one, quality is a matter of opinion. Every prompt tweak becomes a debate, nobody can say whether the change helped, and no one is willing to sign off on a release they can't measure. That's not a model problem; it's the absence of a ruler. **You cannot ship what you cannot score** — which is why we treat this as the gate that comes before all the others, and why we wrote a [whole piece on it](https://gigabit.agency/insights/ai-agent-eval-metrics/).
*You've passed this gate when* you can swap a model or rewrite a prompt and know within minutes whether quality moved, and in which direction.
### Gate 2 — Observability into what the system actually did
Not an aggregate dashboard. Per-run forensics: for any single execution, the input, the context it retrieved, the tools it called, what it returned, what it cost, and how long it took.
Probabilistic systems fail quietly. A deterministic bug throws an error; a model just returns something slightly wrong, confidently, and keeps going. Without traces, those failures are invisible until they accumulate into something a human notices — which is how teams discover a month of bad outputs in a single support escalation.
*You've passed this gate when* someone can answer *why did it do that, on this specific case, last Tuesday* in under five minutes.
### Gate 3 — Guardrails that bound the worst case
Explicit constraints on what the agent can touch, what it can spend, what it can say, and when it must hand off to a person. Scope limits enforced in code — not instructions in a prompt asking it nicely.
This is the gate most pilots actually die at, and it's rarely engineering that kills them. It's the security review, the legal review, the risk conversation that never resolves because nobody can state the blast radius in a sentence. An agent that *probably* behaves cannot be granted write access to a production system, and so it stays a demo forever. The fix is to make the worst case small and explicit rather than to argue that it's unlikely.
*You've passed this gate when* you can state the worst thing the system can do in one sentence, and the person accountable for that risk is comfortable with it.
### Gate 4 — A rollback path for when the ground moves
Pinned model versions, versioned prompts and configuration, and the ability to return to a known-good state on demand. The models underneath you are not stable infrastructure — vendors deprecate, retune, and ship behavior changes on their schedule, not yours.
Teams that skip this discover it the same way every time: something that worked for months degrades over a weekend, and there's no previous configuration to fall back to because the current one was edited in place. Reverting turns into re-deriving.
*You've passed this gate when* rolling back is a deploy, not a rebuild.
### Gate 5 — A named owner accountable for the metric
One person. One number. Not a sponsor, not a working group, not a vendor — a named individual whose job is affected by whether the number moves.
This is the least technical gate and the most predictive. Committees are good at approving AI projects and structurally incapable of shipping them: responsibility spread across eight people is responsibility nobody feels at 6pm on a Thursday when the thing needs one more push. We call the failure mode the [Steering-Committee Tax](https://gigabit.agency/compare/gigabit-vs-consultancies/), and it is the single most reliable signal we see for a program that will stall.
*You've passed this gate when* you can name the person and the number without looking anything up.
### Why most of it goes wrong in the first 90 days
The failure is usually decided long before anyone notices it — inside the first quarter, in three moves that all feel reasonable at the time.
**Month one: the strategy is a list of use cases.** A workshop produces fifteen candidate workflows ranked by excitement, and the most impressive one gets picked instead of the most tractable. Nobody asks which of the fifteen has a measurable cost attached to it today, so there's no number to be accountable to later.
**Month two: nobody owns the metric.** The work has a project sponsor, a vendor, and a steering committee — which is not the same as one named person whose job depends on a number moving. Committees can approve an AI project; they cannot make one work. This is the [Steering-Committee Tax](https://gigabit.agency/compare/gigabit-vs-consultancies/), and it is the single most reliable predictor of a stalled program.
**Month three: the demo lands and quietly redefines success.** It works on the happy path, the room is impressed, and the goal silently shifts from *deployed* to *demonstrated*. From here the project doesn't fail so much as never conclude — which is why **95% of enterprise generative-AI pilots deliver no measurable P&L impact** ([MIT Project NANDA, 2025](https://www.forbes.com/sites/jasonsnyder/2025/08/26/mit-finds-95-of-genai-pilots-fail-because-companies-avoid-friction/)) and Gartner expects **more than 40% of agentic-AI projects to be canceled by the end of 2027** ([Gartner, 2025](https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027)).
The counter-move is unglamorous: pick one workflow that already costs someone money, put a name on the metric, and define *done* as running unattended in production — not as a demo that impressed a room. That's the whole reason our engagements start with a fixed-price [two-week Sprint](https://gigabit.agency/services/ai-transformation/) that ends in something deployed rather than a slide deck. If you're earlier than that, the [AI Readiness Assessment](https://gigabit.agency/tools/ai-readiness-assessment/) is the cheaper first look.
### Which gate are you stuck at?
Stalled pilots feel the same from the inside, but the symptom usually points at exactly one missing gate:
- *We keep tweaking it and can't tell if it's getting better.* — Gate 1. You have no ruler.
- *It works, except when it doesn't, and we can't reproduce it.* — Gate 2. You have no traces.
- *Engineering finished months ago; it's still in review.* — Gate 3. Nobody can bound the worst case.
- *It used to work and something changed.* — Gate 4. You have no known-good state to return to.
- *Everyone agrees it's important and nothing moves.* — Gate 5. No single person owns the number.
That diagnosis matters more than it sounds, because the gates are cheap in the order above and expensive out of it. Retrofitting evals onto a system already in production means reconstructing ground truth from cases nobody recorded. Adding observability after a failure means investigating something you have no trace of. Almost every expensive AI rescue we're called into is a team paying the retrofit price on a gate that would have cost days at the start.
None of this is exotic. It is the same discipline that turned web demos into reliable software a decade ago, applied to a stack that happens to be non-deterministic. The companies stuck in pilot purgatory are not short on ambition. They are short on the gates — which is exactly what a [Gigabit Agents](https://gigabit.agency/services/ai-agents/) build is built to install. If what stalled is a whole app built with AI tools, not one agent, [taking a prototype to production](https://gigabit.agency/services/prototype-to-production/) is the same discipline applied to the codebase.

### FAQ
**Why do most AI pilots fail to reach production?**
Most fail not because the idea was wrong but because nobody built the machinery a probabilistic system needs to run unattended: an eval suite, observability, guardrails, a rollback path, and a named owner accountable for the metric. A demo proves the model can do the task once; production proves it can do it ten thousand times without a human catching every miss.
**What is pilot purgatory?**
Pilot purgatory is the state where an AI project works in a demo but never ships — stuck indefinitely in evaluation because the reliability, ownership, and rollback gates that make a non-deterministic system safe to run in production were never built.
**What separates a demo from a production AI agent?**
Five gates: an eval suite that measures what you actually care about, observability to see failures as they happen, guardrails that contain the worst case, a rollback path for when a model or vendor changes under you, and a named owner accountable for the metric.
**Why do most AI strategies fail in the first 90 days?**
Three moves that each feel reasonable. Month one, the strategy becomes a ranked list of use cases and the most impressive workflow gets picked over the most tractable — so no measurable cost is attached to it. Month two, a committee owns the project instead of a named person owning a metric. Month three, an impressive demo silently redefines success from deployed to demonstrated. After that the project rarely fails outright; it just never concludes.
**How do you avoid stalling in the first quarter?**
Pick one workflow that already costs someone money today, put a single name on the metric it should move, and define done as running unattended in production rather than as a demo that impressed a room. A short fixed-price engagement that ends in something deployed enforces all three better than a strategy phase does.

## AI agent security framework: the 3Cs and beyond
URL: https://gigabit.agency/insights/ai-agent-security-framework/
Published 2026-10-22

An **AI agent security framework** is a set of controls that hold no matter what the model decides. The clearest one published this year is Docker's **3Cs** ([Sekaran and Aleyner, Docker, Feb 2026](https://www.docker.com/blog/the-3cs-a-framework-for-ai-agent-security/)). **Contain** the blast radius. **Curate** the agent's environment before it acts. **Control** each action at runtime, before it completes. It is a good spine. It is also a vendor's framework, written by a company that sells sandboxes, and it leaves out identity, evidence, and testing. If you want these controls tested against the agents you already run, that is the job of our [AI Security Assessment](https://gigabit.agency/services/ai-security/).
This article checks each C against the open work: OWASP, NIST, Google, Meta, the UK NCSC, and the MCP specification. It checks them against the agent incidents of the last eighteen months too. Then it lists what you still have to build yourself. What a security test covers is a separate post, [what an AI security assessment covers](https://gigabit.agency/insights/what-an-ai-security-assessment-covers/). This is the design side: what to build so the test has something to pass.
### Why agents break the security model you already have
Firewalls, VPNs, and secrets managers were built on one assumption: a person sits behind each request. Docker's version of the problem is the unattended laptop. A developer machine already reaches source code, databases, credentials, and production APIs. Teams now hand an agent that reach and let it act at machine speed, in parallel, faster than anyone can review.
Speed is not a metaphor. Anthropic investigated a state-sponsored campaign that ran on Claude Code. It reported that the AI executed **80–90% of tactical operations** on its own. Humans stepped in at **4–6 decision points** per campaign, and requests peaked at multiple operations per second ([Anthropic, Nov 2025](https://www.anthropic.com/news/disrupting-AI-espionage)). That was an attacker using an agent. A defender's agent runs at the same tempo. A review queue that clears tomorrow is not a control on an action that happens in 200 milliseconds.
Adoption has outrun controls. These figures come from different samples, and three of them are vendor-sponsored. Read them as direction, not as one market measurement.
| Signal | Figure | Source |
| --- | --- | --- |
| Developers using AI agents or automated workflows | 26.2% | [Stack Overflow Developer Survey 2026](https://survey.stackoverflow.co/2026/ai) |
| Developers who trust AI output for important decisions | 6.6% (48.0% trust it only when they can verify it) | [Stack Overflow Developer Survey 2026](https://survey.stackoverflow.co/2026/ai) |
| Organisations where every agent went live with full security approval | 14.4% | [Gravitee, Feb 2026](https://gravitee.io/blog/state-of-ai-agent-security-2026-report-when-adoption-outpaces-control) |
| Organisations using shared API keys between agents | 45.6% | [Gravitee, Feb 2026](https://gravitee.io/blog/state-of-ai-agent-security-2026-report-when-adoption-outpaces-control) |
| Organisations that saw agents take unintended actions | 80% (23% saw agents tricked into revealing credentials) | [SailPoint, May 2025](https://www.sailpoint.com/press-releases/sailpoint-ai-agent-adoption-report) |
| Organisations in IBM's breach study that reported a breach of an AI model or app | 13%, and 97% of those lacked proper AI access controls | [IBM Cost of a Data Breach, Jul 2025](https://newsroom.ibm.com/2025-07-30-ibm-report-13-of-organizations-reported-breaches-of-ai-models-or-applications,-97-of-which-reported-lacking-proper-ai-access-controls) |
| Enterprise breaches predicted to trace back to AI agent abuse by 2028 | 25% | [Gartner, Oct 2026](https://www.gartner.com/en/newsroom/press-releases/2026-10-20-gartner-identifies-the-top-strategic-technology-trends-for-2027) |
The reflex fix is more approval prompts, and it fails twice. First, people stop reading them. Docker calls it consent fatigue. Invariant Labs, writing up a GitHub MCP exploit, noted that many users simply click Always Allow ([Invariant Labs, May 2025](https://invariantlabs.ai/blog/mcp-github-vulnerability)). Microsoft's AI Red Team, after a year of testing agents, found human-in-the-loop bypass was the most consistently exploited failure mode ([Microsoft, Jun 2026](https://www.microsoft.com/en-us/security/blog/2026/06/04/updating-taxonomy-failure-modes-agentic-ai-systems-year-red-teaming-taught-us/)). Second, prompts slow an agent back to human speed, which removes the reason to run it. Anthropic measured the alternative. Moving Claude Code to a sandbox **cut permission prompts by 84%** in internal use ([Anthropic, Oct 2025](https://www.anthropic.com/engineering/claude-code-sandboxing)). Docker's line for it is that "human attention is not a scalable control plane."
### The 3Cs at a glance
| C | The question it answers | What it looks like in practice | An incident it would have limited |
| --- | --- | --- | --- |
| **Contain** | If this agent is wrong or hijacked, how far can the damage go? | A sandbox per session with filesystem and network isolation, no standing production credentials, environments you can throw away | An agent deleted a production database during a code freeze |
| **Curate** | What can this agent possibly reach before it starts? | Allowlisted and pinned MCP servers and skills, one scoped credential per task, read-only by default, minimal context | An agent ran with a database key that bypasses row-level security |
| **Control** | Is this specific action allowed right now? | A policy check outside the model on every tool call, file write, network request, and credential use | An agent read private repositories and published them in a public pull request |
The three reinforce each other. Containment limits what a failure costs. Curation narrows what the agent can attempt. Control replaces after-the-fact approval with enforcement at the moment of action. None of them asks the model to behave. That is the point.
### Contain: bound the blast radius
Every execution model before agents came with an isolation boundary. Processes had memory protection, virtual machines had hypervisors, and containers had namespaces. Docker's argument is that agents need the same, so a mistake does not become permanent. It treats containment as the precondition for giving an agent autonomy at all.
The vendors who ship coding agents reached the same design separately. Anthropic's Claude Code sandbox uses two boundaries. Filesystem isolation limits it to named directories, and network isolation limits it to approved servers through a proxy outside the sandbox. Anthropic is explicit that effective sandboxing needs **both** ([Anthropic, Oct 2025](https://www.anthropic.com/engineering/claude-code-sandboxing)). Without the network boundary, a hijacked agent can send your keys out. Without the filesystem boundary, it can escape the sandbox and reach the network anyway. OpenAI's Codex ships three sandbox modes, and its workspace-write mode keeps the network off by default ([OpenAI Codex docs](https://learn.chatgpt.com/docs/agent-approvals-security)). Docker Sandboxes run each session in a microVM with its own kernel. All outbound traffic goes through a host proxy that connects only when a policy rule matches. API keys are injected by that proxy, so they never sit inside the VM ([Docker, Apr 2026](https://www.docker.com/blog/why-microvms-the-architecture-behind-docker-sandboxes/)).
A sandbox is a boundary you configure, not one you get by installing it. Docker's own documentation notes that its clone mode protects the host repository from modification, not from inspection. A .env file in that repository is still readable, and SSH agent forwarding is on by default ([Docker docs](https://docs.docker.com/ai/sandboxes/security/isolation/)). Read the defaults before you call the agent contained.
### Two incidents containment would have capped
**Replit, July 2025.** During a declared code freeze, an agent on Replit deleted a user's production database. Replit's chief executive called it unacceptable, and the fix he announced was automatic separation of development and production databases ([The Register, Jul 2025](https://www.theregister.com/2025/07/22/replit_saastr_response/)). The user had told the agent not to make changes, in capitals, repeatedly. An instruction is not a control. A boundary the agent cannot cross is.
**Nx "s1ngularity", August 2025.** Attackers published malicious versions of the Nx build tool that stayed live for about four hours. The install script tried to use the local AI command-line tools already on developer machines to hunt for secrets ([Nx postmortem, Sep 2025](https://nx.dev/blog/s1ngularity-postmortem)). Wiz found the malware passed the flags that switch those tools' permission checks off. It counted more than **1,000 valid GitHub tokens** and about **20,000 leaked files**, and later more than **5,500 private repositories** made public ([Wiz, Aug 2025](https://www.wiz.io/blog/s1ngularity-supply-chain-attack)). An agent limited to the project directory, with an egress allowlist, could not have read ~/.ssh or pushed the results to GitHub.
### What containment means in practice
- One sandbox per session, rebuilt from a clean image. MITRE ATLAS lists this as mitigation AML.M0032, segmentation of AI agent components ([MITRE ATLAS](https://github.com/mitre-atlas/atlas-data)).
- The filesystem scoped to the working directory, and the agent never running as root. Both are in OWASP's guidance for unexpected code execution, ASI05.
- Network egress denied by default, with an allowlist per task.
- Credentials kept outside the sandbox, injected per request, and short-lived.
- Production behind a role the agent does not hold.
- A kill switch, credential revocation, and quarantine for a suspect agent, as OWASP's ASI10 recommends.
Containment limits damage. It does not stop an agent misusing what is inside the box. If the box holds the customer table, a hijacked agent can still read it. That is the job of the next C.
### Curate: build the agent's world before it acts
Docker puts it in five words: "Curation isn't approval. It is construction." An agent's capability is fixed before it acts, by the tools it can call, the code it can see, the credentials it holds, and the context it reads. Curation is choosing that world deliberately rather than reviewing what the agent wants to do inside an unbounded one.
OWASP reached the same idea from the risk side. Its 2025 entry on **excessive agency** names three root causes: excessive functionality, excessive permissions, and excessive autonomy ([OWASP LLM06:2025](https://genai.owasp.org/llmrisk/llm062025-excessive-agency/)). The first two are curation failures. A document-reading tool that can also delete files is one. So is a read-only task running with UPDATE and DELETE rights. OWASP's agentic list, published in December 2025, extends this into a principle it calls **Least-Agency**: do not give autonomy where the task does not need it ([OWASP Top 10 for Agentic Applications](https://genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-2026/)).
### Tools are a supply chain, and some of it is hostile
An MCP server or an agent skill is code plus natural language the model trusts. Both can be poisoned, and recent incidents show each route.
- **Tool poisoning, April 2025.** Invariant Labs hid instructions in an MCP tool's description, which the model reads and the user usually does not. The agent read ~/.ssh/id_rsa and its own MCP config and passed them out. The same write-up defines a rug pull, where a description changes after you approve it, and shadowing, where one server changes how the agent uses another ([Invariant Labs, Apr 2025](https://invariantlabs.ai/blog/mcp-security-notification-tool-poisoning-attacks)).
- **A backdoored MCP package, September 2025.** An unofficial postmark-mcp package on npm published 15 clean versions. Version 1.0.16 then quietly blind-copied every email it sent to an outside address. Postmark had never published an MCP server to npm ([Postmark, Sep 2025](https://postmarkapp.com/blog/information-regarding-malicious-postmark-mcp-package)).
- **Skills marketplaces, 2026.** Snyk scanned 3,984 public agent skills. **13.4%** had at least one critical issue, and **76** were confirmed malicious ([Snyk, Feb 2026](https://snyk.io/blog/toxicskills-malicious-ai-agent-skills-clawhub)). The OpenClaw marketplace added VirusTotal scanning. Its own announcement says this is not a silver bullet, because a prompt-injection payload written in plain language will not appear in a threat database ([OpenClaw, Feb 2026](https://openclaw.ai/blog/virustotal-partnership)). Unit 42 later found five malicious skills that got past the existing checks. One padded its README with 22 MB of filler to push it past scanner size limits ([Unit 42, Jun 2026](https://unit42.paloaltonetworks.com/openclaw-ai-supply-chain-risk/)).
A marketplace scan is a filter. A curated catalogue is a decision. When anyone with an account can publish, curation has to happen at your boundary, not theirs.
### Credentials are where curation usually fails
In July 2025, General Analysis showed what happens when an agent holds the wrong key. An attacker wrote instructions into a customer support ticket. A developer later asked Cursor to show the latest ticket. The agent was connected through Supabase's MCP server with the service_role key, which bypasses row-level security. It read the integration_tokens table and wrote the contents back into the ticket, where the attacker could see them ([General Analysis, Jul 2025](https://generalanalysis.com/blog/supabase-mcp-blog)). No permission was violated. The agent did exactly what its key allowed. The remedy was a read-only, project-scoped connection: curation, not detection.
### What curation means in practice
- An allowlist of MCP servers and skills, each pinned to a version and verified by hash. A new version gets reviewed before it loads.
- One credential per agent and per task, scoped to the narrowest role. No service-role keys and no shared API keys.
- Read-only by default. Write tools are added per workflow, by name.
- Tools the workflow never calls are removed, not just left unused.
- Context curated too: what the agent reads, what it remembers, and which feeds are labelled untrusted.
Docker's product page says the target state plainly: each session gets the tools, context, and credentials the task needs, and nothing more ([Docker 3Cs framework](https://www.docker.com/products/ai-governance/3cs-framework/)).
### Control: enforce at the moment of action
Curation sets what an agent could ever do. Control decides what it may do now. Docker's requirement is that rules apply as actions happen, on file access, network calls, tool calls, and credential use. A log entry or a warning arrives after the action has started. It explains the damage. It does not prevent it.
This is the C with the most outside support. Google's agent security paper describes the same layer: a deterministic policy engine outside the model that intercepts each action before it runs. It weighs the action's risk and context, including whether the agent has just read untrusted data. It then allows the action, blocks it, or asks for confirmation. Google's example blocks purchases over $500 and asks for confirmation between $100 and $500 ([Díaz, Kern and Olive, Google, May 2025](https://services.google.com/fh/files/misc/google_approach_for_secure_ai_agents.pdf)). OWASP's agentic list calls the same component an intent gate. It treats the planner's output as untrusted and checks it before execution. MITRE ATLAS lists it as AML.M0030: restrict or confirm tool calls once untrusted data is in the context. The MCP specification adds the protocol-level rules. A server must not accept a token that was not issued to it, and a proxy must get consent from each client before forwarding ([MCP Security Best Practices, 2026-07-28](https://modelcontextprotocol.io/docs/2026-07-28/tutorials/security/security_best_practices)).
### Three incidents control would have stopped
- **GitHub MCP, May 2025.** An attacker opened an issue with hidden instructions on a public repository. When the owner asked an agent to look at open issues, the agent read private repositories and published their contents in a pull request on the public one. Invariant Labs noted the server code had no bug to patch. The fix it proposed is a runtime policy: one repository per session ([Invariant Labs, May 2025](https://invariantlabs.ai/blog/mcp-github-vulnerability)).
- **EchoLeak, June 2025.** A single crafted email planted instructions that Microsoft 365 Copilot later followed. Copilot pulled internal data into its answer and sent it out, without a click from the user. Microsoft rated it CVSS 9.3, critical ([MSRC, CVE-2025-32711](https://msrc.microsoft.com/update-guide/vulnerability/CVE-2025-32711)). The chain had to defeat an injection classifier, link redaction, image fetching, and a content security policy ([Reddy and Gujral, 2025](https://arxiv.org/abs/2509.10540)). Each of those is a control on the output channel, and the attack shows why you need several.
- **The agent that approved itself, August 2025.** A prompt injection made GitHub Copilot write chat.tools.autoApprove: true into the project's VS Code settings. That switched off confirmations, and the agent then ran the attacker's commands ([Embrace The Red, Aug 2025](https://embracethered.com/blog/posts/2025/github-copilot-remote-code-execution-via-prompt-injection/)). Policy that the agent can edit is not policy.
### What control means in practice
- Every tool call goes through one gateway that authenticates, authorises, and logs it. Docker's version is an MCP gateway that denies unvetted servers by default.
- Untrusted content taints the session. Once a web page, email, or ticket is in the context, consequential tools need confirmation or are blocked.
- An egress allowlist enforced on the network, not requested in the prompt.
- Rate limits and spend budgets per tool.
- The agent cannot edit its own permissions, settings, or policy files.
- Human confirmation kept for irreversible or high-value actions, with a dry-run diff of what will change. Where to put those stops is its own question, covered in [when an AI agent should stop and ask a human](https://gigabit.agency/insights/ai-agent-human-in-the-loop/).
### Why the boundary has to sit outside the model
The 3Cs make no demands of the model, and that is deliberate. Prompt injection, an instruction hidden in content the agent reads, is not solved. The organisations closest to the problem say so. The UK National Cyber Security Centre calls a language model an "inherently confusable deputy" and says prompt injection "may never be totally mitigated" ([NCSC, Dec 2025](https://www.ncsc.gov.uk/blog-post/prompt-injection-is-not-sql-injection)). OpenAI says it is "unlikely to ever be fully 'solved'" ([OpenAI, Dec 2025](https://openai.com/index/hardening-atlas-against-prompt-injection/)). OWASP's prompt injection entry says it is unclear whether fool-proof prevention exists ([OWASP LLM01:2025](https://genai.owasp.org/llmrisk/llm01-prompt-injection/)).
The measured rates point the same way. Anthropic reports a **1.4%** attack success rate for Claude Opus 4.5 in browser use against an adaptive attacker allowed 100 attempts. That is with its improved safeguards, against **8.0%** for the model alone ([Anthropic, Nov 2025](https://www.anthropic.com/research/prompt-injection-defenses)). That is real progress, and 1.4% is still not a boundary on an agent that can move money. NIST's AI safety institute found that new attacks raised hijacking success on held-out tasks from 11% to 81%. Retrying each attack 25 times raised average success from 57% to 80% ([NIST CAISI, Jan 2025](https://www.nist.gov/news-events/news/2025/01/technical-blog-strengthening-ai-agent-hijacking-evaluations)). A 2025 paper put twelve published defences against adaptive attacks and bypassed most of them more than 90% of the time ([Nasr et al., Oct 2025](https://arxiv.org/abs/2510.09023)). Simon Willison's summary of guardrail products that catch 95% of attacks is blunt: in application security, 95% is a failing grade.
Classifiers and guard models still belong in the stack. Google's paper keeps them as a second layer. They are just not the wall.
### The lethal trifecta and the Rule of Two
Two short rules explain when injection becomes theft. Willison's **lethal trifecta** is an agent with access to private data, exposure to untrusted content, and the ability to communicate externally. Combine all three and the agent can be talked into sending your data to an attacker ([Willison, Jun 2025](https://simonwillison.net/2025/Jun/16/the-lethal-trifecta/)). Meta turned that into a design rule, the **Agents Rule of Two**. Within one session, an agent should have no more than two of three properties: it processes untrustworthy input, it can reach sensitive systems or private data, and it can change state or communicate externally. If a task needs all three, the agent should not run autonomously ([Meta, Oct 2025](https://ai.meta.com/blog/practical-ai-agent-security/)). In its guidance on MCP, Gartner tells software leaders to treat any use case that combines sensitive data, untrusted content, and external communication as a no-go zone ([Gartner, Apr 2026](https://www.gartner.com/en/newsroom/press-releases/2026-04-09-gartner-predicts-25-percent-of-all-enterprise-gen-ai-applications-will-experience-at-least-five-minor-security-incidents-per-year-by-2028)).
| The session has | Example | Run it autonomously? |
| --- | --- | --- |
| Untrusted input + private data | Summarise inbound support email into an internal read-only report | Yes, if nothing leaves and nothing is written |
| Untrusted input + external action | A research agent reads the web and posts to a public channel, with no private data | Yes |
| Private data + external action | An internal agent updates the CRM from your own validated forms | Yes |
| All three | An inbox agent reads outside email, searches the CRM, and sends replies | No. Add human approval, or split it into two sessions |
The examples are ours and illustrative, not a ruling from Meta. The rule maps cleanly onto the 3Cs. Curate decides which two properties a session gets. Contain removes the third, most often by cutting the network. Control enforces the switch Meta allows, from a session that read untrusted content to one that has lost its external channel, and it only switches one way.
### Making the agent itself harder to hijack
The 3Cs treat the agent as a box to put walls around. A second line of work changes what is inside the box. A 2025 paper by Beurer-Kellner and thirteen co-authors describes six design patterns for agents that read untrusted input ([Beurer-Kellner et al., Jun 2025](https://arxiv.org/abs/2506.08837)).
- **Action-Selector:** the model picks from a fixed menu of actions, and results never come back to it.
- **Plan-Then-Execute:** the plan is fixed before untrusted data is read.
- **LLM Map-Reduce:** isolated sub-models each handle one untrusted item.
- **Dual LLM:** a privileged planner never sees untrusted text.
- **Code-Then-Execute:** the model writes a program that then runs over the data.
- **Context-Minimization:** spent content is dropped before it can steer later steps.
Google DeepMind's CaMeL is the best-known build of these ideas. A privileged model writes the plan, a quarantined model parses untrusted data, and capability checks run before every tool call. It solved **77%** of tasks in the AgentDojo benchmark with provable security, against **84%** with no defence ([Debenedetti et al., 2025](https://arxiv.org/abs/2503.18813)). That seven-point cost is the trade. Use these patterns where the workflow allows it, and the 3Cs everywhere.
### How the 3Cs line up with the other frameworks
No standards body has adopted the 3Cs. Its value is as a way to organise controls you will end up labelling in someone else's vocabulary. Here is the crosswalk.
| Framework | Its core idea | Contain | Curate | Control |
| --- | --- | --- | --- | --- |
| [OWASP LLM06 Excessive Agency](https://genai.owasp.org/llmrisk/llm062025-excessive-agency/) (2025) | Damage from too much functionality, permission, or autonomy | — | Functionality and permissions | Autonomy: user approval, authorisation enforced downstream |
| [OWASP Top 10 for Agentic Applications](https://genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-2026/) (Dec 2025) | Ten agent risks, ASI01–ASI10, and Least-Agency | Sandboxed execution (ASI05), quarantine (ASI10) | Supply chain (ASI04), Least-Agency | Intent gate before execution (ASI02) |
| [Lethal trifecta](https://simonwillison.net/2025/Jun/16/the-lethal-trifecta/) (Jun 2025) | Private data + untrusted content + external communication | Network isolation removes the third | Remove one at design time | Block exfiltration at runtime |
| [Agents Rule of Two](https://ai.meta.com/blog/practical-ai-agent-security/) (Oct 2025) | At most two of three properties per session | — | Choose the two | One-way switches, human approval when all three |
| [Google's approach for secure agents](https://research.google/pubs/an-introduction-to-googles-approach-for-secure-ai-agents/) (May 2025) | Defined human controllers, limited powers, observable actions | — | A static ceiling on powers | Runtime policy engine, dynamic least privilege |
| [MCP security best practices](https://modelcontextprotocol.io/docs/2026-07-28/tutorials/security/security_best_practices) (2026-07-28) | Rules for tool servers and their tokens | Run local servers sandboxed | Minimal scopes | Token audience checks, no passthrough, per-client consent |
| [MITRE ATLAS](https://github.com/mitre-atlas/atlas-data) | Attacker techniques and mitigations, including agents | M0032 segmentation | M0026–M0028 permission configuration | M0029 human in the loop, M0030 restrict tools on untrusted data |
| [Singapore IMDA agentic framework](https://www.imda.gov.sg/-/media/imda/files/about/emerging-tech-and-research/artificial-intelligence/mgf-for-agentic-ai.pdf) (Jan 2026) | Bound risks, accountable humans, technical controls, informed users | Bound risks upfront | Bound risks upfront | Technical controls and monitoring |
The line between Curate and Control is blurry, even inside Docker. Its blog files tools and credentials under Curate. Its product page files the MCP server allowlist under Control. The split we use is simple. Curate sets the **ceiling**, everything the agent could ever be allowed. Control enforces the **current allowance** inside that ceiling, for this task, after this input. Google's dynamic least privilege lives exactly on that line.
Mapped against OWASP's agentic list, the 3Cs cover most of the ground and leave specific gaps.
| OWASP agentic risk | Main C | What the 3Cs leave to you |
| --- | --- | --- |
| ASI01 Agent Goal Hijack | Control | Detection stays probabilistic. Limit what a hijacked goal can reach |
| ASI02 Tool Misuse and Exploitation | Control, Curate | Per-tool budgets and rate limits |
| ASI03 Identity and Privilege Abuse | Curate | Agent identity as a principal in its own right |
| ASI04 Agentic Supply Chain Vulnerabilities | Curate | Provenance, signing, and a bill of materials |
| ASI05 Unexpected Code Execution (RCE) | Contain | Little. This is containment's home ground |
| ASI06 Memory & Context Poisoning | Curate, partly | Memory outlives the sandbox |
| ASI07 Insecure Inter-Agent Communication | Control | Agent-to-agent authentication |
| ASI08 Cascading Failures | Contain | Blast radius across several agents, not one |
| ASI09 Human-Agent Trust Exploitation | None | Designing approvals people actually read |
| ASI10 Rogue Agents | Contain, Control | Kill switch, revocation, and the evidence to act on |
### What the 3Cs leave out
### Identity: who acted, and on whose behalf
Credentials appear in the 3Cs. Identity does not. Gartner predicts that by 2028, **25% of enterprise security breaches** will trace back to AI agent abuse, and it blames unmanaged machine privileges and legacy credential practices ([Gartner, Oct 2026](https://www.gartner.com/en/newsroom/press-releases/2026-10-20-gartner-identifies-the-top-strategic-technology-trends-for-2027)). CyberArk's 2025 survey counted **82 machine identities for every human** ([CyberArk, Apr 2025](https://www.silicon.co.uk/press-release/machine-identities-outnumber-humans-by-more-than-80-to-1-new-report-exposes-the-exponential-threats-of-fragmented-identity-security)). Gravitee found only 21.9% of organisations treat agents as identity-bearing entities.
Google's first principle is that every agent needs a well-defined human controller, and agents acting for groups need distinct identities. NIST's NCCoE opened a project on agent identity and authorisation in February 2026 ([NIST, Feb 2026](https://www.nist.gov/news-events/news/2026/02/new-concept-paper-identity-and-authority-software-agents)). The MCP specification forbids token passthrough partly because it breaks the audit trail: downstream logs show the wrong identity. The working rule is one identity per agent. It acts with the user's delegated authority, scoped to the task and set to expire. It never shares a key.
### Evidence: prove what happened
Docker is right that a log does not enforce anything. That makes logging a separate obligation, not a weaker form of control. Google's third principle is that actions and planning must be observable. OWASP calls observability non-negotiable. The NCSC asks for full inputs, outputs, and tool calls, and notes that failed tool calls can reveal an attack early. Docker's newer Agent Baseline work, published with Snyk and Keycard, organises itself around three questions. The third is whether you can prove what happened and stop it ([Docker, Oct 2026](https://www.docker.com/blog/from-baseline-to-secure-software-factory-securing-the-systems-that-build-your-software/)).
Our minimum trail on an agent is four fields: input, tools called, data accessed, output. A reviewer who challenges one decision should be able to replay it from those four.
### Supply chain provenance
Curation decides which tools you use. Provenance checks that the thing you chose is the thing that arrived. Amazon's Q Developer extension shipped version 1.84.0 with malicious code after an attacker used a GitHub token with "inappropriately scoped" access in its build setup. The code failed to run, and AWS found no customer resources affected ([AWS-2025-015, Jul 2025](https://aws.amazon.com/security/security-bulletins/AWS-2025-015/)). Microsoft now recommends bills of materials that include tools and MCP servers. OWASP's ASI04 points to AI bills of materials. Neither appears in the 3Cs.
### Memory and persistent context
A sandbox resets. Memory does not. An injected instruction written to long-term memory, or to a retrieval store, reaches every later session. OWASP lists it as ASI06, and MITRE ATLAS added AI agent context poisoning, AML.T0080, in 2025. Microsoft's mitigations: authenticate every memory write, limit which components can read memory, and constrain what a stored item can look like. Curate has to cover what an agent remembers, not only what it is handed.
### Human oversight that actually works
Consent fatigue is real, and the 3Cs answer it by cutting prompts. But Google and Meta both still require a human when an action is irreversible or a session needs all three risky properties. If people rubber-stamp those prompts, the design fails at the one point it depended on them. Microsoft's red team found that bypass was the most exploited failure mode, and recommends auditing the approval UX as a security control in its own right. Singapore's framework asks organisations to check whether their approvals work at all. Keep approvals few, specific, and shown as a diff of what will change.
### Testing that adapts
A framework is a design. It is only true once someone has tried to break it. NIST's hijacking work found that evaluations must adapt to new attacks and be measured over repeated attempts, because a single pass understates risk. Test against an attacker who gets many tries, and repeat the test whenever a prompt, a tool, or a model changes. The metrics that matter for that loop are covered in [what to measure in an AI agent eval suite](https://gigabit.agency/insights/ai-agent-eval-metrics/).
### Accountability and regulation
The 3Cs are an engineering framework. They do not say who answers for the result. Regulators do. The EU's Digital Omnibus, Regulation (EU) 2026/1744, came into force on 27 July 2026. It moved the AI Act's Annex III high-risk obligations to 2 December 2027 and Annex I to 2 August 2028. Article 50 transparency duties still apply from 2 August 2026 ([European Commission](https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai); [Cooley, Aug 2026](https://cdp.cooley.com/digital-ai-omnibus-delays-key-deadlines-introduces-new-rules/)). [ISO/IEC 42001](https://www.iso.org/standard/81230.html) sets out a management system for AI, the place where accountability gets written down. In May 2026, six cyber agencies from the Five Eyes countries issued joint guidance on agentic AI. It calls for layered defence and strict access controls ([Canadian Centre for Cyber Security, May 2026](https://cyber.gc.ca/en/news-events/joint-guidance-careful-adoption-agentic-artificial-intelligence-services)). Name the person who owns each agent's outcome. Without one, the controls have no owner.
### Score one agent against the framework
Pick one agent in production, or about to be. Answer each row honestly. This is our working checklist, not a standard, and a single fail is enough reason to fix before you scale.
| Check | Pass | Fail |
| --- | --- | --- |
| Contain: where does it run? | A sandbox per session, filesystem and network isolated | The developer's laptop, or a shared server with host access |
| Contain: what does it hold? | No production credentials inside the boundary | A long-lived key in an environment variable |
| Curate: which tools? | A named allowlist, pinned and hash-checked | Whatever MCP servers someone installed |
| Curate: which credentials? | One scoped, expiring credential per task | A service-role or admin key, or a key shared across agents |
| Curate: the trifecta | No session has all three risky properties unsupervised | It reads outside content, holds private data, and sends |
| Control: who checks actions? | A gateway outside the model on every tool call | The system prompt says not to |
| Control: can it change its own rules? | No. Policy sits outside its writable space | It can edit its settings or config |
| Identity | Its own identity, acting for a named user | It runs as a person, or as a shared bot account |
| Evidence | Input, tools called, data accessed, output, retained | Chat history only |
| Oversight | Approvals for irreversible actions, shown as a diff | Every action prompts, or none do |
| Testing | Adversarial test with repeated attempts, rerun after changes | A demo that worked once |
### What to do this week
- **Inventory every agent.** Include coding agents on laptops and in CI. For each, list its tools, MCP servers, skills, and credentials. Most teams find more than they expected.
- **Mark the trifecta.** For each agent, note whether it reads untrusted content, reaches private data, and can act or send externally. Anything with all three goes first.
- **Turn on the sandbox you already have.** Claude Code, Codex, and Docker Sandboxes all ship one. Search your scripts and CI for flags that switch permission checks off, and remove them.
- **Replace broad keys.** Swap service-role, admin, and shared keys for scoped, short-lived credentials, one per agent.
- **Pin the tool list.** Allowlist MCP servers and skills by version and hash, and review updates before they load.
- **Put one gateway in front of tool calls.** Log the four fields: input, tools called, data accessed, output.
- **Test it.** Use an outside attacker with many attempts, not your own demo.
The 3Cs are the right instinct: stop asking the model to behave and build walls it cannot argue with. Contain, curate, and control get you most of the way. Identity, evidence, provenance, memory, oversight, and testing finish it. If you want that tested against the agents you actually run, the [AI Security Assessment](https://gigabit.agency/services/ai-security/) is **from $12,000**, fixed, over **2–4 weeks** ([what moves the price, and what other firms charge](https://gigabit.agency/services/ai-security/#cost)). It threat-models and red-teams the path from untrusted input to tool call, then returns graded findings and a remediation plan. Ongoing guardrails and monitoring sit on [Managed AI Operations](https://gigabit.agency/services/managed-ai-operations/) at **$3,000–$20,000 a month**.
### Sources
- Docker: [The 3Cs: A Framework for AI Agent Security](https://www.docker.com/blog/the-3cs-a-framework-for-ai-agent-security/) (Feb 2026), [3Cs framework page](https://www.docker.com/products/ai-governance/3cs-framework/), [Why microVMs](https://www.docker.com/blog/why-microvms-the-architecture-behind-docker-sandboxes/) (Apr 2026), [Sandboxes isolation docs](https://docs.docker.com/ai/sandboxes/security/isolation/), [Agent Baseline](https://www.docker.com/blog/from-baseline-to-secure-software-factory-securing-the-systems-that-build-your-software/) (Oct 2026)
- OWASP: [LLM01:2025 Prompt Injection](https://genai.owasp.org/llmrisk/llm01-prompt-injection/), [LLM06:2025 Excessive Agency](https://genai.owasp.org/llmrisk/llm062025-excessive-agency/), [Top 10 for Agentic Applications](https://genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-2026/) (Dec 2025)
- Google: [An Introduction to Google's Approach for Secure AI Agents](https://research.google/pubs/an-introduction-to-googles-approach-for-secure-ai-agents/) (May 2025); Debenedetti et al., [Defeating Prompt Injections by Design (CaMeL)](https://arxiv.org/abs/2503.18813) (2025)
- Meta: [Agents Rule of Two](https://ai.meta.com/blog/practical-ai-agent-security/) (Oct 2025). Simon Willison: [The lethal trifecta for AI agents](https://simonwillison.net/2025/Jun/16/the-lethal-trifecta/) (Jun 2025)
- Beurer-Kellner et al.: [Design Patterns for Securing LLM Agents against Prompt Injections](https://arxiv.org/abs/2506.08837) (Jun 2025). Nasr et al.: [The Attacker Moves Second](https://arxiv.org/abs/2510.09023) (Oct 2025)
- Government: [UK NCSC on prompt injection](https://www.ncsc.gov.uk/blog-post/prompt-injection-is-not-sql-injection) (Dec 2025); [NIST CAISI on agent hijacking evaluations](https://www.nist.gov/news-events/news/2025/01/technical-blog-strengthening-ai-agent-hijacking-evaluations) (Jan 2025); [NIST NCCoE agent identity concept paper](https://www.nist.gov/news-events/news/2026/02/new-concept-paper-identity-and-authority-software-agents) (Feb 2026); [Five Eyes joint guidance on agentic AI](https://cyber.gc.ca/en/news-events/joint-guidance-careful-adoption-agentic-artificial-intelligence-services) (May 2026); [Singapore IMDA agentic AI framework](https://www.imda.gov.sg/-/media/imda/files/about/emerging-tech-and-research/artificial-intelligence/mgf-for-agentic-ai.pdf) (Jan 2026); [European Commission AI Act page](https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai)
- Protocol and threat models: [MCP security best practices](https://modelcontextprotocol.io/docs/2026-07-28/tutorials/security/security_best_practices) (2026-07-28); [MITRE ATLAS data](https://github.com/mitre-atlas/atlas-data); [Microsoft taxonomy of agentic failure modes, v2](https://www.microsoft.com/en-us/security/blog/2026/06/04/updating-taxonomy-failure-modes-agentic-ai-systems-year-red-teaming-taught-us/) (Jun 2026)
- Vendors on their own agents: [Anthropic, Claude Code sandboxing](https://www.anthropic.com/engineering/claude-code-sandboxing) (Oct 2025); [Anthropic, prompt injection defenses](https://www.anthropic.com/research/prompt-injection-defenses) (Nov 2025); [Anthropic, AI-orchestrated espionage](https://www.anthropic.com/news/disrupting-AI-espionage) (Nov 2025); [OpenAI, hardening Atlas](https://openai.com/index/hardening-atlas-against-prompt-injection/) (Dec 2025); [OpenAI Codex approvals and security](https://learn.chatgpt.com/docs/agent-approvals-security)
- Incidents: [Invariant Labs, tool poisoning](https://invariantlabs.ai/blog/mcp-security-notification-tool-poisoning-attacks) (Apr 2025) and [GitHub MCP](https://invariantlabs.ai/blog/mcp-github-vulnerability) (May 2025); [General Analysis, Supabase MCP](https://generalanalysis.com/blog/supabase-mcp-blog) (Jul 2025); [The Register, Replit](https://www.theregister.com/2025/07/22/replit_saastr_response/) (Jul 2025); [AWS-2025-015](https://aws.amazon.com/security/security-bulletins/AWS-2025-015/) (Jul 2025); [Nx postmortem](https://nx.dev/blog/s1ngularity-postmortem) and [Wiz](https://www.wiz.io/blog/s1ngularity-supply-chain-attack) (Aug 2025); [Embrace The Red, CVE-2025-53773](https://embracethered.com/blog/posts/2025/github-copilot-remote-code-execution-via-prompt-injection/) (Aug 2025); [MSRC, CVE-2025-32711](https://msrc.microsoft.com/update-guide/vulnerability/CVE-2025-32711) (Jun 2025); [Postmark](https://postmarkapp.com/blog/information-regarding-malicious-postmark-mcp-package) (Sep 2025); [Snyk ToxicSkills](https://snyk.io/blog/toxicskills-malicious-ai-agent-skills-clawhub) (Feb 2026); [OpenClaw](https://openclaw.ai/blog/virustotal-partnership) (Feb 2026); [Unit 42](https://unit42.paloaltonetworks.com/openclaw-ai-supply-chain-risk/) (Jun 2026)
- Surveys and forecasts: [Stack Overflow Developer Survey 2026](https://survey.stackoverflow.co/2026/ai); [IBM Cost of a Data Breach 2025](https://newsroom.ibm.com/2025-07-30-ibm-report-13-of-organizations-reported-breaches-of-ai-models-or-applications,-97-of-which-reported-lacking-proper-ai-access-controls); [Gartner, Apr 2026](https://www.gartner.com/en/newsroom/press-releases/2026-04-09-gartner-predicts-25-percent-of-all-enterprise-gen-ai-applications-will-experience-at-least-five-minor-security-incidents-per-year-by-2028) and [Oct 2026](https://www.gartner.com/en/newsroom/press-releases/2026-10-20-gartner-identifies-the-top-strategic-technology-trends-for-2027); [SailPoint](https://www.sailpoint.com/press-releases/sailpoint-ai-agent-adoption-report) (May 2025); [Gravitee](https://gravitee.io/blog/state-of-ai-agent-security-2026-report-when-adoption-outpaces-control) (Feb 2026); [CyberArk, via press release](https://www.silicon.co.uk/press-release/machine-identities-outnumber-humans-by-more-than-80-to-1-new-report-exposes-the-exponential-threats-of-fragmented-identity-security) (Apr 2025)

### FAQ
**What is the 3Cs framework for AI agent security?**
The 3Cs is a framework Docker published in February 2026. Contain bounds the blast radius, so an agent runs inside an isolation boundary and its mistakes stay reversible. Curate defines the agent's environment before it acts: the tools, code, credentials, and context it can reach. Control enforces rules on each action at runtime, before it completes, rather than logging or asking afterwards. The framework's premise is that approval prompts do not scale, so safety has to be structural.
**Is the 3Cs an industry standard?**
No. It is a vendor framework from Docker, which sells sandboxes and an MCP gateway that implement it. It lines up well with open work such as OWASP's LLM06 Excessive Agency and its Top 10 for Agentic Applications, Google's approach for secure AI agents, Meta's Agents Rule of Two, and MITRE ATLAS mitigations. For a customer or auditor, label your controls in those vocabularies, and the NIST AI RMF, ISO/IEC 42001, and EU AI Act where they apply.
**What is the difference between curating and controlling an AI agent?**
Curation sets the ceiling: every tool, credential, and source of context the agent could ever use, fixed before it starts. Control enforces the current allowance inside that ceiling, for this task and after this input, at the moment of each action. A read-only database role is curation. Blocking a send-email call because the agent has just read an untrusted web page is control. Even Docker's own pages draw the line slightly differently.
**Does a sandbox stop prompt injection?**
No. A sandbox limits what a hijacked agent can damage. It does not stop the hijack. If the sandbox holds private data and has network access, an injected instruction can still move that data out. That is why containment has to come with curation (what is inside the box) and control (what may leave it). Prompt injection itself is not solved. The UK NCSC says it may never be totally mitigated.
**What are the lethal trifecta and the Agents Rule of Two?**
Simon Willison's lethal trifecta is an agent with access to private data, exposure to untrusted content, and the ability to communicate externally. With all three, it can be tricked into sending your data to an attacker. Meta's Agents Rule of Two turns that into a rule. In one session, an agent should have no more than two of these: processing untrustworthy input, reaching sensitive systems or private data, and changing state or communicating externally. If it needs all three, it should not run without supervision.
**What does the 3Cs framework leave out?**
Agent identity and attribution, an audit trail as its own obligation, supply-chain provenance, poisoned memory that outlives the sandbox, human approvals designed so people actually read them, adaptive adversarial testing, and named accountability. Those come from Google's secure-agent principles, OWASP's agentic list, NIST's agent identity and hijacking work, Microsoft's red-team taxonomy, ISO/IEC 42001, and the EU AI Act.

## What does an ERP implementation cost in 2026? A real budget breakdown
URL: https://gigabit.agency/insights/erp-implementation-cost/
Published 2026-10-18, updated 2026-10-22

Here is the part most ERP vendors keep behind a "request a quote" form: the bill for an ERP implementation is mostly *services* — configuring, migrating, integrating, training, and going live — not the software licence. For a mid-market company that services line regularly reaches six figures. A lean rollout for a smaller operation starts far below that: a phased [Odoo implementation](https://gigabit.agency/platforms/odoo/) of a few core apps is **from $15,000** ([our published platform pricing](https://gigabit.agency/pricing/)). The licence everyone fixates on is almost never the line that blows the budget. Knowing where the money actually goes is the difference between an ERP that pays for itself and one that becomes the reason you are still running the business out of spreadsheets.
And the budget does blow — often. Gartner predicts that **by 2027 more than 70% of recently implemented ERP initiatives will fail to fully meet their original business case goals, and as many as 25% will fail catastrophically** ([Gartner](https://www.gartner.com/en/information-technology/glossary/enterprise-resource-planning-erp)). The usual causes are not software problems: weak change management, poor data migration, and implementation teams that have not run this kind of rollout before. An ERP budget is really a budget for the discipline that avoids them.
### How much does an ERP implementation really cost?
The sticker you see — a per-user monthly subscription — is the smallest part. Across platforms the pattern holds: the implementation *services* around the software are three to five times the first-year licence, and the ratio only grows with company size and process complexity. Here is where a typical first-year budget actually goes:
| Budget line | Share of a typical rollout | What it buys |
| --- | --- | --- |
| Software licence / subscription | 15–30% | Per-user, usually annual — the line buyers overestimate |
| Implementation & configuration | 40–60% | Process mapping, setup, workflows, any custom modules |
| Data migration & cleansing | 10–20% | The hidden phase, and the one most often under-scoped |
| Training & change management | 10–15% | The line buyers cut first and regret most |
| Integrations (CRM, payments, BI) | 5–15% | Grows with every system you connect |
Our own figures show the same shape: a phased Odoo implementation starts **from $15,000** and grows with the number of apps, integrations and custom modules, and heavier mid-market platforms such as Microsoft Dynamics 365 Business Central sit higher. The open-source platforms — Odoo Community, ERPNext — remove the licence line but not the implementation line, which is where the real cost lives. That is why two companies on the *same* ERP can pay wildly different amounts: the price tracks your process complexity, your data mess, and your appetite for customization, not the logo on the login screen.
### Why the licence is the cheapest part
ERP pricing pages advertise a per-user, per-month figure because it is the most flattering number they have. For a 50-person company it might total $30,000–$60,000 a year — real money, but a fraction of the first-year total, and the open-source editions charge nothing for the licence at all. What you are actually buying when you hire an implementer is the translation of how your business runs into how the software works: the chart of accounts, the approval workflows, the inventory rules, the tax logic, and the reports your controller needs on day one. Get that wrong and no licence tier saves you — which is exactly why the cheap-licence platforms are not the cheap *projects*.
### Where ERP budgets actually blow up
Not on software — on the human and data work teams defer. Three patterns cause most of the damage:
- **Dirty data moved as-is.** Duplicate suppliers, mismatched units of measure, and an unmapped chart of accounts become duplicate payments, inventory write-offs, and a financial close that will not reconcile — on day one. Data discovery belongs in phase zero, not on cutover weekend.
- **Over-customization.** Every bespoke module is technical debt you pay to maintain through every upgrade. The discipline is to change the process to fit the software wherever the process is not a genuine differentiator.
- **Treating a transformation as a software purchase.** The teams that train users, name an internal owner, and manage the change land inside budget; the ones that "flip the switch" are the ones whose system delivers a fraction of what it promised.
### What a smaller, phased ERP actually costs
You do not need a half-million-dollar program to get the core benefit. For a lean operation, a phased approach keeps the number sane and the risk low. Our own published figures sit at the competitive end of the market: a phased **Odoo implementation from $15,000** (process mapping, configuration of up to four core apps, data migration, training, and go-live support — the Odoo licence is separate), a **CRM set-up on Zoho, HubSpot, or Pipedrive from $6,000**, and a **platform migration from $7,000** with a full redirect map and a rehearsed cutover ([see pricing](https://gigabit.agency/pricing/)). The phasing matters more than the platform: ship finance and the one workflow that hurts most, prove it, then expand — rather than big-banging ten modules and hoping.
### How not to become a statistic
The projects that succeed de-risk before they commit. Three moves do most of the work:
- **Phase it, and fix the price before you start.** A phased [Odoo implementation](https://gigabit.agency/platforms/odoo/) — **from $15,000**, fixed before work begins — ships finance and the one workflow that hurts most first, then expands. You learn the real number on a small scope before you sign the big one.
- **Choose the platform on fit, not hype.** If you are weighing the open-source and SMB options, our [Odoo vs Zoho vs ERPNext comparison](https://gigabit.agency/compare/odoo-vs-zoho-vs-erpnext/) lays out licensing, hosting, and where each one fits — on the vendors' own published facts.
- **Model the return first.** Before you approve a budget, run the hours a workflow burns through our [ROI calculator](https://gigabit.agency/tools/ai-roi-calculator/): hours per week times loaded cost times 52, against a fixed implementation price. If the math does not work at the small scope, it will not work at the large one.
### Is an ERP worth it?
If your team is rekeying the same data across QuickBooks, spreadsheets, and three disconnected tools, usually yes — the payback comes from eliminated double-entry, a faster close, and decisions made on one set of numbers instead of four. But the decision should be made on *your* numbers, not a vendor's ROI brochure. Price the discipline, not just the software: a fixed scope, a real data-migration phase, trained users, and an internal owner cost more up front and are the reason the projects that succeed, succeed. The cheapest ERP is the one you implement once.
For real figures instead of ranges, see [what ten client projects cost](https://gigabit.agency/proof/project-costs/): the build price, the monthly cost after launch, the timeline, and the payback, taken from our published case studies.

### FAQ
**How much does an ERP implementation cost for a mid-sized company?**
Most of the cost is services — configuration, data migration, integration, and training — not the licence, and for a mid-market company that services line regularly reaches six figures. Phased rollouts cost far less: a phased Odoo implementation of a few core apps starts from $15,000. The spread is driven by process complexity, data quality, and how much you customize — not the software brand.
**Why do ERP projects go over budget?**
Because the cost is not in the licence — it is in the services and the risk of a project that misses. Gartner predicts that by 2027 more than 70% of recently implemented ERP initiatives will fail to fully meet their original business case goals. The usual causes — change management, data migration, and inexperienced implementation teams — are organizational, not technical. The expensive outcome is a rushed go-live that has to be reworked.
**Is the ERP software licence the biggest cost?**
No. The licence is usually 15–30% of a first-year budget; implementation, configuration, data migration, and training make up the rest. Open-source platforms like Odoo Community and ERPNext remove the licence line entirely but not the implementation line, which is where most of the money and risk live.
**How do I reduce the risk of an ERP implementation?**
De-risk before you commit: phase the rollout at a fixed price rather than big-banging every module at once, treat data cleansing as a phase-zero deliverable, keep customization to genuine differentiators, and name an internal owner to drive change management. Model the ROI on your own hours and costs before approving the budget.

## How much does an AI agent cost in 2026? A real budget breakdown
URL: https://gigabit.agency/insights/ai-agent-cost/
Published 2026-06-20, updated 2026-10-10

Here is the honest version most vendors won't put in writing: a production AI agent is a flat fee — **from $8,000 for a standard agent up to about $150,000** for a complex, high-reliability build, scoped and priced before work starts — [our published prices](https://gigabit.agency/pricing/) — and the model bill everyone worries about is the *cheapest* line in the whole budget. Processing ten thousand support conversations on a fast model runs roughly **$37** in tokens ([Anthropic's own pricing](https://platform.claude.com/docs/en/about-claude/pricing)). The money goes somewhere else entirely, and knowing where is the difference between an agent that pays for itself and one that joins the graveyard.
Because most AI spend right now is being wasted. **95% of enterprise generative-AI pilots deliver no measurable P&L impact** ([MIT Project NANDA, 2025](https://www.forbes.com/sites/jasonsnyder/2025/08/26/mit-finds-95-of-genai-pilots-fail-because-companies-avoid-friction/)), and Gartner expects **more than 40% of agentic-AI projects to be canceled by the end of 2027** — largely on escalating cost and unclear value ([Gartner, 2025](https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027)). Adoption is nearly universal — **91% of mid-market firms now use generative AI, but only 25% have it integrated into core operations** ([RSM, 2025](https://rsmus.com/newsroom/2025/middle-market-firms-rapidly-embracing-generative-ai-but-expertise-gaps-pose-risks-rsm-2025-ai-survey.html)). The gap between *using* AI and *getting paid by* AI is an engineering problem — and that's where the real cost lives.
### How much does it cost to build an AI agent?
A production AI agent — one agent, one high-value workflow, deployed in your stack with evals and observability — is a flat fee, **from $8,000 for a standard agent up to about $150,000** for the most complex builds, and takes about 90 days. A fixed-price diagnostic that ends in a deployed pilot is **$25,000**. Those are [our published prices](https://gigabit.agency/pricing/), and they track what serious mid-market engagements cost.
The spread between $50K and $150K isn't padding — four things move it:
- **Integration surface.** An agent that reads one system and writes to another is the floor; one orchestrating five tools with human-approval steps is the ceiling.
- **Data readiness.** If the knowledge the agent needs lives in clean, queryable systems, you save weeks. If it lives in inboxes and PDFs, budget for the pipeline — poor data quality is the top obstacle data leaders name for moving pilots to production ([Informatica CDO Insights, 2025](https://www.informatica.com/blogs/cdo-insights-2025-global-data-leaders-racing-ahead-despite-headwinds-to-being-ai-ready-latest-survey-finds.html)).
- **Risk profile.** A customer-facing agent needs the guardrails, evals, and escalation paths an internal tool doesn't.
- **Operating bar.** "Works in a demo" and "runs unattended at 2am" are different products.
### What does it cost to run once it's live?
Less than you'd guess — for a mid-market workflow, inference is usually cents per transaction, not dollars. Current published per-million-token API prices from the major model vendors:
| Model tier | Input / 1M tokens | Output / 1M tokens |
| --- | --- | --- |
| Claude Haiku 4.5 (high-volume) | $1 | $5 |
| Gemini 2.5 Pro | $1.25 | $10 |
| GPT-5.6 (flagship) | $2.50 | $15 |
| Claude Sonnet 5 (production default) | $3 | $15 |
Sources: [Anthropic](https://platform.claude.com/docs/en/about-claude/pricing), [Google](https://ai.google.dev/gemini-api/docs/pricing), and [OpenAI](https://developers.openai.com/api/docs/pricing) pricing pages, 2026. At roughly 3,700 tokens per conversation, processing **10,000 support tickets on Haiku costs about $37** ([Anthropic](https://platform.claude.com/docs/en/about-claude/pricing)) — and two built-in levers cut it further: prompt caching drops repeat-context reads to a tenth of the price, and batch processing halves it.
That line item is also falling off a cliff. By independent measurement, the price to reach a given capability level has dropped **roughly 40× per year** ([Epoch AI, 2025](https://epoch.ai/data-insights/llm-inference-price-trends)). Whatever inference costs today, it costs less next quarter. Which is exactly why budgeting an agent around the token bill is a mistake: **the durable cost is the engineering, integration, and operation around the model** — not the model call.
After launch, models drift, vendors ship breaking changes, and edge cases accumulate. An agent without [managed operations](https://gigabit.agency/services/managed-ai-operations/) degrades quietly until someone notices a month of bad outputs. Budget the retainer — **$3,000–$20,000 a month** — the way you'd budget hosting for software: the cost of *owning* the system, not an optional add-on.
### Where teams actually overspend
Not on tokens — on failed builds. When **95% of pilots return nothing** ([MIT, 2025](https://www.forbes.com/sites/jasonsnyder/2025/08/26/mit-finds-95-of-genai-pilots-fail-because-companies-avoid-friction/)) and **40%+ of agentic projects get canceled** ([Gartner, 2025](https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027)), the dominant cost in the market is money spent on AI that never reached production. The same MIT study found the pattern that avoids it: **buying from or partnering with specialized vendors succeeded about 67% of the time, versus internal builds succeeding roughly a third as often.** The expensive path is the demo-grade build that collapses under real load and gets rebuilt — you pay twice. (We wrote the engineering version of this in [why your AI pilot never reached production](https://gigabit.agency/insights/why-your-ai-pilot-never-reached-production/).)
### Is an AI agent worth it for a mid-sized company?
If a workflow consumes 15+ hours a week of paid time, usually yes — but model it before you commit. Our [AI ROI calculator](https://gigabit.agency/tools/ai-roi-calculator/) turns hours-per-week and loaded cost into an annual savings range and a payback period. A workflow burning 20 hours a week at $75/hour costs about **$78,000 a year**; at 50–70% automatable, a $90K agent pays back inside two years. For scale context, IDC puts the average enterprise return at **$3.70 per $1 invested** in generative AI — though that's a [Microsoft-sponsored study](https://blogs.microsoft.com/blog/2024/11/12/idcs-2024-ai-opportunity-study-top-five-ai-trends-to-watch/), so treat it as an optimistic ceiling, not a floor.
### How to keep the cost down
- **Scope to one workflow.** Cost and failure risk both scale with surface area. One agent, one painful workflow, in production — then expand.
- **Match the model to the job.** Most steps don't need the flagship; a cheaper tier handles high-volume classification and extraction at a fraction of the price (see the table above).
- **Buy the pattern, build the edge.** Partnering beat internal builds roughly 2:1 on success ([MIT, 2025](https://www.forbes.com/sites/jasonsnyder/2025/08/26/mit-finds-95-of-genai-pilots-fail-because-companies-avoid-friction/)); reserve custom engineering for what's genuinely differentiated.
- **Gate on evals.** An eval suite is the cheapest insurance against paying twice — it catches regressions before they reach production ([how we build them](https://gigabit.agency/insights/ai-agent-eval-metrics/)).
- **De-risk with a fixed-price diagnostic first.** A [two-week Sprint](https://gigabit.agency/services/ai-transformation/) tells you the real number before you commit — and credits toward the build it scopes.
The short version: the token bill is a rounding error and getting cheaper; the build is real money; the *failed* build is the most expensive thing you can buy. Price the discipline that avoids it. If AI is the core of your product and you can hire senior AI engineers, building in-house wins long-term — we say so in our [build vs buy vs embed guide](https://gigabit.agency/compare/build-vs-buy-vs-embed/). Otherwise, [one production agent](https://gigabit.agency/services/ai-agents/) on the workflow that's costing you now is the move, and the [ROI calculator](https://gigabit.agency/tools/ai-roi-calculator/) will tell you whether the math works before you spend a dollar.
For real figures instead of ranges, see [what ten client projects cost](https://gigabit.agency/proof/project-costs/): the build price, the monthly cost after launch, the timeline and the payback, taken from our published case studies.

### FAQ
**How much does it cost to build an AI agent?**
A production AI agent — one agent in one high-value workflow, with evals, observability, and guardrails — is a flat fee from $8,000, scoping up to about $150,000 for the most complex builds, and takes about 90 days. A fixed-price two-week diagnostic that ends in a deployed pilot is $25,000. The main cost drivers are integration surface, data readiness, and how high the reliability bar needs to be — not the model bill.
**What does it cost to run an AI agent after launch?**
Inference is usually the cheapest line — cents per transaction (processing ~10,000 support conversations on a fast model runs about $37 at 2026 token prices), and it keeps falling ~40× per year. The real running cost is operations — monitoring, eval regressions, drift management — which runs $3,000–$20,000/month managed, depending on how many systems you operate.
**Why do AI agent projects go over budget?**
Because the money isn't in tokens (cents per transaction) — it's in the build and the risk of a build that fails. 95% of enterprise genAI pilots deliver no measurable P&L impact (MIT, 2025) and Gartner expects 40%+ of agentic projects to be canceled by 2027 on cost and unclear value. The expensive outcome is a demo-grade build that collapses under real load and has to be rebuilt.
**Is an AI agent worth it for a mid-sized company?**
If a workflow consumes 15+ hours a week of paid time, usually yes: at 50–70% automation, payback on a fixed-price build typically lands inside 6–18 months. Model it with real numbers before committing — hours per week × loaded hourly cost × 52, then compare against a fixed engagement price.

## How ChatGPT decides which companies to recommend
URL: https://gigabit.agency/insights/how-ai-assistants-choose-recommendations/
Published 2026-06-22

When someone asks ChatGPT, Perplexity, or Google AI Overviews *"who should I hire for X"*, the assistant doesn't consult a ranking. It **assembles an answer** — retrieving a handful of sources it can defend, extracting claims it can quote, and naming the companies those sources make legible. If your company isn't retrievable, quotable, and verifiable, you're not in the answer. There is no page two.
Three layers decide who gets named. **Retrieval**: the engine's search layer (for ChatGPT, OpenAI's own index plus search partners including Microsoft; Google's index for AI Overviews; Perplexity's own crawl) has to surface a page about you for the query's intent — which is classic search discipline: crawlable pages, clean titles, real topical coverage. **Extraction**: the model lifts claims it can repeat with confidence — named prices, concrete numbers, self-contained answers. Vague marketing copy dies here; a sentence like 'a Gigabit Agents build is a flat fee starting at $8,000, with evals and 30 days of tuning included' survives being quoted. **Verification**: the model cross-checks entities — consistent naming, schema markup, reviews, a coherent footprint across your site, LinkedIn, and directories.
This is why pages built for skimming humans often lose to pages built for citation. The formats that win are consistent: **published pricing** (almost nobody does it — which is why it gets quoted), honest comparison pages, answer-first FAQs where the response is complete in one paragraph, and structured data that tells the machine exactly what you sell, where you operate, and what you charge. We wrote the full playbook in the [GEO field guide](https://gigabit.agency/insights/generative-engine-optimization-field-guide/).
The mistake most teams make is treating this as a copywriting exercise. It's a **measurement discipline**. You can't improve a citation rate you don't track: which buying questions matter in your category, which assistants your buyers use, how often you're named, and who gets named instead. That number is the KPI — everything else is a lever.
You can get your baseline in minutes: the [free AI citation audit](https://gigabit.agency/tools/geo-citation-audit/) tests how the major assistants answer your category's buying queries and shows where you're absent. From there, [Ascent](https://gigabit.agency/services/geo/) is the ongoing motion — structured-data engineering, citation-bait content, entity grounding — measured monthly on the citation rate itself, not on proxies.
The window matters more than the tactics. Most categories still have **no serious GEO players** — the firms that get their pricing, proof, and entities machine-legible in the next year will be the default answers their competitors have to dislodge later. Being early here is structural advantage, the way early SEO was in 2005.

### FAQ
**How do I get my company recommended by ChatGPT?**
Make your company retrievable, quotable, and verifiable: publish pages that directly answer your category's buying questions, name concrete facts (prices, outcomes, specifics) the model can quote, add schema markup, and keep your entity footprint consistent across your site, LinkedIn, and directories. Then measure your citation rate on real buying queries and iterate.
**Is GEO different from SEO?**
They share the retrieval layer — you still need crawlable, relevant pages. GEO adds extraction and verification: the answer engine must be able to lift self-contained claims from your pages and cross-check that your company is who it says it is. SEO competes for position on a list; GEO competes to be quoted inside a single answer.
**How do I measure AI search visibility?**
Define the buying questions in your category, run them across ChatGPT, Perplexity, and Google AI Overviews on a schedule, and record how often your brand is named or cited versus competitors. That percentage is your citation rate — the KPI a GEO program is measured on.

## AI citation rate: what counts as a good one?
URL: https://gigabit.agency/insights/what-is-a-good-ai-citation-rate/
Published 2026-06-25, updated 2026-10-22

**There is no single good AI citation rate, because the number depends on the questions you ask.** We measured one. We put 18 buying questions about AI agencies through Google AI Mode three times each, 54 answers in all. The most-named firm appeared in 20% of them, and no firm held more than 3 of the 18 questions in every run. A rule of thumb that says 30% means you are winning does not describe this category.
AI citation rate is the share of answers on a fixed panel of buying questions that name your company. Six answers naming you out of twenty is 30%. The arithmetic is easy. What makes the number mean anything is the panel, the engine and the number of runs behind it. An earlier version of this page gave three bands for it with no source; the section "What we got wrong before" says what we did about that.
### What we measured
- **Panel:** 18 buying questions about AI agencies: shortlists, offers, prices, regions, platforms and verticals. Questions that contain a brand name are out, because they measure recognition. The full list is at the end.
- **Engines:** Google AI Mode, run three times per question (54 answers), and Bing Copilot Search, which gave the same answer to repeated runs and so counts once per question (15 answers). ChatGPT and Perplexity answer only after sign-in, so they are not in this edition.
- **When and where:** 22 October 2026, logged out, from Bangladesh.
- **What counts:** a firm is named when it appears in the answer. Different spellings of one firm (Intercom, Fin by Intercom) are merged with a published list; nothing else is merged or dropped.
### What the numbers show
| | Google AI Mode | Bing Copilot Search |
| --- | --- | --- |
| Questions that got an AI answer | 18 | 15 |
| Answers counted | 54 | 15 |
| Firms named per answer | median 5, range 0 to 10 | median 4, range 0 to 12 |
| Different firms named | 94 | 66 |
| Firms named in one answer only | 31 (33%) | 61 (92%) |
| Most-named firm | AY Automate, 20% (11 of 54) | 5 firms tied at 13% (2 of 15) |
| Top five firms' share of all mentions | 15% | 14% |
Answers name a lot of firms and almost no firm is named often. On Google, 94 different firms shared 54 answers, and 33% of them were named in a single answer. On Bing, 5 firms tied for first at 2 of 15 answers (Accenture, Deloitte, IBM, LowCode Agency and RTS Labs) and 61 of 66 names appeared once.
| Firm (Google AI Mode) | Answers naming it | Rate | Questions named | Questions held in every run |
| --- | --- | --- | --- | --- |
| AY Automate | 11 of 54 | 20% | 4 | 3 |
| Intercom | 9 of 54 | 17% | 3 | 3 |
| Gigabit (the author) | 9 of 54 | 17% | 3 | 3 |
| Neurons Lab | 6 of 54 | 11% | 3 | 1 |
| Addepto | 6 of 54 | 11% | 2 | 2 |
| Ada | 4 of 54 | 7% | 2 | 1 |
| Lindy | 3 of 54 | 6% | 2 | 0 |
"Questions named" is the number of the 18 questions on which the firm appeared at least once. "Held" means it appeared in every run of that question. The firm list includes software vendors such as Intercom as well as agencies, because the engines named both.
### What good means on this evidence
**Good is relative to the leader on your own panel.** On the Google panel the leader was named in 20% of answers, and the typical firm was named in 3 answers or fewer, which is one question held or a stray mention. Being named in about one answer in five put a firm at the top. A target set in the abstract, before you know who the leader is on your questions, tells you nothing.
**The rate depends on how narrow the panel is.** On 17 of the 18 questions at least one firm was named in every run, and 48 different firms held at least one question. The same firms score 20% or less across the whole panel. A firm that owns one narrow question has a 100% rate on it and a rate of 6% on a panel of 18. So a citation rate quoted without its panel cannot be compared with anything.
**One run proves little.** When we asked the same question three times, the sets of firms named overlapped by 68% on average, and 36 of 107 firm-and-question appearances (34%) showed up in only one run. Bing repeated itself exactly, which is a cache, not stability. Run each question at least three times, spread out in time, and track the questions you hold as well as the rate.
### What we got wrong before
This page, and three other Gigabit articles, used to say that unoptimized brands sit at 0 to 10%, that a deliberate program reaching 30 to 40% is winning its category, and that past 50% you are the default answer. Those bands had no source and no sample. The data above does not support them: on this panel nobody reached 30%, and the leader was at 20%. We have removed the bands from all four pages.
### Where Gigabit stands
Gigabit wrote this panel and is the author of the study, so it is shown but never counted as the leader or in the medians. The panel follows our service lines (pilots, managed operations, embedded engineers, prototype rescue, AI security, GEO, n8n, Odoo, Shopify), which favours us. Gigabit was named on 3 questions (p03, p05, p09) in every run, 9 of 54 answers (17%). That ranks second by answers, behind AY Automate and level with Intercom. It was not named on the other 15 questions. Read that as an example of what a strong result looks like on a panel you chose, not as a ranking of agencies.
### How to report your own number
1. **Write the panel down.** 15 to 25 questions a buyer in your category really asks, with no brand names. Publish it with the rate.
2. **Run each question at least three times per engine,** logged out or on a clean account, and note the date and country.
3. **Report four numbers:** the share of answers that name you, the questions where you appear in every run, the leader's rate on the same panel, and how many runs sit behind it.
4. **Repeat monthly** on the same panel and watch the questions you hold, not any single answer.
The [free citation audit](https://gigabit.agency/tools/geo-citation-audit/) runs this for your category, and the [field guide](https://gigabit.agency/insights/generative-engine-optimization-field-guide/) covers the work that moves it. [Ascent](https://gigabit.agency/services/geo/) is scoped and reported on this basis.
### Limits
- One category (AI agencies), one day, one country (Bangladesh; Bing localises results to the visitor even with a US setting).
- Two engines. Not ChatGPT, Perplexity or the Copilot app, which need a signed-in account that would bias a brand's own results.
- 18 questions and three runs each. Answers to one question are not independent, so the real sample is 18 questions, not 54 answers.
- Google shows source names, not URLs, so we count names in the answer rather than links.
- Bing showed search results instead of an AI answer for 3 questions; those are not counted.
- The panel was written by the firm being measured.
- The next run is planned for late November 2026, on the same panel.
### The panel
| Question | Intent |
| --- | --- |
| Which agencies build production AI agents for mid-market companies at a fixed price? | shortlist |
| What are the best AI agent development companies in 2026? | shortlist |
| Who offers a fixed-price two-week AI pilot or AI transformation sprint? | offer |
| How much does it cost to have an agency build a custom AI customer support agent? | price |
| Which companies offer managed AI operations for production AI agents, and what do they charge? | offer |
| Recommend an agency that can take an AI prototype built in Lovable or Bolt to production. | offer |
| What are the best generative engine optimization (GEO) agencies for getting cited by ChatGPT? | shortlist |
| Who does AI security assessments or LLM red teaming for startups, and how much does it cost? | price |
| Which firms offer embedded AI engineers on a monthly basis instead of staff augmentation? | offer |
| Who are the best n8n agencies or experts for building AI automation workflows? | shortlist |
| Which agencies build HIPAA-compliant AI agents for patient intake? | vertical |
| Which AI development agencies publish their prices? | price |
| What are the best software development companies in Bangladesh for AI projects? | region |
| Is a custom AI support agent cheaper than Intercom Fin? | price |
| Who can build AI agents on Shopify for a DTC brand? | platform |
| Which Odoo implementation partners also do AI automation? | platform |
| Which agencies help SaaS companies automate customer support with AI agents? | vertical |
| Which AI consulting firms for mid-market companies ship working pilots rather than slide decks? | shortlist |
The data is free to reuse under CC BY 4.0: [download every firm named, with its counts](https://gigabit.agency/research/ai-citation-benchmark-2026-10-brands.csv). Please cite it as "Source: Gigabit, AI citation benchmark, October 2026, CC BY 4.0" and link to this page. If you think a row is wrong, write to hello@gigabit.agency.

### FAQ
**What is AI citation rate?**
AI citation rate is the share of answers on a fixed panel of buying questions that name your company in an AI assistant's answer. Six answers naming you out of twenty is 30%. It is only comparable when the panel, the engine and the number of runs are stated with it.
**What is a good AI citation rate?**
It is relative to the leader on your own panel. On our panel of 18 questions about AI agencies, the most-named firm appeared in 20% of 54 Google AI Mode answers and no firm held more than 3 questions in every run. A fixed band such as 30% has no source behind it.
**How many times should I run each question?**
At least three, spread out in time. Running one question three times on Google AI Mode gave firm lists that overlapped by 68% on average, so one run is noise. Bing repeats a cached answer, so space those runs by hours.
**Does a 100% rate on one question count?**
It counts for that question only. On 17 of 18 questions at least one firm was named in every run, so a high rate on a narrow question is common. Report the questions you hold alongside the rate across the whole panel.

## The production AI agent stack: what we actually deploy
URL: https://gigabit.agency/insights/production-ai-agent-stack/
Published 2026-06-18

Every production agent we ship has the same six layers: **a model tier, orchestration, retrieval, evals, observability, and guardrails**. The model gets all the attention; the other five decide whether the system survives contact with real work. Demos have one layer. Production has six.
**Models** are a portfolio decision, not an allegiance. Frontier models where judgment quality is the constraint; smaller, faster models for classification and routing where latency and unit cost dominate. We're deliberately vendor-neutral — the stack is designed so a model swap is a config change plus an eval run, not a rewrite. When a vendor ships a breaking change (they do, regularly), that neutrality is what turns a fire drill into a Tuesday.
**Orchestration** is where the workflow lives: state machines for processes with defined steps and approval gates, agentic loops only where genuine judgment is required. The unfashionable truth is that most business workflows want *less* autonomy than the demos suggest — a reliable eight-step process with two human checkpoints beats an impressive free-roaming agent that's right 80% of the time.
**Retrieval** grounds the agent in your reality: your docs, your tickets, your data, fetched at answer time instead of memorized at training time. This is where [data readiness](https://gigabit.agency/insights/ai-agent-cost/) shows up in the budget — clean sources make this layer a week; inboxes and PDFs make it a month.
**Evals are the layer that separates firms that operate AI from firms that demo it.** Before an agent ships, we build a scored test set from real cases — the awkward ones included — and every change afterward (prompt, model, retrieval) runs against it. No green run, no deploy. It's the same discipline we described in [the five gates](https://gigabit.agency/insights/why-your-ai-pilot-never-reached-production/), and it's the reason 'the model got worse last week' becomes a measurable event instead of a vibe.
**Observability and guardrails** close the loop: every action logged with full context, cost and latency tracked per run, drift surfaced on dashboards; permissions bounded per step, low-confidence states escalating to humans, rollback paths rehearsed. This layer is most of what a [Managed AI Operations](https://gigabit.agency/services/managed-ai-operations/) retainer does all day — because agents don't fail loudly, they fail quietly, and only instrumentation notices quietly.
If you're scoping your first build, the stack above is what 'done' should mean in the statement of work — it's what we deploy in a [Gigabit Agents](https://gigabit.agency/services/ai-agents/) build, and it's why the flat fee includes evals and observability rather than treating them as change orders. Anything less isn't a production agent; it's a demo with your logo on it.

### FAQ
**What does a production AI agent stack include?**
Six layers: a model tier (often multiple models routed by task), orchestration (state machines or agentic loops), retrieval over your own data, an eval suite run on every change, observability (logging, cost, drift), and guardrails (bounded permissions, human escalation, rollback). A demo has the first layer; production needs all six.
**Which LLM is best for AI agents?**
The wrong question — production systems route by task. Frontier models where judgment quality binds, smaller models where latency and cost dominate, evaluated against your own test set rather than public benchmarks. Design for a model swap being a config change, because vendors ship breaking changes regularly.
**Why do AI agents fail in production?**
Silent drift after vendor model updates, compounding errors across multi-step workflows, unhandled low-confidence states, and cost blowouts at volume. All four are engineering problems solved by evals, bounded autonomy, escalation paths, and per-run instrumentation — not by better prompts.

## AI support automation for SaaS: the playbook behind 60%+ autonomous resolution
URL: https://gigabit.agency/insights/ai-support-automation-saas/
Published 2026-06-24, updated 2026-10-22

A well-built support agent resolves **60–70% of tickets autonomously while CSAT goes up, not down**. In one of our deployments, an AI support agent reached [64% autonomous resolution and saved $217K a year](https://gigabit.agency/proof/ai-automation-ecommerce-support/) — with satisfaction rising from 3.9 to 4.5. That result isn't a model capability; it's a system design. Here's the design.
### Why support is usually the first agent
Not every workflow is a good first agent. The ones that pay off fastest in a B2B SaaS company share a shape: high volume, clear success criteria, and a cost someone on the leadership team already feels. Support triage usually wins on all three: the volume is there, the right answer is knowable, and the time saved is easy to measure. Onboarding comes next, because getting a new customer to first value faster moves retention in a way the whole business notices. Renewal-risk signals are higher-leverage but need more data plumbing, so they come after the easier wins have earned trust. Time-to-value is the metric. The exciting build is more fun on day one and far more likely to [stall in pilot purgatory](https://gigabit.agency/insights/why-your-ai-pilot-never-reached-production/).
### The support agent, tier by tier
Start with the **ticket taxonomy, not the technology**. Pull ninety days of tickets and sort them into three tiers. Tier 1: deterministic answers — password resets, billing lookups, how-do-I questions with documented answers. Typically 40–55% of volume. Tier 2: contextual work — account-specific issues needing data from your systems but following knowable patterns. Another 15–25%. Tier 3: judgment calls, angry customers, edge cases — **deliberately routed to humans, fast**. The agent's job on Tier 3 is a clean handoff with full context, not a heroic attempt.
The counterintuitive rule: **the escalation path is the product**. Teams fixate on how many tickets the agent closes; customers remember the one it shouldn't have touched. Confidence thresholds, a visible 'get me a human' exit, and context-rich handoffs are why CSAT rises — people tolerate a bot answering instantly, and they punish a bot that traps them.
Grounding is the second non-negotiable. The agent answers **from your docs, your account data, and your policy pages** — retrieved at answer time — never from the model's general knowledge. Retrieval plus citation of internal sources is what turns 'plausible answer' into 'correct answer', and it's the layer that separates this architecture from a chatbot wrapper, as we detailed in [the production agent stack](https://gigabit.agency/insights/production-ai-agent-stack/).
Then evals, before launch and forever: a scored test set built from real historical tickets — including the nasty ones — run on every prompt, model, or policy change. Resolution rate, accuracy against known-correct answers, escalation precision, and tone. When a model vendor ships an update that subtly changes behavior, the eval run catches it before your customers do. That ongoing discipline is what a [Managed AI Operations](https://gigabit.agency/services/managed-ai-operations/) retainer exists for.
Timeline and economics, honestly: a support agent of this shape is a classic [Gigabit Agents](https://gigabit.agency/services/ai-agents/) build at a flat fee from $8K, with the taxonomy work in weeks 1–2 and the timeline fixed in the written scope before work starts. For a SaaS team spending 60+ hours a week on support, the [ROI math](https://gigabit.agency/tools/ai-roi-calculator/) usually clears in under a year; the [$25K Sprint](https://gigabit.agency/services/ai-transformation/) is the two-week version that proves it on your own ticket data before you commit. More on how we work with product teams on the [AI for SaaS](https://gigabit.agency/industries/saas/) page.

### FAQ
**Which workflow should a SaaS company automate with AI first?**
The one with high volume, clear success criteria, and a cost a leader already feels, which is usually support triage. The volume is there, the right answer is knowable, and the time saved is easy to measure, so the first deployment funds the next. Onboarding and renewal-risk signals come after, because they need more data plumbing and more trust.
**What percentage of support tickets can AI resolve?**
A production-grade support agent typically resolves 60–70% of tickets autonomously: 40–55% deterministic answers plus 15–25% contextual account-specific work. The remaining tier — judgment calls and upset customers — should be deliberately routed to humans with full context, which is why well-built deployments raise CSAT rather than lower it.
**Will AI support automation hurt customer satisfaction?**
Only if the escalation path is an afterthought. Deployments that pair instant autonomous answers with confidence-based escalation and a visible path to a human tend to raise CSAT — in one of our deployments, from 3.9 to 4.5 while resolving 64% of tickets autonomously.
**How long does it take to deploy an AI support agent?**
The timeline is fixed in the written scope before work starts. Ticket taxonomy and scoping come first (weeks 1–2), then build and integration, then eval hardening and a monitored ramp from a small traffic slice to majority coverage. A single narrow workflow can go live in the 3-week, $8,000 Quick Win Sprint.

## HIPAA compliant AI: what it takes, which vendors sign a BAA, and what to check
URL: https://gigabit.agency/insights/hipaa-compliant-ai-agents/
Published 2026-06-26, updated 2026-10-23

An AI system is **HIPAA compliant** when every company that handles protected health information (PHI) for it has signed a business associate agreement (BAA), the model and the people around it see only the PHI the task needs, access is tied to named users and service accounts, every access is logged, PHI is encrypted in transit and at rest, retention is decided on purpose, and a person reviews anything clinical before it reaches a patient or a record. No model is HIPAA compliant on its own; the whole system is. Below: what each piece means in practice, which AI vendors sign a BAA and for which products (checked on 23 October 2026), and a checklist for any vendor or build. This is a buyer's guide written by engineers. Your privacy officer and counsel make the final call.
### What makes an AI system HIPAA compliant
HIPAA does not mention AI. It regulates covered entities (providers, health plans, clearinghouses) and their business associates, whether PHI goes to a billing clerk or a language model. Three parts of the rules do most of the work:
1. **The Privacy Rule** decides who may use or see PHI and how much of it. The minimum necessary standard and de-identification live here.
2. **The Security Rule** sets the technical safeguards: access control, audit controls, integrity, authentication and transmission security, in [45 CFR 164.312](https://www.ecfr.gov/current/title-45/subtitle-A/subchapter-C/part-164/subpart-C/section-164.312).
3. **The business associate rules** require a written agreement with every vendor that creates, receives, maintains or transmits PHI on your behalf.
### BAAs: which AI vendors sign one, and for what
A BAA covers named products and configurations, never a whole company. A vendor can sign one for its API while its chat app stays out of scope. This is what each vendor's own pages said on 23 October 2026:
| Vendor | What a BAA can cover | What stays outside it |
|---|---|---|
| OpenAI | The API on BAA-eligible endpoints, with a standard BAA accepted in the API Platform for organisations with API usage history or by request; ChatGPT for Healthcare, ChatGPT Enterprise with a Regulated Workspace and ChatGPT for Clinicians | ChatGPT Business and consumer plans; API web search with live internet access |
| Anthropic | The Claude API with HIPAA readiness switched on, using the standard BAA in the Claude Console or a negotiated one through sales, for features marked eligible | Claude Free, Pro and Max; PHI typed into the Console; Claude Code; beta features not listed as eligible |
| Google Cloud | Products on its covered list, including Gemini Enterprise Agent Platform (formerly Vertex AI), Generative AI on Agent Platform, Document AI and Speech-to-Text | Any service not on the list. Google's page says that if you cannot disable a service, you must not use it for PHI |
| Microsoft Azure | In-scope Azure services under the HIPAA BAA that is part of the Product Terms and Data Protection Addendum, with no separate signature; models sold by Azure, including Azure OpenAI, are governed by the DPA | Services and models outside the published audit scope. Microsoft has not published a model-by-model list, so get written confirmation for each model |
| AWS | Services on the HIPAA Eligible Services Reference, including Amazon Bedrock, Bedrock AgentCore, Transcribe (with HealthScribe) and Comprehend Medical, once an AWS BAA is in place | Services not on the reference list |
Sources: [OpenAI's data controls guide](https://developers.openai.com/api/docs/guides/your-data) and its help-centre articles on BAAs (the help centre blocked our fetch, so we relied on search results quoting it), [Anthropic's API and data retention docs](https://platform.claude.com/docs/en/manage-claude/api-and-data-retention), [Google Cloud's HIPAA page](https://docs.cloud.google.com/docs/security/compliance/hipaa), [Microsoft's Azure HIPAA page](https://learn.microsoft.com/en-us/azure/compliance/offerings/offering-hipaa-us) and the [AWS HIPAA Eligible Services Reference](https://aws.amazon.com/compliance/hipaa-eligible-services-reference/) (last updated 15 October 2026).
Two details catch teams out. **The same model can sit under different contracts.** Anthropic's docs say that when you use Claude through Amazon Bedrock or Google Cloud, the cloud provider processes the data, so its BAA applies. **Features inside a covered product can fall outside the BAA.** Anthropic's API rejects non-eligible features for HIPAA-enabled organisations, and OpenAI lists live web search as not covered. Read the feature table as well as the product name.
The chain also includes the vector database, the tracing and evaluation tools, the error tracker and any review tool. If PHI reaches it, it needs a BAA or it must stop receiving PHI. Gigabit signs a BAA for engagements that touch PHI, because during a build the engineers are one of those hops.
### The minimum necessary standard
The Privacy Rule requires covered entities and business associates to make reasonable efforts to limit PHI to the minimum necessary to accomplish the intended purpose of a use, disclosure or request ([45 CFR 164.502(b)](https://www.ecfr.gov/current/title-45/subtitle-A/subchapter-C/part-164/subpart-E/section-164.502)). [Section 164.514(d)](https://www.ecfr.gov/current/title-45/subtitle-A/subchapter-C/part-164/subpart-E/section-164.514) adds that you must identify which people or classes of people need access, and to which categories of PHI. The standard does not apply to disclosures to a provider for treatment, to the patient, or under an authorisation, among a few other exceptions.
For an AI system, treat each agent as one more class of user. An intake agent needs demographics, insurance and reason for visit. A prior-authorisation assistant needs the order, the diagnosis codes and the payer rules. Neither needs the full chart. In practice:
1. Build the prompt from named fields pulled by a query you control, instead of pasting the whole chart.
2. Give retrieval indexes only the documents and fields the task needs, and filter by patient and user at query time.
3. Strip identifiers the model does not need before the call, and re-attach them afterwards from your own system.
4. Write the field list into the system's documentation, so the minimum necessary decision is on record.
### De-identifying PHI: Safe Harbor or Expert Determination
Data that meets the de-identification standard in [45 CFR 164.514(b)](https://www.ecfr.gov/current/title-45/subtitle-A/subchapter-C/part-164/subpart-E/section-164.514) is no longer PHI, so it can be used for evaluation sets, prompt testing and analytics without a BAA. The rule gives two methods.
**Safe Harbor** removes 18 kinds of identifiers of the patient and of relatives, employers and household members. They include names, places smaller than a state (the first three ZIP digits may stay if that area holds more than 20,000 people), every date element except the year, ages over 89, contact details, record and account numbers, device identifiers, IP addresses, biometrics and full-face photos. You must also have no actual knowledge that what remains could identify someone.
**Expert Determination** uses a person with appropriate statistical and scientific expertise who determines that the risk of re-identification by the anticipated recipient is very small, and documents the methods and results. HHS's [de-identification guidance](https://www.hhs.gov/hipaa/for-professionals/special-topics/de-identification/index.html) says there is no fixed numerical level that counts as very small for every dataset; the expert sets it for the data and the recipient.
Safe Harbor is hard to apply to free text. Notes and call transcripts mention names, dates and places mid-sentence, and a redactor that catches most of them still leaves the rest. For evaluation sets built from notes, budget for Expert Determination or a human check of every record. You may keep a re-identification code under section 164.514(c) as long as it is not derived from patient information and the key stays with you.
### Access controls and audit logs
The Security Rule's access control standard covers "persons or software programs" that have been granted access rights, which describes an AI agent exactly. Each user needs a unique ID, and so should each agent: one service account per agent and environment, scoped to the actions its task needs, with short-lived credentials. Reviewers' sessions should time out, and provider accounts should use multi-factor sign-in.
The audit controls standard ([164.312(b)](https://www.ecfr.gov/current/title-45/subtitle-A/subchapter-C/part-164/subpart-C/section-164.312)) requires mechanisms that record and examine activity in systems that contain or use ePHI. For AI that means logging which user or agent triggered the call, which records were retrieved, the prompt and model version, the tools called, the output, and who approved or changed it. Keep the log inside your BAA-covered environment with restricted access, because a full prompt log is a copy of the PHI. The [production agent stack](https://gigabit.agency/insights/production-ai-agent-stack/) post shows where logging sits.
### Encryption and retention
In the current rule text, encryption at rest and in transit is an addressable specification: you implement it, or document why an equivalent measure is reasonable. HHS [proposed in January 2025](https://www.federalregister.gov/documents/2025/01/06/2024-30983/hipaa-security-rule-to-strengthen-the-cybersecurity-of-electronic-protected-health-information) to make encryption required, and no final rule had been published by 23 October 2026. Treat it as required anyway: TLS on every hop, encrypted databases, storage and backups, and field-level encryption for the most sensitive columns.
Retention needs two decisions. **Your own records:** HIPAA requires security policies and documentation to be kept for six years ([164.316(b)(2)](https://www.ecfr.gov/current/title-45/subtitle-A/subchapter-C/part-164/subpart-C/section-164.316)); medical record retention is set by state law. **The vendor's copies:** a BAA does not always mean zero retention. Anthropic's docs say HIPAA readiness applies safeguards rather than immediate deletion, and that flagged content can be kept for up to two years under any arrangement. OpenAI's guide says BAA-eligible endpoints can be used for PHI even if data is retained. Azure stores flagged prompts for human abuse review unless you are approved for modified abuse monitoring. Stateful features such as threads, files and vector stores keep data until you delete it. Write down where each copy lives and for how long.
### Human review for clinical outputs
HIPAA does not require a human to check AI output. Patient safety does, and so will most health systems' review boards. Draw the line by consequence: an agent may answer a parking question alone, but a visit summary, a coded claim, a medication message or anything entering the chart should be a draft that a named person approves, with the approval logged. Our [human-in-the-loop guide](https://gigabit.agency/insights/ai-agent-human-in-the-loop/) covers how to place those checkpoints without rebuilding the bottleneck.
Evaluations matter more here because the dangerous failure is a fluent, wrong answer entering a record. Build test sets from de-identified cases, set thresholds that block a release, and re-run them on every model change. Our [Managed AI Operations](https://gigabit.agency/services/managed-ai-operations/) team runs that routine monthly, and the results double as your evidence file.
### How this looks in Gigabit's published builds
A [specialty clinic's patient assistant](https://gigabit.agency/proof/clinic-llm-assistant-front-desk/) answers administrative and prep questions only, with PHI redaction, refusal rules and escalation to staff with a summary. It cut repetitive front-desk inquiries 62% in eight weeks. A [14-location medical group's intake system](https://gigabit.agency/proof/healthcare-unified-intake-ehr-analytics/) syncs to the EHR through an HL7/FHIR bridge with least-privilege credentials, audit logs on every write and encryption throughout, and cut intake and admin time 41%. The [telehealth platform](https://gigabit.agency/proof/healthtech-telehealth-platform/) is not an AI system, but it shows the baseline an AI feature sits on: BAA-covered AWS infrastructure, encryption down to the field, CloudTrail audit logs, role-based access, 15-minute timeouts, MFA and a third-party security assessment before launch. The [patient intake integration guide](https://gigabit.agency/insights/ai-patient-intake-ehr-integration/) covers the intake pattern.
### Buyer checklist for any HIPAA compliant AI vendor or build
| Control | What to ask | What a good answer looks like |
|---|---|---|
| BAA chain | Which companies touch PHI, and does each have a BAA covering the exact product and features used? | A written list of every hop, from model provider to logging tool, each with a BAA reference |
| Model configuration | Which endpoints, features and deployment types are in use? | Only features the vendor marks as BAA-eligible; non-eligible features blocked or unused |
| Minimum necessary | Which PHI fields does each prompt and index receive? | A documented field list per agent, enforced by the query that builds the prompt |
| De-identification | How are evaluation and test sets made? | Safe Harbor or a documented Expert Determination; no production PHI in development |
| Access control | How are users and agents identified and limited? | Unique IDs, one scoped service account per agent, MFA, automatic logoff |
| Audit logs | Can you show who or what accessed which record, and what the model produced? | Logs of retrieval, prompt, model version, tool calls, output and approver, kept in a covered environment |
| Encryption | Where is PHI encrypted? | TLS on every hop; encrypted databases, storage, backups and logs |
| Retention | How long does each vendor and each feature keep data? | A written retention map, including flagged-content and stateful-feature exceptions |
| Human review | Which outputs reach a patient or a record without a person approving them? | None that are clinical; approvals recorded in the audit log |
| Evaluation and monitoring | How do you know quality has not dropped after a model change? | Eval sets with release thresholds, re-run on every model or prompt change |
| Incident response | What happens if PHI goes somewhere it should not? | A runbook that connects to your breach assessment process, with named owners |
### Where to start
Pick a workflow with high volume and bounded clinical risk, such as intake, scheduling, prior-authorisation paperwork or front-desk questions, and design the BAA chain and field list before the first prompt. The [AI Transformation Sprint](https://gigabit.agency/services/ai-transformation/) ($25,000 fixed, 2 weeks) scopes the workflow, the data boundaries and the BAA chain and ends in a working pilot. If AI already runs across teams, [AI governance](https://gigabit.agency/services/ai-governance/) builds the inventory, policies and controls around it, and the [AI Security Assessment](https://gigabit.agency/services/ai-security/) tests an existing agent for prompt injection and data leakage. The [AI for Healthcare](https://gigabit.agency/industries/healthcare/) page sets out how we handle PHI, and the [readiness assessment](https://gigabit.agency/tools/ai-readiness-assessment/) shows where your foundations stand.

### FAQ
**What makes an AI system HIPAA compliant?**
A BAA with every company whose product touches PHI, covering the exact products and features in use; PHI limited to the minimum necessary for each task; unique IDs and scoped access for every user and agent; audit logs of what was accessed and produced; encryption in transit and at rest; a deliberate retention plan; and human review of clinical outputs. No model is compliant on its own; the whole system has to be.
**Is ChatGPT HIPAA compliant?**
Only in specific products under a signed BAA. As of 23 October 2026, OpenAI's help centre lists the API (on BAA-eligible endpoints), ChatGPT for Healthcare, ChatGPT Enterprise with a Regulated Workspace and ChatGPT for Clinicians as HIPAA-eligible, and says it does not offer a BAA for ChatGPT Business. Consumer ChatGPT should never receive PHI.
**Does Anthropic sign a BAA for Claude?**
Yes, for the Claude API with HIPAA readiness enabled. Eligible organisations can accept the standard BAA in the Claude Console, or negotiate one through sales. Claude Free, Pro and Max, PHI entered in the Console, Claude Code and non-eligible beta features are outside it. Claude used through Amazon Bedrock or Google Cloud falls under that cloud provider's BAA instead.
**Can we use de-identified patient data with any AI tool?**
Data de-identified under 45 CFR 164.514(b), by Safe Harbor or Expert Determination, is no longer PHI, so HIPAA does not require a BAA for it. Free-text notes are hard to de-identify reliably with Safe Harbor alone, so check every record or use an expert, and keep any re-identification key under your control.
**Does HIPAA require encryption for AI systems?**
In the current Security Rule, encryption is an addressable specification: you must implement it or document an equivalent measure. HHS proposed making it required in January 2025, and no final rule had been published by 23 October 2026. In practice every AI system handling PHI should encrypt it in transit and at rest, including logs and backups.

## Answer-Market Fit: the metric that replaces rankings in the AI era
URL: https://gigabit.agency/insights/answer-market-fit/
Published 2026-07-11

**Answer-Market Fit is the point where AI assistants and buyers both default to recommending you** when someone asks who's best in your category. It's the answer-engine era's counterpart to product-market fit: not a launch metric or a vanity ranking, but the moment the market's first instinct — human or machine — is to name you. Product-market fit means the market wants what you built. Answer-Market Fit means the new front door to that market, the AI assistant, agrees.
### Why it's the new bar
The recommendation moment has moved. When a buyer asks ChatGPT — **800 million weekly active users as of October 2025** ([TechCrunch](https://techcrunch.com/2025/10/06/sam-altman-says-chatgpt-has-hit-800m-weekly-active-users/)) — or reads a Google AI Overview — **2 billion monthly users** ([TechCrunch](https://techcrunch.com/2025/07/23/googles-ai-overviews-have-2b-monthly-users-ai-mode-100m-in-the-us-and-india/)) — they increasingly act on the single answer they're handed, not a page of ten links to compare. And when Google shows an AI summary, only **8% of users click any traditional link — versus 15% without one** ([Pew Research, 2025](https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/)). The click you used to compete for is being replaced by a citation you now have to earn. Being wanted by the market is no longer enough; you have to be the name the market's assistant reaches for.
### Product-Market Fit vs Answer-Market Fit
| | Product-Market Fit | Answer-Market Fit |
| --- | --- | --- |
| The question | Does the market want what we built? | Does the market's AI default to recommending us? |
| The signal | Retention, word of mouth, pull | Citation rate across category buying queries |
| Who decides | Customers | Customers *and* the assistants they ask |
| How you lose it | A better product | A better-cited competitor |
You can have product-market fit and no Answer-Market Fit — a genuinely great product the assistants never name. In a market where buyers ask a machine first and act on its shortlist, that gap isn't cosmetic; it's pipeline walking to whoever the model does name.
### The path to Answer-Market Fit: three gates
An assistant's recommendation isn't magic; it's a pipeline with three gates, and Answer-Market Fit means clearing all three on the queries that carry buying intent.
- **Retrieval** — the engine's search layer has to surface a page about you for the query's intent. Classic discoverability: crawlable pages, clean titles, real topical coverage.
- **Extraction** — the model has to be able to lift a claim it can repeat with confidence: named prices, concrete numbers, self-contained answers. Vague marketing copy dies here.
- **Verification** — the model cross-checks your entity for consistency: matching names, schema markup, reviews, a coherent footprint across your site, LinkedIn, and directories.
Clear all three and you get cited; get cited on the queries buyers actually ask before they buy, and you get chosen. We break the mechanics down in [how AI assistants choose who to recommend](https://gigabit.agency/insights/how-ai-assistants-choose-recommendations/).
### How you know you have it: citation rate
Product-market fit has retention curves; Answer-Market Fit has **citation rate** — the share of your category's real buying queries where an assistant names you, measured on a fixed panel across ChatGPT, Perplexity, and Google AI Overviews and tracked monthly so the trend, not any single stochastic answer, is the signal. There is no universal good number: in our [measured benchmark](https://gigabit.agency/insights/what-is-a-good-ai-citation-rate/) the most-named firm in a category appeared in 20% of answers, so compare yourself with the leader on your own panel.
### Why the window is open now
The Princeton study that named the GEO field found that adding **citations, quotations, and statistics to your content lifts visibility in generative answers by up to 40%** ([Aggarwal et al., KDD 2024](https://arxiv.org/abs/2311.09735)) — the levers are known and mechanical, and in most categories almost no one is pulling them yet. The traffic is worth more, too: Semrush puts an AI-search visitor at roughly **4.4× the value** of a traditional organic one ([Semrush, 2025](https://www.semrush.com/blog/ai-search-seo-traffic-study/)). Early movers lock in Answer-Market Fit while the channel is uncrowded; late movers optimize to be cited in answers a competitor already owns.
### How to reach it
Answer-Market Fit is exactly what [Ascent](https://gigabit.agency/services/geo/) — our 12-month growth program — is built to install and then measure: brand, website, content, SEO, and GEO run by one team against one number, your citation rate. Start with a baseline. The [free AI Citation Audit](https://gigabit.agency/tools/geo-citation-audit/) runs your category's buying queries through the major assistants and shows where you're named, where you're not, and who's named instead. That score is the starting line; everything after it is closing the cheapest, highest-impact gaps first — and the field mechanics live in the [GEO field guide](https://gigabit.agency/insights/generative-engine-optimization-field-guide/).

### FAQ
**What is Answer-Market Fit?**
Answer-Market Fit is the point where AI assistants and buyers both default to recommending you when someone asks who's best in your category. It's the answer-engine era's counterpart to product-market fit — the market's first instinct, human or machine, is to name you — and it's measured by citation rate across your category's real buying queries.
**How is Answer-Market Fit different from product-market fit?**
Product-market fit asks whether the market wants what you built; Answer-Market Fit asks whether the market's AI assistants default to recommending you. You can have one without the other — a strong product the assistants never name. In a market where buyers ask ChatGPT and Google AI Overviews first, that gap is lost pipeline.
**How do you measure Answer-Market Fit?**
By citation rate: the share of your category's real buying queries where an assistant (ChatGPT, Perplexity, Google AI Overviews) names you, measured on a fixed panel and tracked monthly. There is no universal good number: on our measured panel the most-named firm appeared in 20% of answers, so compare yourself with the leader on your own panel.
**How do you achieve Answer-Market Fit?**
By clearing the three gates that decide AI recommendations — retrieval (be discoverable for the query), extraction (offer quotable, concrete claims like named prices and numbers), and verification (keep a consistent, corroborated entity across the web) — for your category's buying queries. Adding citations, quotations, and statistics to content raised generative-answer visibility up to 40% in Princeton's GEO study; the levers are mechanical and most competitors aren't pulling them yet.

## How much does generative engine optimization (GEO) cost?
URL: https://gigabit.agency/insights/how-much-does-geo-cost/
Published 2026-07-11, updated 2026-10-23

**GEO doesn't have a standard sticker price — it's priced one of three ways: a monthly retainer, a fixed project fee, or performance-based on a metric like citation rate.** For most companies it belongs inside a broader growth program rather than bought as a line item, because getting cited by AI assistants depends on the same brand, content, and technical work that grows the rest of the business. Here's what actually moves the number, what agencies charge, and what we charge.
### Why companies pay for GEO at all
Because the recommendation moment moved. Buyers ask ChatGPT — **800 million weekly active users** ([TechCrunch](https://techcrunch.com/2025/10/06/sam-altman-says-chatgpt-has-hit-800m-weekly-active-users/)) — or read a Google AI Overview — **2 billion monthly users** ([TechCrunch](https://techcrunch.com/2025/07/23/googles-ai-overviews-have-2b-monthly-users-ai-mode-100m-in-the-us-and-india/)) — and act on the answer. When an AI summary appears, only **8% of users click any link, versus 15% without one** ([Pew Research, 2025](https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/)). The click you paid SEO to earn is being replaced by a citation — and the traffic that remains converts better, with an AI-search visitor worth roughly **4.4× a traditional organic one** ([Semrush, 2025](https://www.semrush.com/blog/ai-search-seo-traffic-study/)). GEO is what earns that citation.
### The three ways GEO is priced
- **Monthly retainer** — the most common, because GEO is continuous work: content that answers buying queries, entity and schema cleanup, and re-measurement as competitors move. You're buying an ongoing citation-share program, not a one-off.
- **Fixed project** — a one-time audit plus a batch of fixes. Useful to establish a baseline; weak at holding citation share, because a single pass doesn't compound.
- **Performance-based** — priced against an outcome like citation rate. It aligns incentives best, and it's the direction serious programs are moving.
### What agencies charge
Published GEO prices are scarce, because most agencies quote behind a form. One June 2026 buyer's guide puts GEO agency retainers at **$3,000 to $25,000 a month**: **$3,000–$8,000** for a lean audit-and-content retainer, **$10,000–$25,000** for full-service work with cornerstone content, and **$5,000–$12,000** for a one-time audit ([MarketerHire, June 2026](https://marketerhire.com/blog/geo-agency)). Treat that as one guide's range, not a survey. The spread is wide because the scope behind the word "GEO" is wide, which is the next section.
### What actually drives the cost
Four things, none of them mysterious: **content depth and volume** (how many buying queries you need to own, and how competitive they are), **technical and entity work** (schema, consistent naming, a corroborated footprint the models can verify), **the measurement harness** (tracking citation rate on a fixed query panel across ChatGPT, Perplexity, and Google AI Overviews), and **category difficulty** (a crowded category costs more to win). The levers themselves are known — Princeton's GEO study found that adding citations, quotations, and statistics to content lifted generative-answer visibility **up to 40%** ([Aggarwal et al., KDD 2024](https://arxiv.org/abs/2311.09735)) — so the cost is in doing the work consistently, not in discovering it.
### What GEO costs at Gigabit
We don't sell GEO as a standalone line item, because on its own it underperforms. It's built into [Ascent](https://gigabit.agency/services/geo/), our 12-month growth program: **from $4,000 a month, no upfront cost, over 12 months**, with founding-cohort pricing for the first 10 companies. $4,000 is where Ascent starts, not what every program costs: the monthly fee rises with how many buying queries you need to own and how much of the site and content needs rebuilding, and it is put in writing after the audit, before you commit. At its entry price Ascent sits inside the lean-retainer band above, but the fee covers more than GEO: one senior team owns your brand, site, content, SEO, and GEO against a single number — your citation rate — so every lever reinforces the others toward [Answer-Market Fit](https://gigabit.agency/insights/answer-market-fit/). Everything with a fixed scope, from the $8,000 Quick Win Sprint to the $3,000–$20,000/mo Managed AI Operations tiers, is on [our pricing page](https://gigabit.agency/pricing/). The baseline is free: the [AI Citation Audit](https://gigabit.agency/tools/geo-citation-audit/) runs your category's buying queries through the major assistants and shows where you're cited, where you're not, and who's named instead.
### Cheap GEO vs GEO that compounds
Be wary of the cheapest "we added GEO to our SEO retainer" offers — they tend to chase vanity prompts you already win instead of the category buying queries that move pipeline. Real GEO is measured on [citation rate](https://gigabit.agency/insights/what-is-a-good-ai-citation-rate/) for commercial queries and reported monthly. The mechanics are in the [GEO field guide](https://gigabit.agency/insights/generative-engine-optimization-field-guide/); the honest first step is the [free audit](https://gigabit.agency/tools/geo-citation-audit/), which tells you the size of the gap before you spend anything on closing it.

### FAQ
**How much does GEO cost per month?**
There's no standard monthly rate. GEO is priced as a retainer, a fixed project, or on performance (e.g. citation rate), and one June 2026 buyer's guide puts agency retainers at $3,000 to $25,000 a month. At Gigabit it's built into Ascent, our 12-month growth program, from $4,000 a month with no upfront cost and founding-cohort pricing for the first 10 companies. The baseline citation audit is free.
**Is GEO a one-time project or ongoing?**
Ongoing. AI answers are stochastic and competitors keep moving, so a single pass doesn't hold citation share. Effective GEO is a continuous program — content, entity work, and monthly re-measurement of citation rate on a fixed panel of buying queries.
**How is GEO priced differently from SEO?**
The models are similar (monthly retainers are common in both), but GEO adds work SEO doesn't: entity consistency, extraction-friendly claims, schema, and citation-rate measurement across ChatGPT, Perplexity, and Google AI Overviews. Increasingly it's priced against an outcome — citation rate — rather than deliverables.

## AI agent vs. chatbot: what's the difference (and which do you need)?
URL: https://gigabit.agency/insights/ai-agent-vs-chatbot/
Published 2026-07-11, updated 2026-08-30

**A chatbot responds; an AI agent acts.** A chatbot maps a message to a reply — scripted or generated. An agent takes a goal, decides the steps, uses tools to touch real systems, and completes a multi-step task: booking the appointment, issuing the refund, qualifying and routing the lead — not just describing how. The difference is **autonomy and tool use**, and it's the gap between deflecting a question and finishing the job.
### What does "agentic" actually mean?
Strip the marketing off and it means one thing: **the software decides its own next step.** A traditional program follows a path you wrote; an agentic system is given a goal and works out the sequence itself, calling tools as it goes and adapting when a step fails. That's the whole idea. Everything else — "reasoning," "planning," "autonomous" — describes how well it does that, not whether it's doing it.
The word became noise because vendors applied it to anything containing a language model, including systems that just answer questions. So the useful test isn't whether something is *called* agentic. It's whether it can **take an action in a real system without a human doing it for it.** If a person still has to read the output and go click the thing, you have a very good text generator. That's worth having — it just isn't an agent, and buying it as one is how budgets get spent twice.
### The three things an agent has that a chatbot doesn't
- **Tools** — an agent can act in your systems (your calendar, CRM, helpdesk, database), not just talk about them.
- **State across steps** — it carries context through a multi-step task instead of answering one turn at a time.
- **Bounded judgment** — within guardrails, it chooses the path to the goal, escalating to a human when confidence is low.
These are exactly the layers that separate a demo from production — model, orchestration, retrieval, evals, observability, and guardrails — which we break down in [the production agent stack](https://gigabit.agency/insights/production-ai-agent-stack/).
### Why the distinction decides your ceiling
A chatbot deflects FAQs; an agent resolves the ticket end to end. In practice that's the difference between shaving a few percent off contact volume and reaching **60–70% autonomous resolution** with satisfaction rising, which is what a well-built [support agent](https://gigabit.agency/insights/ai-support-automation-saas/) actually does. Buy a chatbot for a job that needs an agent and you cap your outcome on day one; over-build an agent for pure FAQ deflection and you pay for capability you don't use.
### "We already have a chatbot" — why that isn't AI-readiness
This is the most common reason a company stalls: a chatbot went live, it deflects some percentage of questions, and the box marked *AI* got ticked. But a chatbot is a **content surface**, not an operating capability. It reads from a knowledge base and writes text. Nothing about having one means your systems are reachable by software, your data is queryable, your failure modes are understood, or anyone owns a metric.
Those four things are what readiness actually is, and none of them come free with a chat widget:
- **Integration** — can software authenticate into your helpdesk, CRM, and billing and *write*, not just read?
- **Data** — does the knowledge the work depends on live somewhere queryable, or in inboxes, PDFs, and someone's head?
- **Failure handling** — when the model is wrong, what catches it, and what does it cost you before something does?
- **Ownership** — is there a named person accountable for a number that moves?
The tell is in the numbers: **91% of mid-market firms use generative AI, but only 25% have it integrated into core operations** ([RSM, 2025](https://rsmus.com/newsroom/2025/middle-market-firms-rapidly-embracing-generative-ai-but-expertise-gaps-pose-risks-rsm-2025-ai-survey.html)). That 66-point gap is almost entirely companies with a chatbot and nothing behind it. If your deflection rate plateaued in the teens and never moved, that's not a tuning problem — it's the ceiling of the category you bought. The [AI Readiness Assessment](https://gigabit.agency/tools/ai-readiness-assessment/) scores exactly these four dimensions.
### Where a chatbot is still the right call
If the job is answering common questions from a knowledge base, with no need to touch other systems, a chatbot is cheaper and perfectly adequate. Don't buy autonomy you won't use — the goal is the right tool for the workflow, not the most impressive one.
### Where you genuinely need an agent
When the task requires touching systems, carrying context across steps, and producing a real outcome — a booked appointment, a processed return, a qualified lead handed to sales — you need an agent. Those are the jobs in the [Gigabit Agents](https://gigabit.agency/services/ai-agents/) catalog: Resolver (support), Scheduler (booking), Intake (lead-qual), and the rest, each built for one flat fee.
### How much autonomy do you actually want?
Most business workflows want *bounded* autonomy — a reliable multi-step process with a few judgment points and human checkpoints — not a free-roaming agent that improvises. That's a feature, not a limitation: reliability comes from constraining the agent to the job. It also explains the adoption gap — **91% of mid-market firms use generative AI, but only 25% have it integrated into core operations** ([RSM, 2025](https://rsmus.com/newsroom/2025/middle-market-firms-rapidly-embracing-generative-ai-but-expertise-gaps-pose-risks-rsm-2025-ai-survey.html)) — most are still at chatbot-grade Q&A, not agent-grade execution. If you're not sure which your workflow needs, the [AI Readiness Assessment](https://gigabit.agency/tools/ai-readiness-assessment/) will point you to the right one.

### FAQ
**What's the difference between an AI agent and a chatbot?**
A chatbot responds to a message with a reply; an AI agent takes a goal, uses tools to act in your systems, carries context across multiple steps, and completes a task — booking, refunding, qualifying — rather than just answering. The defining differences are autonomy and tool use.
**Is ChatGPT an agent or a chatbot?**
The plain chat interface is a chatbot — it answers. Given tools and the autonomy to take actions and complete multi-step tasks, the same model becomes agentic. The label depends on whether the system can act in real systems, not just generate text.
**Do I need an AI agent or just a chatbot?**
If the job is answering questions from a knowledge base with no system access, a chatbot is enough and cheaper. If the job is completing tasks that touch your systems — appointments, refunds, lead qualification, order changes — you need an agent. Match the tool to the workflow rather than over-building.
**What does "agentic" actually mean?**
It means the software decides its own next step. A traditional program follows a path you wrote; an agentic system is given a goal, works out the sequence itself, calls tools as it goes, and adapts when a step fails. The practical test is whether it can take an action in a real system without a human doing it for it — if someone still has to read the output and go click the thing, it isn't agentic.
**We already have a chatbot — does that make us AI-ready?**
No. A chatbot is a content surface, not an operating capability: it reads a knowledge base and writes text. Readiness is four different things — whether software can authenticate into your systems and write to them, whether the underlying data is queryable, whether you know what catches the model when it's wrong, and whether a named person owns a metric. 91% of mid-market firms use generative AI but only 25% have it integrated into core operations, and that gap is mostly companies with a chatbot and nothing behind it.

## Software factory, not software agency: how the mid-market ships in the AI era
URL: https://gigabit.agency/insights/software-factory-vs-software-agency/
Published 2026-07-27

**A software factory is an industrialized way to produce software — a standardized, quality-gated production line instead of bespoke, hero-dependent projects.** Senior engineers plus AI agents, a repeatable process, and a QA line that everything passes before it ships. In the AI era it has quietly become how the best software gets built. And for a mid-market company, the real choice isn't *build your own factory* or *hire an agency* — it's **hire a factory that already runs: a software factory delivered as a service.**
### Why “agency” stopped shipping
The traditional software or AI agency is artisanal: a few talented people, a bespoke process reinvented for every project, and a handoff at the end. It works right up until it doesn't — and at scale, it mostly doesn't. **95% of enterprise generative-AI pilots deliver no measurable P&L impact** ([MIT Project NANDA, 2025](https://www.forbes.com/sites/jasonsnyder/2025/08/26/mit-finds-95-of-genai-pilots-fail-because-companies-avoid-friction/)), and Gartner expects **more than 40% of agentic-AI projects to be canceled by the end of 2027** on cost and unclear value ([Gartner, 2025](https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027)). The common thread isn't the models; it's the *method* — hero-dependent builds with no standardized process, no quality gates, and no one who stays to operate what shipped. That's the artisan's ceiling, and it's [why so many pilots never reach production](https://gigabit.agency/insights/why-your-ai-pilot-never-reached-production/).
### What makes it a factory — not a sweatshop
The word “factory” scares people who picture a commodity sweatshop churning out low-craft code. Picture the opposite. A software factory industrializes the *boring* parts so senior people spend their time on the *hard* parts — the way a modern car plant pairs robots with master engineers. Five things make it a factory:
- **A production line, not a project.** A standardized path — Audit → Build → Operate — that every engagement runs, so the process is the asset, not any one hero.
- **Senior engineers and AI agents, not headcount.** The standardized line and its quality gates are the machinery; people and agents run it — agents handle the repeatable, senior engineers handle judgment, architecture, and the edge cases.
- **A QA line.** Evals, observability, and DevSecOps that everything passes before it ships — the reason output is reliable instead of demo-grade.
- **Standardized outputs.** Productized, fixed-price offers with named prices, not an hourly meter that drifts.
- **It keeps running.** The factory operates what it ships; it doesn't hand you the keys and leave.
That's the line between output you can trust and a body shop billing hours — and it's the whole difference between a demo and [a production agent stack](https://gigabit.agency/insights/production-ai-agent-stack/).
### Three things a software factory is *not*
The term is having a moment, and three adjacent things keep getting confused with it.
- **Not an “AI factory.”** That's NVIDIA's term for the GPU *infrastructure* that manufactures intelligence — data centers that turn power into tokens ([NVIDIA](https://www.nvidia.com/en-us/glossary/ai-factory/)). A software factory produces *software and systems* — apps, agents, integrations — not AI models trained and served at scale. Different category entirely.
- **Not a coding-agent tool.** Platforms built around coding agents give *your* engineers a factory to operate — valuable, but you still need the team, the process, and the QA line to run it. A software factory *is* that team and process, delivered as a service.
- **Not a traditional agency or dev shop.** Some agencies do run retainers and managed services — the real line isn't talent, it's the operating model. The typical project shop treats process, quality gates, and post-launch operation as add-ons; a factory makes them the default, systematized and accountable for the system once it's in production.
### Build one, buy tools, or hire the factory
Everyone industrializing software takes one of three paths. **Enterprises and defense build their own** — the term “software factory” has usually meant exactly this: an internal delivery platform with paved-road CI/CD, staffed by its own engineering org. **AI-native startups buy tools** — coding agents and platforms their strong in-house teams operate. **The mid-market can do neither well:** it can't fund a year of platform-building, and it doesn't have the senior bench to run agent tooling at production quality. So it takes the third path — **hire a factory that already runs.** That's what we mean by a **software factory as a service**: the industrialized output — fixed-price, quality-gated, and operated — without building or staffing the plant. It's the same logic as [build vs. buy for agents](https://gigabit.agency/compare/build-vs-buy-vs-embed/), one level up.
### What comes off the line
A factory's value is that different products come off the same line, reliably. Ours runs on one process — [Audit → Build → Operate](https://gigabit.agency/approach/) — and produces standardized offers instead of bespoke quotes: [Ascent](https://gigabit.agency/services/geo/) (the growth line — brand, website, content, SEO, and GEO), [Gigabit Agents](https://gigabit.agency/services/ai-agents/) (production AI agents off the line for a flat fee), and [Managed AI Operations](https://gigabit.agency/services/managed-ai-operations/) (the line that keeps what's live running). Same factory, different outputs — [every one at a named price](https://gigabit.agency/pricing/).
### How to tell a factory from a shop
When you're deciding who builds your next system, five questions separate a factory from an artisan:
- **Fixed price, or an hourly meter?** A factory prices the output; a shop bills the hours.
- **Are evals and observability included, or extra?** A factory has a QA line; a shop treats quality as a change order.
- **Who operates it after launch?** A factory runs what it ships; a shop hands you the keys.
- **Named, productized offers — or a bespoke quote every time?**
- **Senior engineers plus AI leverage — or bodies on a timesheet?**
Answer those and you'll know which one you're actually hiring.
The AI era didn't kill software craft — it industrialized it. The teams that win aren't the ones with the most people; they're the ones with the best factory. For the mid-market, you don't have to build that factory or buy the tools to run one. You can hire ours: start with a [free growth audit](https://gigabit.agency/tools/geo-citation-audit/), or [see the full line](https://gigabit.agency/services/). We own the outcome.

### FAQ
**What is a software factory?**
A software factory is an industrialized, standardized way to produce software — a quality-gated production line of senior engineers and AI agents following a repeatable process, instead of bespoke, hero-dependent projects. It produces reliable, productized output (fixed-price, evals-gated, operated) rather than one-off deliverables.
**Is a software factory the same as an AI factory?**
No. “AI factory” is NVIDIA's term for GPU infrastructure — data centers that manufacture intelligence by training and serving models at scale. A software factory produces software and systems (apps, agents, integrations), not model inference. They are different categories that share the industrial metaphor.
**What's the difference between a software factory and a software agency?**
An agency is artisanal — a bespoke process reinvented per project, hero-dependent, and typically gone after delivery. A software factory is systematized: a standardized production line, quality-gated with evals and observability, productized and fixed-price, and it operates what it ships. The factory model is why output is predictable instead of demo-grade.
**Can a mid-market company use a software factory?**
Yes — by hiring one rather than building it. Enterprises build their own software factories and AI-native startups buy coding-agent tools, but the mid-market usually can't fund or staff either. A software factory as a service gives you the industrialized output — fixed-price, quality-gated, and operated — without standing up the plant yourself.

## Five signs your business is ready for an AI agent
URL: https://gigabit.agency/insights/signs-your-business-is-ready-for-an-ai-agent/
Published 2026-09-01

**Readiness has almost nothing to do with how excited your team is about AI, and almost everything to do with five unglamorous conditions.** We run this checklist before we agree to build, because the projects that fail are rarely the ones with a bad idea — they're the ones that started before the ground under them was solid. Meet four of the five and you should build now. Meet two and building first is the expensive way to discover which three were missing.
That gap is most of the market right now: **91% of mid-market firms use generative AI, but only 25% have it integrated into core operations** ([RSM, 2025](https://rsmus.com/newsroom/2025/middle-market-firms-rapidly-embracing-generative-ai-but-expertise-gaps-pose-risks-rsm-2025-ai-survey.html)). The distance between those two numbers is this checklist.
### 1. A workflow that already costs you real money
Not a workflow that's annoying — one with a number attached. The threshold we use is roughly **15 hours a week of paid time** on a repeated task. Below that, even a perfect agent struggles to clear its own cost. At 20 hours a week and a $75 loaded hourly rate, you're spending about **$78,000 a year** on that one workflow, and a 50–70% automation rate is worth real money against a flat build fee.
The test isn't whether the work is tedious. It's whether you can name the hours and the rate without guessing. If you can't, that's not a disqualification — it's the first thing to go measure. Our [ROI calculator](https://gigabit.agency/tools/ai-roi-calculator/) does the arithmetic once you have the inputs.
### 2. The right answer is knowable
There has to be a correct outcome that a competent person could confirm. *Was this ticket routed to the right team? Was this invoice matched to the right PO? Was this lead qualified correctly?* — all checkable. *Did we write the best possible marketing strategy?* — not checkable, and therefore not something you can hold a system accountable to.
This matters more than it sounds, because a knowable right answer is what makes an [eval suite](https://gigabit.agency/insights/ai-agent-eval-metrics/) possible, and evals are the gate everything else depends on. If nobody can say whether a given output was correct, nobody can say whether the system is working — and you'll be relying on impressions forever.
### 3. Your systems can actually be reached
An agent that can't touch your systems is a [chatbot](https://gigabit.agency/insights/ai-agent-vs-chatbot/) with extra steps. The question is whether software can authenticate into the tools where the work happens — helpdesk, CRM, billing, database — and **write**, not just read.
Most modern SaaS clears this easily. The places it breaks: a homegrown internal system with no API, a vendor whose integration tier costs more than the agent, or a process that runs through a shared inbox and a spreadsheet nobody owns. None of these are fatal, but each one adds weeks, and it's better to find out now than in month two. Data quality is the related trap — poor data readiness is the top obstacle data leaders name for moving pilots to production ([Informatica, 2025](https://www.informatica.com/blogs/cdo-insights-2025-global-data-leaders-racing-ahead-despite-headwinds-to-being-ai-ready-latest-survey-finds.html)).
### 4. One person will own the number
Not a committee, not a sponsor, not your vendor — one named individual whose job is affected by whether the metric moves. This is the least technical item on the list and the one that most reliably predicts the outcome.
If you can't name that person today, you are not ready, and no amount of engineering fixes it. A system with no owner has no one to notice it degrading, no one to defend its budget at renewal, and no one for whom shipping it is genuinely urgent. We've written about how this failure compounds in [why your AI pilot never reached production](https://gigabit.agency/insights/why-your-ai-pilot-never-reached-production/).
### 5. You can live with the worst case
Every autonomous system will be wrong sometimes. Readiness means you've decided in advance what happens when it is: what the agent is allowed to touch, what it must escalate, and what the damage looks like on a bad day.
The useful exercise is to state the worst outcome in one sentence. *It routes a ticket to the wrong queue and a human re-routes it* is a worst case you can ship against. *It issues a refund to the wrong account* needs an approval step before it goes near production. Teams that skip this conversation don't avoid it — they just have it later, in a security review, with a finished build waiting.
### What a partial score means
Most companies we talk to have three of five, and the pattern is consistent: signs 1 and 2 are usually solid, and one of 3, 4, or 5 is the gap. The response depends on which:
- **Missing the workflow (1)** — don't build yet. Go measure two or three candidate workflows properly; the winner is usually not the one people nominate.
- **Missing knowability (2)** — pick a different workflow. This one isn't an agent problem.
- **Missing integration (3)** — buildable, but scope it honestly. This is where budget overruns come from.
- **Missing an owner (4)** — the cheapest and hardest fix. It costs nothing and requires a real decision.
- **Missing risk tolerance (5)** — usually solved by narrowing the agent's authority rather than improving the model.
If you want this scored rather than self-assessed, the [AI Readiness Assessment](https://gigabit.agency/tools/ai-readiness-assessment/) comes at it from the operational side — data, process, tooling, team, and alignment — and returns a score with a recommended first move. The five signs above are the workflow test; that one is the organisational test, and you want both to come back clean. When the score is good, a [Gigabit Agents](https://gigabit.agency/services/ai-agents/) build is a flat fee from $8,000 per agent, deployed in about 90 days — and when it isn't, we'd rather tell you before you spend it.

### FAQ
**Is my business ready for an AI agent?**
Check five conditions: a workflow that already consumes roughly 15+ hours a week of paid time, a right answer a person could confirm, systems software can authenticate into and write to, one named person accountable for the metric, and a worst case you can state in a sentence and live with. Four out of five means build now; two means fix the gap first.
**What makes a workflow a good fit for an AI agent?**
High volume, a knowable correct outcome, and a cost someone already feels. The best first candidate is usually the most tractable workflow rather than the most impressive one — repetitive, measurable, and reachable through systems that have an API.
**What if we only meet three of the five signs?**
It depends which two are missing. No measurable workflow or no knowable right answer means pick a different workflow. Missing integration is buildable but changes the budget. Missing an owner costs nothing to fix and is the most important. Missing risk tolerance is usually solved by narrowing what the agent is allowed to do rather than by better models.
**How long does it take to deploy a first AI agent?**
It depends on the workflow, and the timeline is fixed in the written scope before work starts. A production agent in one workflow, with evals, observability, and guardrails, is a flat fee starting from $8,000 per agent. A shorter fixed-price diagnostic can scope it first if you want the real number before committing.

## Bait-and-Bill: the AI pricing trap to watch for
URL: https://gigabit.agency/insights/bait-and-bill-ai-pricing-trap/
Published 2026-09-01

**Bait-and-Bill is when the number that won the deal was never the price — it was the entry fee for a meter you can't turn off.** A discovery phase comes in suspiciously cheap. The build is quoted as an estimate. Nine months later you've spent three times the figure that anchored the decision, and every increment was individually reasonable. Nobody lied to you. The pricing model did exactly what it was designed to do.
This is worth naming because it's the mechanism behind a lot of the AI money that vanishes. **95% of enterprise generative-AI pilots deliver no measurable P&L impact** ([MIT Project NANDA, 2025](https://www.forbes.com/sites/jasonsnyder/2025/08/26/mit-finds-95-of-genai-pilots-fail-because-companies-avoid-friction/)), and Gartner expects **more than 40% of agentic-AI projects to be canceled by the end of 2027**, largely on escalating cost and unclear value ([Gartner, 2025](https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027)). Those cancellations aren't mostly technical failures. They're budgets that ran out before the thing shipped.
### How the trap is built
It isn't fraud, and it usually isn't even deliberate. It's the natural consequence of three ordinary choices stacked together.
**The cheap front door.** Discovery is priced low — sometimes free — because it isn't the product. It's qualification, and it's where the switching cost gets manufactured. By the time you have a findings document, the vendor knows your systems better than any competitor could, and starting over means paying for that context twice.
**The estimate that isn't a quote.** The build is scoped as a range, with the range anchored on the happy path. AI work is genuinely uncertain, so this sounds like honesty — and in a fixed-fee model, that uncertainty is the vendor's problem to price. In time-and-materials, it becomes yours to fund. Same uncertainty, opposite incentive.
**The change order treadmill.** Every discovery that expands scope becomes billable, and AI projects discover things constantly: the data is messier than the demo suggested, an integration needs an enterprise tier, an edge case needs a human review step. Each change order is defensible on its own. In aggregate they're the actual price.
### Five clauses that tell you which model you're in
| What you see | What it sounds like | What it means |
| --- | --- | --- |
| "Estimated" hours or range | Honest about uncertainty | You absorb the overrun, not them |
| Discovery priced far below build | A low-risk way to start | The switching cost is the product |
| No named deliverable for phase one | Agile, adaptive | Nothing to hold anyone to |
| Rates but no total | Transparent pricing | Transparency about the meter, not the bill |
| Change orders billed at full rate | Standard practice | Scope discovery is a revenue line |
None of these is disqualifying on its own — a serious firm can run time-and-materials honestly, and some genuinely exploratory work should be. The signal is the *combination*: a cheap entry point, an uncapped middle, and no single number anyone will commit to in writing.
### Why almost nobody publishes prices
Because unpublished pricing is worth money. It lets the number be set after the vendor has seen your budget, your urgency, and how many alternatives you have. That's not a conspiracy; it's just how professional services have always worked, and it's why *request a quote* is the industry default.
It's also why a published price is a genuine signal. A firm that puts real numbers on a public page has given up the ability to charge you more than the next buyer, and has to make the fixed fee work — which means the uncertainty gets priced into their model instead of billed to yours. [Our prices are on the site](https://gigabit.agency/pricing/): a Gigabit Agents build is a flat fee from $8,000 per agent, the AI Transformation Sprint is $25,000, and managed operations run $3,000–$20,000 a month. You can compare those to anything else you're considering without booking a call. A few other firms publish too: we put [fourteen fixed-price pilot offers from thirteen firms side by side](https://gigabit.agency/compare/fixed-price-ai-pilots/), ours included.
The related tell is that the [token bill everyone worries about is the cheapest line in the budget](https://gigabit.agency/insights/ai-agent-cost/) — inference is cents per transaction and falling fast. If a proposal spends its pricing section on model costs and stays vague about engineering hours, it's drawing your attention to the rounding error.
### What to ask for instead
- **A named price for a named deliverable.** Not a range for an outcome — a number for a thing, with a date.
- **A fixed-price first phase that ends in something deployed.** If phase one produces a document, you've bought a document.
- **The total, not the rate.** A rate card describes the meter. Ask what the whole engagement costs if nothing goes wrong, and what it costs if the two most likely surprises happen.
- **Who eats the overrun, in writing.** This single question separates the models faster than anything else on the list.
- **What the exit looks like.** Whether you own the code, the prompts, the evals, and the infrastructure — and how long a handover takes.
If a vendor won't answer those in writing before contract, that is the answer. The [nine questions to ask any AI agency](https://gigabit.agency/insights/questions-to-ask-an-ai-agency/) covers the rest of the diligence, and our [comparison against traditional consultancies](https://gigabit.agency/compare/gigabit-vs-consultancies/) lays out where the money actually goes in each model.

### FAQ
**What is bait-and-bill pricing in AI consulting?**
It's when a low headline number — often a cheap or free discovery phase — anchors the buying decision, but the build is quoted as an estimate on time-and-materials so the real total emerges through change orders. Each increment is defensible; the aggregate is several times the number that won the deal.
**How do I avoid hidden costs on an AI project?**
Ask for a named price attached to a named deliverable with a date, insist the first phase ends in something deployed rather than a document, ask for the total rather than the rate card, and get it in writing who absorbs an overrun. A vendor who won't answer those before contract has answered them.
**Why don't AI agencies publish their prices?**
Because unpublished pricing lets the number be set after they've seen your budget, urgency, and alternatives. A published price gives that up and forces the firm to price uncertainty into its own model rather than billing it to yours — which is why it's a meaningful signal when you find one.
**Is time-and-materials always a red flag for AI work?**
No. Genuinely exploratory work is sometimes best run that way, and a serious firm can run it honestly. The warning sign is the combination: a cheap entry point, an uncapped middle, no named deliverable for phase one, and no single total anyone will commit to in writing.

## Nine questions to ask any AI agency before you sign
URL: https://gigabit.agency/insights/questions-to-ask-an-ai-agency/
Published 2026-09-01

Most AI vendor evaluations are decided by a demo, which is close to the worst available signal. **A demo proves a model can do a task once under conditions the vendor chose.** You're buying the thing that does it ten thousand times under conditions they didn't. These are the nine questions that actually separate the two, what each one is testing underneath, and what the evasive version sounds like.
They matter because the base rate is poor: **95% of enterprise generative-AI pilots deliver no measurable P&L impact** ([MIT Project NANDA, 2025](https://www.forbes.com/sites/jasonsnyder/2025/08/26/mit-finds-95-of-genai-pilots-fail-because-companies-avoid-friction/)). The same study found the pattern that beats it — buying from or partnering with specialized vendors succeeded about **67% of the time, roughly three times the rate of internal builds.** Partnering works. Partnering *badly* is most of the 95%.
### 1. What have you put into production, and what does it do now?
Not what they've built — what's *running*, today, with users depending on it. Ask for the workflow, the volume, and how long it's been live.
*Bad answer:* a portfolio of pilots, prototypes, and proofs of concept described in the past tense. Anyone can demo. Shipping and keeping it alive is the whole skill, and a firm that's done it will have specifics. Ours are on the [proof page](https://gigabit.agency/proof/) with the investment and the return on each.
### 2. Who exactly will do the work, and will I meet them before I sign?
The gap between the people in the pitch and the people on the keyboard is where most engagements go wrong. Ask for names, seniority, and whether they're on other accounts.
*Bad answer:* "our team," a org chart, or a partner who'll "stay close to the project." If the people who scope it aren't the people who build it, every piece of context you transfer gets transferred again — which is the entire argument for a [forward-deployed model](https://gigabit.agency/approach/).
### 3. What's the total price, and who absorbs an overrun?
One number, in writing, with the answer to what happens when it goes long. This is the fastest question on the list for sorting engagement models.
*Bad answer:* a rate card, a range, or "it depends on scope." It always depends on scope; the question is who carries that risk. We wrote up the mechanism in [Bait-and-Bill](https://gigabit.agency/insights/bait-and-bill-ai-pricing-trap/), and [our own prices are published](https://gigabit.agency/pricing/) so you can hold us to the same standard.
### 4. How will we know it's working?
You're looking for a metric and a measurement method, ideally agreed before the build. Accuracy on what test set? Resolution rate against what baseline? Measured by whom, how often?
*Bad answer:* qualitative outcomes, adoption numbers, or "we'll define success together in discovery." A firm that builds AI for a living has opinions about how to measure it — that's what an [eval suite](https://gigabit.agency/insights/ai-agent-eval-metrics/) is, and not having a view on yours is a tell.
### 5. What happens when the model gets it wrong?
Every probabilistic system fails. You're testing whether they've thought about the failure mode before you asked: guardrails, escalation paths, human checkpoints, and the blast radius on a bad day.
*Bad answer:* "the model is very accurate," or anything that treats reliability as a property of the model instead of the system around it. The real answer describes constraints in code, not confidence in a vendor's benchmarks.
### 6. What happens when the underlying model changes?
Model vendors deprecate, retune, and ship behavior changes on their own schedule. Ask what breaks, how they'd find out, and how fast they could roll back.
*Bad answer:* a blank look, or a claim of model-agnosticism with nothing behind it. The credible version involves pinned versions, versioned prompts, a regression suite, and a return path to a known-good state — the layers we describe in [the production agent stack](https://gigabit.agency/insights/production-ai-agent-stack/).
### 7. Who owns the code, the prompts, and the evals?
Ask explicitly about all three, plus the infrastructure and the data. Prompts and eval sets are the accumulated knowledge of the engagement, and they're the part most likely to be quietly retained.
*Bad answer:* ownership of "deliverables," or code ownership with the orchestration living on the vendor's platform. If leaving means rebuilding, the price of the engagement includes never leaving.
### 8. What does it cost to run after launch, and who runs it?
Live AI systems drift, and something has to watch them. Ask for the monthly number and who is accountable for it.
*Bad answer:* treating operations as optional or as a support SKU. Inference itself is cheap — cents per transaction, and [falling fast](https://gigabit.agency/insights/ai-agent-cost/) — but monitoring, eval regressions, and drift management are real work. Ours is [$3,000–$20,000 a month](https://gigabit.agency/services/managed-ai-operations/) depending on how many systems are under management, and we'd rather quote it up front than discover it later.
### 9. When would you tell us not to build this?
The most revealing question on the list, and the one almost nobody asks. You're testing whether the firm has a view on fit at all, or whether every problem happens to be solvable by what they sell.
*Bad answer:* enthusiasm. A firm worth hiring will name the conditions under which you shouldn't proceed — the data isn't there, the workflow is too low-volume, the right answer isn't knowable, nobody owns the metric. If they've never talked a client out of something, you're not their client, you're their pipeline.
### How to use the answers
Ask all nine of every vendor, including us, and write the answers down side by side. The pattern matters more than any single response: firms that ship give specific, slightly boring answers with numbers in them, and firms that demo give confident answers about capability. If you want the structural version of this comparison, we lay out the tradeoffs against [traditional consultancies](https://gigabit.agency/compare/gigabit-vs-consultancies/) and [automation agencies](https://gigabit.agency/compare/gigabit-vs-automation-agencies/) — including where each one is genuinely the better choice.

### FAQ
**What questions should I ask an AI agency before hiring them?**
Nine: what have you put into production and what does it do now; who exactly will do the work; what is the total price and who absorbs an overrun; how will we know it is working; what happens when the model gets it wrong; what happens when the underlying model changes; who owns the code, prompts, and evals; what does it cost to run after launch; and when would you tell us not to build this.
**How do I tell a real AI engineering firm from a demo shop?**
Ask what is running in production today, with volume and duration — not what has been built. Demo shops answer in the past tense with pilots and prototypes; firms that ship give specific, slightly boring answers with numbers in them and can describe their failure modes without being prompted.
**What is the most revealing question to ask an AI vendor?**
When would you tell us not to build this? It tests whether the firm has a view on fit at all. A vendor worth hiring will name the conditions under which you shouldn't proceed — thin data, low volume, no knowable right answer, no owner for the metric. Pure enthusiasm is the warning sign.
**Should I ask who owns the prompts and eval suite?**
Yes, explicitly, alongside the code, infrastructure, and data. Prompts and eval sets are the accumulated knowledge of the engagement and the part most often quietly retained. If leaving the vendor means rebuilding from scratch, the real price of the engagement includes never leaving.

## What does an AI MVP actually cost? A build budget breakdown
URL: https://gigabit.agency/insights/ai-mvp-cost/
Published 2026-09-02

First, a distinction that decides your entire budget: **an AI agent is one workflow inside your business; an AI MVP is a product you're putting in front of users.** The agent needs to work. The product needs to work, be signed into, be paid for, be supported, and survive people using it in ways you didn't plan. That's a different cost structure, and conflating the two is why so many AI product budgets are wrong by a factor of three. If you want the one-workflow number instead, that's [what an AI agent costs](https://gigabit.agency/insights/ai-agent-cost/).
For an AI product MVP, the mid-market range lands in the **$50,000–$500,000** window — and where you fall inside it has almost nothing to do with the model you pick.
### What "MVP" has to mean when there's a model in it
In conventional software, an MVP is the smallest thing that delivers value. With AI in the loop, that definition quietly breaks, because the smallest thing that *demos* value and the smallest thing that *survives contact with users* are far apart. A deterministic MVP either works or throws an error. A probabilistic one is wrong 4% of the time, silently, in front of a customer.
So an AI MVP has a floor that a normal MVP doesn't: it needs an [eval suite](https://gigabit.agency/insights/ai-agent-eval-metrics/), traces you can debug from, and guardrails around the worst case. Those aren't scope you can defer to v2 — they're what makes v1 shippable at all. Budget them in from the start or pay the retrofit price, which is materially higher.
### The four cost centres
| Cost centre | Typical share | What drives it |
| --- | --- | --- |
| Product engineering | 40–50% | Auth, billing, admin, the actual UI — the unglamorous majority |
| AI engineering | 20–30% | Prompts, retrieval, orchestration, evals, guardrails |
| Data & integration | 15–25% | Getting the knowledge into a queryable shape; third-party APIs |
| Inference | 1–3% | Token spend. Cents per transaction, and falling |
That last row is the one that surprises people. Teams spend weeks agonising over model pricing and then underfund the data work by a factor of ten. Processing ten thousand conversations on a fast model runs about **$37** in tokens ([Anthropic](https://platform.claude.com/docs/en/about-claude/pricing)), and the cost of reaching any given capability level has been falling roughly **40× per year** ([Epoch AI, 2025](https://epoch.ai/data-insights/llm-inference-price-trends)). Whatever inference costs in your model today, it costs less next quarter. The engineering around it does not.
### What actually moves the number
- **How many users, and are they yours?** An internal tool for 30 colleagues and a product for 3,000 strangers are different builds. Strangers need onboarding, self-service recovery, and support surfaces; colleagues can be trained and can ask you.
- **How wrong is too wrong?** A drafting assistant a human reviews is cheap. A system that acts without review needs the full reliability stack. This single decision can double a budget.
- **Whose data, and where does it live?** Clean data in a queryable system saves weeks. Data in inboxes, PDFs, and someone's head is a pipeline project wearing an AI project's clothes — and poor data readiness is the top obstacle data leaders name for moving pilots to production ([Informatica, 2025](https://www.informatica.com/blogs/cdo-insights-2025-global-data-leaders-racing-ahead-despite-headwinds-to-being-ai-ready-latest-survey-finds.html)).
- **Regulated or not?** Healthcare, finance, and anything touching PHI or payments add compliance surface that is real engineering, not paperwork.
### The line most budgets forget
**What it costs to keep running.** An AI MVP is not a thing you finish. Models get deprecated and retuned on the vendor's schedule, your data distribution shifts, and edge cases accumulate — so a system with nobody operating it degrades quietly until someone notices a month of bad output. That's not a maintenance line item you can defer; it's the difference between a product and a demo with users attached.
Budget it the way you budget hosting. Ours runs **$3,000–$20,000 a month** for [managed operations](https://gigabit.agency/services/managed-ai-operations/) depending on how many systems are under management. Whatever number you use, put it in the model before you commit to the build, not after — we go into the failure modes in [why AI systems break after launch](https://gigabit.agency/insights/why-ai-systems-break-after-launch/).
### Where AI MVPs overspend
Not on tokens. On **surface area** — building four features adequately instead of one properly, because it was hard to say no in the planning meeting. And on **rebuilds**: the demo-grade version that impressed everyone, collapsed under real usage, and got rebuilt from scratch. That's the dominant cost in the market right now — **95% of enterprise generative-AI pilots deliver no measurable P&L impact** ([MIT Project NANDA, 2025](https://www.forbes.com/sites/jasonsnyder/2025/08/26/mit-finds-95-of-genai-pilots-fail-because-companies-avoid-friction/)) — and it's paid in full, twice.
The cheapest insurance is scoping the thing properly before you fund it. A fixed-price [two-week Sprint](https://gigabit.agency/services/ai-transformation/) is $25,000 and ends with a deployed pilot and a real number, which is a considerably better way to discover a $400,000 build was actually a $120,000 one. [Our prices are published](https://gigabit.agency/pricing/) if you want to benchmark whatever you're being quoted, and [why we run this as a software factory](https://gigabit.agency/insights/software-factory-vs-software-agency/) explains the model underneath them.
For real figures instead of ranges, see [what ten client projects cost](https://gigabit.agency/proof/project-costs/): the build price, the monthly cost after launch, the timeline and the payback, taken from our published case studies.

### FAQ
**How much does an AI MVP cost to build?**
For a mid-market AI product MVP, expect the $50,000–$500,000 window. The spread is driven by user count, how wrong the system is allowed to be, the state of your data, and whether you're in a regulated domain — not by which model you choose. Inference is typically 1–3% of the budget.
**What's the difference between an AI agent and an AI MVP?**
An agent automates one workflow inside your business; an MVP is a product you put in front of users, so it also needs auth, billing, admin, support surfaces, and resilience to people using it in unplanned ways. The agent is a flat per-agent fee; the product is a build in the $50K–$500K range.
**How much of an AI product budget is spent on the model?**
Usually 1–3%. Processing ten thousand conversations on a fast model costs about $37 in tokens, and the price of a given capability level falls roughly 40× per year. Product engineering is typically 40–50%, AI engineering 20–30%, and data and integration 15–25%.
**What do AI MVP budgets usually leave out?**
The cost of operating the system after launch. Models get deprecated and retuned on the vendor's schedule and data distributions shift, so an unattended AI system degrades silently. Managed operations typically run $3,000–$20,000 a month depending on how many systems are involved, and it belongs in the model before the build is approved.

## Why AI systems break after launch — and what "operating" one means
URL: https://gigabit.agency/insights/why-ai-systems-break-after-launch/
Published 2026-09-02

**Conventional software fails loudly. AI systems fail quietly.** A broken API throws a 500 and pages someone; a degraded agent keeps returning confident, well-formatted, subtly wrong answers while every dashboard stays green. Uptime is 100%. Latency is fine. Error rate is zero. And quality has been sliding for five weeks.
This is the part of AI work almost nobody prices in, and it's the reason a system that launched well can be worthless by the following quarter. Launch is not the finish line — it's the point at which the system starts drifting away from the conditions it was built for.
### The five ways a live AI system degrades
| Failure | What changed | How it shows up |
| --- | --- | --- |
| Model drift | The vendor retuned or deprecated the model | Behaviour shifts overnight, no deploy on your side |
| Data drift | Your inputs stopped looking like your test set | Accuracy decays gradually, worst on new cases |
| Workflow drift | The humans changed the process | The agent is now correct about a stale reality |
| Integration rot | A downstream API changed shape | Silent partial failures on one path |
| Prompt entropy | Everyone edited the prompt a little | Nobody can say what current behaviour is or why |
**Model drift** is the one people underestimate. You do not control the schedule — vendors retune, deprecate, and ship behaviour changes when it suits them, and a system pinned to nothing inherits every change. Without versioned configuration and a rollback path, "it used to work" becomes an archaeology project.
**Data drift** is the slowest and most expensive. The system was tuned against the cases you had. Your customers change, your product changes, a new segment arrives, and the live distribution walks away from the test set. Nothing breaks. It just gets worse, in a way only measurement catches.
**Workflow drift** is the funniest and the most human. Ops changed a step, added a queue, renamed a status — and the agent is now faithfully executing last quarter's process. It isn't wrong about anything except reality.
### Why nobody notices for weeks
Because the signals people watch are the wrong ones. Uptime, latency, and error rate all describe whether the system *ran* — none of them describe whether it was *right*. Quality is not a property you can infer from infrastructure metrics, and it's the only property that matters here.
Then the volume hides it. If an agent handles 4,000 cases a month and quality drops from 94% to 88%, that's 240 additional wrong outcomes — invisible in aggregate, distributed across the exact customers least likely to write in about it. You typically find out through a downstream complaint, weeks late, and then have to reconstruct what happened without traces because nobody was recording them.
This is also why the adoption gap persists: **91% of mid-market firms use generative AI, but only 25% have it integrated into core operations** ([RSM, 2025](https://rsmus.com/newsroom/2025/middle-market-firms-rapidly-embracing-generative-ai-but-expertise-gaps-pose-risks-rsm-2025-ai-survey.html)). Integrating into core operations means someone has to be accountable for a live system on a Tuesday — and that job has to exist before it can be done.
### What "operating" an AI system actually means
It's a real job with a real definition, not a support contract. Concretely:
- **Eval regressions on a schedule.** The [eval suite](https://gigabit.agency/insights/ai-agent-eval-metrics/) runs continuously, not once at launch, so quality movement is a number someone sees rather than a thing a customer reports.
- **Trace review.** Someone reads real runs — especially the escalations and the low-confidence ones — because that's where drift shows up first.
- **Model migration.** Tracking deprecations, testing the replacement against the eval set *before* the forced cutover, and keeping a rollback path warm.
- **Prompt and config version control.** Changes are reviewed and versioned, so current behaviour is always attributable to a specific decision.
- **Cost and latency watch.** Spend per transaction drifts with usage patterns and model choices, usually upward and usually unnoticed.
- **Workflow liaison.** Somebody who knows when ops changed the process, because ops will not think to tell the AI vendor.
That last one is why we do this [forward-deployed](https://gigabit.agency/approach/) rather than through a ticket queue. The signal that a workflow changed doesn't arrive as a bug report — it arrives as a hallway conversation, and you have to be close enough to the work to hear it. The layers being watched are the ones in [the production agent stack](https://gigabit.agency/insights/production-ai-agent-stack/); operating is what keeps them true after launch.
### What it costs, and what skipping it costs
Ours is **$3,000–$20,000 a month** for [managed AI operations](https://gigabit.agency/services/managed-ai-operations/), scaling with how many systems are under management — one system at the light end, five or more production agent platforms at the heavy end. Whether you buy that or staff it internally matters less than whether the function exists and someone owns it.
The alternative isn't free, it's just unbilled. It shows up as a quarter of degraded output nobody measured, a forced model migration done under deadline, and eventually a rebuild of something that worked at launch — which is how a large share of the **40%+ of agentic-AI projects Gartner expects to be canceled by 2027** ([Gartner, 2025](https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027)) will actually end. Not with a failure. With a slow decline nobody was watching.

### FAQ
**Why do AI agents stop working after launch?**
Five common causes: the model vendor retuned or deprecated the model, your input data drifted away from what the system was tuned on, humans changed the underlying workflow, a downstream integration changed shape, or the prompt accumulated untracked edits. None of them throw an error — the system keeps running and quietly gets less right.
**Why don't standard monitoring tools catch AI degradation?**
Because uptime, latency, and error rate describe whether the system ran, not whether it was right. Quality isn't inferable from infrastructure metrics. Catching decay requires scheduled eval regressions against a known-good test set plus review of real traces, especially escalations and low-confidence runs.
**What does it mean to operate an AI system?**
Running eval regressions on a schedule, reviewing real traces, tracking model deprecations and testing replacements before forced cutovers, version-controlling prompts and configuration, watching cost and latency drift, and staying close enough to the business to know when the underlying workflow changed.
**How much does it cost to maintain an AI system?**
Managed operations typically run $3,000–$20,000 a month depending on how many systems are under management — roughly one system at the light end and five or more production agent platforms at the heavy end. The cost of skipping it is unbilled rather than zero: degraded output nobody measured, and eventually a rebuild.

## How to calculate ROI on an AI project before you build it
URL: https://gigabit.agency/insights/calculate-ai-project-roi/
Published 2026-09-02

**Annual saving = hours per week × loaded hourly cost × 52 × automation rate. Payback = build cost ÷ (annual saving − annual operating cost).** That's the whole model, and it's deliberately boring. The arithmetic has never been what makes AI business cases wrong — the inputs are.
Worth doing before you commit, because the base rate is bad: **95% of enterprise generative-AI pilots deliver no measurable P&L impact** ([MIT Project NANDA, 2025](https://www.forbes.com/sites/jasonsnyder/2025/08/26/mit-finds-95-of-genai-pilots-fail-because-companies-avoid-friction/)). A number you built honestly before starting is the cheapest protection against joining them.
### A worked example
Support triage consumes **20 hours a week**. Loaded cost of the people doing it is **$75/hour**. So the workflow costs about **$78,000 a year**. At a **60%** automation rate, the agent saves roughly **$46,800 a year**. Operating the system runs **$3,000 a month**, or $36,000 — leaving about **$10,800 of net annual benefit** against a build.
And there's the lesson: on a $75,000 build, that's a payback measured in years, which means this is a bad project *as scoped* — despite a 60% automation rate that sounds like a triumph. Most business cases never surface that, because most business cases quietly omit the operating line.
Change one input and it inverts. At 40 hours a week the same workflow costs $156,000, saves $93,600, nets $57,600 after operations, and pays back a $75,000 build in about sixteen months. **Volume is the variable that decides almost everything** — which is why the first question is always which workflow, not which model.
### The four inputs, and how to get them honestly
- **Hours per week.** Measure, don't estimate. Ask the people doing the work, and expect the real number to differ from the manager's estimate in both directions.
- **Loaded hourly cost.** Salary plus benefits, tax, software, and management overhead — typically 1.25–1.4× base salary. Using base salary alone understates the case; using a blended executive rate overstates it.
- **Automation rate.** What share of cases the system closes end to end, with no human touch. Not deflection, not assistance — completion.
- **Annual operating cost.** Monitoring, eval regressions, drift management, model migrations. Ours is [$3,000–$20,000 a month](https://gigabit.agency/services/managed-ai-operations/); whatever you use, it is not zero.
### The three numbers people inflate
**Automation rate.** The most common failure. Teams model 90% because the demo handled 9 of 10 curated cases; production sees the long tail and lands at 55%. Model **50–70%** for a well-scoped workflow and treat anything above that as upside you have to earn. A [support agent reaching 60–70% autonomous resolution](https://gigabit.agency/insights/ai-support-automation-saas/) is a good outcome, not a floor.
**Hours saved as money saved.** Freeing 12 hours a week across a team of eight doesn't remove a salary — it removes ninety minutes each from eight people, who fill it with other work. That's real value, but it is capacity, not cash, and a CFO will make that distinction even if the business case doesn't. Say which one you're claiming.
**Time to value.** Savings do not start on the contract date. A build takes weeks to months and ramps after that, so year one typically captures a fraction of the annual figure. Model it from go-live, not from kickoff.
### What to do with a long payback
If payback lands past 18 months, don't tune the spreadsheet — change the project. Three moves, in order of how often they work:
- **Pick a higher-volume workflow.** Volume dominates every other input. The second-choice workflow with twice the throughput usually beats the favourite.
- **Narrow the scope.** A tighter agent with a higher completion rate on a smaller case set often beats a broad one that half-handles everything.
- **Check the non-cash benefits are real.** Response time, consistency, after-hours coverage, and reduced churn can carry a project — but only if someone will actually commit to a number for them. If nobody will, they aren't benefits, they're hopes.
The counter-case is worth stating too: IDC puts the average enterprise return at **$3.70 per $1 invested** in generative AI, though that's a [Microsoft-sponsored study](https://blogs.microsoft.com/blog/2024/11/12/idcs-2024-ai-opportunity-study-top-five-ai-trends-to-watch/), so treat it as an optimistic ceiling rather than a planning input. Your own arithmetic on your own workflow beats any benchmark.
### Presenting it internally
Bring three scenarios, not one. Conservative (40% automation), expected (60%), and optimistic (75%) — with payback for each. A single confident number invites someone to attack the assumption; a range shows you know which assumption matters and pre-empts the question. Name the automation rate as the sensitive variable before anyone else does.
Then state what makes the number real: one named owner, a defined metric, and a go-live date. That's the part of the [Audit → Build → Operate](https://gigabit.agency/approach/) model that turns a projection into something anyone is accountable for.
The [ROI calculator](https://gigabit.agency/tools/ai-roi-calculator/) runs this arithmetic if you'd rather not build the spreadsheet, and [what an AI agent costs](https://gigabit.agency/insights/ai-agent-cost/) gives you a real build number to divide by. If you're not yet sure the workflow qualifies at all, the [five signs](https://gigabit.agency/insights/signs-your-business-is-ready-for-an-ai-agent/) come first.

### FAQ
**How do you calculate ROI on an AI project?**
Annual saving = hours per week × loaded hourly cost × 52 × automation rate. Payback = build cost ÷ (annual saving − annual operating cost). The arithmetic is simple; the accuracy comes from measuring hours rather than estimating them, using a fully loaded hourly cost, modelling a realistic automation rate, and including the cost of operating the system.
**What automation rate should I assume for an AI agent?**
Model 50–70% end-to-end completion for a well-scoped workflow. Demos routinely suggest 90% because they run on curated cases; production meets the long tail. Treat anything above 70% as upside to be earned rather than a planning assumption.
**What do most AI business cases get wrong?**
Three things: an inflated automation rate taken from demo performance, counting freed hours as cash when spreading small time savings across a team is capacity rather than salary removal, and starting the savings clock at kickoff instead of go-live roughly 90 days later. Omitting annual operating cost is the fourth and most common.
**What is a good payback period for an AI project?**
Inside 18 months is generally defensible for a mid-market build. If your model says longer, change the project rather than the spreadsheet — pick a higher-volume workflow, narrow the scope to raise the completion rate, or confirm that any non-cash benefits are ones someone will commit to a number for.

## How to check your own AI visibility in 30 minutes
URL: https://gigabit.agency/insights/check-your-ai-visibility-in-30-minutes/
Published 2026-09-02

**You can get a defensible baseline in half an hour with a spreadsheet and four browser tabs.** It won't be as rigorous as a repeated panel, but it will tell you whether you have a problem, how big it is, and who's winning instead of you. Here's exactly how we'd run it if we were doing it by hand.
One thing to fix before you start: what you're measuring is **citation rate** — the share of your category's buying queries where an assistant names you. If you want the definition and what counts as a good score, that's [covered separately](https://gigabit.agency/insights/what-is-a-good-ai-citation-rate/). This piece is the runbook.
### Minutes 0–10: write the query set
Write **fifteen questions a real buyer would type**, in their words, not yours. Pull them from sales calls, your support inbox, and the search terms already in your analytics. Cover four shapes:
- **Category** — *best X for Y*, *top X companies for Z*
- **Cost** — *how much does X cost*, *X pricing*
- **Comparison** — *X vs Y*, *alternatives to X*
- **Problem-first** — the symptom your buyer has before they know your category exists
Two rules. **Don't include your brand name** — *what is [your company]* measures recognition, not acquisition, and it will flatter you into thinking you're visible. And use the words buyers use: if your team says *generative engine optimization* and buyers say *getting found on ChatGPT*, test the second.
### Minutes 10–25: run the panel
Open four tabs: **ChatGPT, Perplexity, Google (with AI Overviews), and Claude.** Run each question in each, and for each one record five columns:
| Column | What to record |
| --- | --- |
| Query | The question, verbatim |
| Engine | Which assistant |
| Named? | Were you mentioned at all |
| Cited? | Was your site linked as a source |
| Who else | Every competitor named, in order |
Use a **logged-out or temporary session**. This is the mistake that ruins most DIY audits: your own account has months of history and memory about your company, so it will name you in ways a stranger's session never would. A fresh incognito window is the difference between measuring your visibility and measuring your own footprint in your own account.
Also record **who else** rigorously. The competitor column is the most actionable data in the whole exercise, and it's the one people skip.
### Minutes 25–30: score it
Three numbers:
- **Citation rate** — appearances ÷ total query-engine combinations. Fifteen questions across four engines is sixty cells; named in twelve is 20%.
- **Competitive share** — of the cells where *anyone* in your category was named, what fraction was you. This is the more useful number early, because it separates *the engines don't answer this* from *the engines answer it with someone else*.
- **Engine spread** — your rate per assistant. A brand strong in one and invisible in another usually has an entity or third-party footprint problem rather than a content problem.
For calibration, see the [measured benchmark](https://gigabit.agency/insights/what-is-a-good-ai-citation-rate/): the most-named firm on a category panel appeared in 20% of answers, so judge your number against the leader on your own panel, not against a fixed band.
### The three mistakes that make this useless
**Using your own logged-in account.** Covered above, and it's the big one — it can turn a 5% into an apparent 40%.
**Running it once.** Answers are stochastic; the same question asked twice can name different companies. A single run is an anecdote. Re-run the panel monthly and track the trend — that's the actual signal, and it's why the query set has to be *fixed* rather than rewritten each time.
**Testing queries you already win.** It's a natural instinct and it wastes the exercise. Deliberately include the questions you expect to lose, because those are the ones with pipeline attached.
### What to do with the result
If you're under 10% on category queries, you're in the normal starting position for a company that hasn't done this work — not in trouble, just invisible. The next question is *which* gap, and [what gets a company excluded from AI answers](https://gigabit.agency/insights/what-gets-a-company-excluded-from-ai-answers/) walks the causes and how to tell them apart.
If you'd rather not build the harness by hand — or you want the panel run repeatedly rather than once — our [free citation audit](https://gigabit.agency/tools/geo-citation-audit/) does exactly this across the major assistants and sends back the scored result, including who's being named instead of you. That report is also how [Ascent](https://gigabit.agency/services/geo/) gets scoped, and the number it produces is what the program is held to afterwards.

### FAQ
**How do I check if ChatGPT knows my business?**
Write fifteen questions a real buyer would ask — category, cost, comparison, and problem-first — then run each through ChatGPT, Perplexity, Google AI Overviews, and Claude in a logged-out session, recording whether you were named, whether you were cited as a source, and which competitors appeared. Appearances divided by total query-engine combinations is your citation rate.
**Why should I use a logged-out session to test AI visibility?**
Because your own account carries history and memory about your company, so it will name you in ways a stranger's session never would. Testing while logged in is the single most common reason a DIY audit reports good visibility that doesn't exist — it can turn a 5% citation rate into an apparent 40%.
**How many queries do I need to test AI visibility?**
Fifteen buyer-phrased questions across four engines gives sixty data points, which is enough for a defensible baseline in about thirty minutes. Keep the set fixed so later runs are comparable, exclude your own brand name, and deliberately include questions you expect to lose.
**Is a one-time AI visibility check enough?**
It's enough to tell you whether you have a problem and roughly how big it is. It isn't enough to manage against, because assistant answers are stochastic and a single run is an anecdote. Re-run the same fixed query set monthly and track the trend rather than any individual answer.

## What should a mid-market company budget for AI in 2026?
URL: https://gigabit.agency/insights/mid-market-ai-budget/
Published 2026-09-02

Most AI budget questions get answered at the wrong altitude. *What does an agent cost* has [a specific answer](https://gigabit.agency/insights/ai-agent-cost/). *What should our AI line item be for the year* is a different question, and it's the one a CFO is actually asking.
For a 20-to-500-person company getting serious in 2026, a realistic first-year AI budget lands between **$75,000 and $300,000 all-in** — and the way it's split matters more than the total, because the split is what determines whether year two builds on year one or starts over.
### Why the total is smaller than you'd fear
The number that scares people is the model bill, and it's a rounding error: processing ten thousand conversations on a fast model costs about **$37** ([Anthropic](https://platform.claude.com/docs/en/about-claude/pricing)), and the price of a given capability level has been falling roughly **40× per year** ([Epoch AI, 2025](https://epoch.ai/data-insights/llm-inference-price-trends)). You are not buying compute. You are buying engineering, integration, and someone to operate the result.
The number that *should* scare you is waste. **95% of enterprise generative-AI pilots deliver no measurable P&L impact** ([MIT Project NANDA, 2025](https://www.forbes.com/sites/jasonsnyder/2025/08/26/mit-finds-95-of-genai-pilots-fail-because-companies-avoid-friction/)), so the dominant risk isn't overspending — it's spending correctly-sized amounts on things that never reach production.
### A first-year split that works
| Line | Share | Typical first year |
| --- | --- | --- |
| Diagnose — find the workflow, prove it | 10–15% | A fixed-price sprint |
| Build — one or two production systems | 50–60% | Flat-fee per system |
| Operate — keep them working | 25–30% | Monthly, starts at go-live |
| Contingency | 10% | The integration you didn't know about |
Two things about this split surprise people.
**Operate is a quarter of the budget, not an afterthought.** A system nobody watches degrades silently — models get deprecated, data drifts, someone changes the workflow — so an unoperated build is a depreciating asset with a roughly one-year half-life. We put real numbers on this in [why AI systems break after launch](https://gigabit.agency/insights/why-ai-systems-break-after-launch/). Budget it from go-live or you'll fund it from next year's build money.
**Diagnose is small but load-bearing.** Ten percent spent proving which workflow deserves the other ninety is the cheapest insurance available. Our [AI Transformation Sprint](https://gigabit.agency/services/ai-transformation/) is $25,000 and ends in a deployed pilot rather than a document, which is the version of diagnosis worth paying for.
### Sizing it against your revenue
A useful sanity check rather than a rule: mid-market companies getting real value are typically committing **1–3% of revenue** to AI in the first serious year, weighted toward the build. Below about 1% you generally can't fund a production system plus its operations, and you end up with a pilot — the most expensive outcome available, because it costs real money and returns nothing.
Above 3% in year one usually means buying more surface area than the organisation can absorb. The constraint is rarely money; it's how many workflows you can change at once without the change management collapsing. **91% of mid-market firms use generative AI but only 25% have it integrated into core operations** ([RSM, 2025](https://rsmus.com/newsroom/2025/middle-market-firms-rapidly-embracing-generative-ai-but-expertise-gaps-pose-risks-rsm-2025-ai-survey.html)) — that gap is an absorption problem, not a budget one.
### What year two looks like
If year one worked, year two is cheaper per unit of value and shaped differently: operate grows because you're running more systems, build stays flat or drops because the integration groundwork is done, and diagnose shrinks to near zero because you now know your own workflow inventory.
If year one didn't work, year two is identical to year one — which is the actual cost of getting it wrong. Not the wasted budget: the wasted *year*, spent re-learning what a properly scoped first engagement would have told you in two weeks.
### Build the number for your business
Start from the workflow, not the budget. Find the one already costing you 15+ hours a week, [model the return](https://gigabit.agency/tools/ai-roi-calculator/), and let the build cost fall out of that rather than allocating a round number and looking for somewhere to spend it. Then add operations from day one and a real contingency.
[Our prices are published](https://gigabit.agency/pricing/) — a Gigabit Agents build is a flat fee from $8,000 per agent, the Sprint is $25,000, managed operations run $3,000–$20,000 a month — so you can assemble a defensible annual figure without a single sales call. If the arithmetic on your first workflow doesn't clear, that's a useful answer too, and [how to calculate the ROI honestly](https://gigabit.agency/insights/calculate-ai-project-roi/) will tell you before you commit.

### FAQ
**How much should a mid-market company budget for AI?**
For a 20-to-500-person company getting serious in its first year, roughly $75,000–$300,000 all-in, or about 1–3% of revenue. Split it approximately 10–15% diagnosis, 50–60% build, 25–30% operations, and 10% contingency. Below about 1% of revenue you usually can't fund a production system plus its operations and end up with a pilot instead.
**What share of an AI budget goes to model and token costs?**
Almost none — typically low single-digit percentages. Processing ten thousand conversations on a fast model costs about $37, and the price of a given capability level falls roughly 40× per year. The budget is engineering, integration, and operations, not compute.
**Should operating costs be in the first-year AI budget?**
Yes, from go-live, and at roughly a quarter of the total. AI systems degrade silently as models are deprecated, data drifts, and workflows change, so an unoperated build loses much of its value inside a year. Budgeting it later means funding it out of next year's build money.
**What does a second-year AI budget look like?**
If year one worked, operations grow because more systems are live, build stays flat or falls because the integration groundwork exists, and diagnosis shrinks because you know your workflow inventory. If year one didn't reach production, year two tends to look identical to year one — which is the real cost of getting it wrong.

## Vendor Roulette: why switching AI vendors keeps you at square one
URL: https://gigabit.agency/insights/vendor-roulette-switching-ai-vendors/
Published 2026-09-02

**If you're evaluating your third AI vendor in two years, the problem is probably not that you keep picking badly.** It's that each engagement ended without leaving anything behind, so every new vendor starts from the same blank page — and you pay for the same discovery, the same integration spelunking, the same context transfer, for the third time.
This is Vendor Roulette. It looks like bad luck and it's structural.
### What each switch actually costs you
Not just the fees. The expensive part is the reset:
- **Context.** Months of learning about your edge cases, your data's quirks, and why that one workflow has an exception — gone with the team that learned it.
- **Calendar.** A new vendor needs 4–8 weeks before they're productive. Three vendors is half a year of ramp with nothing shipped.
- **Organisational patience.** The most under-counted cost. After two visible failures, the internal appetite to try again collapses, and the third attempt gets less access, less budget, and more scepticism — which makes it likelier to fail too.
It compounds. **95% of enterprise generative-AI pilots deliver no measurable P&L impact** ([MIT Project NANDA, 2025](https://www.forbes.com/sites/jasonsnyder/2025/08/26/mit-finds-95-of-genai-pilots-fail-because-companies-avoid-friction/)), and every one of those is a company whose next attempt starts with less credibility than the last.
### Why the last engagement left nothing behind
Something should have survived it. Usually four things didn't:
**The evals.** If nobody built a test set with known-good outcomes, the next vendor has no way to know whether they're better than the last one — and neither do you. This is the single most valuable artifact an engagement can leave, and it's almost never in the deliverables list.
**The integration work.** Authenticating into your helpdesk, mapping your data, handling the exceptions — that's weeks of work that should be reusable. If it lives inside a vendor's platform rather than your repo, it leaves when they do.
**The decisions.** Why this workflow and not that one. Why the human checkpoint is there. Which approach was tried and abandoned. Undocumented, the next vendor re-derives it — often re-making the same abandoned choice.
**The owner.** If the accountable person was on the vendor's side, accountability left with the invoice. This is the one that guarantees a repeat.
### The pattern underneath
Vendor Roulette is usually a *scoping* failure wearing a vendor's face. If each engagement was scoped as a project with a delivery date rather than a capability with an owner, then finishing means handing over and leaving — and nothing about that arrangement is designed to leave you more capable than it found you.
That's also why the fix isn't "find a better vendor." It's changing what you're buying. The [MIT study](https://www.forbes.com/sites/jasonsnyder/2025/08/26/mit-finds-95-of-genai-pilots-fail-because-companies-avoid-friction/) found buying from or partnering with specialised vendors succeeded about **67% of the time versus roughly a third that rate for internal builds** — so partnering works. It's the *shape* of the partnership that determines whether you keep starting over.
### How to make this one the last
- **Demand the artifacts by name.** Evals, integration code, prompts, decision log, runbook — in your repo, in your accounts, from week one. Not at handover. Handover never happens on the timeline you imagined.
- **Name your owner, on your side.** Someone whose job is affected by the metric, who will still be there when the vendor isn't.
- **Ask what the exit looks like before you sign.** If leaving means rebuilding, you haven't bought a capability — you've rented one. The [nine questions](https://gigabit.agency/insights/questions-to-ask-an-ai-agency/) cover this in diligence detail.
- **Buy something that ends in production, not a phase that ends in a document.** A deployed system is a thing that survives a relationship. A strategy deck is not.
- **Carry the last engagement's wreckage in.** Whatever exists — a half-built pipeline, an abandoned prototype, notes from the failed pilot — is worth real weeks to the next team. A vendor who doesn't want to look at it is planning to bill you for rebuilding it.
This is most of the argument for the [forward-deployed model](https://gigabit.agency/approach/): the engineers sit inside your operation, the work lands in your systems as it's built, and the [Audit → Build → Operate](https://gigabit.agency/approach/) sequence is designed so the operate phase has something real to operate. Where that differs from a traditional consultancy is laid out in [our comparison](https://gigabit.agency/compare/gigabit-vs-consultancies/) — including the cases where a consultancy is genuinely the better call.
If you're mid-roulette right now, the most useful thing you can do before signing anything is inventory what your previous vendors left behind. Usually it's less than you assumed, and knowing exactly how much less is what makes the next scope honest.

### FAQ
**Why do companies keep switching AI vendors?**
Usually because each engagement ends without leaving anything reusable behind — no eval suite, no integration code in your own repo, no record of the decisions, and no accountable owner on your side. Every new vendor then starts from the same blank page, so the pattern repeats regardless of which vendor you pick.
**What should an AI engagement leave behind?**
Four things: an eval suite with known-good outcomes so the next team can tell whether they're better, the integration code and prompts in your own repository and accounts, a written record of which approaches were tried and why choices were made, and a named owner on your side who remains after the vendor leaves.
**What does switching AI vendors actually cost?**
More than the fees. You lose the accumulated context about your edge cases and data, you spend another 4–8 weeks of ramp before the new team is productive, and — most expensively — you spend organisational patience. After two visible failures the third attempt typically gets less access, less budget, and more scepticism.
**Should I tell a new AI vendor about the failed previous attempt?**
Yes, and show them everything that exists — the half-built pipeline, the abandoned prototype, the notes. It's worth real weeks to a competent team. A vendor who doesn't want to look at prior work is planning to bill you for rebuilding it.

## What is an AI readiness assessment, and do you actually need one?
URL: https://gigabit.agency/insights/what-is-an-ai-readiness-assessment/
Published 2026-09-02

**An AI readiness assessment is a scored diagnostic of whether your organisation can actually ship and run AI — not whether AI could theoretically help you.** It looks at five operational dimensions, produces a score per dimension, and names the gap to close first. It is a triage instrument, not a strategy document, and that distinction is most of what separates a useful one from an expensive one.
The reason it exists: **91% of mid-market firms use generative AI, but only 25% have it integrated into core operations** ([RSM, 2025](https://rsmus.com/newsroom/2025/middle-market-firms-rapidly-embracing-generative-ai-but-expertise-gaps-pose-risks-rsm-2025-ai-survey.html)). That 66-point gap is almost never a shortage of ideas. It's one of five things being missing, and an assessment's whole job is telling you which.
### The five dimensions
- **Data** — does the knowledge the system needs live somewhere queryable, or in inboxes, PDFs, and people's heads? Poor data readiness is the top obstacle data leaders name for moving pilots to production ([Informatica, 2025](https://www.informatica.com/blogs/cdo-insights-2025-global-data-leaders-racing-ahead-despite-headwinds-to-being-ai-ready-latest-survey-finds.html)).
- **Process** — is the workflow stable and documented enough that a correct outcome is definable? A process that changes weekly can't be automated; it can only be chased.
- **Tooling** — can software authenticate into your systems and *write*, not just read? This is where homegrown internal tools and expensive integration tiers surface.
- **Team** — is there anyone who will own the system after launch, and enough technical capacity to be a real counterpart during the build?
- **Alignment** — do the people funding it, running it, and being changed by it agree on what success is? This is the quietest failure and the most fatal.
You'll notice none of these are about the model. That's the point. Model capability is not your constraint in 2026 — it's the least of your problems and it improves without you doing anything.
### What a good assessment produces
Three things, and nothing else:
- **A score per dimension**, so the weak one is visible rather than averaged away. A single composite number is worse than useless — a company strong on four dimensions and broken on one has a specific problem, not a mediocre overall grade.
- **A named first move.** One thing to do next, chosen because it unblocks the most.
- **An honest verdict on timing** — including *not yet*, when that's the answer.
What it should *not* produce is a roadmap, a maturity curve with your logo on it, or a list of fifteen use cases ranked by excitement. Those are the deliverables of an assessment sold as the front end of a longer engagement, and they're the reason "readiness assessment" has a slightly bad reputation.
### Readiness of the organisation vs. readiness of the workflow
These are two different tests and you want both to come back clean.
The five dimensions above are the **organisational** test — can this company operate a live AI system at all. The **workflow** test is separate: is this specific piece of work worth automating, and is it automatable? That one turns on volume, whether a right answer is knowable, and whether anyone owns the metric — we cover it in [five signs your business is ready for an AI agent](https://gigabit.agency/insights/signs-your-business-is-ready-for-an-ai-agent/).
A company can pass one and fail the other in both directions. Great data and tooling but no workflow that costs enough to justify a build. Or an obviously painful, high-volume workflow inside an organisation with no owner and no alignment — which fails despite the business case being excellent on paper.
### When you don't need one
Skip it if you can already answer these without discussion: which workflow, what it costs you today, who owns the number, and whether software can reach the systems involved. If those four are genuinely settled, an assessment will confirm what you know and cost you two weeks. Go scope the build.
Skip it too if you've shipped production software before and have engineers who can evaluate the integration surface themselves. Readiness assessments are most valuable to organisations that haven't done this — which is most of the mid-market, and no criticism of anyone.
**Don't** skip it if two or more of the five dimensions make you hesitate, if a previous AI attempt stalled and nobody agreed on why, or if the people funding this and the people whose work changes have never been in the same conversation.
### Free triage vs. a paid diagnostic
They answer different questions, and conflating them is how people overpay.
A **free assessment** is triage. Ours is [six questions, one for each of these five dimensions plus what the problem costs you today](https://gigabit.agency/tools/ai-readiness-assessment/), scored, with a recommended first move, in a few minutes. It tells you roughly where you stand and what's weakest. It cannot tell you what to build, because it hasn't looked at your systems.
A **paid diagnostic** looks at the actual workflow and data and ends with something real. Ours is the [AI Transformation Sprint](https://gigabit.agency/services/ai-transformation/) at $25,000 — two weeks, ending in a deployed pilot and a costed build plan rather than a slide deck, and it credits toward the build it scopes. That's the version worth money: a diagnosis that leaves working software behind, which is the [Audit phase](https://gigabit.agency/approach/) of how we run everything.
The rule of thumb: use the free one to find out whether you have a problem. Pay for a diagnostic only when it ends in something that runs.

### FAQ
**What is an AI readiness assessment?**
A scored diagnostic of whether your organisation can ship and run AI, across five operational dimensions — data, process, tooling, team, and alignment. A good one returns a score per dimension, one recommended first move, and an honest verdict on timing, including 'not yet'. It is triage, not a strategy document.
**Do I actually need an AI readiness assessment?**
Not if you can already say which workflow, what it costs today, who owns the number, and whether software can reach the systems involved — go scope the build instead. Do take one if two or more dimensions make you hesitate, if a previous attempt stalled without agreement on why, or if the funders and the people whose work changes haven't been in the same conversation.
**What does an AI readiness assessment measure?**
Data (is the knowledge queryable), process (is the workflow stable enough that a correct outcome is definable), tooling (can software authenticate and write to your systems), team (will anyone own it after launch), and alignment (do funders, operators, and affected staff agree on success). Notably, none of them concern model capability.
**What's the difference between a free assessment and a paid AI diagnostic?**
A free assessment is self-reported triage — it tells you roughly where you stand and what's weakest, in minutes, without looking at your systems. A paid diagnostic examines the actual workflow and data and should end in something that runs: a deployed pilot and a costed build plan, not a slide deck. Pay only for the second kind.

## What an AI operations runbook actually contains
URL: https://gigabit.agency/insights/ai-operations-runbook/
Published 2026-09-02

A runbook exists so that the person holding the pager at 7am doesn't have to reconstruct your architecture from first principles. For conventional software that's well-trodden. For an AI system it's harder, because **the failures are quieter, the causes are often outside your codebase, and "is it broken?" is a question with a numeric answer rather than a yes or no.**
Here's what actually needs to be in one. Seven entries — each answering a question somebody will ask at an inconvenient hour.
### 1. What this system does, and what it must never do
Two paragraphs at the top: the workflow it owns, the systems it can write to, and the explicit list of actions it is not permitted to take. The negative list matters more than the positive one during an incident, because the first question in any AI incident is *what's the worst this could have done?* and nobody should be inferring the answer from source code under pressure.
### 2. The eval suite and its schedule
Where the test set lives, how to run it, what the current baseline scores are, and how often it runs unattended. Include the last several runs, so drift is visible as a trend rather than a single reading.
This is the entry that makes the rest possible. Without a scored baseline, "is it degraded?" is an opinion — which is why we treat [the eval suite as the first gate](https://gigabit.agency/insights/ai-agent-eval-metrics/), not a nice-to-have.
### 3. Alert thresholds — including the quality one
Latency and error rate, obviously. But the entry that distinguishes an AI runbook is a **quality threshold**: the eval score below which someone gets woken up. Most teams never set one, which is precisely why [degradation runs for weeks unnoticed](https://gigabit.agency/insights/why-ai-systems-break-after-launch/) while infrastructure dashboards stay green.
Set the number in advance, when you're calm. Deciding what counts as "too degraded" during an incident guarantees you'll rationalise upward.
### 4. How to read a trace
For any single run: where the traces live, and how to reconstruct what happened — the input, the retrieved context, the tool calls in order, the output, the cost, the latency. Include one worked example of a real failure walked through end to end.
That worked example is the highest-value paragraph in the document. It's the difference between having observability and knowing how to use it at speed.
### 5. Rollback: what to revert, and to what
Which prompt versions, model versions, and configuration are currently pinned; where the last known-good state is; and the literal command sequence to return to it. Rolling back should be a deploy, not an investigation.
Include the model version explicitly. When a vendor ships a behaviour change, "revert to yesterday's config" is meaningless if the config didn't pin the model.
### 6. The vendor-change watch
Which model versions you depend on, their announced deprecation dates, where those announcements get posted, and who is subscribed. Model vendors deprecate and retune on their schedule, not yours, and a forced migration discovered at the cutover is an emergency; discovered eight weeks out it's a sprint task.
### 7. Ownership and escalation
Who owns this system by name. Who to call when they're unavailable. Who can authorise turning it off, and — importantly — what the business fallback is while it's off. If there is no manual process to fall back to, that's worth knowing before the day you need one.
### The whole thing, on one page
| Entry | The question it answers at 7am |
| --- | --- |
| Scope and prohibitions | What is the worst this could have done? |
| Eval suite and baseline | Is it actually degraded, or does it just feel wrong? |
| Alert thresholds | Should someone have been woken up sooner? |
| Trace reading | What did it do on this specific case? |
| Rollback | How do we stop the bleeding right now? |
| Vendor-change watch | Did something change underneath us? |
| Ownership and escalation | Who decides, and what do we do meanwhile? |
### What a bad runbook looks like
Three failure modes, all common. **Architecture diagrams instead of procedures** — beautiful, and useless at 7am; a runbook is a set of instructions, not an explanation. **No numbers** — "monitor for degradation" is not a threshold. And **written by someone who has never been paged**, which produces a document describing the system as designed rather than as it fails.
The test is simple: hand it to a competent engineer who didn't build the system and ask them to diagnose a degradation. If they can't get to a hypothesis in fifteen minutes, it isn't a runbook yet.
### Who writes it, and when
It gets written during the build, not after — the knowledge is freshest then, and requiring it forces the observability to actually exist. In our [Build phase](https://gigabit.agency/approach/) the runbook and the eval suite ship with the system as part of the deliverable, alongside the code, because a production system handed over without them isn't finished.
If you'd rather someone else hold the pager, that's what [Managed AI Operations](https://gigabit.agency/services/managed-ai-operations/) is — continuous evals, drift and quality monitoring, prompt and model versioning, vendor-change response, and a quarterly roadmap against your KPIs, from $3,000 a month for a single system. Either way, the runbook has to exist. The only question is whose desk it sits on.

### FAQ
**What should an AI operations runbook contain?**
Seven entries: what the system does and what it must never do, where the eval suite lives with its current baseline and schedule, alert thresholds including a quality threshold, how to read a trace for a single run with a worked example, the rollback procedure and what versions are pinned, the model deprecation watch, and named ownership with escalation and a business fallback.
**How is an AI runbook different from a normal software runbook?**
Because AI failures are quiet and often originate outside your codebase. A conventional runbook assumes broken is a yes/no state; an AI runbook needs a numeric quality threshold from the eval suite, a way to reconstruct a single run's reasoning from traces, and a watch on model vendor deprecations that can change behaviour with no deploy on your side.
**What alert threshold should an AI system have?**
Beyond latency and error rate, set a quality threshold — the eval score below which someone is paged. Choose the number in advance while you're calm, because deciding what counts as too degraded during an incident reliably drifts upward. Most teams never set one, which is why degradation typically runs for weeks while infrastructure dashboards stay green.
**When should the runbook be written?**
During the build, not after. The knowledge is freshest then and requiring the document forces the observability to actually exist. A production system handed over without a runbook and an eval suite isn't finished — both should be in the deliverable alongside the code.

## How to hand over an AI system — to a new vendor or in-house
URL: https://gigabit.agency/insights/ai-system-handover/
Published 2026-09-02

**A handover has worked when the receiving team can break the system and fix it without calling anyone.** That's the bar, and almost every handover that gets called complete falls short of it — which is why companies end up rebuilding a working system rather than inheriting it.
We've written about the pattern this creates in [Vendor Roulette](https://gigabit.agency/insights/vendor-roulette-switching-ai-vendors/): each engagement ends leaving nothing behind, so each new team starts from a blank page. That piece names what should survive. This one is how to actually collect it.
### The inventory
Everything below should be in *your* accounts and *your* repository before the outgoing team's last day — not shared from theirs, not promised at handover.
| Asset | Why it's the one that matters |
| --- | --- |
| Application code | Table stakes. Rarely the thing that's missing |
| Prompts, versioned | The accumulated tuning. Often lives outside the repo |
| Eval suite + test data | Without it the new team can't tell if they've made things worse |
| Data pipelines | The unglamorous majority of the real work |
| Infrastructure config | As code, not as someone's console clicks |
| Credentials and accounts | In your name, with your billing, before the last day |
| Decision log | Why this approach, what was tried and abandoned |
| Runbook | So somebody can be on call from day one |
The two that go missing most often are the **eval suite** and the **decision log**, and they're the two that cost the most. Without evals, the incoming team has no way to know whether their first change improved or degraded the system — so they either freeze it or break it. Without a decision log, they re-derive the same choices, frequently re-making an option the last team tried and abandoned for a reason nobody wrote down.
### The acceptance test
Don't accept a handover on a walkthrough. Accept it on a demonstration, with the outgoing team present but silent:
- **Reproduce the baseline.** The new team runs the eval suite from a clean checkout and gets the documented score. If they can't, the environment isn't actually transferred.
- **Ship a trivial change.** A prompt tweak, through the full pipeline to production, and roll it back. This proves the deploy path, the versioning, and the rollback are real rather than described.
- **Diagnose a seeded failure.** The outgoing team breaks something deliberately; the new team finds it from traces alone. This is the one that exposes missing observability.
- **Answer three "why" questions.** Why this model, why that human checkpoint, why this workflow was scoped the way it was — from the decision log, not from memory in the room.
Pass all four and the handover is real. Anything less and you have a document transfer.
### Build in an overlap, and make it work
Two weeks of paid overlap where the **new team operates and the old team advises** — not the reverse. If the outgoing team is still driving, nothing has transferred; you've just extended the engagement.
The most valuable thing to spend that overlap on is a real incident. If one doesn't occur naturally, seed one. Nothing surfaces an incomplete handover faster than something genuinely being wrong while the people who built it are still reachable.
### Taking it in-house instead
Same inventory, one extra question: **who is on call?** Teams routinely underestimate this because the build looks like the hard part. It isn't — the build ends, and operations don't. Before committing, confirm someone owns the eval baseline, someone watches model deprecations, and someone is genuinely reachable when quality drops, [because it will](https://gigabit.agency/insights/why-ai-systems-break-after-launch/).
In-house makes clear sense when AI is core to your product and you can hire senior AI engineers — we lay out the tradeoff in [build vs. buy](https://gigabit.agency/compare/build-vs-buy-vs-embed/). It makes less sense when you're absorbing one production system into a team with no capacity to operate it, which is how systems quietly stop being maintained without anyone deciding to stop maintaining them.
### Ask for this before you sign, not at the end
Every item above is dramatically cheaper to obtain as a contractual condition at kickoff than as a request during offboarding, when goodwill is lowest and the team is already reassigned. Put the inventory and the four-part acceptance test in the agreement — it's one of the [nine questions worth asking any vendor](https://gigabit.agency/insights/questions-to-ask-an-ai-agency/), and the answer tells you a lot about the engagement you're about to enter.
If the answer is that you'd rather not hold the pager at all, [Managed AI Operations](https://gigabit.agency/services/managed-ai-operations/) is the alternative — but even then, insist the inventory lands in your accounts. **Being able to leave is what makes staying a decision rather than a condition.**

### FAQ
**What should be included in an AI system handover?**
Application code, versioned prompts, the eval suite and its test data, data pipelines, infrastructure config as code, credentials and accounts in your own name, a decision log explaining what was tried and abandoned, and an operations runbook. All of it in your repository and accounts before the outgoing team's last day.
**How do I know an AI handover actually worked?**
Test it rather than accepting a walkthrough. The receiving team should reproduce the documented eval baseline from a clean checkout, ship and roll back a trivial change through the full pipeline, diagnose a deliberately seeded failure from traces alone, and answer three 'why' questions from the decision log. Pass all four and it's real.
**What usually goes missing in an AI vendor handover?**
The eval suite and the decision log. Without evals the incoming team can't tell whether a change improved or degraded the system, so they either freeze it or break it. Without a decision log they re-derive past choices and often re-make an option the previous team abandoned for an unrecorded reason.
**Should we take our AI system in-house?**
It makes sense when AI is core to your product and you can hire senior AI engineers. It makes less sense when you're absorbing one production system into a team with no capacity to operate it. The question people underestimate isn't whether they can maintain the code — it's who is on call when quality drops.

## The hidden costs of AI projects nobody puts in the proposal
URL: https://gigabit.agency/insights/hidden-costs-of-ai-projects/
Published 2026-09-02

There are two ways an AI project costs more than the number on the proposal. One is the vendor's pricing model — the estimate that becomes a meter, which we covered in [Bait-and-Bill](https://gigabit.agency/insights/bait-and-bill-ai-pricing-trap/). **The other is entirely your own side of the ledger, and it's the one nobody quotes because no vendor can.**
These are the seven internal costs that reliably surprise people. None of them are anyone's fault. All of them are predictable, which means all of them are budgetable.
### 1. Your own team's time
The largest hidden line, by a wide margin. Any real AI build consumes your people: subject-matter experts explaining the workflow and its exceptions, IT provisioning access, a data person answering questions about tables nobody documented, and someone senior making decisions.
Budget **10–20% of a technical person's time and a meaningful share of one domain expert's** for the duration. On a 90-day build that's real money, and it comes out of work those people were otherwise doing. Vendors don't quote it because they can't, and buyers don't count it because it isn't cash leaving the building. It's still cost.
### 2. Getting the data usable
Almost every proposal assumes the data is roughly where it should be. Frequently it's in a shared inbox, a spreadsheet with three versions, or a system whose export is a PDF. Poor data readiness is the top obstacle data leaders name for moving pilots to production ([Informatica, 2025](https://www.informatica.com/blogs/cdo-insights-2025-global-data-leaders-racing-ahead-despite-headwinds-to-being-ai-ready-latest-survey-finds.html)) — and the cleanup usually falls to you, because it requires knowing which of the three spreadsheet versions is correct.
### 3. The integration tier you didn't know existed
Your helpdesk has an API. It's on the enterprise plan. This is a per-seat licence increase discovered in week three, and it's occasionally larger than a chunk of the build itself. Check the API tier of every system in scope *before* the build is approved — it takes an afternoon and it's the single highest-yield hour of diligence available.
### 4. Security and compliance review
A new system touching customer data needs a review, and reviews take calendar time even when they pass. In regulated environments — healthcare, finance — budget weeks, not days, and budget the *internal* effort of assembling evidence, not just the reviewer's time.
The mitigation is to start it in parallel with the build rather than after. Teams that sequence it last routinely finish engineering on time and ship two months late.
### 5. Change management
The cost of people actually using the thing. Training, documentation, the productivity dip while a team learns a new process, and the political work of persuading people whose jobs change. This is why **91% of mid-market firms use generative AI but only 25% have it integrated into core operations** ([RSM, 2025](https://rsmus.com/newsroom/2025/middle-market-firms-rapidly-embracing-generative-ai-but-expertise-gaps-pose-risks-rsm-2025-ai-survey.html)) — the gap is adoption, not capability.
### 6. The parallel-run period
You will not switch off the manual process on day one, nor should you. For some weeks you run both: the agent handles cases and humans check them. That's a real, temporary cost increase — you're paying for the system and the old process simultaneously — and it belongs in the model as a defined period rather than appearing as an unpleasant surprise in month four.
### 7. Operating it forever
The one that isn't hidden so much as deferred. Monitoring, eval regressions, drift management, and model migrations don't stop, and an unoperated system [decays quietly](https://gigabit.agency/insights/why-ai-systems-break-after-launch/). Managed, that's $3,000–$20,000 a month depending on how many systems you run. Staffed internally, it's a fraction of an engineer indefinitely. It is never zero.
### The whole picture
| Cost | Who pays it | When it surfaces |
| --- | --- | --- |
| Your team's time | You | Throughout |
| Data cleanup | Usually you | Week 2–4 |
| Integration licence tiers | You | Week 3, painfully |
| Security and compliance review | You | Late, unless started early |
| Change management | You | At rollout |
| Parallel run | You | Weeks 8–16 |
| Operations | Either | Forever, from go-live |
A workable rule of thumb: **add 30–50% to the vendor's number** to get the true first-year cost of the change. That isn't a criticism of the quote — the quote is for the software. This is the cost of the business absorbing it.
### What to do about it
Price them, don't discover them. Before approving anything, ask your own team four questions: *how many hours of ours does this consume, what state is the data actually in, what API tier does each system need, and who runs it after launch?* Four answers turns most of this list from a surprise into a line item.
Then build the total into the business case rather than the vendor's figure alone, so the [ROI arithmetic](https://gigabit.agency/tools/ai-roi-calculator/) is honest — a project that pays back on the quote and not on the true cost is a project that will disappoint someone in month nine. [What a mid-market AI budget should look like](https://gigabit.agency/insights/mid-market-ai-budget/) puts the whole first-year number together, and [our prices](https://gigabit.agency/pricing/) are published so at least the vendor half of it needs no discovery.

### FAQ
**What are the hidden costs of an AI project?**
Seven internal ones that no vendor quotes: your own team's time, getting the data into usable shape, API or integration licence tiers you discover mid-build, security and compliance review, change management and training, the parallel-run period where you pay for both the system and the old process, and ongoing operations after launch.
**How much should I add to a vendor quote for the true cost?**
As a rule of thumb, 30–50% for the first year. The quote covers the software; the addition covers your organisation absorbing it — your people's time, data cleanup, licence tier increases, review cycles, training, and the parallel-run period.
**What's the biggest unbudgeted cost in an AI project?**
Your own team's time. A real build consumes subject-matter experts explaining workflow exceptions, IT provisioning access, a data person answering questions, and someone senior deciding. Budget 10–20% of a technical person plus a meaningful share of a domain expert for the duration. It isn't cash leaving the building, which is exactly why it goes uncounted.
**How do I avoid being surprised by AI project costs?**
Ask four questions before approving anything: how many hours of our own people does this consume, what state is the data actually in, what API tier does each system in scope require, and who operates it after launch. Checking integration tiers alone takes an afternoon and prevents the most common week-three surprise.

## Solution looking for a problem: how to spot a doomed AI project
URL: https://gigabit.agency/insights/solution-looking-for-a-problem/
Published 2026-09-02

Most doomed AI projects are recognisable in the first meeting. Not because anyone is incompetent — usually the opposite — but because the project is shaped wrong from the start, and the shape is visible if you know what to look for.
This is the *pre-commitment* view. [Why your AI pilot never reached production](https://gigabit.agency/insights/why-your-ai-pilot-never-reached-production/) is the post-mortem: the engineering gates that separate a demo from a deployed system. This piece is earlier and cheaper — five tells, each with a question you can ask out loud before anyone writes code.
The stakes justify the paranoia. **95% of enterprise generative-AI pilots deliver no measurable P&L impact** ([MIT Project NANDA, 2025](https://www.forbes.com/sites/jasonsnyder/2025/08/26/mit-finds-95-of-genai-pilots-fail-because-companies-avoid-friction/)), and Gartner expects **more than 40% of agentic-AI projects to be canceled by the end of 2027** ([Gartner, 2025](https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027)). Those are mostly projects that never should have been funded in the shape they were funded.
### Tell 1 — It starts with the technology
The project began as *we should be using AI* and worked backwards to a use case. You can hear it in how it's described: capability first, problem second, and the problem sounds slightly retrofitted because it is.
**Ask:** *what were we going to do about this before AI existed as an option?* If the honest answer is "nothing, it wasn't a priority," you're looking at a technology in search of a justification. That's not automatically fatal — but it means there's no baseline urgency to carry the project through the hard middle, and every AI project has a hard middle.
### Tell 2 — Nobody can name whose day gets better
Healthy projects have a person attached. *Priya in support stops spending her mornings triaging.* Doomed ones have an abstraction: efficiency, the customer experience, the organisation.
**Ask:** *whose specific job changes on the day this ships, and have we talked to them?* If nobody's day changes, nobody will adopt it. And if the person whose day changes hasn't been consulted, expect to discover in month four that the workflow has exceptions the plan didn't account for — because it always does, and they're the only one who knows them.
### Tell 3 — Success is defined as launch
Listen to how done is described. *Deployed by Q3.* *Live in the support flow.* Those are milestones, not outcomes — and a project whose definition of success is its own existence cannot fail, which means it also cannot succeed.
**Ask:** *what number will be different in ninety days, and who reports it?* A project that can't answer will produce something that launches, gets a congratulatory email, and quietly stops being used. That's the modal outcome, and it's worse than a visible failure because nobody learns anything.
### Tell 4 — The sponsor can't say what stops
If the AI does the work, something a human does now must stop, shrink, or change. A sponsor who hasn't thought about that hasn't thought about the project — they've thought about the demo.
**Ask:** *what do we stop doing when this works?* Hesitation here predicts the parallel-run that never ends: the agent handles cases, humans check every one forever, and the saving never materialises because the old process was never actually retired. The answer can legitimately be "nothing stops, we absorb more volume with the same team" — that's a fine answer. Having no answer is not.
### Tell 5 — It's a demo hunting for a budget
Someone built something impressive, showed it around, and now the organisation is looking for a problem worthy of it. The tell is that the project's champion is whoever built the prototype rather than whoever owns the workflow.
**Ask:** *if this prototype didn't exist, would this be in our top five priorities?* If not, the prototype is doing the arguing, and prototypes are extremely persuasive about their own necessity.
### The five, as questions
| Tell | Ask | Bad answer sounds like |
| --- | --- | --- |
| Technology first | What were we doing about this before AI? | Nothing, it wasn't a priority |
| No named beneficiary | Whose job changes, and have we asked them? | It helps the whole org |
| Launch as success | What number changes in 90 days? | We'll define KPIs post-launch |
| Nothing stops | What do we stop doing when this works? | Hesitation, then "good question" |
| Demo-driven | Would this be top five without the prototype? | But have you seen the demo? |
### What to do when you spot one
Killing it is rarely the right move, and rarely politically available. Reshaping it usually is:
- **Re-anchor on a costed workflow.** Find the work already consuming 15+ hours a week and point the project at it. Most doomed projects are aimed at the wrong target rather than being fundamentally bad ideas — the [five signs](https://gigabit.agency/insights/signs-your-business-is-ready-for-an-ai-agent/) are the qualification test.
- **Attach a name and a number.** One person, one metric, before more money is spent. This is the cheapest intervention available and the most predictive.
- **Redefine done as running unattended,** not as deployed. It changes what gets built, starting immediately.
- **Shrink it until it's fundable on evidence.** A two-week fixed-price diagnostic that ends in a [deployed pilot](https://gigabit.agency/services/ai-transformation/) tells you more than another quarter of planning, and it's cheap enough to be wrong about.
If you'd rather diagnose the organisation than the project, the [AI Readiness Assessment](https://gigabit.agency/tools/ai-readiness-assessment/) scores the five operational dimensions underneath all of this. And the reason our engagements start with a paid [Audit phase](https://gigabit.agency/approach/) is precisely this: the most expensive AI project is the well-executed one aimed at the wrong problem.

### FAQ
**How can I tell if an AI project is doomed?**
Five tells visible in the first meeting: it started from the technology rather than a problem, nobody can name whose specific job changes, success is defined as launching rather than a number moving, the sponsor can't say what the business stops doing when it works, and the project's champion is whoever built the prototype rather than whoever owns the workflow.
**What is a solution looking for a problem in AI?**
A project that began as 'we should be using AI' and worked backwards to a use case. The test is asking what you were going to do about this before AI was an option — if the honest answer is nothing, there's no baseline urgency to carry the project through its hard middle, and every AI project has one.
**What question best predicts whether an AI project will succeed?**
What do we stop doing when this works? A sponsor who hasn't considered it has thought about the demo rather than the project, and it predicts a parallel run that never ends — the agent handles cases, humans check every one indefinitely, and the saving never materialises because the old process was never retired.
**Should I cancel a badly shaped AI project?**
Usually reshape rather than cancel. Re-anchor it on a workflow that already costs measurable time, attach one named owner and one metric, redefine done as running unattended rather than deployed, and shrink it to a short fixed-price diagnostic that ends in something running. Most doomed projects are aimed at the wrong target rather than being bad ideas.

## Rules-based automation vs. AI agents: when you need the upgrade
URL: https://gigabit.agency/insights/rules-based-automation-vs-ai-agents/
Published 2026-09-02

**Rules-based automation is deterministic, auditable, essentially free to run, and does exactly the same thing every time. Those are advantages, not limitations.** If your workflow fits a rules engine, an AI agent is a downgrade — slower, costlier, and non-deterministic in exchange for judgment you don't need.
The upgrade question isn't which technology is more advanced. It's whether your workflow has an exception problem.
### What rules do well
Anything expressible as *if this, then that* with a knowable, finite set of conditions. Routing on field values, moving data between systems, triggering on thresholds, scheduled batches. Zapier, Make, n8n, and RPA tools handle these excellently and will keep doing so.
They also have properties an agent can't match: you can read the logic, an auditor can verify it, it costs approximately nothing per execution, and it behaves identically on the ten-thousandth run. In regulated contexts that determinism is sometimes the requirement, not a preference.
**Don't replace working rules with an agent because agents are newer.** That's the most common expensive mistake in this category.
### The signal that you've outgrown them
Not complexity — **exception rate**. Rules handle the cases you anticipated. The upgrade question is what fraction of real volume falls outside them and lands on a human.
Measure it for a month: of everything entering the workflow, what percentage does the automation route correctly with no human touch? Roughly:
| Exception rate | What it means | Move |
| --- | --- | --- |
| Under 10% | Rules fit the work | Keep them. Tune the edges |
| 10–25% | Rules are straining | Try more rules first; measure again |
| Over 25% | The work needs judgment | An agent is likely justified |
| Rules count keeps growing | You're encoding judgment as branches | Agent, or a redesign |
That last row is the real tell. When a rules system accumulates dozens of conditions that only one person understands, you haven't automated judgment — you've encoded it in a form nobody can maintain. The maintenance burden becomes the cost, not the licence.
### The three differences that matter
- **Unstructured input.** Rules need fields. If the work arrives as free text, an email thread, a PDF, or a photo, rules can only act on what someone else has already structured — and that someone is usually a person doing manual triage.
- **Judgment under ambiguity.** *Is this complaint urgent?* has no threshold. Rules approximate it with proxies (keywords, sender domain) that work until they don't.
- **Multi-step work with state.** An agent carries context across steps and adapts when a step fails. A rules chain executes its branch and stops.
If none of those describe your workflow, you don't need an agent. That's a good outcome — it's cheaper.
### The pattern that usually wins
**Both.** The best production designs we build are rules-first with an agent on the exceptions: deterministic logic handles the 70–80% it handles perfectly and cheaply, and the agent takes only what falls out.
This is better than either alone. You keep determinism and auditability where they're available, you pay model costs only on genuinely hard cases, and — usefully — the exception queue is already a labelled dataset showing exactly what the agent needs to handle. It also shrinks the blast radius, because the agent touches a minority of volume rather than all of it.
### What changes when you upgrade
An agent is not a drop-in replacement, and the operational shape is different. Rules either work or throw an error; an agent can be wrong quietly and confidently. So you inherit obligations rules never had: an eval suite to score correctness, traces to reconstruct decisions, guardrails to bound the worst case, and someone to [operate it after launch](https://gigabit.agency/insights/why-ai-systems-break-after-launch/).
That's the real cost of the upgrade — not the tokens, which are cents per transaction. It's why the exception rate needs to genuinely justify it. If you're deciding between extending your rules and commissioning an agent, [build vs. buy](https://gigabit.agency/compare/build-vs-buy-vs-embed/) covers the sourcing question, and the [five signs](https://gigabit.agency/insights/signs-your-business-is-ready-for-an-ai-agent/) test whether the workflow qualifies at all. When it does, [Gigabit Agents](https://gigabit.agency/services/ai-agents/) builds are a flat fee from $8,000 per agent.

### FAQ
**When should I replace rules-based automation with an AI agent?**
When the exception rate — the share of real volume your rules route incorrectly or hand to a human — exceeds about 25%, or when the rule count keeps growing because you're encoding judgment as branches. Under 10%, keep the rules: they're cheaper, deterministic, auditable, and behave identically every run.
**What's the difference between RPA and an AI agent?**
RPA and rules engines execute predetermined logic on structured inputs and do the same thing every time. An agent handles unstructured input, exercises judgment where no threshold exists, and carries state across multiple steps, adapting when one fails. The trade is determinism and near-zero run cost for flexibility.
**Can rules-based automation and AI agents work together?**
Yes, and it's usually the best design: rules handle the 70–80% they handle perfectly and cheaply, and an agent takes only the exceptions. You keep auditability where it's available, pay model costs only on hard cases, shrink the agent's blast radius, and your existing exception queue is already a labelled dataset of what the agent must handle.
**Are AI agents better than Zapier or n8n?**
Not for work those tools fit. If a workflow is expressible as if-this-then-that on structured fields with a finite set of conditions, a rules tool is faster, cheaper, auditable, and fully predictable — an agent would be a downgrade. Replacing working rules simply because agents are newer is the most common expensive mistake in this category.

## When should an AI agent stop and ask a human?
URL: https://gigabit.agency/insights/ai-agent-human-in-the-loop/
Published 2026-09-02, updated 2026-10-10

**An AI agent should stop and ask a human in three situations: the action is hard to undo, a mistake would not be noticed quickly, or the agent is not confident in its own answer.** Everything else it can do alone. The rest of this piece is how to apply that, action by action.
The question is usually asked as though autonomy were a dial with one setting per agent. It isn't. **A well-built agent runs at different supervision levels for different steps of the same workflow**, and getting that mapping right is most of what makes an agent safe enough to deploy and useful enough to be worth deploying.
Two questions decide the level for any given action: **how reversible is it**, and **how detectable is a mistake**. Everything else follows from those.
### The four supervision patterns
| Pattern | The agent… | Right when |
| --- | --- | --- |
| Full autonomy | Acts, nobody reviews | Reversible and low-value per case |
| Sampled review | Acts; a percentage is audited after | High volume, errors tolerable and detectable |
| Approve-before-act | Proposes; a human confirms | Irreversible or high-value |
| Escalation-only | Acts when confident, hands off when not | Most real production workflows |
**Full autonomy** suits work where a mistake is cheap and self-correcting — tagging, routing, drafting an internal summary. Nobody should review every ticket classification; the cost of review exceeds the cost of the error.
**Sampled review** is the workhorse for volume. The agent acts on everything, and 2–5% is audited by a human whose real job is watching for drift rather than catching individual errors. This is how you keep the throughput of autonomy and still have a human signal on quality.
**Approve-before-act** is mandatory wherever an action moves money, contacts a customer irreversibly, changes an entitlement, or writes to a system of record you can't easily undo. The agent prepares the action and a person clicks. Slower — and still far faster than a human doing the whole task from scratch.
**Escalation-only** is what most production workflows actually look like: the agent handles what it's confident about and routes the rest to a person with its reasoning attached. This is what produces the [60–70% autonomous resolution](https://gigabit.agency/insights/ai-support-automation-saas/) that a well-built support agent reaches — not 100%, and 100% was never the target.
### The two questions, applied
**Reversibility.** Can you undo it in one step, at low cost? Re-routing a mis-routed ticket is trivial. Un-sending an email to a customer is impossible. Refunding a wrongly-charged card is possible but expensive in trust. Sort your actions by how hard the undo is, and the approval gates place themselves.
**Detectability.** If it goes wrong, will anyone find out — and how fast? This is the one teams under-weight. An error that surfaces immediately is far safer than a quiet one, even at the same severity, because a quiet error compounds. An agent silently mis-categorising 8% of records for three months does more damage than one that fails loudly on day one.
Where detectability is poor, add supervision even if the individual action seems low-stakes. That's the same reason a live system needs a quality threshold in its [runbook](https://gigabit.agency/insights/ai-operations-runbook/) — the failures that hurt are the ones nobody sees.
### The escalation matrix
Put the two questions together and each kind of action lands in one row. Use it as a starting point and tighten or loosen a row as the evidence comes in.
| The action is… | Example | The agent should |
| --- | --- | --- |
| Easy to undo, and a mistake shows up fast | Tagging, routing, drafting an internal summary | Act alone |
| Easy to undo, high volume, errors tolerable | Classifying every incoming ticket | Act alone, with 2–5% audited afterwards |
| Hard to undo or high-value | Issuing a refund above a threshold, a bespoke customer email, changing an entitlement, writing to a system of record | Prepare it, then wait for a person to approve |
| Any action where the agent isn't confident | A case that doesn't match what it has seen | Hand off, with its reasoning attached |
| Low-stakes, but a mistake would stay quiet | Categorising records nobody re-checks | Act, under added review until errors are visible |
| Wrong would be both unnoticed and unrecoverable | A decision with no way to verify or reverse it | Not act at all; this stays with a person |
### Map it per step, not per agent
Take a refund agent. Reading the ticket, retrieving the order, checking the policy, and drafting a decision are all fully autonomous — reversible, cheap, no reason for a human. Issuing the refund is approve-before-act above a threshold and autonomous below it. Emailing the customer is autonomous for templated confirmations and approve-before-act for anything bespoke.
One agent, three supervision levels, chosen per action. Teams that set a single level per agent end up either **approving everything** — at which point the agent saves nothing and adoption collapses — or **approving nothing**, at which point the security review never clears and it stays a demo. Both failure modes come from treating autonomy as a property of the agent instead of the action.
### What an agent genuinely can't do unsupervised
Independently of risk appetite, some things aren't delegable:
- **Decisions where being wrong is undetectable and unrecoverable.** No supervision pattern helps if you'll never know.
- **Novel judgment with no precedent.** Agents generalise from patterns. A genuinely unprecedented case is exactly where confidence estimates are least reliable.
- **Anything requiring accountability a person must hold.** Regulatory sign-off, clinical or legal decisions, employment actions. The constraint isn't capability — it's that responsibility can't be assigned to software.
- **Work whose correct answer nobody can define.** If a human expert can't say whether an output was right, you can't evaluate it, and an unevaluable system can't be trusted unattended. That's the first of the [five gates](https://gigabit.agency/insights/why-your-ai-pilot-never-reached-production/).
### Autonomy should increase with evidence
Start more supervised than you'll end. Run approve-before-act on a step for a few weeks, record how often the human changes the agent's proposal, and let that number decide. Under 2% overrides on a few hundred cases is a good argument for dropping to sampled review. A rising override rate is drift, and it's a signal you'd otherwise have missed.
That's the honest version of autonomy: not a setting chosen at design time, but a level earned per action against measured evidence — and lowered again when the evidence changes. The [layers that make this measurable](https://gigabit.agency/insights/production-ai-agent-stack/) are the same ones that make an agent deployable at all, which is why [our builds](https://gigabit.agency/services/ai-agents/) install them from day one rather than retrofitting them once someone asks for a number.

### FAQ
**When should an AI agent stop and ask a human?**
In three situations: when the action is hard to undo, such as moving money, contacting a customer or writing to a system of record; when a mistake would not be noticed quickly; and when the agent is not confident in its answer. Reversible, low-value actions whose errors show up fast can run without review.
**Can AI agents work without a human in the loop?**
For some actions, yes — those that are reversible, low-value per case, and where errors are detectable. Full autonomy suits tagging, routing, and drafting. Actions that move money, contact customers irreversibly, or change entitlements should stay approve-before-act. Most production workflows run escalation-only: the agent acts when confident and hands off when not.
**How do I decide how much autonomy to give an AI agent?**
Ask two questions per action: how reversible is it, and how detectable is a mistake. Irreversible actions get approval gates. Poorly detectable errors get supervision even when each individual action seems low-stakes, because quiet errors compound while loud ones get fixed.
**Should autonomy be set per agent or per action?**
Per action. A single refund agent might read tickets and draft decisions fully autonomously, require approval to issue refunds above a threshold, and send templated confirmations without review. Teams that set one level per agent either approve everything — so it saves nothing — or approve nothing, so it never clears security review.
**What can an AI agent never do unsupervised?**
Decisions where being wrong is both undetectable and unrecoverable, genuinely novel judgment with no precedent to generalise from, anything requiring accountability a person must legally hold such as regulatory or clinical sign-off, and work whose correct answer nobody can define — because an output no expert can evaluate can't be trusted unattended.

## In-house, freelancer, or agency: who should build your AI agent
URL: https://gigabit.agency/insights/in-house-freelancer-or-agency-ai/
Published 2026-09-02

This question comes after a different one. *Should we build custom at all, or buy a product?* is [build vs. buy](https://gigabit.agency/compare/build-vs-buy-vs-embed/), and it's worth settling first. What follows assumes you've decided to build something specific to your business, and the remaining question is **who does it**.
Four options, and the honest answer turns on two things: whether AI is core to what you sell, and who is operating the system in eighteen months.
### The four models
| Model | Ships in | Best when | Fails when |
| --- | --- | --- | --- |
| Hire in-house | 4–9 months | AI is core to the product | You need it this quarter |
| Freelancer | Weeks | Scoped, self-contained build | It becomes load-bearing |
| Agency / consultancy | Weeks–months | Defined project, clear handover | Nothing is left behind |
| Embedded team | ~2 weeks to start | You want it built *and* transferred | You want to stay hands-off |
### Hiring in-house
The right answer when AI is genuinely core to your product — when the models and the pipelines *are* the thing customers pay for. Then you can't outsource your differentiation, and you shouldn't try.
The cost isn't just salary. It's the **4–9 month reality**: writing a spec you may not be qualified to evaluate, competing for scarce senior people, then ramp. And the first hire is the hardest, because a strong AI engineer wants to work near other strong AI engineers — so hire one and you often have to hire two.
It's the wrong answer when you need something in production this quarter, when this is one system rather than an ongoing capability, or when nobody currently on staff can technically evaluate the candidates.
### Freelancers
Genuinely good for scoped, self-contained work — a proof of concept, a well-defined integration, a specific model fine-tune. Fast to start, no overhead, and the ceiling on individual talent is higher than people assume.
The failure mode is predictable and it's about **continuity**, not competence. One person means no code review, a single point of failure, and no coverage when they're unavailable. When the thing they built becomes load-bearing, you discover the operating burden was never anyone's job. If you go this route, insist from day one on the [artifacts that make it transferable](https://gigabit.agency/insights/ai-system-handover/) — evals, a decision log, a runbook — because a freelancer engagement ends more abruptly than any other kind.
### Agencies and consultancies
The mainstream answer, and often right: capacity now, a team rather than a person, and someone who has done this before. **95% of enterprise generative-AI pilots deliver no measurable P&L impact, but buying from or partnering with specialised vendors succeeded about 67% of the time — roughly three times the rate of internal builds** ([MIT Project NANDA, 2025](https://www.forbes.com/sites/jasonsnyder/2025/08/26/mit-finds-95-of-genai-pilots-fail-because-companies-avoid-friction/)). Partnering works.
The risk is structural rather than moral: a project-shaped engagement ends at a delivery date, and if nothing transfers, you're back at the start when you next need a change. That's [Vendor Roulette](https://gigabit.agency/insights/vendor-roulette-switching-ai-vendors/), and it's the reason to judge an agency by what it leaves behind rather than by what it demos. Ask [the nine questions](https://gigabit.agency/insights/questions-to-ask-an-ai-agency/) before signing anything.
### Embedded teams
The hybrid, and the one we run: senior engineers working inside your systems, on your sprint board, alongside your people — building the thing *and* transferring the capability as they go. [Embedded AI Teams](https://gigabit.agency/services/embedded-ai-teams/) are $3,000–$7,500 per engineer per month, and a pod assembles in about two weeks rather than the months a hire takes.
It suits the common middle case: you want production software soon, you expect to own it eventually, and you don't have the internal bench yet. Your team learns by working next to people who've done it, so capability accrues to you rather than to a vendor relationship — which is the [forward-deployed argument](https://gigabit.agency/approach/) in one sentence.
It's the wrong choice if you want to be genuinely hands-off. Embedding requires your people's time, and if nobody on your side is participating, you're paying a premium for something a project engagement does more cheaply.
### The two questions that decide it
**Is AI core to what you sell?** If yes, you will hire eventually — the only question is whether you hire before or after your first production system. Building first with an embedded team and hiring into a working system beats hiring into a blank page, because your first hire inherits evals, a runbook, and a codebase instead of a mandate.
**Who operates it in eighteen months?** Answer this before you choose, not after. Every model can build something; they differ almost entirely in what happens next. A freelancer leaves, an agency's engagement closes, an in-house team is already there, an embedded pod is designed to hand over. If the honest answer is *nobody yet*, then either budget [managed operations](https://gigabit.agency/services/managed-ai-operations/) or pick a model that leaves your team able to run it.
Most mid-market companies get this wrong in the same direction: they choose on **build cost** when the decision is really about **who holds it afterwards**. Our full comparison of the three shapes — [build vs. buy vs. embed](https://gigabit.agency/compare/build-vs-buy-vs-embed/) — lays out the tradeoffs including where each one is genuinely the better call.

### FAQ
**Should I hire AI engineers in-house or use an agency?**
Hire in-house when AI is core to what you sell — you can't outsource your differentiation — but budget 4–9 months to spec, hire, and ramp. Use an agency or embedded team when you need production software sooner, when it's one system rather than an ongoing capability, or when nobody on staff can technically evaluate AI candidates yet.
**Is a freelancer a good choice for building an AI agent?**
For scoped, self-contained work like a proof of concept or a defined integration, yes — fast to start with no overhead. The risk is continuity rather than competence: one person means no code review, a single point of failure, and no owner for operations once the system becomes load-bearing. Insist on evals, a decision log, and a runbook from day one.
**What is an embedded AI team?**
Senior engineers working inside your systems and on your sprint board alongside your staff, building the system while transferring the capability. A pod assembles in about two weeks rather than the months a hire takes, at $3,000–$7,500 per engineer per month. It suits companies that want production software soon and expect to own it eventually but lack the internal bench.
**What's the most important question when choosing who builds your AI?**
Who operates it in eighteen months. Every model can build something; they differ almost entirely in what happens afterwards — a freelancer leaves, an agency engagement closes, an in-house team is already there, an embedded pod is designed to hand over. Most mid-market companies choose on build cost when the decision is really about who holds it later.

## Your buyers are asking ChatGPT for a shortlist. Are you on it?
URL: https://gigabit.agency/insights/are-you-on-the-ai-shortlist/
Published 2026-09-02, updated 2026-10-22

Somewhere in the last two years, a step got inserted into your sales process that you can't see. Before a buyer fills in a form, before they read a case study, before they've heard of you — a good number of them now ask an assistant *who should we look at for this*, and take the three or four names it gives them.
**That shortlist gets drafted without you in the room.** If you're on it, you get the call. If you're not, you don't lose the deal — you never learn it existed.
### This is not a small channel any more
ChatGPT reached **800 million weekly active users** in October 2025 ([TechCrunch](https://techcrunch.com/2025/10/06/sam-altman-says-chatgpt-has-hit-800m-weekly-active-users/)), and Google's AI Overviews serve **2 billion monthly users** ([TechCrunch](https://techcrunch.com/2025/07/23/googles-ai-overviews-have-2b-monthly-users-ai-mode-100m-in-the-us-and-india/)). The behaviour has already normalised in your buyers' personal lives, and professional research follows personal habit — it always has.
What's different from a search result is that there's no page two. A results page is a menu of ten options a buyer can scan past the first entry. An answer is a recommendation with three or four names in it. The long tail of "also ranked on page one" doesn't exist any more.
### Why your analytics can't see it
This is the part that makes it easy to miss. When an assistant recommends you and the buyer later searches your brand name and clicks through, that visit arrives in your analytics as **direct or branded organic traffic**. It looks like someone who already knew you.
So the channel is invisible in exactly the reports you'd use to notice a channel. Meanwhile the visits you *can* attribute to AI sources tend to convert better — Semrush found an AI-search visitor is roughly **4.4× more valuable** than a traditional organic one ([Semrush, 2025](https://www.semrush.com/blog/ai-search-seo-traffic-study/)), which makes sense: someone arriving after a recommendation is further along than someone who clicked a link.
The uncomfortable implication is that you can be losing this channel steadily without a single metric moving in a way anyone would investigate.
### Where the shortlist comes from now
For two decades B2B discovery ran through intermediaries. A buyer searched *best [category] software*, landed on a review grid or a *top 10* roundup, and picked three names from it. Those intermediaries were powerful because they **owned the click**. When Google shows an AI summary, only **8% of users click any traditional link, versus 15% without one, and just 1% click a source inside the summary itself** ([Pew Research, 2025](https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/)). The roundup still gets *read*, by the model, but it stops being *visited*. **Its role shifts from destination to source.**
So the job changes. You used to want to be listed somewhere a buyer would visit; now you want to be described somewhere a model will read. Three things feed that, in order of how much control you have: your own pages, if they answer real buying questions with quotable specifics; corroboration across the open web that a model can cross-check; and the directories and roundups themselves, still worth being in for their testimony rather than their traffic. Judge a listing partly on whether it is the kind of source an assistant would cite, not only on the referral clicks it sends, and move some spend from renting placement to owning statements.
### Check it in five minutes
You don't need a tool or a budget to find out where you stand. Right now:
- Open a **private or logged-out** browser window. This matters more than anything else here — your own account has months of context about your company and will name you in ways a stranger's session never would.
- Ask the question a buyer would actually type. Not your brand name — something like *best [your category] for [your customer type]*, or *how much does [your service] cost*.
- Read who gets named. Write the names down.
Three outcomes. **You're named** — good, now find out how consistently, because a single answer is stochastic. **A competitor is named and you aren't** — that's the common one, and it's specific enough to act on. **Nobody in your category is named** — the engines don't have a confident answer for this question yet, which is the most interesting result of the three, because it means the position is unclaimed.
### What to do with the answer
Don't over-react to one query. Assistant answers vary between runs, so a single result is an anecdote rather than a measurement — if it's worth knowing properly, it's worth running a real panel, which takes about half an hour and is [written up step by step here](https://gigabit.agency/insights/check-your-ai-visibility-in-30-minutes/).
If a competitor came back and you didn't, the cause is usually one of five specific things — an answer-shaped page you don't have, facts a model can't quote, a thin third-party footprint, inconsistent entity data, or a query nobody actually asks. [Which one is yours](https://gigabit.agency/insights/what-gets-a-company-excluded-from-ai-answers/) determines whether this is a two-week fix or a two-quarter one.
And if nobody in your category was named, that's worth sitting with. Most B2B categories still have no company deliberately working on this. The firm that becomes the default answer over the next year is the one competitors then have to dislodge — which is a much harder job than being first. We call reaching that position [Answer-Market Fit](https://gigabit.agency/insights/answer-market-fit/), and the [field guide](https://gigabit.agency/insights/generative-engine-optimization-field-guide/) is how it's done in practice.
The five-minute version above will tell you whether you have a problem. The [free citation audit](https://gigabit.agency/tools/geo-citation-audit/) runs your category's real buying questions across the major assistants and sends back the scored result — where you appear, where you don't, and who's being recommended instead. Either way: check. Most companies never have, and the ones being recommended today mostly got there by accident.

### FAQ
**Do B2B buyers really use ChatGPT to find vendors?**
Increasingly, yes — often before any form fill or first contact, to draft an initial shortlist of three or four names. ChatGPT reached 800 million weekly users in October 2025 and Google's AI Overviews serve 2 billion monthly, and professional research habits tend to follow personal ones.
**How can I tell if ChatGPT recommends my company?**
Open a private or logged-out browser window, ask the question a buyer would type — best [category] for [customer type], or how much does [service] cost — and note who gets named. Use a logged-out session: your own account carries context about your company and will name you when a stranger's session wouldn't.
**Why doesn't AI search traffic show up in my analytics?**
Because when an assistant recommends you, the buyer typically searches your brand name afterwards and clicks through — arriving as direct or branded organic traffic that looks like someone who already knew you. The channel is effectively invisible in the reports you'd use to notice a channel.
**What if no company in my category gets recommended?**
That's the most interesting result, because the position is unclaimed. Most B2B categories still have no company deliberately working on AI visibility, so the firm that becomes the default answer over the next year is the one competitors then have to dislodge — a much harder job than being first.
**Are review sites like G2 still relevant for B2B discovery?**
Yes, but their role is shifting from destination to source. As AI answers absorb more of the click (only 8% of users click any traditional link when Google shows an AI summary, versus 15% without), directories and roundups get read by models more than visited by buyers. They still shape who gets recommended; they just collect less of the traffic.
**Should I still pay for directory listings and review placements?**
Often yes, but judge them differently. Valuing a listing purely on referral clicks now under-counts one that models read and cite. The useful question is whether a placement is the kind of source an assistant would retrieve from, not just how much traffic it sends.

## llms.txt explained: the new file most sites don't have yet
URL: https://gigabit.agency/insights/llms-txt-explained/
Published 2026-09-02

**`llms.txt` is a single Markdown file at the root of your site that describes, in plain language a model can parse, what your organisation is and where its important pages are.** It's a proposed convention rather than a ratified standard, it takes about an hour to write, and it's frequently oversold. Here's the accurate version.
### What it is
One file, at `https://yourdomain.com/llms.txt`, in Markdown. The conventional shape is a title, a one-paragraph summary of what you do, and then linked sections pointing to the pages that matter, each with a short description of what's there.
Ours opens with a paragraph naming what Gigabit is, the model we run, where our offices are, and how long we've been shipping — then a **Key facts** block with every price stated outright, then linked sections for pricing, services, industries, and proof. You can [read it](https://gigabit.agency/llms.txt); it's about forty lines.
The design intent is a summary that survives being read without a browser: no navigation to traverse, no JavaScript to execute, no marketing copy to wade through. Just what you do, what it costs, and where the substance lives.
### What it isn't
Three corrections, because this file attracts a lot of overclaiming:
**It is not robots.txt.** `robots.txt` grants or denies access; `llms.txt` offers a summary. They solve different problems and one doesn't substitute for the other. If your `robots.txt` blocks the AI crawlers, an `llms.txt` changes nothing at all.
**It is not a ranking factor.** No major assistant has publicly committed to reading it as an input to recommendations. Anyone telling you it will lift your citation rate is describing a hope, not a documented behaviour.
**It is not a substitute for the pages themselves.** The claims in it need to be true and verifiable on your actual site. A summary asserting things your pages don't support is worse than no summary — the verification step is exactly where an inconsistent entity gets quietly dropped from an answer.
### So does it actually do anything?
The honest answer: **not much yet, on its own, and it's still worth doing.**
The case for adoption is cost, not evidence. It's an hour of work and a static file. The convention is being adopted by developer-tooling companies and documentation sites at a reasonable clip, several agent frameworks fetch it when pointed at a domain, and a machine-readable summary is trivially useful to anything programmatically inspecting your site — including tools you haven't heard of yet.
The case against overinvesting is that **the things that demonstrably move citation rate are elsewhere**: content in the raw HTML, pages that answer real buying questions, quotable facts, and consistent entity data across the web. Those are documented mechanisms. `llms.txt` is a cheap bet on a convention that may become significant. Treat it as the last item on the checklist, not the first — the [five things that actually matter](https://gigabit.agency/insights/generative-engine-optimization-field-guide/) come before it.
### How to write a good one
- **Lead with what you are, in one paragraph.** A stranger with no context should finish it knowing what you sell and to whom.
- **Put the facts in, literally.** Prices, dates, locations, numbers. This is the highest-value section, because specifics are what a model can quote and verify. If you publish pricing, state it here in figures.
- **Link the pages that matter, with a description each.** Not your whole sitemap — the ten or fifteen pages that carry the substance.
- **Say what you are not.** Ours notes we're tool-agnostic and not affiliated with any model vendor's deployment arm. Negative statements disambiguate, and disambiguation is the point.
- **Keep it current.** A stale file asserting last year's prices actively hurts you at the verification step.
Keep it short. This is a summary for a machine, not a brochure — ours is forty lines and that's about right.
### The neighbouring files
Two others worth knowing, since they're often confused with it. **`robots.txt`** controls crawler access — check that the named AI agents are allowed, because an inherited blanket disallow silently negates everything else. **`ai.txt`** is another proposed convention stating your usage and citation policy; ours says citation is encouraged, names the attribution we'd like, and points at the `llms.txt`.
None of the three is a growth strategy. Together they take an afternoon and remove a category of dumb, invisible blockers — which is a reasonable definition of table stakes.
### What to do instead, if you only have an afternoon
Write the file, then go spend the rest of the day on something with evidence behind it: check your key pages render their content in raw HTML, confirm the AI crawlers aren't blocked, and make sure your company name and description match across your site, your schema, and LinkedIn.
Then measure. The [free citation audit](https://gigabit.agency/tools/geo-citation-audit/) tells you where you actually stand across the major assistants, which is a far better guide to what to fix next than any checklist — including this one. The [full field guide](https://gigabit.agency/insights/generative-engine-optimization-field-guide/) covers the mechanisms that do the heavy lifting.

### FAQ
**What is llms.txt?**
A single Markdown file at the root of your site — /llms.txt — that describes in plain, machine-parseable language what your organisation is, what it charges, and where its important pages are. It's a proposed convention rather than a ratified standard, and typically takes about an hour to write.
**What's the difference between llms.txt and robots.txt?**
robots.txt controls access — which crawlers may fetch what. llms.txt offers a summary — what you are and where the substance lives. They solve different problems, and one doesn't substitute for the other: if robots.txt blocks the AI crawlers, an llms.txt changes nothing.
**Does llms.txt improve AI search visibility?**
There's no documented evidence that it does on its own. No major assistant has publicly committed to reading it as an input to recommendations. It's worth adding because it costs about an hour and some agent frameworks and tools do fetch it — but the things that demonstrably move citation rate are content in raw HTML, answer-shaped pages, quotable facts, and consistent entity data.
**What should an llms.txt file contain?**
A one-paragraph statement of what you are and who you serve, a block of concrete facts including prices and dates stated literally, links to the ten or fifteen pages that carry real substance with a short description each, and disambiguating statements about what you are not. Keep it short — around forty lines — and keep it current, since stale facts hurt at the verification step.

## Where AI agents pay off first in ecommerce ops
URL: https://gigabit.agency/insights/where-ai-agents-pay-off-first-in-ecommerce-ops/
Published 2026-09-21

The first AI agents that pay for themselves in ecommerce ops are **returns processing, inventory reconciliation, and order exceptions** — not a customer-facing chatbot. In one [multi-channel e-commerce deployment](https://gigabit.agency/proof/ai-workflow-automation-ecommerce-ops/), five automations cut a four-channel ops team's manual workload from **32 hours per week to 4.5** (−86%), processed **85% of returns autonomously**, lifted inventory accuracy from **94.2% to 99.1%**, and delivered roughly **$87,000** in Year 1 impact against **$74,000** in Year 1 investment. Support tickets are a different product; this post is about the work behind the storefront.
### Rank workflows by hours × exception rate
Start with a time audit, not a model demo. Pull 30 days of ops activity and score each workflow on two axes: **hours per week of paid labor** and **how often a judgment call is required**. High hours + low judgment = automate first. High hours + high judgment = agent drafts, human decides. Low hours = leave alone.
In the published case, an 8-person ops team across Shopify Plus, Amazon, Walmart Marketplace, and B2B wholesale was burning time like this:
| Workflow | Hours / week | Pattern |
| --- | --- | --- |
| Returns processing | ~9 | Mostly ruleable; free-text reasons need classification |
| Inventory reconciliation | ~7 | Nightly sync; true exceptions are rare |
| Manual reporting | ~6 | Same four reports, always stale on delivery |
| Vendor follow-up | ~5 | Inconsistent reply formats; chase logic is repetitive |
| Order exception handling | ~5 | Prep is automatable; final action needs a human |
| **Total** | **~32** | ~$78,000/year in direct labor at loaded cost |
That ranking is the product roadmap. Teams that skip it usually automate the flashiest workflow (a chat widget) and leave the 9-hour returns queue untouched. When the decision is rules vs agent, use the exception-rate test from [rules-based automation vs. AI agents](https://gigabit.agency/insights/rules-based-automation-vs-ai-agents/): if most cases follow a known path and only a slice need language understanding or messy input parsing, start with rules and put the model on the exceptions — not the other way around.
### Returns first — why 85% can go autonomous
Returns win the first slot because volume is high, policy is documented, and the happy path is mechanical: **check eligibility → create the return record → generate the label → confirm with the customer → refund on delivery**. The failure mode that kills amateur bots is free-text return reasons and defect flags — exactly where a small classifier earns its keep.
### What the autonomous path should do
- Validate the order against your return window, restocking rules, and channel-specific policy.
- Create the return in your ERP or OMS (NetSuite in the published build).
- Generate the shipping label and send the confirmation email.
- Classify free-text reasons so defect SKUs surface for QA, not for every "changed my mind" return.
- Auto-process the refund when the carrier marks delivery — with a hard stop if confidence or policy checks fail.
In the live deployment, **85% of returns** ran that path with no human touch; per-return handling fell from 12–18 minutes to zero on the autonomous slice. The remaining 15% — defects, edge policies, VIP exceptions — stayed with people on purpose. Autonomy is a property of the *action*, not a slogan for the whole queue.
### Inventory and exceptions — keep humans on judgment
Inventory reconciliation and order exceptions look similar from a spreadsheet: both are "fix the mismatch." They are not the same product.
### Inventory: reconcile nightly, surface only true exceptions
A nightly job that compares channel stock to the system of record and auto-corrects *known* discrepancy causes removes most of the 7-hour grind. Humans should only see true exceptions — shortages, allocation conflicts, channel-specific holds. That is how accuracy moved from **94.2% to 99.1%** in the published build: not by trusting a model to invent stock levels, but by eliminating manual copy-paste and routing the residue.
### Exceptions: prep one-click actions, do not auto-refund blind
Order exceptions (address failures, payment holds, split-ship conflicts) are where blind autonomy loses money. The agent pattern that works: **classify the exception, assemble the resolution packet, present one-click actions** — keep a human approval in the loop because the stakes are cash and customer trust. In the same deployment, exception resolution time fell from **25 minutes to 6 minutes** without handing the final click to the model.
Vendor follow-up sits in the same family: send the PO, parse inconsistent vendor reply emails into ship dates and tracking, chase delays automatically, escalate only when the date slips past the threshold you set. Reporting is often pure rules plus a short narrative summary — ship it early for morale, but do not let it outrank returns on the ROI list.
### The payback math operators actually need
Model the year before you buy a build. The published commerce-ops program landed roughly like this:
- **Year 1 investment:** $74,000 total ($6,000 assessment, $38,000 build, $2,500/month operation).
- **Year 1 impact:** ~$87,000 ($40,040 direct labor from 27.5 hours/week recovered, plus estimated inventory and faster-exception gains).
- **Payback:** about **10.2 months**, with ROI accelerating in Year 2 once the build is sunk cost.
Map that shape to [published prices](https://gigabit.agency/pricing/), not a "request a quote" fog:
- A fixed-price [AI Transformation Sprint](https://gigabit.agency/services/ai-transformation/) is **$25,000** for two weeks — workflow map, ROI model, and a scoped pilot path that credits toward a build.
- [Gigabit Agents](https://gigabit.agency/services/ai-agents/) start **from $8,000** flat per agent for a production workflow; complex multi-system builds scope toward the upper end of the catalog.
- [Managed AI Operations](https://gigabit.agency/services/managed-ai-operations/) runs **$3,000–$20,000/month** so nightly jobs and classifiers do not silently drift after launch.
If your time audit shows 15+ hours a week on one ops workflow at a mid-market loaded cost, the Sprint usually clears the "is this real?" question before you commit six figures. Use the [AI ROI calculator](https://gigabit.agency/tools/ai-roi-calculator/) with your own hours and rates; do not trust a vendor's slide-deck payback.
### What to do this week
1. **Export 30 days of returns and order exceptions** from your OMS/helpdesk — counts, average handle time, and the top ten reason codes.
2. **Time-box a one-hour audit** with the ops lead: hours per workflow across channels, not opinions about which tool is trendy.
3. **Pick one workflow** — almost always returns if volume is high — and write the happy path in ten bullets, including where a human must stay in the loop.
4. **Score rules vs agent** on that path: if free-text or messy vendor email is the blocker, that is the model slice; the rest can be deterministic.
5. **Price the decision** against a fixed diagnostic: book the [$25,000, 2-week AI Transformation Sprint](https://gigabit.agency/services/ai-transformation/) to validate ROI on your stack, or start from the [Commerce industry page](https://gigabit.agency/industries/commerce/) for the vertical patterns we already ship.
Support automation can wait until the warehouse and returns queues stop eating a full headcount. The operators who get paid by AI in commerce are the ones who automate the boring 32 hours first — then put humans on allocation, negotiation, and the exceptions that actually need judgment. The [full proof write-up](https://gigabit.agency/proof/ai-workflow-automation-ecommerce-ops/) has the week-by-week build order if you want the engineering detail.

### FAQ
**Where should AI agents pay off first in ecommerce operations?**
Start with returns processing, then inventory reconciliation and order-exception prep. Those workflows combine high weekly hours with mostly standardized paths. In one multi-channel deployment, automating that set cut manual ops work from 32 hours per week to 4.5 and processed 85% of returns without a human.
**Is an ecommerce ops agent the same as an AI support agent?**
No. Support agents resolve customer tickets (order status, sizing, policy questions). Ops agents run behind the storefront — returns records, labels, refunds, inventory sync, vendor chase, and exception routing. They share patterns like confidence gates and evals, but the systems they write to and the failure modes are different.
**How much does ecommerce ops AI automation cost, and how long does payback take?**
In a published multi-channel program, Year 1 investment was $74,000 and Year 1 impact about $87,000, with roughly 10-month payback. Gigabit publishes a $25,000 two-week Transformation Sprint, production agents from $8,000 flat, and managed operations at $3,000–$20,000 per month — model your own hours before you commit.
**Which ecommerce ops steps should stay human-in-the-loop?**
Keep humans on final exception actions that move money or allocate scarce inventory, on defect and VIP return edge cases, and on vendor negotiation. Let the agent validate eligibility, create records, generate labels, reconcile known inventory discrepancies, classify free-text, and prep one-click resolutions — not silently approve every refund.

## Patient intake agents + EHR: the pattern that works
URL: https://gigabit.agency/insights/ai-patient-intake-ehr-integration/
Published 2026-09-22

The AI patient intake EHR integration that pays for itself first is **one path from demographics and eligibility through consent into the chart** — not a chatbot bolted onto a PDF form. In a [14-location provider network](https://gigabit.agency/proof/healthcare-unified-intake-ehr-analytics/), unifying intake for web, tablet, and kiosk with an HL7/FHIR bridge cut median intake/admin time per patient **41%**, lifted visits per clinician session **16%**, and dropped form error/rework **32%** over 12 weeks. Diagnosis is the wrong first workflow. Double entry at the desk is the right one.
### Rank intake by hours × rework, not by demo appeal
Start with a time audit, not a model pitch. Pull 30 days of front-desk and intake activity and score each step on two axes: **minutes of paid labor per patient** and **how often staff re-key or fix what the patient already typed**. High minutes + high rework = automate and integrate first. High minutes + clinical judgment = agent drafts, clinician signs. Low volume = leave alone.
In the published multi-site build, the bottleneck was **disjointed web forms plus double data entry into the EHR and practice management system**. Staff re-keyed demographics and insurance, incomplete submissions bounced back, and the desk gated throughput before any clinician saw the patient.
| Step | Typical failure | Pattern |
| --- | --- | --- |
| Demographics capture | Incomplete fields, typos | Validation on submit; picklists over free-text |
| Insurance | Eligibility checked too late | Run 270/271 before the visit, not at the window |
| Consent / e-sign | Paper or separate portal | Same session as intake; audit-ready artifact |
| Chart write | Staff re-keys into EHR | Event-driven HL7/FHIR sync with retries |
| Prep / FAQs | Phone pile-up | Narrow assistant for admin topics only |
That ranking is the roadmap. Teams that skip it ship a waitlist widget and leave the 41% admin-time opportunity on the table. Compliance architecture — BAAs, minimum necessary, audit logs — is day-one design; see [HIPAA-compliant AI agents](https://gigabit.agency/insights/hipaa-compliant-ai-agents/). This post is the **integration pattern** that makes those controls usable in production.
### The integration pattern that survives review
The shape that worked across 14 locations is boring on purpose: **capture → verify → consent → write → observe**.
### Capture once, on every channel
One intake experience for web, tablet, and kiosk — demographics, insurance, consent, and prep in a stepped flow. Validation guardrails and document upload cut incomplete submissions before they become desk rework. Patients should not fill three forms that disagree with each other.
### Verify eligibility before the visit
270/271 insurance eligibility checks belong in the intake path, not as a surprise at check-in. When coverage fails early, staff fix it on a scheduled callback — not while a provider waits.
### Consent and e-sign in the same session
E-signature and consent artifacts must travel with the intake record. A separate portal that patients abandon is how you get paper at the window again.
### Write to the EHR with an auditable bridge
An HL7/FHIR bridge (vendor APIs / Redox-style intermediaries are common) eliminates double entry with **event-driven sync** from intake through appointment to encounter: field mapping, least-privilege credentials, retries with a dead-letter queue, and an audit log on every write. Encryption in transit and at rest is table stakes. The model — when you use one — sees **minimum-necessary fields** for the task (demographics and insurance for intake, not the full chart). That is the same minimum-necessary rule from the HIPAA post, applied to the pipe rather than the policy deck.
### What the agent owns vs what stays human
Autonomy is a property of the *action*, not a slogan for the clinic.
### Let the system own mechanical volume
- Collect and validate demographics, insurance carrier, and visit type.
- Run eligibility and flag failures before arrival.
- Capture consent and e-sign with an immutable trail.
- Sync structured fields into the EHR/PM so staff do not re-key.
- Send procedure-specific prep instructions after intake completes.
A [specialty clinic front-desk assistant](https://gigabit.agency/proof/clinic-llm-assistant-front-desk/) shows the adjacent win: corpus limited to hours, directions, insurance, forms, and prep — with refusal rules and escalation — cut repetitive inquiries **62%**, median time-to-answer **68%** (4:10 → 1:20), lifted CSAT **18%**, and self-served **40%** of after-hours interactions. It collects visit type, insurance, and allergy/medication checklists (**no diagnoses**) and hands staff a summary on escalation — not a clinical decision system.
### Keep humans on judgment and chart finalization
Clinicians own diagnosis, treatment, and chart truth. Staff own eligibility edge cases, VIP exceptions, and dead-letter clears. An intake summary can be **drafted** and **confirmed**; a note can be **proposed**, not auto-filed. Blind auto-write into the chart ends pilots.
When no-shows are the bleed, booking sits next to intake on the ROI list. A [multi-location dental group](https://gigabit.agency/proof/dental-online-booking-kept-appointments/) with self-serve booking and two-way reminders lifted kept appointments **27%** and cut no-shows **29%** — same idea: capture synced to the system of record, humans on high-value exceptions.
### Failure modes that kill healthcare pilots
Four patterns show up every time an intake demo never reaches production:
**Demo first, BAA chain second.** If the model provider, cloud, vector store, and partner are not on BAA-covered, zero-retention paths before PHI flows, you have a breach report waiting — not a pilot. Design the hop list on day one; see the [HIPAA field guide](https://gigabit.agency/insights/hipaa-compliant-ai-agents/).
**Auto-file into the chart.** Treating system output as chart truth without a human gate is how confidently wrong text becomes a medical record. Propose and confirm; do not silent-write clinical content.
**Full-chart context to a general-purpose model.** Minimum necessary means the intake path gets the fields the workflow needs. Dumping the chart “for better answers” expands blast radius and fails review.
**EHR writes without retries or a dead-letter queue.** Partial syncs mean staff trust the bridge, then find missing fields at the visit. Every write needs an audit log; every failure needs a clear path for humans.
### The payback math operators actually need
Model hours before you buy a build. The published intake program leads with operational results, not a dollar ROI line: **41% less intake/admin time per patient**, **16% more visits per clinician session**, **32% less form rework**, across 14 locations in 12 weeks. Translate those percentages against your loaded front-desk and MA cost — that is the business case.
Map the decision to [published prices](https://gigabit.agency/pricing/), not a quote wall:
- A fixed-price [AI Transformation Sprint](https://gigabit.agency/services/ai-transformation/) is **$25,000** for two weeks — workflow map, data boundaries, BAA hop list, and a scoped pilot path that credits toward a build.
- [Gigabit Agents](https://gigabit.agency/services/ai-agents/) start **from $8,000** flat per agent for a production workflow; multi-system EHR bridges scope toward the upper end of the catalog.
- [Managed AI Operations](https://gigabit.agency/services/managed-ai-operations/) runs **$3,000–$20,000/month** so eligibility jobs, sync retries, and guardrails do not drift after launch.
If double entry or incomplete forms eat a material share of desk hours, the Sprint clears the “is the EHR path real?” question before six figures. Use the [AI ROI calculator](https://gigabit.agency/tools/ai-roi-calculator/) with your own minutes-per-patient and rates.
### What to do this week
1. **Export 30 days of intake metrics** — median admin minutes per patient, incomplete-form rate, eligibility failures at the window, and no-show rate.
2. **Walk one patient journey** with ops and IT: which systems get re-keyed today, and where does the chart get written?
3. **Draw the hop list** — every system that would touch PHI on a unified intake path, and whether a BAA exists today.
4. **Pick one workflow** — usually demographics + insurance + consent sync — and write the happy path in ten bullets, including where a human must confirm.
5. **Price the decision**: book the [$25,000, 2-week AI Transformation Sprint](https://gigabit.agency/services/ai-transformation/) to validate ROI and the EHR bridge, or start from the [AI for Healthcare](https://gigabit.agency/industries/healthcare/) page for patterns we already ship.
Front-desk deflection and booking expand after the chart write path is trustworthy. Kill double entry first; keep humans on clinical judgment and real exceptions. The [full intake proof write-up](https://gigabit.agency/proof/healthcare-unified-intake-ehr-analytics/) has the 12-week build shape.

### FAQ
**Where should AI patient intake EHR integration start?**
Start with unified demographics, insurance eligibility, consent, and structured chart write — the path that kills double entry at the desk. In a published 14-location deployment, that pattern cut median intake/admin time 41%, lifted clinician-session throughput 16%, and dropped form rework 32%. Diagnosis and treatment decisions are the wrong first workflow.
**How does intake sync to the EHR without staff re-keying?**
Use an event-driven HL7/FHIR bridge with field mapping, least-privilege credentials, retries plus a dead-letter queue, and an audit log on every write. Capture once on web, tablet, or kiosk — including 270/271 eligibility and e-sign — then sync from intake through appointment to encounter so the desk is not a transcription layer.
**How much does healthcare intake automation cost?**
Gigabit publishes a $25,000 two-week AI Transformation Sprint to scope the workflow, BAA hop list, and EHR path; production agents from $8,000 flat; and managed operations at $3,000–$20,000 per month. Translate your own minutes-per-patient and rework rate into dollars — published proof leads with operational percentages, not a universal payback claim.
**Which intake steps should stay human-in-the-loop?**
Keep humans on clinical judgment, chart finalization, eligibility edge disputes, and anything the sync dead-letter queue cannot clear safely. Let the system validate fields, run eligibility, capture consent, sync structured data, send prep instructions, and deflect repetitive admin FAQs — with escalation summaries — not auto-file diagnoses or silently approve chart notes.

## What to measure in an AI agent eval suite
URL: https://gigabit.agency/insights/ai-agent-eval-metrics/
Published 2026-09-23, updated 2026-10-22

The AI agent eval metrics that matter are the ones that **gate a deploy** — intent accuracy, response accuracy against known-correct answers, policy compliance, escalation precision, and tone — scored on a golden set of real cases before traffic moves. Public leaderboards and demo vibes do not. In a [published D2C support deployment](https://gigabit.agency/proof/ai-automation-ecommerce-support/), a suite built from **2,000** historical tickets with known resolutions hit **94.2%** intent classification accuracy, **91.7%** response accuracy, **99.1%** policy compliance, and a **4.3/5.0** tone score before staged rollout; confidence gates of **0.85** (intent) and **0.80** (retrieval) kept low-certainty tickets with humans. That is the ruler. Without it you are debating feelings.
### Why the suite comes before the agent
Ask a team how they will know their agent is working, and the quiet ones are the ones in trouble. *It seems good in the demo* is not a measurement, and it does not survive a model upgrade, a prompt edit, or the long tail of inputs a real workflow throws at you. So build the eval suite first: collect the inputs that matter, define what a correct output looks like, and write the checks that score it. The suite becomes the spec. It tells you when you have shipped, it catches regressions when a vendor changes the model under you, and it turns *trust me* into *here is the score*, which is what lets a non-deterministic system live in production at all. You do not have an AI strategy until you have one.
### Which metrics gate a deploy — and which are vanity
Score the behavior that can hurt the business. Everything else is commentary.
| Metric | What it answers | Gate it? |
| --- | --- | --- |
| Intent / task classification accuracy | Did the agent route the case correctly? | Yes — wrong route = wrong tools |
| Response / action accuracy vs known-correct | Did it do the right thing on labeled cases? | Yes — the core quality score |
| Policy / compliance hit rate | Did it break a documented rule? | Yes — refunds, PHI, contract language |
| Escalation precision & recall | Did it hand off when it should — and only then? | Yes — silent wrong answers and over-escalation both cost money |
| Tone / brand match (scored) | Would your best human send this? | Yes for customer-facing; lighter for internal |
| Public benchmark scores | How does the base model rank? | No — not your workflow |
| Demo "feels good" | Did stakeholders smile in the room? | No — dies on day-two traffic |
Resolution rate and CSAT belong in the **operating** dashboard after launch — they are outcomes, not pre-deploy gates. In the published support build, **64%** autonomous resolution and CSAT moving **3.9 → 4.5** showed up after the suite passed and traffic ramped. Do not put outcome KPIs in the SOW as if they were eval metrics; put the accuracy and policy numbers that predict them.
The failure mode is mixing layers. Teams that skip the suite and watch live CSAT discover regressions from customers, not from a red CI run. **95% of enterprise generative-AI pilots deliver no measurable P&L impact** ([MIT, 2025](https://www.forbes.com/sites/jasonsnyder/2025/08/26/mit-finds-95-of-genai-pilots-fail-because-companies-avoid-friction/)) — the operational gap is almost always measurement and ownership, not the model card.
### Build the golden set from real cases, not invented prompts
A golden set is a scored collection of **representative inputs, defined correct outputs, and automated checks**. Pull from production history — tickets, calls, form submissions — not from marketing scenarios. Label what "correct" means for each case: the right classification, the right answer or action, the right escalation.
### Size and composition
Aim for hundreds to low thousands for a first production workflow, stratified by volume and by failure mode. The published support agent tested against **2,000** historical tickets with known resolutions — enough to cover the top categories and the nasty tail. Under-sampling awkward cases is how you ship 90% accuracy on happy path and blow up on the 10% that create refunds and chargebacks.
### What every case must carry
- **Input** as the system will see it (raw text, structured fields, attached context).
- **Known-correct label** — intent, expected answer or tool call, or "must escalate."
- **Policy tags** — which rules apply (return window, discount eligibility, medical-advice refusal).
- **Difficulty flag** — easy / edge / adversarial, so you can report accuracy by band, not one blended number that hides the tail.
### Who labels
Domain owners label; engineers wire the runner. If only the vendor labels "correct," you bought a vibe check. The suite becomes the [spec for the production agent stack](https://gigabit.agency/insights/production-ai-agent-stack/) — evals sit next to retrieval, observability, and guardrails as a required layer, not a nice-to-have appendix.
### Pass thresholds and confidence gates belong in the SOW
Write numbers into the statement of work before build week one. Vague "we'll tune quality" is how pilots never exit.
### Thresholds that matched a live deploy
Use the published support suite as a concrete pattern — not a universal law for every vertical:
- Intent classification ≥ **94%** on the golden set (they hit **94.2%**).
- Response accuracy ≥ **90%** against known-correct answers (**91.7%**).
- Policy compliance ≥ **99%** (**99.1%**).
- Tone ≥ **4.0/5.0** against your best human baseline (**4.3/5.0**).
Your floors may differ for clinical-adjacent or payments workflows — raise them; do not invent softer ones after a bad demo week.
### Confidence gates at inference time
Evals score the model offline. Confidence gates decide **online** whether this case may go autonomous. In the same deployment, the agent replied without a human only when intent confidence exceeded **0.85** and retrieval relevance exceeded **0.80**; below either threshold the ticket routed to a human with the draft and context pre-loaded. That split is the difference between "64% autonomous" and "64% confidently wrong."
Put both in the SOW: **offline pass criteria** for the suite, **online gates** for runtime. No green suite run, no promote. No gate breach without escalation.
### What changes after go-live
Launch is not the finish line. Vendor model updates, prompt edits, and retrieval corpus drift all move scores. The suite must run on **every** change — prompt, model, retrieval config — before production sees it.
### Regression rhythm
- Re-run the full golden set on every candidate change; block deploys on red.
- Sample live traces weekly — especially escalations and low-confidence paths — and promote hard cases into the suite.
- When a model vendor deprecates a version, score the replacement on the suite *before* the forced cutover; keep rollback warm.
That ongoing discipline is what [Managed AI Operations](https://gigabit.agency/services/managed-ai-operations/) is for: **$3,000–$20,000/month** so quality movement is a number someone sees, not a ticket from an angry customer. The same eval layer that gates first launch is the layer that keeps the agent honest at month six.
### What to do this week
1. **Export 90 days of real cases** for one workflow — tickets or forms — and pick 200–500 to label first; expand toward ~2,000 if volume supports it.
2. **Define correct** with the domain owner: classification labels, acceptable answers/actions, mandatory escalations, and policy tags.
3. **Write four gate numbers** into a one-pager: intent accuracy, response accuracy, policy compliance, and the confidence thresholds for autonomy.
4. **Refuse a demo-only milestone** — if the vendor cannot run your golden set, they cannot ship your agent.
5. **Price the diagnostic**: book the [$25,000, 2-week AI Transformation Sprint](https://gigabit.agency/services/ai-transformation/) to build the suite and a scoped pilot path that credits toward a build, or start from [Gigabit Agents](https://gigabit.agency/services/ai-agents/) (**from $8,000** flat) when the workflow and metrics are already clear.
You cannot ship what you cannot score. Put the metrics in the contract, build the golden set from your own ugly cases, and treat every model change as a deploy that needs a green run. The [full support proof write-up](https://gigabit.agency/proof/ai-automation-ecommerce-support/) shows how those numbers looked before traffic moved — and what autonomous resolution looked like after.

### FAQ
**What AI agent eval metrics should gate a deploy?**
Gate on intent or task classification accuracy, response or action accuracy against known-correct labels, policy compliance, escalation precision, and (for customer-facing agents) a scored tone match. Public model benchmarks and demo impressions are not deploy gates. Outcome metrics like resolution rate and CSAT belong in the operating dashboard after launch.
**How large should an AI agent golden set be?**
Start with hundreds of real, labeled cases stratified by volume and failure mode; expand toward low thousands for a high-volume production workflow. A published support deployment scored against 2,000 historical tickets with known resolutions before staged rollout. Invented demo prompts are not a golden set.
**What pass thresholds and confidence gates did a live support agent use?**
In a published D2C support build, offline evals hit 94.2% intent accuracy, 91.7% response accuracy, 99.1% policy compliance, and 4.3/5.0 tone. Online, the agent went autonomous only when intent confidence exceeded 0.85 and retrieval relevance exceeded 0.80 — otherwise a human got the draft and context.
**When should you re-run AI agent evals after launch?**
On every prompt, model, or retrieval change before it reaches production, and on a schedule against live-trace additions to the suite. Vendor model updates are a forced regression event — score the replacement on your golden set before cutover. Managed operations retainers exist to keep that rhythm from slipping.
**What is an AI eval suite?**
An eval suite is a set of representative inputs, defined correct outputs, and automated checks that score an AI system's behavior. It turns "it seems good in the demo" into a number you can put in front of a board, and it catches regressions when a vendor changes the model underneath you.
**Why build the eval suite before the agent?**
Because the suite becomes the spec. It defines what "correct" means for your workflow, tells you when you have actually shipped, and is what lets a non-deterministic system live in production. Build the agent first and you have no way to know when it is working.

## What gets a company excluded from AI answers
URL: https://gigabit.agency/insights/what-gets-a-company-excluded-from-ai-answers/
Published 2026-09-29, updated 2026-10-22

A company gets **excluded from AI answers** when an assistant can find a page about you and still refuses to name you. The crawler never saw the text, two of your URLs disagree on a fact, or the engine cannot tell you are one company. Rank does not prevent that. When Google shows an AI summary, only **8% of users click any traditional link — versus 15% without one — and just 1% click a source inside the summary** ([Pew Research, 2025](https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/)). Only **38% of AI Overview citations now come from the top-10 organic results, down from about 76% a year earlier** ([Ahrefs, 2026](https://ahrefs.com/blog/ai-overview-citations-top-10/)). Page one is not the answer. This post is the drop list: conditions that remove you after the page already exists. The checklist that makes you citable is in [the GEO field guide](https://gigabit.agency/insights/generative-engine-optimization-field-guide/). Why a named competitor beats you is the last section.
### Why a top rank still gets you dropped
Answer engines assemble a short answer. They do not hand the buyer ten links. Retrieval can surface your URL, and verification can still drop the name. The model has to lift a claim it can repeat and confirm that the claim matches the rest of what it can read. Fail either test and you are absent, even if you were in the candidate set.
Princeton's GEO study found that adding **citations, quotations, and statistics lifted visibility in generative-engine answers by up to 40%** ([Aggarwal et al., KDD 2024](https://arxiv.org/abs/2311.09735)). The inverse is the exclusion. A page of adjectives gives the model nothing it can defend, so it names the company whose numbers it can stand behind. Specificity is the condition for staying in the answer, not a writing style.
That is why a [published support proof](https://gigabit.agency/proof/ai-automation-ecommerce-support/) is quotable and a slogan page is not. The proof states one set of figures: **64%** of tickets resolved with no human, **94.2%** intent accuracy, **91.7%** response accuracy, **99.1%** policy compliance, first response from **4.2 hours to 6 minutes**, and **$217,200** saved in year one against **$92,000** of Year 1 investment. Those sentences can be lifted. They are not a citation-rate result. They show the shape of a page an engine can quote without guessing. If a second URL on the same site stated a different resolution rate, the safer move for the model is to name neither page.
### The four hard exclusions
Score these before you commission another page. Each one removes you from the candidate set. More copy does not override them.
| Exclusion | What the engine sees | How to test it |
| --- | --- | --- |
| Crawler never reads the page | Empty HTML, or a robots.txt block | View source on the page you want quoted; read robots.txt for named agents |
| Your URLs disagree | Two prices, two names, or two outcomes | Diff the pricing page, the proof page, and llms.txt |
| The entity does not resolve | Two companies that might be you | Compare legal name and address on the site, schema, and LinkedIn |
| No second source | Only your domain states the claim | Search the figure off-site. If nothing else says it, verification fails |
### The crawler never reads the page
GPTBot, ClaudeBot, and PerplexityBot fetch HTML. They largely do not run JavaScript. If the price, the proof, and the answer render only in the browser, the crawler gets a shell and you are excluded from AI answers before extraction starts. A disallow you inherited from a CDN preset or a security review does the same thing. Check `robots.txt` for `GPTBot`, `OAI-SearchBot`, `ChatGPT-User`, `ClaudeBot`, `Claude-User`, `PerplexityBot`, `Perplexity-User`, and `Google-Extended`. Blocking them on purpose is a real choice. Inheriting the block is an accident.
`llms.txt` does not repair a block. It offers a map. It does not grant access. The test order — raw HTML, then crawler access, then answer-shaped copy — is the [technical checklist in the GEO field guide](https://gigabit.agency/insights/generative-engine-optimization-field-guide/). If view-source is empty, stop. A citation rate does not move until the numbers are in the response.
### Your own pages contradict each other
This exclusion is easy to miss because each URL looks fine alone. Say the pricing page states a production agent starts at **$8,000** flat, an older article still says the fee is a custom quote, and `llms.txt` lists last year's number. A model that retrieves all three has no figure it can defend. Naming you would mean picking one page and contradicting the others. Skipping you is safer. Two numbers for one metric is an exclusion.
Treat three URLs as one fact set: the pricing page, the proof page you want quoted, and [llms.txt](https://gigabit.agency/llms.txt) if you publish one. If any repeated figure differs, fix the stale URL before you publish a new one. A stale summary is worse than no summary — that limit is in [what llms.txt is and is not](https://gigabit.agency/insights/llms-txt-explained/). Put a date next to the figure. "As of September 2026, Gigabit Agents start at $8,000 flat" is a claim a model can check. "Affordable production agents" is not.
### The engine cannot resolve one company
Entity collision is invisible from the inside. The website uses one legal name, LinkedIn uses a product name, schema uses a third, and a directory listing still has the old address. The model will not guess which entity deserves the recommendation. It leaves the name out. Matching the name, matching the address, and stating both in Organization schema is the fix. When a competitor is named on a question you already have a page for, use the competitor diagnosis below instead of adding another adjective to the page.
### Nobody else says the same thing
A company describing itself is the weakest evidence an answer engine has. If the only URL that states your price, outcome, or category is your own domain, verification fails on the questions a buyer asks before they know your name. Roundups, comparison pages, directories, and reviews are corroboration — testimony, not a click source. If no second source repeats a specific claim, do not expect to be named for it.
### Soft gaps that look like exclusions
Two measurement mistakes produce the same symptom — you are not in the answer — and neither is an exclusion.
**You tested your own name.** "What is Gigabit" measures recognition. The buying query does not contain the brand: how much a production agent costs, who builds a support workflow, what an intake agent has to clear before PHI moves. [AI citation rate](https://gigabit.agency/insights/what-is-a-good-ai-citation-rate/) is the share of a fixed panel of those questions where an assistant names you. Six appearances on a 20-query panel is **30%**. Branded hits do not count.
**You tested a sentence nobody types.** "Best forward-deployed engineering firm for mid-market healthcare intake" can return nothing because the phrasing is yours, not the buyer's. Absence on a vanity prompt is not a disqualification. Write the panel from sales calls and from the questions already on the site. Run it more than once. One answer is an anecdote. The monthly trend is the signal.
If you are absent on real category questions and the four hard exclusions are clean, you are in the competitor diagnosis: they answer the question, they publish a fact, or someone else vouched for them. That is a coverage gap. Publish the missing page. Do not treat it as a block you cannot clear.
### When a competitor is named instead of you
Being absent is one problem. Watching a competitor get named on a question you also serve is a narrower one, and it almost always comes down to one of three asymmetries. **They answer the question; you describe your services.** Their page exists to answer *how much does X cost* or *best X for Y*; yours describes capabilities and answers nothing anyone typed. **They publish facts; you publish adjectives.** *A production agent is a flat fee from $8,000* survives extraction; *bespoke AI solutions that drive transformative outcomes* does not. **Someone else vouched for them.** If roundups, comparison posts, and directories mention them and only your own domain mentions you, the model has one source for you and eight for them.
- Competitor named, you absent, on a question you **do** have a page for: the page is not quotable or the engine cannot verify you. Run the four hard exclusions above.
- Competitor named, and you have **no page** for that question: a coverage gap, and the cheapest fix here.
- You appear for your brand name but never for category questions: missing pages plus a thin third-party footprint.
- Cited by one assistant, invisible in another: usually corroboration or entity data, because each engine retrieves from a different index.
Most categories still have no serious GEO players, so the company being cited today is usually the one that happened to publish specifics. That is a lead measured in months, not years, and it is reversible.
### What to do this week
1. **View source** on the pricing page and one proof page. If the numbers are not in the raw HTML, server-render them. Stop until they are.
2. **Read robots.txt** for the named agents above. Delete a disallow you did not choose.
3. **Diff three URLs** — pricing, the proof you want quoted, and llms.txt. Make every repeated figure identical, including the **$8,000** agent floor and any proof metric you cite.
4. **Check the entity** in one sitting: site title, Organization schema, LinkedIn, and one directory. Any two that disagree are an exclusion.
5. **Run the free** [AI Citation Audit](https://gigabit.agency/tools/geo-citation-audit/). It scores category buying questions across the major assistants and shows where you are absent and who is named instead. Use the report to pick the first exclusion. Do not treat the score as a trophy.
The audit is the baseline. [Ascent](https://gigabit.agency/services/geo/) is the 12-month GEO program that keeps the pages, the entity, and the measurement honest after that, from $4,000 a month with no upfront fee. The exact monthly number is set on a scoping call after the audit. If three URLs disagree, fix the conflict this week. Do not buy a content program to paper over it. The [support proof](https://gigabit.agency/proof/ai-automation-ecommerce-support/) is the pattern for a page worth quoting: one outcome, one number, the same number wherever you repeat it.

### FAQ
**What gets a company excluded from AI answers?**
A company is excluded from AI answers when an assistant can retrieve a page and still will not name it. Four hard conditions cause that. The crawler never reads the text, because it is rendered in the browser or blocked in robots.txt. Two of your own URLs disagree on a price, a name, or an outcome. The engine cannot resolve your site, schema, and directory listings into one entity. Or the claim appears only on your domain, with no second source to check. Rank does not override any of those. A top-10 URL can be retrieved and still dropped when the model cannot defend the claim. Fix the conflict before you add another page.
**Can you rank on page one and still be excluded from AI answers?**
Yes. Only 38% of Google AI Overview citations now come from the top-10 organic results, down from about 76% a year earlier, according to Ahrefs. When an AI summary is on the results page, Pew Research found that 8% of users click any traditional link, versus 15% when no summary appears, and 1% click a source inside the summary. The engine retrieves candidates, then keeps the claims it can quote and verify. A page-one URL with empty HTML, two conflicting prices, or no third-party corroboration fails that second step. Exclusion is a verification drop, not a rank miss. Publish one defensible figure and make every other URL on the site agree with it.
**Do contradictory prices on your own site get you excluded from AI answers?**
They can. If the pricing page, a proof page, and llms.txt state different figures for the same offer, the model has no number it can defend. Naming you would commit it to one of the conflicting claims. Skipping you is safer. Gigabit publishes one floor: production agents start at $8,000 flat. A published support proof states one outcome set — 64% of tickets resolved with no human, 94.2% intent accuracy, and $217,200 saved in year one against $92,000 of Year 1 investment. Repeat a figure everywhere or do not repeat it. Two numbers for one metric is an exclusion, not a nuance a buyer should average.
**How do you tell a real exclusion from a bad test query?**
Ask the question a buyer types, not your brand name. A prompt like "what is [your company]" measures recognition. Category questions — how much an agent costs, who should build the workflow — are the citation-rate panel. Six names across a 20-query panel is a 30% citation rate, and [what counts as a good one](/insights/what-is-a-good-ai-citation-rate/) depends on the panel. Also drop prompts nobody says out loud. Absence on a sentence you invented is not an exclusion. If you are missing on real category questions and the four hard checks are clean — HTML in the raw response, crawlers allowed, one set of facts, one entity — you have a coverage gap, not a disqualification. The free AI Citation Audit scores that panel across the major assistants.
**Why does ChatGPT recommend my competitor instead of me?**
Usually one of five asymmetries: they have a page that directly answers the buying question and you do not, they publish quotable facts while you publish adjectives, they have third-party corroboration across the web and you only appear on your own site, the engine cannot resolve your company into one confident entity because your naming is inconsistent, or the query you tested is not one real buyers use.
**Why am I cited by one AI assistant but not another?**
Because each assistant retrieves differently: ChatGPT search uses OpenAI's own index plus search partners including Microsoft, AI Overviews use Google's index, and Perplexity runs its own crawler. Uneven visibility across engines usually points to a thin third-party footprint or inconsistent entity data rather than a content problem on your own site.

## When a multi-step AI agent fails in production
URL: https://gigabit.agency/insights/when-a-multi-step-ai-agent-fails-in-production/
Published 2026-09-30

A multi-step AI agent fails in production when each step looks fine and the chain does not. If every step is right **95%** of the time and the misses are independent, eight steps finish right about **66%** of the time — 0.95 to the eighth is 0.663. Put that number in the review, not the model card. Shorten the chain, split it, or put a confidence gate on any step that writes money, a record, or a customer-visible message. What an agent costs is [a separate budget](https://gigabit.agency/insights/ai-agent-cost/). The layers we deploy are [the production stack](https://gigabit.agency/insights/production-ai-agent-stack/). Why a pilot stalls is [the five gates](https://gigabit.agency/insights/why-your-ai-pilot-never-reached-production/). This post is only the step math.
### Per-step accuracy is not the success rate
Quote the step, not the demo. A classifier at 95% on a labeled set is a one-step score. A workflow that classifies, retrieves, extracts a field, calls a tool, checks a policy, and writes back is six chances to be wrong. When those misses are independent, end-to-end accuracy is the product of the step rates. Four steps at 95% are already about **81%**. Twelve are about **54%**.
Independence is the assumption under the table, and it is not free. It holds when each step can miss for its own reason: a stale field, a tool timeout, a threshold that does not fire. It does not hold when one upstream miss — usually a bad retrieval — poisons every later step. In that case the product understates how often the whole run dies together, and a gate on the weak step beats another prompt on the last one. Say which case you are in before you multiply. These figures are rounded arithmetic. They are not a client measurement.
| Steps | 99% each | 95% each | 90% each |
| --- | --- | --- | --- |
| 1 | 99% | 95% | 90% |
| 4 | 96% | 81% | 66% |
| 8 | 92% | 66% | 43% |
| 12 | 89% | 54% | 28% |
**95% of enterprise generative-AI pilots deliver no measurable P&L impact** ([MIT Project NANDA, 2025](https://www.forbes.com/sites/jasonsnyder/2025/08/26/mit-finds-95-of-genai-pilots-fail-because-companies-avoid-friction/)). Chain length is one mechanical reason a three-turn demo falls over on a twelve-call workflow. The pilot never had a product to fail. It had a slide that said 95%.
### Count every step that can be wrong
A step is any place the run can pick the wrong thing. If you cannot list them, you do not have a success rate.
- Classify the request.
- Retrieve the policy, the order, or the record.
- Extract a field from messy text.
- Call a tool.
- Decide whether to proceed.
- Write back — a reply, a refund, a label, an address, a chart note.
The write-back is the step teams forget to score. They grade the sentence and ignore the tool that changes the system of record. That tool is the one with the blast radius.
The published counterexample is short on purpose. In a [D2C support deployment](https://gigabit.agency/proof/ai-automation-ecommerce-support/), offline scoring on **2,000** historical tickets hit **94.2%** intent accuracy, **91.7%** response accuracy, and **99.1%** policy compliance. If intent and response both had to be right, and if those two rates were independent, the joint figure would be about **86%** (0.942 × 0.917). The case study does not publish that joint rate. It publishes the gates that kept the agent off the weak tail: an autonomous reply only when intent confidence exceeded **0.85** and retrieval relevance exceeded **0.80**. Under that rule the agent resolved **64%** of tickets with no human — **3,149 a month** — not 94%. First response on those AI replies fell from **4.2 hours to 6 minutes**. Year 1 savings were **$217,200** against **$92,000** of Year 1 investment. The 64% is the share that cleared the gates. It is not an eight-step product.
Copy the shape. A long chain without a gate ships the product of its misses. A short chain with a gate ships the slice it can defend. Which scores belong in the suite is [what to measure before a deploy](https://gigabit.agency/insights/ai-agent-eval-metrics/). Use that list on the steps you just counted. Do not invent a second metric taxonomy here.
### Three moves that raise the end-to-end number
You have three levers. More model calls is not one of them.
### Shorten the chain
Most of the path does not need a model. Eligibility checks, field validation, return-window tests, and label generation are deterministic. Put the model on the step that needs language — free-text classification, a messy vendor email — and keep the rest a state machine. Every model step you delete multiplies what remains by something closer to 1. The layer list for that state machine lives in [the production stack](https://gigabit.agency/insights/production-ai-agent-stack/).
### Gate the irreversible step
Irreversible means a customer sees it, money moves, or a system of record changes. Those steps get a floor and a human. In the support build the floors were **0.85** on intent and **0.80** on retrieval. Below either one, a person received the draft and the context. Reversible reads — look up an order, draft a reply, classify a reason — can run hotter. The question per action is the one in [what an agent can do without a person in the loop](https://gigabit.agency/insights/ai-agent-human-in-the-loop/): how reversible is the action, and how fast would you notice a miss. A quiet miss that runs for a quarter costs more than a loud failure on day one, because nobody stops it.
### Split one loop into two scored workflows
A twelve-step loop has one blended score and no owner. The average hides the last step. Two workflows — classify and retrieve, then act — each with a golden set, fail in a place you can see. The handoff is a typed object: intent, record id, confidence, the fields the next step is allowed to touch. If the second workflow cannot name its inputs, you still have one loop with extra branding.
[Gigabit Agents](https://gigabit.agency/services/ai-agents/) are **from $8,000** flat per agent for one production workflow. A twelve-step "one agent" is often two workflows and should be priced that way. Paying once for a chain that finishes right about **54%** of the time, at 95% per step, is the expensive version.
### Failure modes that survive the demo
Five patterns show up after the room claps.
**The model card becomes the workflow rate.** A slide says 95%. The workflow has eight tool calls. You bought about 66% and reported 95%.
**One loop, no checkpoint.** The agent may call tools until it decides it is finished. There is no step list, so there is no product to compute and no place to insert a gate.
**A blended score.** Say intent is 98% and the write-back is 80%. The average looks fine. The customer sees the write-back.
**Autonomy set to the classification rate.** 94.2% intent accuracy is not permission to auto-send 94% of traffic. The support proof auto-sent the slice above the gates and left the rest with people. Copy the gate, not the headline accuracy.
**No labels, so no product.** If you cannot put a measured rate on each step, stop multiplying. A vendor benchmark is not your step rate. Label a sample from real tickets first. That sample is also the start of the eval suite that separates a pilot from a system you can keep.
### What to do this week
1. **Write the chain** for one workflow as a numbered list. Include retrieval, each tool call, the proceed-or-stop decision, and the write-back.
2. **Put a rate on each step** from a labeled sample of real cases. If you have no labels, that is the work — not another demo.
3. **Multiply only if the misses are independent.** If one upstream miss causes the rest, fix that step before you add length. The table above is the independent case, rounded.
4. **Mark irreversible steps** and set a floor. If you have no measured floor, do not auto-execute them. **0.85** intent and **0.80** retrieval are a published support example, not a default for payments or clinical notes.
5. **Price the cut.** Book the [$25,000, two-week AI Transformation Sprint](https://gigabit.agency/services/ai-transformation/) to map the chain, the gates, and a scoped build that credits toward an agent. If the workflow is already one clear step with a golden set, buy [Gigabit Agents](https://gigabit.agency/services/ai-agents/) **from $8,000** flat and skip the diagnostic.
A 95% step is not a 95% workflow. Count the steps, delete the ones a rule can do, and gate the ones that write. The [published support proof](https://gigabit.agency/proof/ai-automation-ecommerce-support/) is the reference: a short chain, two numeric gates, **64%** of tickets handled with no human, and the rest handed over with a draft. The [$25,000 Sprint](https://gigabit.agency/services/ai-transformation/) is the two weeks to run that test on your own queue before you fund a loop that cannot clear 66%.

### FAQ
**When does a multi-step AI agent fail in production?**
A multi-step AI agent fails in production when each step looks accurate and the full chain does not. If every step is right 95% of the time and the misses are independent, eight steps finish right about 66% of the time, because 0.95 to the eighth power is 0.663. Four steps land near 81%. Twelve land near 54%. That arithmetic assumes each step can miss on its own. It does not hold when one bad retrieval poisons every later step — then the whole run fails together, and the product understates the damage. Either way, the fix is the same: shorten the chain, split it into two scored workflows, or put a confidence gate on steps that write money, records, or customer-visible messages. Do not report the model card as the workflow rate.
**How do you calculate end-to-end accuracy for a multi-step AI agent?**
List every step that can be wrong, including retrieval, tool calls, the proceed-or-stop decision, and the write-back. Put a measured rate on each step from a labeled sample of real cases, not a vendor benchmark. If the misses are independent, multiply the rates. At 95% per step the products are about 81% after four steps, 66% after eight, and 54% after twelve. Round only after you multiply, and say the independence assumption out loud. If one upstream miss causes the rest, do not multiply — fix that step. A published support agent scored 94.2% intent and 91.7% response accuracy as separate numbers. A joint 86% (0.942 × 0.917) is arithmetic, not a figure the case study reports.
**How many steps can a multi-step AI agent run before it needs a gate?**
There is no universal step cap. The limit is the product of the step rates and whether a miss is reversible. At 99% per step, eight steps still finish near 92%. At 90% per step, four steps are already about 66% and eight are about 43%. Gate the irreversible steps no matter how short the chain: a customer-visible send, a payment, a refund, or a write into the system of record. In a published support deployment the agent replied on its own only when intent confidence exceeded 0.85 and retrieval relevance exceeded 0.80. That kept autonomy at 64% of tickets (3,149 a month), not at the 94.2% intent score. Treat 0.85 and 0.80 as that workflow's floors, not as a default for every vertical.
**What should you do when a multi-step AI agent is right on the demo and wrong in production?**
Stop adding steps. Write the live chain as a numbered list and mark the write-back. Score each step on real cases. Delete model calls that a rule can handle, and split a long loop into two workflows with a typed handoff if the blended score is hiding the last step. Do not raise autonomy to match the classification rate. The support proof auto-sent only the slice above its gates and left the rest with a person, draft attached. Gigabit prices the diagnostic as a $25,000, two-week AI Transformation Sprint that credits toward a build, and a single production agent from $8,000 flat when the workflow and the golden set are already clear. Book the Sprint when you cannot yet list the steps.

## AI for consulting firms: start with the archive
URL: https://gigabit.agency/insights/ai-for-consulting-firms/
Published 2026-10-01, updated 2026-10-22

AI for consulting firms should start with the engagement archive, not a proposal bot. In one [management consulting deployment](https://gigabit.agency/proof/ai-transformation-professional-services/), a 140-person firm recovered a published **1,200 hours per quarter** and **$480,000 a year**, at **129% Year 1 ROI** and a **5.3-month** payback, against **$210,000** of Year 1 investment. The sequence was knowledge search, then proposals, time capture, deliverable formatting, and the remaining internal work. The failure mode is a draft generator with nothing sourced behind it: partners rewrite every page. The workflow menu lives on the [professional services page](https://gigabit.agency/industries/professional-services/).
### Rank the leak before you pick a tool
Pull one quarter of where senior time actually goes before anyone picks a product. The published assessment covered a firm of **95 consultants** and **45 operations staff**, **$32 million** in annual revenue, growing **20%** year over year, with overhead scaling in a straight line with headcount. Consultants billing **$250–$400 an hour** were spending **35%** of their time on work that did not require a consultant. The case states that slice as roughly **$4.8 million** a year in wasted billable capacity.
Rank each leak on two axes: **hours of paid labour** and **the rate of the person doing it**. A consultant at **$250–$400 an hour** formatting a slide is a different loss from an operations coordinator doing similar work at the **$45 an hour** rate the case used for ops. High hours at a high rate, on a path you can write down, is the first build. Low hours stay manual.
The same assessment put the admin leaks on a clock:
| Leak | Hours / quarter | What the time was |
| --- | --- | --- |
| Proposal generation | 320 | 4–8 hours each, across 40+ proposals |
| Knowledge retrieval | 280 | SharePoint, Outlook, Teams, and a partner's memory |
| Deliverable formatting | 240 | Structure and presentation, not the analysis |
| Internal operations | 200 | Work that grew with the firm, not with the advice |
| **Listed total** | **1,040** | Before the hours nobody entered |
A second **$4.8 million** figure appears in the same case, and it is a different pool. Memory-based Friday time entry left **15–20%** of billable time unrecorded. On **$32 million** of revenue, the case calls that roughly **$4.8 million** in potential unbilled work. Wasted capacity and unrecorded billings are both real. Adding the two figures double-counts the story. Keep them in separate lines when you build the business case.
Under the hours sat the file pile: **8 years** of deliverables, proposals, and methodologies across **47,000** SharePoint files, no naming convention, no tags, search limited to the filename. That pile is why the archive is the first system. A proposal assistant that cannot retrieve a prior engagement writes a confident draft a partner will throw out. You have paid for a rewrite.
### The archive comes before the proposal
The readiness assessment ran two weeks and cost **$12,000**: 15 interviews across partners, consultants, and operations, plus shadowing a full proposal and a delivery cycle. It set the sequence before any build. Knowledge base first. Then proposals. Then time entry. Then deliverables. Then operational automation. Reverse that order and the later tools have nothing true to stand on.
### What phase 1 actually shipped
Phase 1 ran weeks 1–8. It ingested **12,000** of **47,000** files — the case calls that slice the most relevant — into vector search over the archive. Access followed practice area and client authorization. The same index served a Slack bot and a web page. A consultant asks in Slack and gets a sourced answer in **4 seconds**, with links to the underlying deliverables. Embedding every draft and personal note brings the obsolete file back with the same confidence as the one a partner still stands behind.
### Proposals after the index exists
The proposal assistant came after search worked. It generates a full draft from a structured brief. Creation time fell from **4–8 hours** to **45 minutes** of review. The 45 minutes is the product. The draft is the input to a person who still owns the recommendation, the pricing, and the scope. In the first full quarter with everything live, that assistant recovered **260 hours**, worth **$78,000** at the case's **$300 an hour** average.
If retrieval is wrong, those 45 minutes become a rewrite and the 260 hours disappear. Gate the draft on sources the reviewer can open. A brief that cites three prior engagements is a proposal. A brief that cites none is a writing exercise.
Phase 2, weeks 9–14, added report and presentation drafts: structure, formatting, and data presentation, with the consultant supplying the analysis. Phase 3, weeks 15–24, covered client intake, monthly partnership reporting, and compliance documentation. The program ran six months across three phases. Those week numbers are that deployment's schedule.
### Time capture is the cash line
Search returns hours to the day. Time capture returns money the firm already earned and failed to bill. The assistant runs every morning. It suggests entries from calendar, email, document, and channel activity. A person approves with one tap. It does not invent a code and post it.
Capture moved from **80–85%** to **94%**. The case values that lift at **$95,000** in the measured quarter. The managing partner estimates **$380,000 a year** in additional billings. Keep that yearly figure labeled as the partner's estimate. The measured number is the quarter.
The failure mode is a suggestion people accept without reading. Then Friday is reconstructed wrong, and the invoice is wrong. Keep the approval tap, because the entry hits a bill. If reviewers change most suggestions over a month, fix the activity feed — calendar titles that say "busy," mail that is not client work — before you rewrite the prompt.
Fund the archive first when it is the larger hour pile. Here, retrieval was **280 hours** a quarter and proposals were **320**. Time capture showed up as recovered billings, separate from those hours. A firm already capturing in the mid-90s has little to gain from the morning suggestion.
### The payback math operators should copy
Read the first full quarter with all three phases live, then read the headline. The case publishes these line items for that quarter:
| Workflow | Hours | Value |
| --- | --- | --- |
| Knowledge base | 280 | $84,000 at a $300/hour average |
| Proposal assistant | 260 | $78,000 |
| Deliverable tools | 240 | $72,000 |
| Operational automation | 200 | $36,000 at a $45/hour ops rate |
| Monthly reporting | 40 | $12,000 |
| Time capture | — | $95,000 from the capture-rate lift |
| **Quarter, as itemized** | **1,020+** | **$377,000** |
The published headline metric on the same proof is **1,200 hours per quarter** and **$480,000 a year**, with a **5.3-month** payback and **129% Year 1 ROI**. The itemized quarter is **1,020+ hours** and **$377,000**. Both figures are in the case study. This post does not explain the difference, and it does not annualize the quarter by multiplying $377,000 by four. The case annualizes, with continued optimization, to **$480,000 a year**.
Year 1 investment was **$210,000**: a **$12,000** assessment, a **$156,000** build, and **$4,500 a month** to operate. That operate line is inside the published band for [Managed AI Operations](https://gigabit.agency/services/managed-ai-operations/), **$3,000–$20,000 a month**. It is one engagement's retainer, not a price-list SKU.
- The [AI Transformation Sprint](https://gigabit.agency/services/ai-transformation/) is **$25,000** for two weeks: workflow map, ROI model, build-versus-buy, and a scoped pilot that credits toward a build.
- [Gigabit Agents](https://gigabit.agency/services/ai-agents/) start **from $8,000** flat per agent for one production workflow.
- A multi-workflow year — assessment, several builds, and an operate retainer — is the **$80,000–$500,000** Year 1 range. This proof landed at **$210,000**, inside that range, because it was three phases.
An **$8,000** agent is the right purchase when one workflow is already scoped and the source of truth exists. It is the wrong purchase for **47,000** unsorted files plus proposals plus time capture. Buy the Sprint when you cannot yet rank the leaks. Buy the agent when the rank is done and one path is ready to ship.
### What stays with a person
The partner still edits and sends the proposal. Consultants still supply the analysis; the tools handle structure, formatting, and data presentation. Time-entry suggestions need a one-tap approval, because the entry becomes a bill. Access to the archive follows practice area and client authorization, not an open dump of **47,000** files into one prompt. What can run without that pass is the search itself, a sourced answer in about **4 seconds**, and the draft that exists so the partner has something to mark up. The 45-minute review is the control. Remove it and a wrong file leaves the building.
This is a different list from a software company's. A B2B SaaS support agent resolves customer tickets in a helpdesk; a consulting-firm agent works inside the firm, on the archive, the calendar, and the billing tool. Where agents pay off first in a product company is covered in the [support automation playbook for SaaS](https://gigabit.agency/insights/ai-support-automation-saas/).
### What to do this week
1. **Export one quarter of proposals.** Count them, the hours on each, and where the precedent lives.
2. **Time five retrievals.** If finding a prior deliverable takes longer than writing the paragraph, the archive is the first build.
3. **Pull last quarter's capture rate.** Near **80–85%**, keep the unrecorded slice as its own dollar line. Leave it off the wasted-capacity figure.
4. **Write the order:** index the files you would quote, then the draft that cites them, then morning suggestions a person approves.
5. **Price the diagnostic.** The [$25,000, 2-week AI Transformation Sprint](https://gigabit.agency/services/ai-transformation/) ranks the leaks and scopes phase 1. Vertical patterns are on the [professional services page](https://gigabit.agency/industries/professional-services/). The phase result is the [consulting proof](https://gigabit.agency/proof/ai-transformation-professional-services/).
The deployment left the consultants in place. They open an engagement by asking what the firm has already done, and the largest gains showed up with the people who used to wait on someone else's memory. The published return is **1,200 hours a quarter**, **$480,000 a year**, payback in **5.3 months**, on **$210,000** in Year 1. If proposals still start from a blank page, book the [$25,000 Sprint](https://gigabit.agency/services/ai-transformation/). If one workflow is already scored, a [Gigabit Agent from $8,000](https://gigabit.agency/services/ai-agents/) is the smaller fixed purchase.

### FAQ
**Where should AI for consulting firms start?**
Start with the engagement archive, then proposals that cite it, then time capture. In a published consulting deployment, the assessment ranked proposal generation at 320 hours a quarter, knowledge retrieval at 280, deliverable formatting at 240, and internal operations at 200. The first build ingested 12,000 of 47,000 files into searchable form. Proposal drafts only paid off after that index existed: creation time fell from 4–8 hours to 45 minutes of review. A proposal bot with nothing to retrieve gets rewritten.
**How many hours can AI return at a consulting firm?**
A published management-consulting program reports 1,200 hours per quarter and $480,000 a year in recovered value, with 129% Year 1 ROI and a 5.3-month payback on $210,000 of Year 1 investment. The same case itemizes the first full quarter at 1,020+ hours and $377,000 across knowledge search, proposals, deliverables, operations, reporting, and a time-capture lift. Both figures are in the proof. They are not reconciled here, and the quarter is not multiplied by four to invent the annual number.
**What is a realistic AI time-capture gain for a consulting firm?**
In that deployment, capture moved from 80–85% to 94%. The case values the lift at $95,000 in the measured quarter. The managing partner estimates $380,000 a year in additional billings. The quarterly figure is the measurement. The yearly figure is the partner's estimate. Suggestions came from calendar, email, document, and channel activity, and a person approved each entry. The system did not post time on its own.
**How much does a consulting-firm AI program cost?**
The published program was $210,000 in Year 1: a $12,000 assessment, a $156,000 build, and $4,500 a month to operate. Gigabit lists a $25,000, two-week AI Transformation Sprint, production agents from $8,000 flat, and Managed AI Operations at $3,000–$20,000 a month. A multi-workflow year is scoped in the $80,000–$500,000 range. One agent from $8,000 fits a single scoped workflow. It does not fit an unsorted archive plus proposals plus time capture.
**Is an AI agent for a consulting firm the same as a SaaS support agent?**
No. A B2B SaaS support agent resolves customer tickets in the product's helpdesk. A consulting-firm agent works inside the firm: search across past deliverables, a draft proposal from a brief, and a morning timesheet suggestion the consultant approves. The systems of record are the document archive, the calendar, and the billing tool. The failure mode differs too. A wrong support reply annoys a customer. A proposal that leaves the firm without a partner review puts last year's advice under this year's name.
**Which consulting-firm steps should stay with a person?**
The partner still edits and sends the proposal. Consultants still supply the analysis; the tools handle structure, formatting, and data presentation. Time-entry suggestions need a one-tap approval, because the entry becomes a bill. Access to the archive follows practice area and client authorization. What can run without that pass is the search itself, a sourced answer in about 4 seconds with links to the underlying deliverables, and the draft that exists so the partner has something to mark up.

## Build an AI agent eval set from real tickets
URL: https://gigabit.agency/insights/build-ai-agent-eval-set-from-tickets/
Published 2026-10-06

To **build an AI agent eval set from real tickets**, export production history for one workflow, stratify by volume and failure mode, and have a domain owner label what "correct" means before any model change reaches traffic. Invented prompts are not a suite. In a [published D2C support deployment](https://gigabit.agency/proof/ai-automation-ecommerce-support/), testing against **2,000** historical tickets with known resolutions produced **94.2%** intent classification accuracy, **91.7%** response accuracy, **99.1%** policy compliance, and a **4.3/5.0** tone score before staged rollout. Which numbers should gate a deploy is [a separate checklist](https://gigabit.agency/insights/ai-agent-eval-metrics/). This post is the labeling workflow that produces those numbers — export, sample, define correct, and know when the set is good enough to block a ship.
### Start from history, not from a brainstorm
Pull the last **90 days** of tickets, forms, or call transcripts for **one** workflow. Keep the raw text the system will see: subject, body, structured fields, and the resolution that already happened. Do not rewrite tickets into "cleaner" demo language. The agent will meet the messy version in production.
A published support build began with classification of **28,400** tickets from **6 months** of helpdesk history, then scored the agent against a **2,000**-ticket slice with known resolutions. That sequence matters. Assessment maps the distribution. The eval set is the scored subset you refuse to ship without. If you only have a dozen happy-path scripts from a workshop, you have a demo script — not a ruler.
### What to export on day one
- **Input as seen** — raw ticket text, channel, and any fields the agent will receive.
- **Human resolution** — the answer sent, the tool action taken, or the escalation reason.
- **Category / intent** — whatever your team already tags, even if noisy.
- **Outcome flags** — refund issued, return opened, policy exception, reopened ticket.
Strip customer PII before it leaves your environment. Keep enough structure that a labeler can still decide intent and the correct action. If legal blocks export, label inside the helpdesk with a locked view — do not invent substitute tickets to "move faster."
### Stratify by volume and by the cases that hurt
A random sample of the top three intents will pass a demo and fail on the refund that creates a chargeback. Stratify on two axes: **share of volume** and **cost of a miss**.
| Bucket | How to sample it | Why it belongs |
| --- | --- | --- |
| Top intents by volume | Oversample the categories that burn hours | Protects the common path |
| Money / policy edge | Returns, discounts, cancellations, medical-advice refusal | Misses here cost cash or compliance |
| Escalation-required | Cases a senior always takes | Tests whether the agent knows to stop |
| Adversarial / ugly | Truncated text, rage, mixed languages, attachments | Catches the tail demos skip |
| Recent policy changes | Tickets after a policy edit | Stops the suite from scoring yesterday's rules |
Aim for hundreds first — **200–500** labeled cases — then expand toward low thousands when volume supports it. The published support suite used **2,000** historical tickets before traffic moved. Under-sampling the awkward **10%** is how you report **90%** accuracy on the happy path and discover the rest from customers. For multi-step agents, score each irreversible step on its own labeled slice — the failure math is in [when a multi-step AI agent fails in production](https://gigabit.agency/insights/when-a-multi-step-ai-agent-fails-in-production/).
### Define "correct" before anyone scores a model
The suite is a specification. Write the rubric while the model is still offline. Domain owners decide; engineers wire the runner. If only the vendor labels correct, you bought a vibe check.
### Fields every labeled case must carry
- **Intent / task label** — the route the agent must choose.
- **Known-correct answer or action** — the reply a senior would send, or the tool call that should fire.
- **Must-escalate flag** — cases where autonomy is wrong even if the draft looks fine.
- **Policy tags** — return window, discount eligibility, refusal rules.
- **Difficulty band** — easy / edge / adversarial, so you never hide the tail inside one blended score.
### Who labels, and how disagreements die
Pick one accountable owner per workflow — support lead, ops manager, clinical ops — not a rotating Slack channel. Have a second reviewer dual-label a **10–20%** sample. Where they disagree, write the rule into the rubric and relabel. Unresolved disagreements become "must escalate" until the business picks a side. That disagreement rate is a product signal: if seniors cannot agree, the agent cannot either.
Do not let engineers invent the gold answer from memory of a demo. The [eval suite is the strategy](https://gigabit.agency/insights/ai-agent-eval-metrics/) only when correct is a business decision with a name next to it.
### Know when the set is good enough to gate a deploy
A set is ready when three conditions hold — not when someone is bored of labeling.
1. **Coverage** — top volume intents and the expensive failure modes each have enough labeled cases to move a percentage point.
2. **Rubric stability** — dual-label disagreements on the sample are resolved, and new tickets rarely invent a new "correct" mid-week.
3. **Runner exists** — the same cases score automatically on every prompt, model, and retrieval change, and a red run blocks promote.
Then write the floors into the SOW. The published support pattern — not a universal law — hit **94.2%** intent, **91.7%** response accuracy, **99.1%** policy compliance, and **4.3/5.0** tone, with online autonomy only above **0.85** intent confidence and **0.80** retrieval relevance. Put your floors next to those numbers. Softening them after a bad demo week is how pilots never exit. After launch, promote hard live traces into the set weekly — especially escalations and low-confidence paths — so the suite tracks the traffic you actually get.
### Failure modes that fake a suite
**Workshop prompts.** Marketing scenarios with perfect grammar. They pass every vendor demo and miss the truncated mobile ticket.
**Vendor-only labels.** The builder marks its own homework. Scores rise; production does not.
**Happy-path only.** Ninety percent of volume, zero chargeback cases. The blended score looks green.
**One blended metric.** Intent is fine; the write-back is wrong. Without per-step labels you ship the product of the misses.
**Frozen set.** Policy changed in July; the suite still scores June. Green CI, wrong answers.
If a vendor cannot run **your** labeled tickets, they cannot ship **your** agent. Refuse a demo-only milestone.
### What to do this week
1. **Export 90 days** of one workflow's tickets with resolutions. Keep raw input; strip PII.
2. **Stratify** — top intents by volume, plus money/policy edges, mandatory escalations, and ugly cases.
3. **Label 200–500** with a named domain owner: intent, correct action, must-escalate, policy tags, difficulty.
4. **Dual-label 10–20%** and resolve disagreements into the rubric before you score a model.
5. **Price the diagnostic** — book the [$25,000, 2-week AI Transformation Sprint](https://gigabit.agency/services/ai-transformation/) to turn that export into a suite and a scoped pilot that credits toward a build, or start from [Gigabit Agents](https://gigabit.agency/services/ai-agents/) (**from $8,000** flat) when the workflow and labels are already clear.
You cannot gate what you have not labeled. Build the eval set from the tickets your team already resolved, put a name on "correct," and treat every model swap as a deploy that needs a green run. The [support proof](https://gigabit.agency/proof/ai-automation-ecommerce-support/) shows the **2,000**-ticket suite and the scores before traffic moved — and the [metrics checklist](https://gigabit.agency/insights/ai-agent-eval-metrics/) is what those scores should look like in the contract.

### FAQ
**How do you build an AI agent eval set from real tickets?**
Export production tickets for one workflow with the raw input and the human resolution, stratify by volume and by costly failure modes, and have a domain owner label intent, the known-correct answer or action, must-escalate cases, and policy tags. Start with 200–500 labeled cases and expand toward low thousands when volume supports it. A published support deployment scored against 2,000 historical tickets with known resolutions before staged rollout. Invented workshop prompts are not an eval set.
**Who should label correct answers on an AI agent eval set?**
A named domain owner for that workflow — support lead, ops manager, or equivalent — not only the vendor and not a rotating Slack channel. Engineers wire the scoring runner. Dual-label a 10–20% sample; resolve disagreements into the rubric or mark the case must-escalate until the business decides. If seniors cannot agree on correct, the agent cannot either.
**How do you know an AI agent eval set is ready to gate a deploy?**
When top volume intents and expensive failure modes are covered, the labeling rubric is stable after dual-review, and an automated runner can score every prompt, model, and retrieval change with a red run that blocks promote. Then write pass floors into the SOW. In a published D2C support build, offline evals hit 94.2% intent accuracy, 91.7% response accuracy, 99.1% policy compliance, and 4.3/5.0 tone before traffic moved.
**Why not invent prompts instead of labeling real tickets?**
Invented prompts match the demo, not production. They miss truncated text, policy edges, mandatory escalations, and the ugly 10% that creates refunds and chargebacks. Teams that skip real tickets discover regressions from customers instead of from a red CI run. Pull history, label what already happened, and promote hard live traces into the set after launch.

## Calibrate an LLM-as-judge for AI agents
URL: https://gigabit.agency/insights/calibrate-llm-as-judge-for-ai-agents/
Published 2026-10-07

**LLM-as-judge for AI agents** is usable only after you calibrate it against a human-labeled holdout — not after the first rubric prompt looks plausible. Auto-scoring without that check is how teams greenlight a regression and call it quality. The ruler those labels produce is concrete: in a [published D2C support deployment](https://gigabit.agency/proof/ai-automation-ecommerce-support/), a suite scored on **2,000** historical tickets with known resolutions hit **94.2%** intent classification accuracy, **91.7%** response accuracy, **99.1%** policy compliance, and a **4.3/5.0** tone match before staged rollout. Which metrics belong in that suite is covered in [what to measure in an AI agent eval suite](https://gigabit.agency/insights/ai-agent-eval-metrics/). Why the suite comes before strategy slides is covered in the same piece. This post is the scoring layer in between: when an LLM may judge, when a human must, and when disagreement blocks the deploy.
### Exact match vs LLM-as-judge
Split the suite by how "correct" is defined. If a domain owner can write a binary check, use the binary check. If the answer is free text with more than one acceptable phrasing, you need a judge — and a calibration loop before you trust it.
| Score type | Example | Prefer |
| --- | --- | --- |
| Exact / deterministic | Intent label equals `return_status`; refund amount matches policy table; required escalation fired | Code check, not a judge |
| Structured field match | Tool call name and args match the labeled action | Schema compare + exact args |
| Free-text quality | Reply matches known-correct meaning; tone matches brand baseline | Calibrated LLM-as-judge |
| Irreversible judgment | Money moved, clinical note filed, contract term waived | Human label; judge drafts only |

Policy compliance in the support proof is a **99.1%** hit rate against documented rules. That is closer to exact match than to vibes: the rule either held or it did not. Response accuracy at **91.7%** is the free-text band — different correct phrasings still count as correct. Put the first band in code. Put the second band behind a judge only after agreement on a holdout. Treating every metric as free text is how you pay for a second model to restate what a string compare already knew.
### Calibrate on a human-labeled holdout
Calibration means a fixed set of cases where a human already scored the agent output, then you measure how often the judge agrees. Promote auto-scoring only when that agreement clears a floor you wrote down. Until then, the judge is a draft scorer, not a deploy gate.
### Build the holdout like a ruler, not a demo
Pull real agent outputs — including failures — not invented chat. Use the same domain owners who labeled the golden set. Keep the holdout frozen while you tune the judge prompt; if you keep editing labels to make the judge look better, you measured nothing.
### An instructional agreement bar
Assume this pattern unless your risk profile is stricter: do not promote LLM-as-judge as the primary score until judge–human agreement on the holdout clears about **85%**, and dig every disagreement before you raise the bar. That **85%** is an operator assumption for this guide, not a client measurement and not a substitute for the offline floors in the support proof (**94.2%** / **91.7%** / **99.1%**). Those floors are agent quality against human-known-correct labels. Judge agreement is whether your auto-scorer tracks those humans. Confusing the two numbers is the failure mode.
### What to log on every disagreement
- The case id and the human score.
- The judge score and the one-line reason it gave.
- Whether the rubric was ambiguous, the human was wrong, or the judge hallucinated a rule.
- The fix: rubric edit, holdout re-label, or a hard exact check that should never have gone to a judge.
Re-run the frozen holdout after each rubric change. A green agent suite with a red judge–human agreement is not a pass — it is a broken measuring stick.
### What the judge may score vs what humans must
Give the judge the work that is high volume and recoverable. Keep humans on the work that is rare, irreversible, or undefined.
### Safe for a calibrated judge
Tone match against a written brand baseline (the support suite used **4.3/5.0** against the team's best agent). Semantic match of a reply to a known-correct answer when several phrasings are fine. Flagging likely policy language for a second pass — not the final compliance call when the rule is binary.
### Keep humans on the holdout forever
Irreversible actions: refunds above a threshold, payment release, anything that writes money or legal status. Clinical-adjacent or regulated wording where a miss is silent until audit. Cases where two senior operators disagree — if humans cannot define correct, [you are not past the first production gate](https://gigabit.agency/insights/why-your-ai-pilot-never-reached-production/).
Online confidence gates are a separate layer from judging. In the same support deployment, the agent went autonomous only when intent confidence exceeded **0.85** and retrieval relevance exceeded **0.80**; below either threshold a human got the draft. Offline scoring and online gating both need numbers. Neither replaces the other. The [production AI agent stack](https://gigabit.agency/insights/production-ai-agent-stack/) puts evals next to retrieval, observability, and guardrails for that reason.
### Block the deploy on disagreement
Wire three checks into the same CI path that already runs the golden set:
1. **Agent suite green** — intent, response, policy, and any other gate metrics you put in the SOW.
2. **Judge–human holdout green** — agreement at or above the floor you chose; sample disagreements reviewed.
3. **No silent rubric drift** — judge prompt and rubric version pinned; a change is a deploy event.
If the agent suite passes but the holdout agreement drops, fail the build. The model under test is not the only thing that can regress — the scorer can too. Vendor model updates that retune judge behavior are a forced recalibration, the same way a production agent change is a forced golden-set run.
After launch, promote hard live traces into both the golden set and the judge holdout. That rhythm is what [Managed AI Operations](https://gigabit.agency/services/managed-ai-operations/) covers at **$3,000–$20,000/month**: regressions as numbers, not as customer tickets. In the published support build, **64%** of tickets ran autonomous (**3,149/month**) only after the offline suite passed and confidence gates held — not after a judge prompt was pasted into a notebook.
### What to do this week
1. **Split your suite** into exact checks vs free-text scores. Move every binary policy and schema match out of the judge.
2. **Carve a frozen human-labeled holdout** — start with dozens of scored outputs if volume is low; expand with the nasty tail. Do not invent chat.
3. **Write the agreement floor** on one line (use ~**85%** as a starting assumption, or raise it for irreversible workflows) and the agent pass floors next to it.
4. **Fail CI** when holdout agreement drops or the agent suite goes red. Pin judge prompt versions like any other deploy artifact.
5. **Price the diagnostic**: book the [$25,000, 2-week AI Transformation Sprint](https://gigabit.agency/services/ai-transformation/) to build the suite, the holdout, and a scoped pilot path that credits toward a build — or start from [Gigabit Agents](https://gigabit.agency/services/ai-agents/) (**from $8,000** flat) when the workflow and labels are already clear.
You cannot auto-score what you have not calibrated. Put human labels on a holdout, measure judge agreement, and keep exact checks in code. The [full support proof](https://gigabit.agency/proof/ai-automation-ecommerce-support/) shows what a human-known-correct suite looked like before traffic moved — **94.2%** / **91.7%** / **99.1%** on **2,000** tickets, then **64%** autonomous with gates at **0.85** and **0.80**. Use that shape. Do not skip the calibration step and call the judge the truth.

### FAQ
**When should you use an LLM-as-judge for AI agents?**
Use a calibrated LLM-as-judge for free-text quality — semantic match to a known-correct answer, tone against a brand baseline — after a human-labeled holdout shows the judge agrees with domain owners. Prefer exact or schema checks for intent labels, tool-call args, and binary policy rules. Do not put irreversible money or regulated sign-off behind an uncalibrated judge.
**What judge–human agreement rate should gate auto-scoring?**
Write a floor before you promote auto-scoring. An instructional starting bar is about 85% agreement on a frozen human-labeled holdout, with every disagreement reviewed — raise it for irreversible workflows. That 85% is an operator assumption, not a published client metric. Keep it separate from agent accuracy floors such as the published support suite's 94.2% intent, 91.7% response, and 99.1% policy scores.
**How is LLM-as-judge calibration different from AI agent eval metrics?**
Eval metrics define what to measure and which numbers gate a deploy — intent, response accuracy, policy, escalation, tone. LLM-as-judge calibration defines whether an auto-scorer is allowed to produce those free-text scores at scale. You still need the golden set and pass thresholds; the judge is only the measuring stick for the free-text band after it tracks humans on a holdout.
**Should we book the Sprint or Gigabit Agents to set this up?**
Book the $25,000, 2-week AI Transformation Sprint when the workflow, golden set, and judge holdout are not yet defined — the Sprint maps the suite and a scoped pilot that credits toward a build. Buy Gigabit Agents from $8,000 flat when one production workflow already has labeled cases and clear gates. Managed AI Operations at $3,000–$20,000/month keeps regression runs and holdout refresh from slipping after launch.

## What is a forward deployed engineer?
URL: https://gigabit.agency/insights/what-is-a-forward-deployed-engineer/
Published 2026-10-23

A **forward deployed engineer (FDE)** is a software engineer who works embedded with one customer, on site or inside the customer's own tools, and builds and ships software against that customer's real problem until it works in production. A product engineer builds one product for many customers. An FDE does the reverse: many capabilities for one customer, with the outcome as the job. Palantir, which made the title famous, describes the split the same way: its product engineers build one capability for many customers, and its forward deployed engineers enable many capabilities for one customer ([Palantir, A Day in the Life of a Forward Deployed Software Engineer](https://blog.palantir.com/a-day-in-the-life-of-a-palantir-forward-deployed-software-engineer-45ef2de257b1)). If you are hiring for the role, read on. If you are buying the work rather than the headcount, the last section covers hiring a forward-deployed team instead of an FDE.
### Where the role came from
The title comes from Palantir. Its two largest engineering roles are Software Engineers and Forward Deployed Software Engineers, known internally as "Devs" and "Deltas". Palantir's own engineering blog says the name Delta is a leftover from its early days, when each business development team was named after a letter of the NATO alphabet ([Palantir, Dev versus Delta](https://blog.palantir.com/dev-versus-delta-demystifying-engineering-roles-at-palantir-ad44c2a6e87)). Devs build the core platforms. Deltas sit with customers in defense, government, healthcare and industry, and turn those platforms into working systems for one organisation's problem.
The idea is older than the title. Marty Cagan of SVPG argues that sending a technical person to embed with the target customer, to understand the environment well enough to find a solution that delivers the outcome, is good product practice in general, and credits Steve Blank and Bill Campbell with similar arguments ([SVPG, Forward Deployed Engineers](https://www.svpg.com/forward-deployed-engineers/)). His preferred form sends one FDE to several customers so the patterns they find can become a single product.
### What a forward deployed engineer does
The work is software engineering, done in someone else's building. A typical week covers five kinds of task:
1. **Discovery.** Sitting with the people who do the work today, reading the data, and finding the one workflow where software will change a number the customer cares about.
2. **Scoping.** Turning that into a build the customer can approve: what ships, what it connects to, and how success is measured.
3. **Building.** Writing production code against the customer's systems, data and security rules. Integrations take more of the time than the model or the interface.
4. **Shipping.** Getting the system into production and into daily use, which includes access reviews, change control and training.
5. **Feeding back.** Telling the product team what customers keep asking for, so the next customer gets it as a feature.
Job postings show the same shape. OpenAI's FDE listings ask the engineer to own discovery, technical scoping, system design, build and production rollout, with success judged by production adoption and measurable impact on the customer's workflow ([OpenAI careers, Forward Deployed Engineer](https://openai.com/careers/forward-deployed-engineer-%28fde%29-nyc-new-york-city/)). Anthropic's Applied AI team says its FDEs embed directly with its most strategic customers to drive AI adoption, building production applications with Claude inside customer systems and travelling to customer sites 25–50% of the time ([Anthropic, Forward Deployed Engineer, Applied AI](https://jobs.accel.com/companies/anthropic/jobs/62104998-forward-deployed-engineer)).
### Forward deployed engineering in AI
An **AI forward deployed engineer** does the same job with large language models in the stack. Bloomberry's analysis of about 1,000 FDE job postings found that the most requested skills were Python (66% of postings), TypeScript (35%) and AWS (32%), with AI agents (35%) and LLM experience (31%) leading the AI skills. Working directly with customers appeared in 55% of postings ([Bloomberry, I analyzed 1,000 forward deployed engineer jobs](https://bloomberry.com/blog/i-analyzed-1000-forward-deployed-engineer-jobs-what-i-learned/)). The model itself is rarely the hard part. Connecting it to the customer's data, adding guardrails, writing evaluations and getting a nervous compliance team to sign off is the job.
### FDE vs solutions engineer vs consultant vs staff augmentation
The four roles get confused because all of them put a technical person in front of a customer. The difference is what they are paid to deliver.
| Role | Main job | Writes production code? | Measured on | Usually paid by |
| --- | --- | --- | --- | --- |
| Forward deployed engineer | Build and ship a working system for one customer's problem | Yes, most of the week | The system in production and the customer's number moving | The vendor, as part of the deal |
| Solutions or sales engineer | Show that the product fits, before the sale | Demos and proofs of concept | Deals won | The vendor's sales budget |
| Management or IT consultant | Diagnose, recommend, sometimes implement | Sometimes | Deliverables in the statement of work | The customer, by the day or project |
| Staff augmentation contractor | Add capacity to the customer's own team | Yes, on tickets someone else wrote | Hours billed | The customer, by the hour |

The useful test is ownership. A solutions engineer is finished when the contract is signed. A consultant is finished when the deliverable is accepted. A staff augmentation contractor is finished when the hours run out. A forward deployed engineer is finished when the system works in production and people use it. How the forward-deployed model differs from staff augmentation and a fixed statement of work, with prices, is on [our comparison page](https://gigabit.agency/compare/embedded-vs-staff-augmentation/).
### Why AI companies are hiring FDEs now
Demand for the title has grown fast. Bloomberry counted postings titled "forward deployed engineer" from January to October 2025 and found **1,165%** more than in the same months of 2024, with October 2025 the highest month on record at the time ([Bloomberry](https://bloomberry.com/blog/i-analyzed-1000-forward-deployed-engineer-jobs-what-i-learned/)). Different counts give different multiples, so treat the exact figure as one data point. The direction is not in doubt.
Three things explain it:
- **AI products do not install themselves.** A model with an API is a capability. A support agent that resolves tickets inside one company's help desk, with its policies and its edge cases, is a project. Somebody has to do that project, and AI vendors have decided it should be their own engineers.
- **The value sits in the customer's data and workflow.** Generic demos are easy to copy. The deployment that works on a customer's real records is hard to copy, and it is where renewals come from.
- **Discovery has to happen in the field.** Cagan's point applies directly: startups building AI agents adopt the model because they need deep customer context to find what is worth building ([SVPG](https://www.svpg.com/forward-deployed-engineers/)).
### Skills and pay
The role asks for a full-stack engineer who can also run a meeting. Postings ask for production experience with Python or TypeScript and a major cloud, recent LLM work, and comfort working directly with customer executives. OpenAI's London FDE listing asks for 5+ years of engineering or technical deployment experience ([OpenAI careers](https://openai.com/careers/forward-deployed-engineer-london/)). On pay, Bloomberry puts the median at **$173,816** across postings that disclosed a salary range ([Bloomberry](https://bloomberry.com/blog/i-analyzed-1000-forward-deployed-engineer-jobs-what-i-learned/)), and one Anthropic Applied AI posting listed a base salary of **$200,000–$300,000** ([Anthropic posting](https://jobs.accel.com/companies/anthropic/jobs/62104998-forward-deployed-engineer)). The frontier-lab posting sits well above that median.
Looking for an FDE job rather than an FDE? Gigabit hires engineers who want to work this way. See [open roles](https://gigabit.agency/careers/).
### Hiring a forward-deployed team instead of an FDE
Most mid-market companies do not need to employ an FDE. They need the result an FDE produces: a system built inside their business, against their real problem, shipped to production and owned until it works. Hiring one senior engineer with that profile is slow and expensive, and one person cannot cover design, back end, integrations, evaluations and operations alone.
That is the gap Gigabit fills. Gigabit is an AI-native software factory, and forward deployment is how it works: senior engineers embed in your team, your repos, your Slack and your board, and build against your problem with AI agents doing the repetitive work. You get the forward-deployed model as a team, priced by the month, instead of a single hire you have to find, pay and keep. [How the model runs](https://gigabit.agency/approach/) is written up step by step.
| | Hire one FDE | Embed a forward-deployed team |
| --- | --- | --- |
| Time to start | A senior search, often months | Assembled in about **2 weeks**, first pull requests by **Day 5** |
| Coverage | One person's skills | A pod matched to the work: full-stack, AI, QA, data |
| Cost | A US senior salary plus equity and benefits | **$3,000–$7,500 per engineer per month** |
| Commitment | Permanent headcount | **2–4 week** trial, no lock-in |
| Who owns the outcome | You, through the hire | Shared: your owner sets the bar, the pod ships to it |

### What it looked like for one client
A Series B HR tech company had two senior engineering seats open for four months and had lost three offers. Three embedded Gigabit engineers cost **$14,400 a month**. Their first pull requests landed in **5 days**, sprint velocity rose **2.3×** (34 to 78 points) within the quarter, and the company saved **$392,400 (65%)** a year against the US hires it had priced ([full case study](https://gigabit.agency/proof/team-augmentation-hrtech/)). That is one engagement, not a promise. It shows what to demand from any forward-deployed team: named engineers against a written spec, a pull request in your repo in the first week, and a trial before you scale.
### When to hire your own FDE instead
Hire the FDE yourself when you are a software vendor and forward deployment is part of your product: your engineers need to sit with your customers, learn what they need, and carry it back to your roadmap. That knowledge should stay in your company. Embed a team when you are the customer: you have a problem to solve, a backlog longer than your team can finish, and no time to spend a quarter recruiting. If you are weighing the two, [when to hire vs embed AI engineers](https://gigabit.agency/insights/hire-vs-embed-ai-engineers/) walks through the numbers.
### The short version
A forward deployed engineer is an engineer who works inside a customer's business and owns the result until the software runs in production. Palantir made the title. AI companies grew demand for it more than tenfold in 2025. If you are buying the outcome rather than the headcount, [Embedded AI Teams](https://gigabit.agency/services/embedded-ai-teams/) give you the forward-deployed model at **$3,000–$7,500 per engineer per month**, with first pull requests by Day 5 and no lock-in.

### FAQ
**What is a forward deployed engineer?**
A forward deployed engineer (FDE) is a software engineer who works embedded with one customer, on site or inside the customer's tools, and builds and ships software against that customer's real problem until it works in production. A product engineer builds one capability for many customers; an FDE builds many capabilities for one customer and is measured on the outcome.
**Where does the term forward deployed engineer come from?**
Palantir popularised it. Its two largest engineering roles are Software Engineers and Forward Deployed Software Engineers, called Devs and Deltas internally. Palantir says the name Delta dates from its early days, when business development teams were named after letters of the NATO alphabet. AI companies including OpenAI and Anthropic now hire FDEs under the same name.
**How is a forward deployed engineer different from a solutions engineer?**
A solutions or sales engineer proves the product fits before a sale and is measured on deals won. A forward deployed engineer builds and ships the working system after the sale and is measured on that system running in production and moving the customer's metric. FDEs spend most of their week writing production code.
**How much does a forward deployed engineer earn?**
Bloomberry's analysis of about 1,000 FDE job postings put the median at $173,816 among postings that disclosed a salary range. Frontier AI labs pay more: one Anthropic Applied AI FDE posting listed a base salary of $200,000 to $300,000.
**Can I hire a forward-deployed team instead of an FDE?**
Yes. Gigabit embeds forward-deployed teams of senior engineers in your repos, Slack and board. Teams are assembled in about two weeks, ship first pull requests by Day 5, and cost $3,000 to $7,500 per engineer per month with a 2 to 4 week trial and no lock-in. In one published HR tech engagement, three engineers at $14,400 a month raised sprint velocity 2.3x and saved $392,400 a year against US hires.

## When to hire vs embed AI engineers
URL: https://gigabit.agency/insights/hire-vs-embed-ai-engineers/
Published 2026-10-08

**Hire vs embed AI engineers** when the open seat is already costing the roadmap — not when a recruiter sends the first shortlist. In a [published Series B HR Tech engagement](https://gigabit.agency/proof/team-augmentation-hrtech/), two senior positions sat open for **4 months**: **200+** applications, **40** phone screens, **3** offers extended, and all three lost to counter-offers or competitors. Fully loaded cost per hire in that market was **$195K–$220K/year**. The internal team was **7** engineers against a **14-month** P0/P1 roadmap that their velocity would only ship in **8 months**. The pod that replaced the failed search was three engineers at **$14,400/month** (**$207,600** a year) versus **$600,000** for the US-equivalent hires — **$392,400 (65%)** saved — with first pull requests in **5 days**. Four months open is this client's miss, used here as the worked example, not a universal recruiting law. How forward-deployed differs from staff augmentation and a fixed statement of work is in [forward-deployed vs staff aug vs a fixed SOW](https://gigabit.agency/compare/embedded-vs-staff-augmentation/). What the model means is on [our approach page](https://gigabit.agency/approach/).
### The recruiting miss that should stop the search
The workforce platform served **600+** mid-market companies. At **$8M ARR** and **45%** year-over-year growth, seven engineers were stretched across too many priorities. Two hundred applications and forty screens are a process. Three lost offers are a result. **14 months** of P0/P1 work sat against a velocity that would deliver **8 months** of it. The missing six months do not return when an offer letter finally lands.
### Treat four months as this case, then write your own line
Do not copy "four months" onto a different market and call it a rule. Write the line for your req: how long the seat has been open, how many offers you have lost, and how many months of committed roadmap your current velocity can actually finish. In this proof those three numbers were **4 months**, **3 lost offers**, and **14 vs 8 months**. If yours rhyme with that shape, more phone screens are not the plan.
Fully loaded **$195K–$220K** per US senior is the comparison the buyer was already using. Two seats, if both had closed, land at **$390K–$440K** before the next req. That product is our arithmetic on the published range (2 × $195K and 2 × $220K), not a third figure in the case study. Use it to size the alternative. Do not average the range into a single "typical" salary.
### What the first two weeks must prove
Week 1 was assembly against a written spec, not a résumé blast. The VP of Engineering asked for **2** senior full-stack engineers (React, TypeScript, NestJS, PostgreSQL) plus **1** QA automation engineer (Cypress). **5** candidate profiles arrived within **4 business days**. He interviewed all five in **30-minute** architecture conversations and selected three.
Week 2 was Sprint Zero. Access was provisioned on days 1–2. Each engineer traced a core workflow and wrote a technical summary — deep enough in **48 hours** to produce documentation the internal team had never written. First pull requests were in by end of day Friday. The published timeline is **5 days** to first PR and full sprint velocity in **3 weeks**.
### The buying test is a PR in your repo
Judge the pod on artifacts you can open. Named seniors matched to a spec you wrote. Interviews you ran. A written trace of your codebase. A pull request inside the first week, reviewed on your bar. If week two ends in onboarding checklists and no PR, you bought a ramp. [Embedded AI Teams](https://gigabit.agency/services/embedded-ai-teams/) are priced for that test: assembled in about two weeks, first PRs by Day 5, a **2–4 week** trial, no lock-in, at **$3,000–$7,500 per engineer per month**.
From week 3 they ran the client's cadence: biweekly sprints, the same Linear board, repos, and Slack. Async updates posted at **5:30 PM Dhaka** landed at **7:30 AM Eastern**. Sync standups ran twice weekly in a **2-hour** overlap. Pull requests were reviewed within **24 hours**. A pod that cannot name the overlap window and the review SLA is still a staffing pitch.
Quarter-later, sprint velocity rose **2.3×** (34 → 78 points). That result, and how it differs from staff aug or a fixed SOW, belongs to the [comparison page](https://gigabit.agency/compare/embedded-vs-staff-augmentation/). Use it as confirmation that Day-5 PRs were not a stunt. Do not shop for a vendor on the 2.3× alone.
### Dollar math on one page
Put the published totals next to each other before you extend another offer.
| Path | What the proof records |
| --- | --- |
| Keep recruiting two US seniors | Fully loaded **$195K–$220K** each; seats already open **4 months**; hiring cycle the case contrasts is **4–6 months** |
| Two seats if both had closed | **$390K–$440K** — our product of the published range, not a case-study line |
| Three-engineer pod | **$14,400/month**, **$207,600** for the year |
| US-equivalent hires the case priced | **$600,000** |
| Difference | **$392,400 (65%)** saved annually; productivity in **2 weeks** instead of that hiring cycle |

Dividing **$14,400** by three is about **$4,800** per engineer per month. That division is ours. The case publishes the pod total, not a per-seat rate card. **$4,800** sits inside the catalog band of **$3,000–$7,500** per engineer per month. Cite the band when you budget a new pod. Cite **$14,400** when you cite this engagement.
A pod is the wrong SKU when the work is one clear workflow. Then the buy is [Gigabit Agents](https://gigabit.agency/services/ai-agents/) **from $8,000** flat, or the [$25,000, 2-week AI Transformation Sprint](https://gigabit.agency/services/ai-transformation/) when you still need the diagnostic before anyone staffs a seat. Staffing three engineers at a fuzzy problem is how embedded work turns into a body shop.
### When you should still hire
Embed does not retire the staff role that sets the spec. In this engagement the VP of Engineering wrote the stack, ran the architecture interviews, and picked three of five. The pod joined a team that already had an owner. If you have no one who can reject a pull request, you do not have a capacity problem. You have a management problem, and a pod will not invent the owner.
### Hire the seat
Hire when the role is long-term architecture and people management, you can close in weeks rather than a lost quarter, and a senior already on staff can evaluate the work. Hire when the roadmap gap is judgment, not hands. A **7-person** team that cannot review AI changes should not add three external committers and hope taste appears.
### Embed the seats
Embed when reqs have already slipped the way this one did — months open, offers lost, committed months of roadmap larger than the months velocity can finish — and the work spans more than one system. Demand the week-1 spec match and the Day-5 PR. Scale only after the **2–4 week** trial. In the proof the engagement extended and the client expanded from **3 to 5** engineers, with two more planned in **Q3 2026** ahead of a Series C. Expansion followed production, not a kickoff deck.
### What to do this week
1. **Write three numbers** for each open senior seat: days open, offers lost, fully loaded range. Use **$195K–$220K** only if that is your market; otherwise use your own range and label it as yours.
2. **Write the roadmap gap** the way the proof does: months of P0/P1 committed versus months current velocity will finish. Theirs was **14 vs 8**.
3. **Stop adding screens** when those lines match a failed search. Ask for named profiles against a written spec inside a few business days, interviews you run, and a pull request by Day 5.
4. **Price the trial** on [Embedded AI Teams](https://gigabit.agency/services/embedded-ai-teams/) at **$3,000–$7,500/engineer/mo**, **2–4 week** trial, no lock-in. Read the [HR Tech proof](https://gigabit.agency/proof/team-augmentation-hrtech/) for the **$14,400/month** pod, the **$392,400** gap, and the Day-5 PR — not as a promise that your sprint velocity will match theirs.
5. **Switch SKUs** if the job is one workflow. Book the [$25,000 Sprint](https://gigabit.agency/services/ai-transformation/) when the workflow is still fuzzy, or [Gigabit Agents from $8,000](https://gigabit.agency/services/ai-agents/) when it is already clear.
A senior search that has already burned a quarter is a delivery decision. Start [Embedded AI Teams](https://gigabit.agency/services/embedded-ai-teams/) at **$3,000–$7,500 per engineer per month** when open reqs look like this proof: empty seats, lost offers, and a roadmap longer than velocity can finish. Hire when you can close the owner. Do not buy another month of screens to avoid writing the spec.

### FAQ
**When should you embed AI engineers instead of hiring them?**
Embed when senior seats have already sat open for months, offers are being lost, and committed roadmap months exceed what current velocity can finish. In a published HR Tech engagement that shape was two seats open four months, three lost offers, and 14 months of P0/P1 work against velocity that covered 8. Four months is that client's miss, not a universal cutoff. Hire when you can close an owner in weeks and someone on staff can review the work.
**How much does an embedded pod cost versus a US senior hire?**
Embedded AI Teams are $3,000–$7,500 per engineer per month, with a 2–4 week trial and no lock-in. In the published HR Tech pod, three engineers cost $14,400/month ($207,600 for the year) versus $600,000 for the US-equivalent hires — $392,400, or 65%, less. Fully loaded US seniors in that market were $195K–$220K each. Dividing $14,400 by three is our arithmetic, not a published per-seat rate.
**What should the first two weeks of an embedded team prove?**
A written spec, named seniors, interviews you run, a codebase trace, and a pull request in your repo by about Day 5. In the HR Tech proof, five profiles arrived in four business days, the VP of Engineering picked three after 30-minute architecture interviews, and first PRs landed in five days, with full sprint velocity in three weeks. Onboarding checklists without a PR are a ramp, not a trial.
**Should we hire a full-time owner if we embed a pod?**
Yes, if you do not already have one. The pod in the proof joined a team whose VP of Engineering wrote the spec and selected the engineers. Embed adds hands and senior delivery inside your board, Slack, and repos. It does not replace the person who can reject a pull request. If nobody on staff can review the work, fix that before you add committers.

## When ChatGPT and Perplexity disagree
URL: https://gigabit.agency/insights/when-chatgpt-and-perplexity-disagree/
Published 2026-10-09

When **ChatGPT and Perplexity disagree**, score each assistant before you average them. A blended citation rate can read **30%** while the engine your buyer actually opens names you on zero queries. The site's panel definition is already fixed: twenty category buying questions, and six names is a **30%** citation rate. Brands with no GEO work typically sit at **0–10%**. What counts as good, and why a single run is an anecdote, is in [what is a good AI citation rate](https://gigabit.agency/insights/what-is-a-good-ai-citation-rate/). This post is the split inside that number. Worked example, our arithmetic, not a client measurement: ChatGPT names you on **12 of 20** queries (**60%**) and Perplexity names you on **0 of 20**. Pool the cells and you get **12 of 40 = 30%**. Average the two rates and you also get **30%**, because both panels are the same length. Either blend hides a total miss. Do not ship the blend as the KPI.
### The blend that hides a zero
A citation rate is a fraction. The numerator is queries where an assistant names you. The denominator is the panel. Collapse two assistants into one fraction and a zero can sit inside a number that looks like a program is working. **30%** is the rate the citation-rate post calls a category win when it is real on the queries that matter. **30%** produced by 12 hits and 20 misses is a different fact. One engine is carrying the whole score. The other is invisible.
The coincidence is the trap. **(60% + 0%) / 2 = 30%**, and **12/40 = 30%**. Those match only because each engine was asked the same twenty questions. If the panels had different lengths, the unweighted average of the rates and the pooled-cell rate would diverge. Write the denominator down. A slide that says "citation rate: 30%" without the engine column is not a measurement.
### What to put on the slide instead
Report one row per engine, then the pool if you still want a single line. The single line is a summary. It is not the decision.
| Engine | Named | Panel | Rate |
| --- | --- | --- | --- |
| ChatGPT | 12 | 20 | 60% |
| Perplexity | 0 | 20 | 0% |
| Pooled | 12 | 40 | 30% |

Those three rows are the worked example above. Replace the 12 and the 0 with your own counts. Keep the shape. If a third assistant — Google AI Overviews, or Claude — is on the panel, give it its own row. Do not fold it into the 40 until you have looked at its rate alone. The [30-minute visibility check](https://gigabit.agency/insights/check-your-ai-visibility-in-30-minutes/) already records an engine column, a named column, and a who-else column. Use those columns. Do not sum them before you read them.
### Score each engine before you average
Run the same verbatim questions in each assistant. A fresh session, not an account that already knows your company — that mistake is in the 30-minute runbook. Then score three things per cell, not one:
- **Named.** Your company appears in the answer.
- **Cited.** Your URL is a source, which is a different event from being named.
- **Who else.** The competitors named, in order.
A cell where Perplexity cites a competitor's proof page and never says your name is a miss, even if ChatGPT names you in the first sentence. A cell where both name you and neither links you is a mention, not a source. Track them separately. Averaging "named" across engines answers a different question from "are we a source on the engine the buyer uses."
How a company becomes the source an assistant quotes is the [GEO field guide](https://gigabit.agency/insights/generative-engine-optimization-field-guide/). This post only stops the measurement error that makes that work look further along than it is. Answers also move. The citation-rate post already requires a monthly trend, because one run is noise. A split that is still there next month — Perplexity at zero, ChatGPT carrying the pool — is the finding.
### What a split usually means
ChatGPT and Perplexity do not share one index. The citation-rate post already notes the practical result: a brand routinely shows up in one assistant and stays invisible in another, because each assistant retrieves differently. A split is the ordinary case. It is not proof that the brand is broken, and it is not proof that the zero engine is "wrong."
### A retrieval miss is not an exclusion
If you are absent everywhere — empty HTML, two prices, two legal names — you are excluded, and the blend is not the diagnosis. That drop list is [what gets a company excluded from AI answers](https://gigabit.agency/insights/what-gets-a-company-excluded-from-ai-answers/). Fix the conflict before you tune copy for one assistant. If you are named on ChatGPT and absent on Perplexity for the same buyer question, you have a coverage gap on one retriever. The page may already be citable. The other engine has not picked it up, or it is quoting someone who answered the question more directly.
### The page they can quote
A slogan does not travel across indexes. A page with one outcome and one set of figures does. The [published D2C support proof](https://gigabit.agency/proof/ai-automation-ecommerce-support/) is that shape, and it is not a citation-rate result: **64%** of tickets resolved with no human (**3,149** a month), **94.2%** intent accuracy, **91.7%** response accuracy, **99.1%** policy compliance, first response from **4.2 hours to 6 minutes**, **$217,200** saved in year one against **$92,000** of Year 1 investment. Those sentences can be lifted. If a second URL on your site states a different resolution rate, both engines have a reason to skip you. Agreement between your own pages is a precondition. It does not guarantee both assistants will name you this month.
When a named competitor shows up on the engine where you are absent, use [the competitor diagnosis](https://gigabit.agency/insights/what-gets-a-company-excluded-from-ai-answers/) for that engine's answer, not for the blended rate. The five asymmetries — they answer the question, they publish a fact, someone else vouched, the entity is unclear, you tested a prompt nobody types — apply per assistant. A competitor who wins only on Perplexity is not a sitewide failure.
### Which engine to fix first
Fix the assistant your buyers open. That rule is an operator assumption, not a measured share of your pipeline. Ask sales which tool prospects mention. If they paste Perplexity answers into deals and your Perplexity row is zero, the 60% ChatGPT rate is not the work. If they live in ChatGPT, a Perplexity zero is still real and still second.
Do not "fix" a disagreement by prompting a logged-in account until your name appears. That measures your own session. The 30-minute check uses a logged-out window for that reason. Do not add your brand to the query. Branded prompts measure recognition. And do not brief a content program against the pooled **30%** while one row is zero — you will publish for the engine that already names you and call the average an improvement. Dropping the zero engine to raise the rate is the same cheat. Shrinking the denominator is not a gain.
After the per-engine baseline is honest, the program that keeps pages and measurement current is [Ascent](https://gigabit.agency/services/geo/): a 12-month program covering brand, website, content, SEO, and GEO, from $4,000 a month with no upfront fee. What GEO costs, and what it does not, is [a separate post](https://gigabit.agency/insights/how-much-does-geo-cost/). This one does not reprice it.
### What to do this week
1. **Keep the twenty category questions fixed.** Same wording in every assistant. No brand name in the prompt. The definition of the rate lives in the [citation-rate post](https://gigabit.agency/insights/what-is-a-good-ai-citation-rate/).
2. **Add a row per engine** before you calculate a pool. If ChatGPT is 12/20 and Perplexity is 0/20, write both. The **30%** pool is the summary, and in this example it is our arithmetic.
3. **Read who-else on the zero row.** If the same competitor is named on the queries that pay, that engine is the backlog. The [30-minute spreadsheet](https://gigabit.agency/insights/check-your-ai-visibility-in-30-minutes/) already has the column.
4. **Check exclusions only if every engine misses you.** Crawler, contradictory figures, and a split entity are in the [exclusion post](https://gigabit.agency/insights/what-gets-a-company-excluded-from-ai-answers/). A one-engine miss is not that list.
5. **Run the free** [AI Citation Audit](https://gigabit.agency/tools/geo-citation-audit/). It scores category buying questions across the major assistants and shows where you are named and who is named instead. Read the report by engine. Do not average it into a trophy.
ChatGPT and Perplexity disagree because they retrieve different pages. A pooled **30%** can be a **60%** on one assistant and a **0%** on the other. Score the rows. Fix the engine your buyers use. Start with the free [AI Citation Audit](https://gigabit.agency/tools/geo-citation-audit/).

### FAQ
**Why do ChatGPT and Perplexity disagree on the same query?**
They retrieve different pages. A brand can be named in one assistant and absent in the other on the same buying question. That split is ordinary. It is not, by itself, proof the brand is broken or that the zero engine is wrong. Score each assistant. A blended citation rate can read 30% while one engine is at zero.
**Should you average ChatGPT and Perplexity into one citation rate?**
Report each engine first. A pooled rate is a summary, not the decision. In a worked example — our arithmetic, not a client measurement — ChatGPT at 12 of 20 (60%) and Perplexity at 0 of 20 pool to 12 of 40, which is 30%. The unweighted average of the two rates is also 30% only because both panels have 20 queries. Either figure hides the zero.
**What citation rate should you act on when engines disagree?**
Act on the engine your buyers open. That priority is an operator assumption, not a measured share of pipeline. If prospects paste Perplexity answers into deals and Perplexity names you on none of the category queries, the ChatGPT rate is not the backlog. The site's panel definition stays the same: twenty category questions, six names is 30%, and brands with no GEO work often sit at 0–10%. Those benchmarks are per honest panel, not a license to drop the zero engine.
**Is a one-engine miss the same as being excluded from AI answers?**
No. Exclusion is when an assistant can find a page and still will not name you: the crawler never read the HTML, your own URLs disagree, or the entity does not resolve. A miss on one assistant and a hit on another is a coverage gap on that retriever. Fix exclusions only when every engine misses you. The free AI Citation Audit shows the split by assistant.

## AI security review vs a pentest: what each one covers
URL: https://gigabit.agency/insights/what-an-ai-security-assessment-covers/
Published 2026-10-19, updated 2026-10-22

You need an **AI security review** when a customer or an auditor has asked whether the AI you are about to ship is safe, and a network penetration test will not answer them. The fixed-price way to test that layer is the [AI Security Assessment](https://gigabit.agency/services/ai-security/): **from $12,000**, over **2–4 weeks**. It threat-models and red-teams the agents, large language models, and RAG pipelines you already run, then returns findings graded by severity and a remediation plan. A pilot that dies in a security review is a different article — [the five gates that stop a pilot](https://gigabit.agency/insights/why-your-ai-pilot-never-reached-production/). This post is one of those gates: what the assessment covers, what a pentest leaves out, and what you can hand the reviewer.
### A penetration test leaves the AI path untested
A traditional penetration test covers network, application, and infrastructure vulnerabilities. It finds an open port, a broken login, a missing patch. It does not walk an untrusted sentence through a model, into a tool, and out as a business action. Prompt injection, jailbreaks, RAG poisoning, tool-use abuse, and system-prompt leakage sit on that path. A clean pentest report can be true and still say nothing about them.
Most teams shipping AI need both tests. The pentest still belongs on the network and the app. The assessment belongs on the AI layer and the application the model can reach. Treating the pentest PDF as coverage of the model is how a feature ships, then fails the questionnaire that asks how you tested prompt injection.
The failure mode is a questionnaire row marked done because a pentest PDF exists. The reviewer who knows the OWASP LLM Top 10 asks for the path from untrusted input to a tool call. Without that path, the launch waits.
### Six layers, and the model is only the first
An AI feature inherits a new set of boundaries around the product you already run. Score all six. A finding on layer one with no look at layer three is a model demo, not an assessment.
| Layer | What has to hold |
| --- | --- |
| LLMs and prompts | Jailbreaks, system-prompt leakage, insecure output handling, and the guardrails that have to hold when a user gets creative |
| RAG pipelines | Poisoning, context manipulation, data leakage, and whether tenant identity survives every retrieval and inference boundary |
| AI agents and tool use | Agent identity, least-privilege authorisation, scoped credentials, tool allowlists, and a human checkpoint before a high-stakes action |
| Data and PHI | Every path patient or customer data takes, including model APIs and tracing tools that capture it in logs |
| AI supply chain | Model provenance, fine-tuning data, embeddings, and third-party dependencies a tampered component can enter through |
| Cloud, secrets, and identity | Secrets handling, API permissions, MCP servers, and the blast radius when one credential leaks |
Name only the model vendor and you have scoped one layer. The assessment starts from the other five: what is deployed, which systems it touches, and what done means for this review.
### The tests follow one real execution path
Adversarial testing here means one path, repeated until it breaks or it holds: untrusted input, then model context, then an agent decision, then a tool, then a business action. The cases below are the ones every deployment shares. They are derived from the OWASP LLM Top 10 and then aimed at your system, not scored as a generic benchmark.
- **Prompt injection, direct and indirect.** A user types an instruction, or a document, email, or web page the model reads hides one. The model obeys it. In healthcare the high-risk variant is an instruction sitting inside a clinical note.
- **Sensitive-data disclosure.** The model returns a secret, an internal detail, or another tenant's data.
- **RAG and embedding poisoning.** Tampered content in the knowledge base bends what the system retrieves, or carries data out through the retrieval path.
- **Excessive agency and tool misuse.** An agent with more access than the task needs is talked into turning "summarise this ticket" into "export the customer table and send it."
- **System-prompt leakage.** The hidden prompt — rules, logic, sometimes keys — is coaxed out, and the attacker has a map of the guardrails.
- **Insecure output handling.** Model output goes straight into a browser, a shell, or a database, and a response that carries a script or a query becomes an ordinary injection in your own app.
- **Model and supply-chain risk.** A backdoored model, a compromised dependency, or a third-party integration is reviewed for provenance and for what it can execute in production.
- **Cross-tenant isolation.** A user authenticated for one tenant reaches another's data once the request crosses an embedding, an inference call, an agent tool, or a downstream API.
Authorisation decides whether an intended action may run. An assessment that stops at the sentence, and never calls the tool, misses the step where the damage happens.
### What you hand the reviewer
The deliverable is a graded finding list and a prioritised remediation plan. If you want the fixes implemented, the same team that builds these systems can do that work. The assessment itself is the test and the plan. Implementation is a separate decision, named before it starts.
Reviewers ask for evidence, not a slogan. For healthcare the work is HIPAA-aware: data stays on your infrastructure, and a BAA is signed when the review requires one. For SaaS the work is on a SOC 2 path: audit logging, access control, and a trail a customer security questionnaire can point at. Neither phrase is a certification. We do not claim a HIPAA certification or a SOC 2 report from this engagement. We test and document to the controls the review expects, and we say what is actually in place.
The trail on an agent is four fields: input, tools called, data accessed, output. A reviewer who challenges one decision should be able to replay it. Alignment language on the report maps findings to the frameworks auditors cite — the OWASP LLM Top 10, NIST AI RMF, ISO 42001, and EU AI Act expectations. "Aligned with" means the findings are labelled in that vocabulary. It does not mean those bodies certified the system.
A published build shows what a pre-launch security pass looks like when it is not an AI red team. In month 5 of a [telehealth platform](https://gigabit.agency/proof/healthtech-telehealth-platform/), a third-party security assessment produced two medium findings, and both were remediated before launch. That was application and infrastructure security on a HIPAA platform. It was not a test of prompt injection or agent tool use. Use it as the pattern — findings graded, closed before go-live — and buy the AI-layer test separately when the system has a model in the path.
### What the $12,000 buys, and what stays on a retainer
The [AI Security Assessment](https://gigabit.agency/services/ai-security/) is **from $12,000**, fixed, over **2–4 weeks**. Larger or multi-system scopes are quoted before work starts; [what moves the price, and what other firms publish](https://gigabit.agency/services/ai-security/#cost), is on the service page. The number is named up front. There is no open meter hiding behind the floor.
Inside that window: a scoping call, an attack-surface map, adversarial testing along the real path, and a graded report with a remediation plan. Runtime guardrails, eval regressions, and monitoring sit on [Managed AI Operations](https://gigabit.agency/services/managed-ai-operations/) at **$3,000–$20,000 a month**. Prompts, tools, and models change after the report. The assessment fee does not include that retainer.
An unpatched host is a pentest. An agent that can call tools is this assessment. Where identity verification or specialist tooling is the right control, integrate it and say so.
### What to do this week
1. **List what is deployed.** Models, agents, retrieval stores, and the systems they can reach. If the list is one vendor name, you have not scoped the work.
2. **Mark the path.** Any place untrusted text — a user, a document, an email, a web page — can become a tool call or a write. That path is the test.
3. **Read the pentest against that path.** If the report never mentions prompt injection, retrieval, or tool use, it does not cover this review.
4. **Ask which control the reviewer will cite.** OWASP LLM Top 10, a SOC 2 control, a HIPAA safeguard, or a customer questionnaire row. The report should use that label.
5. **Price the test.** The [AI Security Assessment](https://gigabit.agency/services/ai-security/) is **from $12,000**, fixed, over **2–4 weeks**. Put monitoring on [Managed AI Operations](https://gigabit.agency/services/managed-ai-operations/) at **$3,000–$20,000 a month** after the findings are closed.
A penetration test answers a different question. An AI security assessment answers whether the model, the retrieval layer, and the tools can be talked into doing something your policy forbids — and it answers it at a fixed price, **from $12,000**. The [telehealth proof](https://gigabit.agency/proof/healthtech-telehealth-platform/) shows the shape of a pre-launch pass: two medium findings, both remediated before launch. That pass was not an AI red team. When the system you are shipping has an agent or a RAG pipeline in it, start with the [AI Security Assessment](https://gigabit.agency/services/ai-security/).

### FAQ
**What does an AI security review test that a pentest skips?**
An AI security assessment maps the AI you have deployed, builds a threat model for that architecture, and red-teams it. The tests follow one path: untrusted input, model context, agent decision, tool, business action. Coverage includes direct and indirect prompt injection, sensitive-data disclosure, RAG and embedding poisoning, excessive agency and tool misuse, system-prompt leakage, insecure output handling, model supply-chain risk, and cross-tenant isolation. You receive findings graded by severity and a prioritised remediation plan. The fixed-price entry is from $12,000 over 2–4 weeks.
**How is an AI security assessment different from a penetration test?**
A penetration test covers network, application, and infrastructure vulnerabilities. An AI security assessment tests the path a pentest does not walk: prompt injection, jailbreaks, RAG poisoning, tool-use abuse, and system-prompt leakage. Most teams shipping AI need both. A clean pentest report does not answer a reviewer who asks how you tested prompt injection.
**Does an AI security assessment make a system HIPAA or SOC 2 certified?**
No. The work is HIPAA-aware for healthcare and on a SOC 2 path for SaaS: audit trails on agent actions, data kept on your infrastructure, and NDAs or BAAs when the review requires them. Findings are labelled against the OWASP LLM Top 10, NIST AI RMF, ISO 42001, and EU AI Act expectations. That is alignment in the report, not a certification from those bodies. A published telehealth build had a third-party security assessment with two medium findings remediated before launch. That assessment was not an AI red team.

## n8n pricing in 2026: every plan, and what self-hosting really costs
URL: https://gigabit.agency/insights/n8n-pricing/
Published 2026-10-23

**n8n pricing** starts at nothing: the self-hosted Community edition is free, and you pay only for the server it runs on. n8n's hosted cloud starts at **€20 a month billed annually** for 2,500 workflow executions (Starter), then **€50** for 10,000 (Pro). The self-hosted **Business** plan is **€667 a month billed annually** for 40,000 executions, and **Enterprise** is quoted by n8n's sales team. Those are the prices on [n8n's pricing page](https://n8n.io/pricing/), checked on 23 October 2026. n8n bills per execution, which is one run of a whole workflow however many steps it has. That single fact is why n8n is often far cheaper than Zapier or Make for complex workflows, and the rest of this guide shows by how much.
### Every n8n plan and its price
n8n shows its prices in euros or US dollars depending on where you view the page. Starter and Pro carry the same number in both currencies; Business does not. Here is what each page showed on 23 October 2026, per month and billed annually.
| Plan | Euro page | Dollar page | Executions a month | Runs on |
| --- | --- | --- | --- | --- |
| Community edition | Free | Free | Unlimited (your server's limit) | Your server |
| Cloud Starter | €20 | $20 | 2,500 | n8n's cloud |
| Cloud Pro | €50 | $50 | 10,000 | n8n's cloud |
| Cloud Pro, larger tier | Not shown | $120 | 50,000 | n8n's cloud |
| Business | €667 | $800 | 40,000, with larger tiers | Your server only |
| Enterprise | Quoted | Quoted | Custom | Your server or n8n's cloud |
Billed monthly, the dollar page showed **$24** for Starter, **$60** for Pro and **$960** for Business. n8n says annual billing saves 17%. Every plan includes unlimited users and unlimited active workflows, so you are never paying per seat.
### What each tier adds
**Starter** is for a first workflow: 5 concurrent executions, one shared project and 1,600 Assistant credits a month for n8n's AI building assistant. **Pro** adds admin roles, workflow history, evaluations and up to 9,600 Assistant credits; concurrency is 20 on the 10,000 tier and 50 on the 50,000 tier, and anything over the limit waits in a queue. **Business** is self-hosted only and adds SSO and LDAP, multiple environments, Git version control and six shared projects. **Enterprise** adds external secret stores, log streaming, extended data retention, 200 or more concurrent executions and support with an SLA, on your server or n8n's.
Two discounts are worth knowing. Companies with fewer than 20 employees can apply for n8n's Start-up Plan, which takes 50% off Business. And there is a free cloud trial with 1,000 executions and no card, plus a 14-day trial of Business.
### n8n enterprise pricing and overages
n8n does not publish Enterprise prices; you get a quote. It does publish how overages work on Business and Enterprise: workflows keep running when you pass your quota, the sales team contacts you, and unresolved overages are invoiced 45 days later. On Business the overage price is **€4,000 for each extra bucket of 300,000 executions**. The dollar page's Business tiers run from $800 for 40,000 executions to $5,245 for 500,000, billed annually.
### What an execution is, and why it changes the maths
n8n's definition is short: an execution is a single run of your entire workflow, and it does not matter how many steps it has or how much data it processes. A workflow that receives an order, loops over 40 line items, calls three systems and an AI model, then posts to Slack, is one execution.
The other two big tools count differently. [Zapier](https://zapier.com/pricing) bills **tasks**: every successful action step is one task, while the trigger, failed actions and built-in tools such as Filter, Formatter, Paths and Looping are free. An action inside a loop costs a task on every pass. [Make](https://www.make.com/en/pricing) bills **credits** (it replaced operations with credits in August 2025, one for one): almost every module action costs a credit, including the trigger reading data, while routers and error handlers are free, and Make's own AI Provider can cost more than one credit per action.
So the same workflow converts into very different numbers. A five-step workflow (one trigger, four actions) that runs once is 1 n8n execution, 4 Zapier tasks and 5 Make credits. Add a loop and the gap widens with every item.
### Worked examples: what n8n costs at 5,000 and 50,000 runs
Three workloads, priced on each vendor's page on 23 October 2026, per month and billed annually. The conversions into tasks and credits, and the tier each lands on, are our arithmetic. n8n figures are the same in euros and dollars except where marked.
| Workload | n8n | Zapier | Make |
| --- | --- | --- | --- |
| 5,000 runs of a five-step workflow | 5,000 executions: Cloud Pro, **50** | 20,000 tasks: Professional, **$189** | 25,000 credits: Core on the 40,000 tier, **$29** |
| 50,000 runs of the same workflow | 50,000 executions: Cloud Pro, **$120** | 200,000 tasks: Professional, **$769** | 250,000 credits: Core on the 300,000 tier, **$182.16** |
| 2,000 orders a month, 20 line items each, two actions per item | 2,000 executions: Cloud Starter, **20** | About 82,000 tasks: Professional on the 100,000 tier, **$489** | At least 86,000 credits: Core on the 150,000 tier, **$99** |
### Example 1: a team running 5,000 executions a month
Starter's 2,500 executions are too few, so this team is on **Pro at €50 a month billed annually**, or $60 billed monthly. The same workload is $189 a month on Zapier and $29 on Make Core. At this size Make is the cheapest hosted option. If you choose n8n here, it is for its code steps, AI nodes or the option to self-host later.
### Example 2: a team running 50,000 executions a month
On the cloud, **Pro's 50,000 tier is $120 a month billed annually**. Business, at 40,000 executions, is too small, and its 50,000 tier is $985 a month on the dollar page, so Business is a purchase for SSO, Git and environments. The same workload is $769 a month on Zapier and $182.16 on Make Core. Self-hosted on the free Community edition, n8n's licence costs nothing and the bill is your hosting.
### Example 3: loops
The third row is where per-execution billing matters most. Each order is one n8n execution, so 2,000 orders fit inside Starter at 20 a month. On Zapier each of the 40 line-item actions is a task, so one order costs 41 tasks and the month lands on the 100,000-task tier at $489. On Make it costs at least 43 credits per order, which lands on the 150,000-credit tier at $99. Count your loop passes before you compare plans.
### The real cost of self-hosting n8n
The Community edition costs nothing to license. Running it costs four things: a server, a database, backups and somebody's time.
**Server and database.** As one example of list prices on 23 October 2026, a [DigitalOcean](https://www.digitalocean.com/pricing/droplets) droplet with 4 GB of memory and 2 vCPUs is $24 a month, weekly backups add 20% of the droplet price, and a [managed PostgreSQL](https://www.digitalocean.com/pricing/managed-databases) database starts at $15.15 a month. By our arithmetic, a single production instance is therefore about **$44 a month** in infrastructure.
**Scaling.** n8n stores data in SQLite by default. For volume you run [queue mode](https://docs.n8n.io/hosting/scaling/queue-mode/): a main instance receives triggers, worker instances run executions, Redis passes jobs between them, and PostgreSQL holds the data, because n8n does not support queue mode on SQLite. A main instance and two workers at $24 each, a $12 droplet for Redis, backups and the managed database come to roughly **$100 a month** by our arithmetic. One catch: in queue mode, workflows that handle files need S3 external storage, and [n8n lists external binary storage](https://docs.n8n.io/hosting/community-edition-features/) as a paid feature.
**What the free edition leaves out.** Community has no SSO, no Git version control, no environments, no external secrets, no log streaming beyond basic logging, and no projects. Workflows and credentials can be seen only by the instance owner and the person who created them. Registering for a free key unlocks folders, debugging in the editor and custom execution data.
**Someone's time.** This is the real line item. Upgrades should be tested on a copy before production, the encryption key for stored credentials must be backed up, execution data needs pruning, and an error workflow should alert a person. n8n's own guidance is plain that self-hosting needs technical expertise and that you provide and maintain the infrastructure. As an illustration, if an engineer spends four hours a month on it at $100 an hour, that is $400 a month, more than Cloud Pro. Self-host when volume, data control or custom nodes justify that time, or when the time is already paid for.
### What the licence lets agencies and SaaS products do
n8n is fair-code under its Sustainable Use License, and the free licence covers self-hosted use only. The [licence FAQ](https://docs.n8n.io/n8n-community-license/community-license/license-faq) draws the line by who controls the workflow logic.
**Allowed:** using n8n for your own business, personal projects and research; an agency or consultant building, running and maintaining workflows for a client and charging for that work, as long as the client cannot create or edit workflows; and using n8n behind the scenes in your own product, where users trigger automations you built.
**Not allowed without a commercial licence:** hosting n8n as a service where clients build their own workflows; letting outside users build or configure workflows through your product, which rules out an embedded workflow builder in a SaaS app; forking n8n to launch your own automation product; white-labelling it; and using enterprise features that ask for a licence key. For those, n8n points you to its sales team for an Enterprise licence.
### When to pay Gigabit, and when not to
Do not pay anyone if one person can build what you need on Starter or Pro and keep it running. n8n is designed for that. Pay for help when a workflow touches money, stock or customers and has to keep working when an API changes at 2 a.m. That is a build problem more than a licence problem: retries, idempotent steps, validation between stages, alerts and a human approval where a mistake is expensive.
- **One workflow:** our [AI Quick Win Sprint](https://gigabit.agency/platforms/n8n/) is $8,000 fixed over 3 weeks. One workflow scoped, built and deployed on your instance, with up to three tool integrations, error handling and 30 days of light support.
- **An AI agent in a workflow:** from $8,000, flat per agent.
- **Several workflows, a migration from Zapier or Make, or a self-hosted set-up:** an embedded pod at $3,000–$7,500 per engineer per month.
- **Running them afterwards:** Managed AI Operations from $3,000 a month for one system, covering monitoring, fixes and vendor-change response.
We build on your n8n instance under your licence, so the workflows and the relationship with n8n stay yours. If you are still choosing a tool, our [n8n vs Zapier vs Make comparison](https://gigabit.agency/compare/n8n-vs-zapier-vs-make/) puts the three side by side, and our [Zapier and Make page](https://gigabit.agency/platforms/zapier-make/) covers when to stay where you are. All of our prices are on the [pricing page](https://gigabit.agency/pricing/).

### FAQ
**How much does n8n cost?**
Self-hosted, the Community edition is free and you pay for the server. On n8n's cloud, Starter is €20 a month billed annually for 2,500 executions and Pro is €50 for 10,000. The self-hosted Business plan is €667 a month billed annually for 40,000 executions on the euro page ($800 on the dollar page), and Enterprise is quoted. Prices checked on n8n.io/pricing on 23 October 2026.
**Is n8n free?**
The self-hosted Community edition is free under n8n's Sustainable Use License for your own business use, and an agency may build and maintain workflows on your instance. It lacks SSO, Git version control, environments, external secrets and sharing of workflows and credentials. n8n Cloud has no free plan, only a trial with 1,000 executions and no card required.
**What counts as an execution in n8n?**
One run of an entire workflow, from trigger to finish, however many steps it has or how much data it processes. A workflow that loops over 40 items and calls five systems is still one execution. Zapier, by contrast, bills each action step as a task, and Make bills almost every module action as a credit.
**How much does it cost to self-host n8n?**
The licence is free. As an example, using DigitalOcean list prices on 23 October 2026, a single instance with a managed Postgres database and weekly backups is about $44 a month by our arithmetic, and a queue-mode set-up with two workers and Redis about $100. The larger cost is the engineer time for upgrades, backups and monitoring, which is why low-volume teams are often better off on n8n Cloud.
**Is n8n cheaper than Zapier?**
For workflows with several steps or loops, usually by a wide margin. By our arithmetic on 23 October 2026 prices, 50,000 runs a month of a five-step workflow cost $120 a month on n8n Cloud Pro and $769 on Zapier Professional, both billed annually. For a few two-step Zaps, Zapier's entry plans can cost less, and its app catalogue is larger.

## Webflow pricing in 2026: plans, seats, add-ons and what a site really costs
URL: https://gigabit.agency/insights/webflow-pricing/
Published 2026-10-23

**Webflow pricing** in 2026 has two layers. Each site needs a site plan: Starter is free, Basic is **$15 a month** and Premium is **$25 a month** when billed yearly, or $25 and $39 billed monthly. Then each person who edits needs a seat: **$39 a month** for a Full seat that designs, **$15** for a Limited seat that edits content, and nothing for reviewers who only comment. Team is **$2,500 a month** on an annual contract and Enterprise is quoted. Languages, extra bandwidth and analytics are add-ons. All figures are US dollars per site, before tax, from [Webflow's pricing page](https://webflow.com/pricing), checked on 23 October 2026.
### Webflow site plans in 2026
A site plan pays for hosting and unlocks features on one site. The figures below are per month.
| Plan | Billed yearly | Billed monthly | What it is for |
| --- | --- | --- | --- |
| Starter | Free | Free | A webflow.io address, 2 static pages, 1 GB bandwidth, 50 form submissions, 50 CMS items |
| Basic | **$15** | $25 | A custom domain, 300 static pages, 10 GB bandwidth, unlimited form submissions. **No CMS** |
| Premium | **$25** | $39 | Everything in Basic plus the CMS (20,000 items in 40 collections), 50 GB bandwidth that can be raised to 2.5 TB, code components and site search |
| Team | **$2,500** | Annual contract only | Premium plus 5 Full and 5 Limited seats, Localize with 2 locales, publishing workflows, page branching, priority support |
| Enterprise | Quoted | Quoted | Custom limits, SSO, custom roles, a 99.99% hosting uptime agreement |

The jump that matters for most companies is Basic to Premium. Basic has no CMS at all, so the day you want a blog, a case-study grid or a job board, you are on Premium. Starter is for building and testing; it cannot use your own domain.
### What changed in 2026: goodbye CMS and Business plans
Webflow [announced in its community](https://community.webflow.com/updates/post/updates-to-webflow-plans-and-pricing-mmTMW66ihrRXZ0o) that the old **CMS** and **Business** site plans are merged into one **Premium** plan, and added the **Team** plan for companies that have outgrown self-serve but do not need Enterprise. CMS limits went up and CMS item add-ons were removed. New purchases moved to the new plans straight away. Most existing sites move at their next renewal or billable change on or after **29 June 2026**; sites in Freelancer or Agency Workspaces and sites on legacy pricing move on or after **16 November 2026**. Switching to annual billing before your date keeps your current plan for one more year.
If you are reading an older guide that quotes a CMS plan or a Business plan price, it describes plans you can no longer buy.
### Workspace plans and seat pricing
Every site lives in a Workspace, and the Workspace is where people and seats are billed. Every account gets a free Starter Workspace with **one Full seat** for the owner. Anyone else who edits needs a seat of their own.
| Seat | Billed yearly | Billed monthly | What the person can do |
| --- | --- | --- | --- |
| Full | **$39** | $45 | Design whole sites and manage admin settings |
| Limited | **$15** | $19 | Edit content and build pages from approved components |
| Free | $0 | $0 | Review and comment, no edit access |

Paid Workspace plans add staging capacity and collaboration controls, not seats. **Core** is $19 a month yearly ($28 monthly) and adds 10 staging sites, custom code on staged sites, code export and Shared Libraries. **Growth** is $49 yearly ($60 monthly) and adds unlimited staging, site-level roles and publishing permissions. Freelancers and agencies get their own versions: **Freelancer** at $16 and **Agency** at $35 a month yearly, with free client seats per site.
Webflow AI runs on credits pooled across the Workspace: 200 a month on Starter, 300 on Core and 400 on Growth. Another 2,000 credits a month cost **$20** billed yearly or $25 monthly.
### CMS items, bandwidth and the auto-upgrade rule
Premium includes **20,000 CMS items** in **40 collections**. The free Starter plan allows 50 items in 20 collections, enough to build a blog template and nothing more. Team keeps the 20,000-item allowance but raises collections to 100 and sells extra items for a fee. Above that you are on Enterprise.
Bandwidth is where the bill grows. Starter includes 1 GB a month, Basic 10 GB and Premium 50 GB. On Premium you can buy more in steps, for example:
| Extra bandwidth on Premium | Billed yearly | Billed monthly |
| --- | --- | --- |
| +50 GB (100 GB total) | +$20 a month | +$30 a month |
| +150 GB (200 GB total) | +$60 | +$90 |
| +450 GB (500 GB total) | +$174 | +$260 |
| +950 GB (1 TB total) | +$374 | +$560 |
| +2.45 TB (2.5 TB total) | +$974 | +$1,460 |

Paid plans include surge protection for short spikes. The rule to plan around is in the plan table: **two consecutive months over the limit triggers an automatic plan upgrade**. One viral month is covered; a site that has simply grown will be moved up and billed for it. Webflow Cloud apps hosted beside the site draw on the same bandwidth pool.
### Webflow Localize pricing
Translation is an add-on on Basic and Premium, priced per added locale (your primary language is free).
| Localize plan | Per locale, billed yearly | Billed monthly | Limits |
| --- | --- | --- | --- |
| Essential | **$9** | $12 | Up to 3 locales, 10,000 AI-translated words per locale a month |
| Advanced | **$29** | $35 | Up to 10 locales, 50,000 words per locale a month, localized URLs, per-locale images, automatic visitor routing |

Both plans translate CMS and static content, localize SEO titles and descriptions, serve locales from subdirectories such as /es/ and add hreflang tags to the sitemap. Team includes two locales; Enterprise sets its own limit.
### Webflow ecommerce pricing and transaction fees
A store needs an Ecommerce plan in place of a regular site plan.
| Ecommerce plan | Billed yearly | Billed monthly | Products | Webflow fee |
| --- | --- | --- | --- | --- |
| Standard | **$29** | $42 | 500 | **2%** per transaction |
| Plus | **$74** | $84 | 5,000 | 0% |
| Advanced | **$212** | $235 | 15,000 | 0% |

The Webflow fee is charged on top of Stripe or PayPal processing fees. Webflow lists no cap on sales volume on any of the three. Ecommerce plans carry smaller CMS allowances than Premium (2,000 items on Standard, up to 10,000 on Plus and Advanced) and 50 to 100 GB of bandwidth.
The 2% fee decides the plan. Plus costs $45 a month more than Standard on yearly billing, and 2% of **$2,250** is $45, so once a store sells more than about $2,250 a month, Plus is the cheaper plan. That is our arithmetic on Webflow's published prices.
### Three worked examples
These use yearly billing and prices checked on 23 October 2026. They leave out tax, your domain and any paid apps.
### A five-person marketing team with a blog
One designer owns the Workspace, a second designer builds pages, and three marketers write posts and update copy.
| Line | Per month |
| --- | --- |
| Premium site plan | $25 |
| Owner's Full seat | Included |
| Second Full seat | $39 |
| Three Limited seats (3 × $15) | $45 |
| **Total** | **$109 a month, $1,308 a year** |

Billed monthly the same team costs $39 + $45 + $57 = **$141 a month**. If the blog stays under 20,000 posts and the site under 50 GB of traffic, nothing else is required. Add Core ($19) if you want code export or a shared component library across sites.
### A site in three languages
An English site translated into German and French, edited by one designer and one marketer.
| Line | Per month |
| --- | --- |
| Premium site plan | $25 |
| Localize Essential, 2 locales (2 × $9) | $18 |
| One Limited seat | $15 |
| **Total** | **$58 a month, $696 a year** |

If you need translated URL slugs, different images per language or automatic routing by visitor language, Advanced costs $58 for the two locales in place of $18, which brings the total to **$98 a month**. Machine translation is a first draft; budget for a native speaker to review it.
### A small store
A shop with 200 products, run by an owner with one Limited seat for a part-time assistant.
| Line | Per month |
| --- | --- |
| Ecommerce Standard | $29 |
| One Limited seat | $15 |
| 2% Webflow fee on $5,000 of monthly sales | $100 |
| **Total before Stripe or PayPal fees** | **$144** |

On Plus the same month is $74 + $15 with no Webflow fee, **$89**. At $5,000 a month in sales, Plus saves $55 a month. Webflow suits a store where the brand and content lead and the catalog is small. For subscriptions, wholesale, many sales channels or a large catalog, a commerce platform is the better fit; see [our Shopify work](https://gigabit.agency/platforms/shopify/).
### The costs no pricing page shows
The subscription is usually the smallest number in a Webflow budget. Three others decide it.
**Building the site.** Design, the CMS structure, templates, forms wired into your CRM, analytics and redirects from the old site. Our [Website Build](https://gigabit.agency/pricing/) starts at **$6,000 fixed** for up to eight page templates, from 4 weeks. A template from the Webflow marketplace is cheaper and fine for a simple site, as long as someone adjusts it to your content rather than the reverse.
**Moving to Webflow.** Content has to be mapped to collections, every old URL redirected and metadata carried over, or search traffic drops after launch. A [platform migration](https://gigabit.agency/pricing/) with us starts at **$7,000 fixed**, including the redirect map, a rehearsed cutover and 30 days of monitoring.
**Keeping it.** Webflow runs the servers, so there are no plugin or server updates. What remains is change: new page types, custom code components, integrations and the occasional redesign. Teams without a developer either buy that as projects or keep an engineer on hand; our embedded pods are $3,000–$7,500 per engineer per month.
Leaving has a cost too. Paid Workspaces can export HTML and CSS, but Webflow's pricing page says CMS items and pages are left out of the code export (collections come out separately, one at a time), forms stop working after export and exported code cannot be imported back. Moving off Webflow later means rebuilding the CMS elsewhere.
### When to pay Gigabit, and when not to
Do it yourself when the site is a handful of pages on Basic, a designer on your team already knows Webflow, or a good template covers 90% of what you need. The plan prices above are the whole cost in that case.
Bring us in when the site has a real CMS model (case studies, locations, a resource library), several languages, forms that must land in HubSpot or Salesforce, or custom code components. We build on [Webflow](https://gigabit.agency/platforms/webflow/) from $6,000 and move sites onto it from $7,000. If you are still choosing a platform, read [Webflow vs WordPress](https://gigabit.agency/compare/webflow-vs-wordpress/) and [Framer vs Webflow](https://gigabit.agency/compare/framer-vs-webflow/) first; we build on all three and will say when Webflow is the wrong pick. For a store that has outgrown Webflow Ecommerce, a [Shopify store build](https://gigabit.agency/platforms/shopify/) starts at $9,000.
Webflow changes its plans often. Check [webflow.com/pricing](https://webflow.com/pricing) before you sign, and count seats and bandwidth along with the site plan.

### FAQ
**How much does Webflow cost per month?**
Per site, billed yearly: Starter is free, Basic is $15 a month with no CMS, and Premium is $25 a month with the CMS. Billed monthly, Basic is $25 and Premium is $39. Each person who edits needs a seat on top: $39 a month for a Full seat, $15 for a Limited seat, free for reviewers. Team is $2,500 a month on an annual contract. Prices are from webflow.com/pricing, checked on 23 October 2026.
**What happened to the Webflow CMS plan and Business plan?**
Webflow merged the CMS and Business site plans into one Premium plan in 2026 and added a Team plan above it. Premium includes 20,000 CMS items in 40 collections and 50 GB of bandwidth, extendable to 2.5 TB. Existing sites move to the new plans at their next renewal or billable change on or after 29 June 2026, or 16 November 2026 for Freelancer and Agency Workspaces and legacy-priced sites.
**How does Webflow seat pricing work?**
Seats are billed per Workspace. The owner's Full seat is included. Additional Full seats, for people who design sites or manage settings, are $39 a month billed yearly or $45 monthly. Limited seats, for people who edit content or build pages from components, are $15 yearly or $19 monthly. Free seats let reviewers comment without editing. Team plans include 5 Full and 5 Limited seats.
**What happens if a Webflow site goes over its bandwidth or CMS limit?**
Paid plans include surge protection for short spikes, but two consecutive months over the bandwidth limit trigger an automatic plan upgrade. On Premium, extra bandwidth is sold in steps from +50 GB for $20 a month billed yearly up to +2.45 TB for $974. Premium's CMS limit is 20,000 items; Team sells more items for a fee, and above that is Enterprise.
**How much are Webflow's ecommerce plans and fees?**
Billed yearly, Standard is $29 a month for 500 products with a 2% Webflow transaction fee, Plus is $74 for 5,000 products with no Webflow fee, and Advanced is $212 for 15,000 products with no fee. Billed monthly they are $42, $84 and $235. Stripe or PayPal processing fees apply on top. Above roughly $2,250 a month in sales, Plus costs less than Standard.

## Odoo pricing in 2026: plans, hosting, implementation and what a rollout really costs
URL: https://gigabit.agency/insights/odoo-pricing/
Published 2026-10-23

**Odoo pricing** in the US, checked on 23 October 2026 on [odoo.com/pricing](https://www.odoo.com/pricing): One App Free costs nothing for one app and unlimited users. **Standard** is **US$31.10 per user per month billed yearly** (US$38.90 billed monthly) and covers every app on Odoo's cloud. **Custom** is **US$61.00 per user per month billed yearly** (US$76.20 monthly) and adds Studio, multi-company, the external API and the choice of Odoo.sh or your own servers. New subscriptions get 20% off for the first 12 months, on the users ordered at the start. Odoo Community, the open-source edition, has no licence fee at all. The licence is rarely the biggest line, though. Implementation, data migration and upgrades usually cost more in year one, and this guide prices those too.
Odoo sets prices by country, so check the page from where you will buy. The same page viewed from Bangladesh showed Standard at US$8.95 and Custom at US$16.40, billed yearly. Every figure below is the US price.
### Odoo editions and plans
Odoo ships as two editions of the same software. **Community** is open source under the LGPLv3 licence ([source](https://github.com/odoo/odoo)), free to download, run and change. **Enterprise** is the licensed edition, and the three plans on the pricing page are ways of buying it.
| Plan | US price per user per month | What you get |
| --- | --- | --- |
| Community | $0 (self-hosted) | Open-source core: CRM, sales, inventory, manufacturing, eCommerce, invoicing. No full accounting, Studio, payroll, mobile apps, Odoo support or upgrade service |
| One App Free | $0 | One Enterprise app (plus the apps it depends on) for unlimited users, on Odoo Online |
| Standard | $31.10 yearly, $38.90 monthly | All apps, on Odoo Online only |
| Custom | $61.00 yearly, $76.20 monthly | All apps, Studio, agentic AI, multi-company, external API; Odoo Online, Odoo.sh or on-premise |

First-year prices are 20% lower: $24.90 yearly or $31.10 monthly on Standard, $49.00 or $61.10 on Custom. Plan for the list price from month 13. Odoo also lists a **Light User** at US$8.90, for employees who only log time off, expenses, attendance and similar self-service tasks; you get it through an Odoo advisor.
The Community feature list comes from Odoo's [editions comparison](https://www.odoo.com/page/editions). The gaps that matter most: Community has invoicing but not the full accounting app, and no barcode, PLM, payroll or Studio. That is why most companies that want their books in Odoo end up on Enterprise.
### Who counts as a paying user
A paying user is anyone with backend access who creates, views or edits documents, plus any employee record with no linked user. Customers and suppliers who use the portal to see invoices or track orders are free, as are website shoppers. An employee who only submits expenses or timesheets is a paying user, unless you license them as a Light User.
### What Custom adds over Standard
Odoo describes Custom as the plan for companies that run several companies in one database or customise Odoo through Studio, agentic AI, custom development or the API. In practice four things push you onto it:
1. **Odoo Studio**, the no-code tool for custom fields, screens, reports and automations.
2. **Multi-company** in one database. Adding a second company moves you to Custom automatically.
3. **External API.** Outside software calling Odoo's JSON-RPC or XML-RPC API counts. Calls Odoo makes outward, to Stripe, a carrier or a bank feed, stay on Standard.
4. **Hosting choice.** Standard runs only on Odoo Online. Custom can also run on Odoo.sh or on your own servers, and only those accept custom Python modules.
That last point decides more budgets than the per-user price does. If you need one custom module, you need Custom, and the licence roughly doubles.
### Hosting options and what they cost
**Odoo Online** is included in every plan, with storage unmetered up to 100 GB. Odoo runs the servers, backups and upgrades. The catch is in [Odoo's own documentation](https://www.odoo.com/documentation/20.0/administration/odoo_online.html): Odoo Online is incompatible with custom modules and with modules from the Odoo Apps Store.
**Odoo.sh** is Odoo's platform for hosting Enterprise with custom code, with GitHub integration, staging branches, shell access and backups in three locations. It is priced separately from the licence. [Odoo.sh pricing](https://www.odoo.sh/pricing), checked on 23 October 2026:
| Odoo.sh item | Billed yearly (per month) | Billed monthly |
| --- | --- | --- |
| Worker | $57.60 each | $72.00 each |
| Storage | $0.20 per GB | $0.25 per GB |
| Staging environment | $14.40 each | $18.00 each |
| Dedicated server | $480 | $600 |

The smallest set-up, one worker and 1 GB, is $57.80 a month billed yearly. Odoo recommends one extra worker per 25 backend users and about 1 GB of storage per user to start.
**On-premise** means downloading Enterprise and running it on your own servers or cloud account. Odoo charges nothing for hosting here, so the cost is whatever your host charges plus the time to run it: patches, backups you have actually restored, monitoring. **Community** is always self-hosted, at the same running cost and with no licence fee.
### Implementation cost
Odoo's pricing page is clear that implementation is not included. It also says who usually does it: companies under 50 employees tend to buy **Success Packs** from Odoo, companies above 50 usually work with a partner, and companies over 250 should talk to Odoo directly.
[Success Packs](https://www.odoo.com/pricing-packs) are blocks of consultant hours, US prices checked on 23 October 2026:
| Pack | Consultant hours | Price | New customer (15% off first pack) |
| --- | --- | --- | --- |
| Starter | 4 | $580 | $493 |
| Basic | 25 | $3,600 | $3,060 |
| Standard | 50 | $7,000 | $5,950 |
| Custom | 100 | $12,500 | $10,625 |
| Pro | 200 | $25,000 | $21,250 |

The hours cover analysis, configuration, training and data import help. Odoo recommends the Custom and Pro packs when customisation is involved, and notes that extra monthly charges apply to maintain extra modules or lines of code. For a small company with standard processes, a Success Pack bought straight from Odoo is often the cheapest sensible route, and you do not need a partner.
A partner makes sense when the work runs past configuration: several apps that must agree with each other, data from an old ERP, integrations with a store or a warehouse system, or custom modules. Gigabit is an Odoo Partner. Our [Odoo implementation](https://gigabit.agency/platforms/odoo/) is **from $15,000 fixed**, from 8 weeks, delivered in phases: process mapping, configuration of up to four core apps, data migration, user training and go-live support. Manufacturing, multi-company set-ups and custom modules are quoted higher, as a fixed price before work starts. Odoo's licence is separate and paid to Odoo. One integration on its own, such as store orders into Odoo with stock synced back, is $8,000 fixed over 3 weeks.
### The costs beyond the licence
Four lines catch people out, roughly in this order.
**Data migration.** Customers, products, open orders, stock levels and opening balances have to be cleaned and mapped before they load. It is the phase most often under-scoped, and it is where a cheap quote usually gets expensive.
**Customisation.** Studio changes are cheap to make and Odoo upgrades them for you while Studio stays installed and subscribed. Custom Python modules are more powerful and cost more to own. Write every change as a proper module; databases edited by hand are the ones that cannot be upgraded later.
**Training.** Budget time for the people who will run Odoo every day, plus a person inside the business who owns it after go-live.
**Upgrades.** Odoo ships one major version a year, and each gets three years of standard support ([supported versions](https://www.odoo.com/documentation/20.0/administration/standard_extended_support.html)); Odoo 20 came out in September 2026. On Odoo Online a major-version upgrade is mandatory every two years ([upgrade policy](https://www.odoo.com/documentation/20.0/administration/upgrade.html)). Upgrading standard apps and Studio work is included in the Enterprise subscription. Upgrading custom modules is not, unless they sit under a maintenance contract, and a database with custom modules cannot be upgraded until those modules are ported to the new version. Plan for that work every year or two.
Our [ERP implementation cost guide](https://gigabit.agency/insights/erp-implementation-cost/) breaks these lines down across platforms. Custom modules and version upgrades after go-live are an embedded pod with us at $3,000–$7,500 per engineer per month.
### Worked examples
These are our arithmetic on the US list prices above, checked on 23 October 2026. Taxes are extra.
### 10 users on Standard, Odoo Online
A distributor running sales, inventory, purchasing and accounting with no custom code.

- Licence, billed yearly: 10 × $24.90 × 12 = **$2,988** in year one, then 10 × $31.10 × 12 = **$3,732** a year.
- Billed monthly instead: $3,732 in year one, then $4,668 a year.
- Hosting: $0 on Odoo Online.
- Implementation: Odoo's Standard Success Pack, 50 hours, $5,950 as a first pack.
- Year one is about **$8,940**; year two about **$3,730**.
This is the case where Odoo is hard to beat on price, and where buying hours from Odoo directly is usually enough.
### 25 users on Custom with Odoo.sh
A manufacturer with two companies in one database, a store integration and two custom modules.

- Licence, billed yearly: 25 × $49.00 × 12 = **$14,700** in year one, then 25 × $61.00 × 12 = **$18,300** a year.
- Odoo.sh, billed yearly: two workers ($115.20), 25 GB ($5.00) and one staging environment ($14.40) = $134.60 a month, about **$1,615** a year.
- Implementation: a partner project. Ours starts at $15,000; multi-company and custom modules put a project like this above that from-price.
- Year one is at least **$31,315** before the custom modules are priced; year two is about $19,900 plus upkeep of the custom code.
### Community, self-hosted
A workshop with its own developer runs inventory and manufacturing on Community. The licence is $0. The costs are a server, the developer's time, and every upgrade, which Odoo's upgrade service does not cover for Community. Accounting stays in another tool, because Community has invoicing only. This works when you have the engineer. Without one it is the most expensive option of the three.
### How to keep the bill down
Configure before you customise. Every Studio field you use instead of a custom module is code you never have to port. Count real backend users; portal customers are free, and Light Users cost far less than full ones. Bill yearly if you are sure of the headcount, because monthly billing costs 25% more. And pick hosting last: if nothing needs custom code, Odoo Online on Standard is the cheapest Enterprise there is.
Weighing Odoo against bigger systems? See [Odoo vs NetSuite](https://gigabit.agency/compare/odoo-vs-netsuite/) for the comparison with Oracle's cloud ERP, and [Odoo vs Zoho vs ERPNext](https://gigabit.agency/compare/odoo-vs-zoho-vs-erpnext/) for the open-source and suite alternatives.
When you know which apps, how many users and what data has to move, the rest of the budget can be priced. Our [Odoo implementation](https://gigabit.agency/platforms/odoo/) starts at $15,000 fixed, with Odoo's licence paid to Odoo. If Odoo's own Success Pack will do the job, we will say so on the first call.

### FAQ
**How much does Odoo cost per user?**
In the US, checked on 23 October 2026: Standard is US$31.10 per user per month billed yearly or US$38.90 billed monthly, and Custom is US$61.00 or US$76.20. New subscriptions get 20% off for the first 12 months on the users ordered at the start, so year one on Standard is US$24.90 per user per month billed yearly. Odoo sets prices by country, so the figure on odoo.com/pricing where you are may differ.
**Is Odoo free?**
Two ways. Odoo Community is open source under LGPLv3 with no licence fee; you host it yourself and it lacks the full accounting app, Studio, payroll and Odoo's support and upgrade service. One App Free gives you one Enterprise app, plus the apps it depends on, for unlimited users on Odoo Online at no charge. Installing a second app moves you to the paid Standard plan.
**What is the difference between Odoo Standard and Custom?**
Standard covers all apps on Odoo Online. Custom adds Odoo Studio, agentic AI, several companies in one database, outside software calling Odoo's API, and the option to host on Odoo.sh or your own servers, which is the only way to run custom Python modules. In the US, Custom costs US$61.00 per user per month billed yearly against US$31.10 for Standard.
**How much does an Odoo implementation cost?**
Odoo's own Success Packs run from US$580 for 4 hours to US$25,000 for 200 hours of consultant time, with 15% off a new customer's first pack (US prices, 23 October 2026). Odoo suggests them for companies under 50 employees. Larger or more complex projects usually go to a partner. As an Odoo Partner, Gigabit's fixed-price Odoo implementation starts at $15,000, with the licence paid separately to Odoo.
**How much does Odoo.sh cost?**
Odoo.sh is charged on top of the Custom licence. Billed yearly, it is US$57.60 per worker per month, US$0.20 per GB of storage and US$14.40 per staging environment, so the smallest set-up of one worker and 1 GB is US$57.80 a month. Billed monthly, a worker is US$72. Dedicated hosting starts at US$480 a month billed yearly. Prices checked on odoo.sh/pricing on 23 October 2026.

## Composable commerce: what it is, what it costs, and when it pays
URL: https://gigabit.agency/insights/composable-commerce/
Published 2026-10-23

**Composable commerce is a way of building an online store from separate services that talk to each other through APIs**: one for the storefront, one for cart and checkout, one for the product catalog or PIM, one for search, one for content (the CMS) and one for orders and fulfilment (the OMS). Each piece can come from a different vendor and be replaced without rebuilding the rest. The opposite is a single platform such as Shopify or BigCommerce running all of those jobs in one product, which is what most stores use and what most stores should keep using.
commercetools, one of the vendors that sells it, describes composable commerce as building commerce platforms from "independent, best-of-breed components" ([commercetools](https://commercetools.com/composable-commerce)). Shopify's own definition is close: composable commerce lets a business "assemble independently replaceable services" ([Shopify](https://www.shopify.com/enterprise/blog/headless-commerce)). The rest of this guide covers how it differs from headless, what MACH means, which parts make up a stack, what it costs to run, and where the line sits between worth it and overkill.
### Composable vs headless vs monolithic
The three terms describe how much of the store is split apart.
| Architecture | What is separate | Example | Who keeps it running |
| --- | --- | --- | --- |
| Monolithic | Nothing. Storefront, checkout, catalog and admin are one product | Shopify with a Liquid theme | Mostly the platform vendor |
| Headless | The storefront only. One commerce engine behind it, reached through its API | Shopify Hydrogen on the Storefront API | You run the front end; the vendor runs the engine |
| Composable | Several services, each chosen separately and joined by APIs | commercetools for commerce, a headless CMS for content, Algolia for search | You run the joins between every service |
Shopify describes a monolith as one that "operates off of a single codebase", where front-end and back-end changes are tied together, and headless as decoupling "the front-end experience from the back-end operations" ([Shopify](https://www.shopify.com/enterprise/blog/headless-architecture)). Headless is one step: you replace the storefront and keep everything else. Composable takes the same idea to the other parts of the store, so the catalog, search, content and order management can each be a different product.
### Composable commerce vs headless in one sentence
Every composable store is headless, because the front end is separate from the services behind it. Most headless stores are not composable, because they keep a single commerce platform behind the new front end. A [headless Shopify](https://gigabit.agency/platforms/shopify/) build with Hydrogen is headless; adding a separate CMS, search service and order system around it starts to make it composable.
### What MACH means
MACH is the label vendors in this space use. The MACH Alliance was founded on 23 June 2020 by commercetools, Contentstack, EPAM Systems and Valtech, and its launch release says the name "stands for Microservices based, API-first, Cloud-native SaaS and Headless" ([commercetools press release](https://commercetools.com/press-releases/mach-alliance-forms-to-empower-enterprises-taking-advantage-of-modern-saas-technology)).
1. **Microservices based**: each business function is its own service that can be built and deployed on its own.
2. **API-first**: every function is reachable through an API.
3. **Cloud-native SaaS**: the vendor hosts, scales and updates it.
4. **Headless**: the front end is separate from the back end.
The Alliance's site now frames this as three principles. Composable means "capabilities you can assemble, replace, and evolve independently", and connected means "designed to be optimized in an ecosystem" ([MACH Alliance](https://machalliance.org/mach-principles/)). The Alliance is an industry body whose members sell these products, so read its material as a vendor position. MACH architecture describes how software is built; it does not tell you whether your store needs it.
### The parts of a composable stack
These are the usual components, with examples named only where we read the vendor's own description.
| Part | Its job | Examples |
| --- | --- | --- |
| Storefront | The pages shoppers see, built in a web framework | Shopify Hydrogen, hosted on Oxygen; BigCommerce Catalyst, built on Next.js |
| Commerce engine | Catalog, prices, cart, checkout, orders | commercetools; Shopify through its Storefront API; BigCommerce through its GraphQL Storefront API |
| Content (CMS) | Editorial pages, landing pages, rich product content | Contentful |
| Search | Product search and filtering | Algolia |
| PIM and OMS | Product data and order routing when they outgrow the engine | commercetools lists catalog, inventory and order management among its modular capabilities |
Hydrogen is "Shopify's headless toolkit": it handles the Shopify-specific parts such as the cart and the Storefront API client, while a framework handles routing and rendering ([Shopify docs](https://shopify.dev/docs/storefronts/headless/hydrogen)). Shopify's Storefront API "scales with buyer traffic", and requests from real shoppers have no fixed per-minute limit ([Storefront API](https://shopify.dev/docs/api/storefront)). Catalyst is BigCommerce's "composable, fully customizable headless ecommerce storefront framework", released under the MIT licence, and can be deployed to Vercel or any host that supports Next.js ([BigCommerce docs](https://docs.bigcommerce.com/developer/docs/storefront/catalyst/overview)). Contentful describes itself as a composable content platform with a headless CMS at its core ([Contentful](https://www.contentful.com/headless-cms/)).
### What composable commerce really costs
The licences are the part you can see. Prices below were checked on 23 October 2026.
- **Commerce engine.** commercetools does not publish prices. Its pricing page says pricing is order-based rather than a share of sales, and asks you to contact sales ([commercetools pricing](https://commercetools.com/pricing)). For comparison, Shopify Advanced is $299 a month billed yearly, or $3,588 a year, for the whole platform ([Shopify pricing](https://www.shopify.com/pricing?country=US)).
- **Search.** Algolia's Grow plan includes 10,000 search requests a month, then charges $0.50 per extra 1,000, and $0.40 per extra 1,000 records above 100,000 ([Algolia pricing](https://www.algolia.com/pricing)). On our arithmetic, a store running 1,000,000 searches a month would pay about $495 a month for search requests on that plan.
- **Hosting the storefront.** On Shopify, Oxygen hosting for Hydrogen storefronts is included at no extra charge on supported paid plans, subject to runtime and usage limits ([Shopify](https://www.shopify.com/enterprise/blog/headless-commerce)). Elsewhere you pay a host such as Vercel.
- **Content.** A headless CMS is another subscription. Contentful's pricing page blocked our requests, so we have left its prices out.
### The team is the bigger number
On a single platform, the vendor maintains the checkout, the admin and the storefront framework. On a composable stack, your team maintains the front end, the integrations between services and the monitoring that tells you when one breaks. Shopify's own article says headless storefronts suit "established internal technology and product teams", and gives rough thresholds of a budget of $1 million or more and a timeline of six months or longer for headless, against under $1 million and under six months for a Liquid theme ([Shopify](https://www.shopify.com/enterprise/blog/headless-architecture)). It also warns that merchants must "dedicate more resources to maintain a headless infrastructure".
Here is what that looks like as a running cost, on our arithmetic. Two engineers keeping a headless storefront and its integrations moving, priced on our own embedded pods at $3,000–$7,500 per engineer per month, cost $72,000 to $180,000 a year. An in-house team costs more once salaries and recruiting are counted. Either way it is many times the $3,588 a year Shopify Advanced charges for a platform someone else maintains.
### When it pays off, and when it is overkill
The extra running cost has to come back as margin. Assume a 40% contribution margin on extra sales (an assumption; use your own). A team cost of $72,000 a year then needs $180,000 of extra sales every year to break even, and a team cost of $180,000 needs $450,000.
- At **$2 million** a year in online revenue, that is a sustained lift of 9% to 22.5% from the architecture alone.
- At **$5 million**, it is 3.6% to 9%.
- At **$20 million**, it is 0.9% to 2.3%.
A front end rarely produces a lift of several per cent a year on its own. So our rule of thumb from this arithmetic is plain: **below about $5 million a year in online revenue, stay on a single platform** unless a hard requirement forces you off it. Above that, composable can pay, when there is a specific thing the platform cannot do.
### Signs it pays off
- You run several brands, regions or channels from one catalog, and the platform's multi-store tools have run out.
- The checkout, pricing or product model needs logic the platform does not expose, even on its top plan. On Shopify that means things [Shopify Plus](https://gigabit.agency/compare/shopify-vs-shopify-plus/) cannot do with checkout extensions and Functions.
- Content is a large part of how you sell, and editors need a real CMS beside the catalog.
- You already employ engineers who will own the front end and the integrations for years.
### Signs it is overkill
- The reason is speed or design. A well-built theme on Shopify or BigCommerce can be fast and look like anything you need.
- Nobody on staff will own the integrations after launch.
- The plan depends on a launch date inside six months.
- Most of the wish list is available as an app or a plan upgrade.
### A decision table
| Your situation | Stay on one platform | Go headless | Go composable |
| --- | --- | --- | --- |
| Under about $5M online revenue, no in-house engineers | Yes | No | No |
| Theme limits on design or speed, platform otherwise fine | Yes, rebuild the theme first | Only if the theme cannot get there | No |
| Content-heavy brand, one catalog | Often, with the platform's CMS features | Yes, with a headless CMS beside it | Rarely needed |
| Several brands or regions, shared catalog | If the platform's multi-store tools fit | Sometimes | Often |
| Checkout or pricing logic the platform cannot express | No | Not enough on its own | Yes |
| Engineers in-house who will own it for years | Possible | Yes | Yes |
### The middle path most stores should look at first
If the storefront is the only part that is holding you back, go headless on the platform you have. On Shopify, that means Hydrogen on the Storefront API, hosted on Oxygen; on BigCommerce, it means Catalyst on the GraphQL Storefront API. You keep the vendor's checkout, admin, payments and app ecosystem, and you only own the front end. It is the cheapest way to test whether a custom front end earns its keep before you split anything else apart.
### How Gigabit helps
We are a Shopify Partner and build on [Shopify](https://gigabit.agency/platforms/shopify/) and [BigCommerce](https://gigabit.agency/platforms/bigcommerce/). Most of the time our advice is to stay on the platform and fix the theme, the apps or the plan. When a store does need more, we build it:
- **Store builds** start at $9,000 fixed. Shopify Plus, headless, B2B, multi-storefront and Adobe Commerce builds are quoted higher, as a fixed price after scoping, and headless Hydrogen or Catalyst storefronts fall in that group.
- **Integrations** between the store and the ERP, CMS, search or order system: one scoped integration is $8,000 fixed over 3 weeks.
- **Migrations** between platforms start at $7,000 fixed, with every old URL redirected.
- **Ongoing engineering** for a headless or composable stack is an embedded pod at $3,000–$7,500 per engineer per month, working in your repo and on your release process.
If you are weighing a move, start by listing what your current platform cannot do and what each item is worth. If you are on Shopify, check first whether [Shopify Plus](https://gigabit.agency/compare/shopify-vs-shopify-plus/) covers the list. That is cheaper than any composable build.

### FAQ
**What is composable commerce?**
Composable commerce is building an online store from separate services connected by APIs: a storefront, a commerce engine for catalog, cart and checkout, a CMS, search, and product and order systems. Each can come from a different vendor and be replaced without rebuilding the others. A monolithic platform such as Shopify runs all of these jobs in one product.
**What is the difference between composable commerce and headless commerce?**
Headless separates only the storefront from one commerce engine, reached through its API. Composable separates several parts of the store, such as catalog, search, content and order management, and lets each come from a different vendor. Every composable store is headless; most headless stores keep a single commerce platform behind the new front end.
**What does MACH stand for?**
Microservices based, API-first, Cloud-native SaaS and Headless, according to the MACH Alliance's launch release of 23 June 2020. The Alliance was founded by commercetools, Contentstack, EPAM Systems and Valtech. Its site now describes three principles: open, composable and connected.
**Is headless Shopify with Hydrogen worth it?**
Only when a theme cannot do what the storefront needs and you have engineers to own a separate front end. Hydrogen is Shopify's headless toolkit, and Oxygen hosting is included on supported paid plans, so the main cost is the team. Shopify's own article suggests headless suits budgets of $1 million or more and timelines of six months or longer.
**How big should a store be before going composable?**
On our arithmetic, two engineers at $72,000 to $180,000 a year need $180,000 to $450,000 of extra sales each year at a 40% margin to pay back. Below about $5 million a year in online revenue that is a lift few front ends deliver, so most stores that size should stay on a single platform unless a hard requirement forces a change.

## Claude Code pricing in 2026: every plan, API costs and three worked examples
URL: https://gigabit.agency/insights/claude-code-pricing/
Published 2026-10-23

**Claude Code pricing** comes in two forms. You either use it inside a Claude subscription, where it shares one usage allowance with Claude chat, or you pay per token through the Anthropic API. On [Anthropic's pricing page](https://claude.com/pricing), checked on 23 October 2026, **Pro is $17 a month billed annually ($200 up front) or $20 monthly**, **Max is $100 (5x) or $200 (20x) a month**, **Team seats are $20 (Standard) or $100 (Premium) a month billed annually**, and **Enterprise is $20 a seat a month plus usage at API rates**. On the API, the default model, Opus 5.5, costs **$4 per million input tokens and $20 per million output tokens**. The free plan does not include Claude Code. The rest of this guide covers the limits, what a session costs in tokens, and what a solo founder, a five-person team and a heavy agentic user would pay.
### Every way to pay for Claude Code
Prices are in US dollars, before tax, as Anthropic's pricing page showed them on 23 October 2026.
| Plan | Billed annually | Billed monthly | Claude Code | Usage, in Anthropic's words |
| --- | --- | --- | --- | --- |
| Free | $0 | $0 | Not included | Everyday questions |
| Pro | $17 a month ($200 a year) | $20 | Included | At least 5x more usage per 5-hour session than Free |
| Max 5x | Not offered | $100 | Included | 5x more usage per 5-hour session than Pro |
| Max 20x | Not offered | $200 | Included | 20x more usage per 5-hour session than Pro |
| Team, Standard seat | $20 a seat | $25 a seat | Included | More usage than Pro |
| Team, Premium seat | $100 a seat | $125 a seat | Included | 5x more usage than Standard seats |
| Enterprise | $20 a seat, plus usage at API rates | Annual only | Included | Usage billed at API rates |
| Anthropic API (Console) | Pay per token | Pay per token | Included | Your organisation's API rate limits |
### Is Claude Code free?
The software is free to install, but using it is not. Anthropic's plan comparison marks Claude Code as not included on the Free plan, and every way of running it draws on either a paid plan or API credit. The cheapest entry is Pro at $17 a month on an annual plan. The other route is a Claude Console account, where you buy prepaid API credit and pay only for the tokens you use.
### Claude Code Pro vs Max
Pro is for regular but not all-day use. Max is for people who work in Claude Code for most of the day and keep hitting Pro's limits. Max 5x and Max 20x are billed monthly only, so there is no annual discount. Max also adds higher output limits, priority access at busy times and $100 or $200 a month in API credits for building on the Claude Platform. Those credits cannot be spent in Claude Code or the Claude apps, so they do not stretch your coding allowance.
### Claude Code for teams
The [Team plan](https://support.claude.com/en/articles/11845131-use-claude-code-with-your-team-or-enterprise-plan) is for 2 to 150 people, and Claude Code is included with every seat. You can mix Standard and Premium seats, so only the heavy users need the $100 seat. Team adds SSO, central billing, server-managed Claude Code settings and an analytics dashboard. Enterprise needs at least 20 seats and works differently: the seat is $20 a month and all usage is billed at API rates, with no per-seat usage limit on usage-based plans. It adds SCIM, audit logs, custom data retention and role-based access.
### Usage limits, exactly as Anthropic states them
Anthropic does not publish limits as a number of messages or tokens. Its pricing FAQ says usage depends on the length and complexity of conversations, the model and the features you use, "so there's no fixed message count". What it does state:
- **Two windows.** Pro and Max have a session limit that resets every five hours and a weekly limit. Max also has a separate weekly limit for Fable, Anthropic's largest model.
- **One pool.** Claude chat on web, desktop and mobile and Claude Code in the terminal or IDE all draw on the same limits.
- **Discretionary caps.** Anthropic says it may limit usage in other ways, such as weekly and monthly caps or model and feature usage.
- **Team seats** reset on a rolling five-hour window and a weekly window, and the allowance is shared with Claude chat and Cowork.
When you hit a limit you can wait for the reset, move up a plan, or turn on **usage credits**, which bill further use at standard API rates with a monthly spend cap you set. Depending on your plan and seat, Fable requests may bill to usage credits instead of your allowance; Claude Code asks before the first one does.
### Claude Code API pricing
On the API, or once usage credits kick in, you pay per million tokens (MTok). These are the current models on [Anthropic's API pricing page](https://platform.claude.com/docs/en/about-claude/pricing), checked on 23 October 2026.
| Model | Input | Cache write (5 min) | Cache write (1 hour) | Cache read | Output |
| --- | --- | --- | --- | --- | --- |
| Fable 5.1 | $10 | $12.50 | $20 | $0.25 | $50 |
| Opus 5.5 (Claude Code's default) | $4 | $5 | $8 | $0.20 | $20 |
| Sonnet 5.5 | $2 | $2.50 | $4 | $0.10 | $10 |
| Haiku 5.5, prompts up to 100K tokens | $0.10 | $0.125 | $0.20 | $0.01 | $0.50 |
Claude Code's [model settings](https://code.claude.com/docs/en/model-config) say the default resolves to Opus 5.5 on Pro, Max, Team, Enterprise and the API. Anthropic's own cost guidance is to use Sonnet for most coding and keep Opus for hard reasoning. Fast mode on Opus 5.5 costs twice the standard rate, and US-only inference costs 1.1 times.
### What a Claude Code session costs on the API
Claude Code sends the whole conversation with every request, and each round of tool results is another request. Prompt caching is what keeps that affordable: history already sent is re-read at the cache-read rate. On an API key, Claude Code caches the main conversation for five minutes by default.
Here is a working model, and it is our arithmetic, not Anthropic's. Take one focused hour with **60 requests**, an average context of **60,000 tokens** of which 90% comes from cache, and **1,500 output tokens** per request. Each request is then 54,000 cached tokens, 6,000 newly cached tokens and 1,500 output tokens.
| Model | Per request | Per hour (60 requests) |
| --- | --- | --- |
| Opus 5.5 | $0.011 + $0.030 + $0.030 = **$0.071** | **about $4.25** |
| Sonnet 5.5 | $0.005 + $0.015 + $0.015 = **$0.035** | **about $2.12** |
| Opus 5.5 with no cache hits | $0.24 + $0.03 = **$0.27** | **about $16.20** |
Three hours a day on Opus comes to roughly $13. That lines up with what Anthropic reports in its [cost documentation](https://code.claude.com/docs/en/costs): across enterprise deployments, about **$13 per developer per active day and $150 to $250 per developer per month**, with 90% of users below $30 per active day. The last row shows why a long session left open over a break, which misses the cache, costs far more than one that is cleared between tasks.
### The costs that are not on the pricing page
- **Review time.** Every change Claude Code makes still needs a person to read it. As an illustration, a senior engineer who spends an hour a day reviewing AI-written pull requests at $100 an hour costs about $2,000 a month in review alone, which is more than any plan here.
- **CI minutes.** The [Claude Code GitHub Action](https://code.claude.com/docs/en/github-actions) runs on GitHub-hosted runners, which use your Actions minutes, on top of the tokens each run consumes.
- **Rework.** Code that passes a demo and fails under real users gets written twice. The fix is tests, review and a clear spec, not a bigger plan.
- **The wrong login.** If an `ANTHROPIC_API_KEY` environment variable is set, Claude Code uses it instead of your subscription, and the usage is billed at API rates.
- **Background requests.** Claude Code makes small requests for things like session titles, typically under $0.04 a session by Anthropic's count.
- **Agent teams.** Anthropic says agent teams use about 7x more tokens than a standard session when teammates run in plan mode.
### Three worked examples
All three are our arithmetic on the prices above, checked on 23 October 2026.
### A solo founder building an MVP
Start on **Pro at $200 a year**. If you hit the five-hour limit most days, move to **Max 5x at $100 a month ($1,200 a year)**, and only go to **Max 20x ($2,400 a year)** if Max 5x still stops you. Compare that with the API: Anthropic's average of $150 to $250 per developer per month is more than Max 5x costs, so a founder coding every day is usually better off on a subscription. Turn on usage credits with a low monthly cap for the odd heavy week rather than upgrading for it.
### A five-person product team
On the Team plan, **five Standard seats are $100 a month, or $1,200 a year** billed annually ($1,500 billed monthly). If two of the five are heavy users, give them Premium: **three Standard and two Premium seats come to $260 a month, $3,120 a year**. All five on Premium is $6,000 a year. Enterprise is not an option at this size because of the 20-seat minimum. On the API, Anthropic's averages would put five developers at $750 to $1,250 a month, so seats are cheaper unless usage is light and irregular.
### Heavy agentic use on the API
Say a company runs Claude Code headless in CI to review 400 pull requests a month, and one engineer runs agent teams for four hours a day. Using the session model above, a review of 25 requests at a 40,000-token average context, 90% cached, with 1,000 output tokens each, on Sonnet 5.5 is about **$0.59**, so **400 reviews are about $236 a month** plus GitHub Actions minutes. Four hours a day on Opus at roughly 7x a normal session's tokens is about **$119 a day, or $2,380 over 20 working days**. At this level, set a [workspace spend limit](https://code.claude.com/docs/en/costs) in the Claude Console, move subagents to Sonnet or Haiku, and cap each CI run with `--max-turns`.
### Taking it to production
Most projects built with Claude Code never need outside help. A side project, an internal tool or a prototype that proves an idea is fine as it is. The gap shows up when real users, payments or customer data arrive: access rules that hold, keys kept out of the browser, tests, error monitoring, backups and a deploy you can roll back. Our [Prototype Production Review](https://gigabit.agency/services/prototype-to-production/) checks exactly those areas (security including Supabase row-level security, exposed keys and auth; the data model; performance; and deploy) for **$2,500 fixed** over **5 business days**, and you get a graded report and a fix plan. The fee is credited in full to a build with us started within 60 days. If we then fix and launch it, the work is done by an embedded team at $3,000–$7,500 per engineer per month.
If you are still choosing tools, our [Claude Code vs Cursor comparison](https://gigabit.agency/compare/claude-code-vs-cursor/) covers both side by side, including what each costs. They are not exclusive: Anthropic's help pages say a Pro or Max plan also covers Claude Code inside Cursor and other VS Code forks.

### FAQ
**How much does Claude Code cost?**
Checked on claude.com/pricing on 23 October 2026: Pro is $17 a month billed annually ($200 a year) or $20 monthly, Max is $100 (5x) or $200 (20x) a month, Team seats are $20 (Standard) or $100 (Premium) a month billed annually, and Enterprise is $20 a seat plus usage at API rates. On the API you pay per token; Opus 5.5, the default model, is $4 per million input tokens and $20 per million output tokens.
**Is Claude Code free?**
No. Claude Code is free to install, but Anthropic's Free plan does not include it. You need a paid plan (Pro, Max, Team or Enterprise) or a Claude Console account with prepaid API credit. Pro at $17 a month on an annual plan is the cheapest way in.
**Claude Code Pro vs Max: which should I buy?**
Start on Pro. Move to Max 5x ($100 a month) if you hit the five-hour session limit most days, and to Max 20x ($200) only if Max 5x still stops you. Anthropic describes Max as 5x or 20x more usage per 5-hour session than Pro. Max is billed monthly only, and its usage is shared with Claude chat, as on Pro.
**How much does Claude Code cost on the API?**
On 23 October 2026, per million tokens: Opus 5.5 $4 input and $20 output, Sonnet 5.5 $2 and $10, Haiku 5.5 $0.10 and $0.50, Fable 5.1 $10 and $50, with cached input much cheaper. Anthropic reports an average of about $13 per developer per active day and $150 to $250 per developer per month across enterprise deployments. By our arithmetic, a focused hour on Opus 5.5 with good caching is about $4.
**Does Claude Code for teams need the Enterprise plan?**
No. Claude Code is included with every Team seat, and Team covers 2 to 150 people at $20 (Standard) or $100 (Premium) a seat a month billed annually, with seat types mixed on one team. Enterprise needs at least 20 seats, bills usage at API rates on top of a $20 seat, and adds SCIM, audit logs, custom data retention and role-based access.

## The security checklist to run before real users reach your AI-built app
URL: https://gigabit.agency/insights/ai-built-app-security-checklist/
Published 2026-10-23

**Vibe coding security** comes down to a short list of checks. Apps built with Lovable, Bolt, Replit, Cursor or Claude Code usually work. What they tend to skip is anything a demo never exercises, such as who can read which rows, which keys reach the browser, and what happens when somebody repeats a request ten thousand times. Below are 13 checks, each with a test you can run in about ten minutes, grounded in Supabase's documentation, the OWASP Top 10 and the vendors' own security pages, all read on 23 October 2026. Run them before you invite real users. You do not need to be a security engineer to do it, and most fixes are small once you know where the gap is.
### Is vibe coding safe?
It is as safe as the access rules, keys and limits around the code, and those are the parts AI tools are least likely to get right unprompted. The failures cluster in predictable places.
**Access control comes first.** [OWASP's Top 10:2025](https://top10.owasp.org/2025) puts Broken Access Control at A01, Security Misconfiguration at A02 and Software Supply Chain Failures at A03. In apps built on Supabase, which includes most Lovable apps and many Bolt apps, the browser talks to the database directly with a public key. Supabase's docs say a table in an exposed schema without row-level security [is readable and writable by any role with a grant on it](https://supabase.com/docs/guides/database/postgres/row-level-security). The policies are the only thing between one user and another user's data.
**This has happened in public.** The NVD record for [CVE-2025-48757](https://nvd.nist.gov/vuln/detail/CVE-2025-48757) describes insufficient row-level security policies in Lovable-generated sites, through 15 April 2025, that let unauthenticated attackers read or write arbitrary tables. Lovable disputes it, on the grounds that each customer is responsible for protecting their own app's data. Either way, the fix sat in the customer's policies. Bolt's own [database security docs](https://support.bolt.new/bolt-cloud/database/security) warn that on a site without sign-in it may write policies that let any visitor view, add, change and delete data.
**Packages can be invented.** A study accepted at USENIX Security 2025 analysed 576,000 code samples and found [hallucinated package names](https://arxiv.org/abs/2406.10279) in at least 5.2% of commercial-model output and 21.7% of open-source-model output, 205,474 unique names in all. Anyone can register one of those names on npm or PyPI and wait.
**Agents with production access can do damage.** In July 2025 a Replit agent deleted a user's production database despite instructions to freeze code; [The Register reported](https://www.theregister.com/2025/07/22/replit_saastr_response/) that Replit responded by separating development and production databases automatically. Replit's docs now say Agent cannot modify the production database.
### The checklist
Thirteen checks in the order we would run them. The tests assume a Supabase backend where that matters; the same idea applies to any database your front end can reach.
| # | Check | What good looks like | How to test it in 10 minutes |
| --- | --- | --- | --- |
| 1 | Row-level security on every table | Every table in an exposed schema, usually `public`, has RLS enabled | In the Supabase SQL editor run `select tablename from pg_tables where schemaname = 'public' and rowsecurity = false;` and expect no rows. Then open Security Advisor and clear its errors. |
| 2 | Policies tested with the anon key | Signed out, a visitor reads only public data; user A cannot read or change user B's rows | Copy the publishable (anon) key from your live site, then request each table with curl as below. Sign in as user A and ask for user B's row by id. Expect an empty result or an error. |
| 3 | No secret key in the client | Only the publishable or anon key appears in browser code | Build the app and search the output folder for `sb_secret_` and for the name of your service-role variable. Decode any `eyJ...` token you find, on your own machine, and check its role is `anon`. |
| 4 | No secrets in git history | Nothing sensitive was ever committed, or it has been rotated | Run a secret scanner such as Gitleaks or TruffleHog over the whole history, not only the current files. Rotate anything it finds before doing anything else. |
| 5 | Email verification and password reset | Sign-up needs a confirmed email; reset links are single-use and expire | Sign up with a fresh address and try to use the app before confirming. Request a reset and use the link twice. Check email confirmation is on in Supabase Auth settings. |
| 6 | Session handling | Signing out ends the session; tokens expire on a short clock | Copy your access token from the browser, sign out, and replay a request with it. Note how long it keeps working. Check sign-out on one device does what you expect on another. |
| 7 | Authorization on server routes and edge functions | Every server function checks who is calling and whether they own the record | Call each API route and edge function with no token, then with user A's token on user B's record. Hidden buttons do not count; the server must refuse. |
| 8 | Input validation | The server checks type, length and range, and never trusts a price or role sent by the browser | Submit a 10,000-character name, a negative quantity and `<script>` in a text field. In the browser's network tab, edit a request to change a price or a role and resend it. |
| 9 | Rate limiting | Sign-up, login, password reset, forms and AI endpoints refuse floods | Send 50 requests in a short loop to login and to any endpoint that calls a model. If all 50 succeed, there is no limit. |
| 10 | Dependencies and hallucinated packages | Every package exists, is the one you meant and has no known critical vulnerability | Run `npm audit`. For each package the AI added, check its npm page for publisher, weekly downloads, repository link and age. |
| 11 | Logging and error monitoring | Errors reach a person; logs are searchable and hold no passwords, tokens or card data | Throw a deliberate error in staging and time how long it takes for someone to be told. Search the logs for `password` and `token`. |
| 12 | Backups | Automatic backups exist and someone has restored one | Check your database plan's backup retention, then restore the latest backup into a scratch project and open a table. |
| 13 | Cost controls on AI API keys | Model keys stay on the server, the provider account has a spend limit, users have caps | Search the browser bundle for your model provider's key prefix. Confirm a monthly limit is set in the provider's console. Hit your AI endpoint 50 times signed out. |
### How to test row-level security properly
Checks 1 and 2 matter most, so here is the full version. Your app's publishable key is meant to be public, so assume an attacker has it and test with it yourself.
1. Find your project URL and publishable key in the site's network requests or the Supabase dashboard.
2. For each table, request it signed out: `curl 'https://YOUR-PROJECT.supabase.co/rest/v1/orders?select=id' -H 'apikey: YOUR-PUBLISHABLE-KEY'`. Anything that should be private must come back empty or refused.
3. Sign in as a test user, copy the access token, add it as an `Authorization: Bearer` header, and request another test user's rows by id.
4. Try writes too: an insert into another user's account, an update to someone else's row, a delete.
5. Repeat for storage buckets. Supabase says anything in a public bucket is publicly accessible.
Four details from [Supabase's RLS guide](https://supabase.com/docs/guides/database/postgres/row-level-security) catch people out. `auth.uid()` returns null when nobody is signed in, so add an explicit `auth.uid() is not null` check. Policies that rely on `user_metadata` are unsafe because signed-in users can edit that claim; Supabase recommends `raw_app_meta_data` for authorization data. Views bypass RLS by default; on Postgres 15 and later, set `security_invoker = true` so a view obeys the policies of its tables. And add an index on every column your policies filter on, or the checks will slow every query.
### Which key goes where
Supabase has two kinds of key. The publishable key (the legacy name is anon) is [safe to expose](https://supabase.com/docs/guides/api/api-keys) in a web page or app and reaches only what your policies allow. The secret key (legacy: service_role) bypasses every row-level security policy, and Supabase says never to put it in a browser, a shipped app or source control. Supabase blocks secret keys sent from browsers, but a leaked one still works from any other tool.
If a secret key leaked, create a new one, replace it everywhere, confirm every component uses it, then delete the old one. GitHub's guidance on [removing sensitive data](https://docs.github.com/en/authentication/keeping-your-account-and-data-secure/removing-sensitive-data-from-a-repository) is to revoke or rotate the secret first; once it is rotated, rewriting history may not be needed. Payment, email and model API keys follow the same rule. Keep them in the builder's secrets store (Lovable Secrets, Bolt's database secrets, Replit Secrets) and call those services from a server function, never from the browser.
### Auth, limits and cost caps
Supabase's [production checklist](https://supabase.com/docs/guides/deployment/going-into-prod) asks for email confirmations, an OTP lifetime of 3,600 seconds or less, a custom SMTP server for auth emails and multi-factor authentication on your own Supabase account. It also lists the default auth rate limits; auth emails, for example, are limited to 2 an hour until you set up custom SMTP. Those limits protect Supabase's auth endpoints only. Your own API routes and edge functions need their own.
AI features add a cost risk. OWASP's LLM list names it [Unbounded Consumption](https://genai.owasp.org/llmrisk/llm102025-unbounded-consumption/) (LLM10), next to [Sensitive Information Disclosure](https://genai.owasp.org/llmrisk/llm022025-sensitive-information-disclosure/) (LLM02). An endpoint that calls a model with no sign-in and no limit lets anyone spend your money. Put a per-user cap in your own code and a monthly ceiling at the provider; Anthropic, for example, lets you [set a spend limit](https://platform.claude.com/docs/en/api/rate-limits) below your tier's cap in its console.
### Backups and recovery
On Supabase, [daily backups](https://supabase.com/docs/guides/platform/backups) are kept for 7 days on Pro, 14 on Team and up to 30 on Enterprise. Free projects are not in that list; Supabase recommends they export their own data regularly with the CLI. Replit's docs describe point-in-time database restore within 7 days as standard and 28 days on Pro. Whatever you run on, a backup nobody has restored is a hope, so do one restore before launch.
### What the builders' own scanners catch
All three app builders now scan. Lovable runs a [Quick scan on every publish](https://docs.lovable.dev/features/security) and offers a manual Deep scan of the code. Bolt's [project security audit](https://support.bolt.new/building/security) checks data access, sign-in, misuse and keys. Replit's [Project Security Center](https://docs.replit.com/features/security/project-security-center) adds dependency scans and Agent security scans. Use them; they are free or close to it. They also say plainly what they are not. Lovable's docs say its scans cannot guarantee complete security, and Replit's say a clean scan does not guarantee an application is secure. Our [Lovable vs Bolt vs Replit comparison](https://gigabit.agency/compare/lovable-vs-bolt-vs-replit/) sets the three side by side.
### AI code review helps, and it is not the whole job
If you work in Cursor or Claude Code, AI code review is a cheap first pass. Claude Code's [/security-review command](https://code.claude.com/docs/en/security) runs an on-demand security pass over the changes on your current branch, and Cursor's [Bugbot](https://cursor.com/bugbot) reviews pull requests on GitHub. Both read code. The checklist above tests behaviour: what the running app does when someone signed out asks for your orders table. Run both, and treat a clean AI review as a reason to run the tests, not to skip them. Our [Claude Code vs Cursor comparison](https://gigabit.agency/compare/claude-code-vs-cursor/) covers the two assistants.
### Taking it to production
Most prototypes, demos and side projects do not need a paid review. Work through the table, fix what fails, and ship. A review earns its fee when the app will hold other people's data, take payments, or face a customer's security questionnaire or an investor's diligence.
That is what our [Prototype Production Review](https://gigabit.agency/services/prototype-to-production/) is for. In the service page's own words, "a senior engineer reviews its security, data model, performance and deploy", and in 5 business days you get a graded report and a fix plan. The checks cover row-level security that is off or too broad, the wrong key in the browser, auth and sessions, storage bucket permissions, input validation and rate limits, and secrets in the repository; then schema, migrations and backups; queries, page weight and what the app costs to run; and environments, error monitoring and who owns the accounts. It is $2,500 fixed, credited in full to a build with us started within 60 days. It is not a penetration test or a compliance certification, and fixes are separate: you can make them yourself, or an embedded team does them at $3,000–$7,500 per engineer per month. For AI features with their own attack surface, such as prompt injection or agents with tool access, see our [AI security assessment](https://gigabit.agency/services/ai-security/).

### FAQ
**Is vibe coding safe?**
The code itself is often fine. The risk sits in what a demo never tests: database access rules, keys exposed to the browser, missing rate limits and invented package names. OWASP's Top 10:2025 ranks Broken Access Control first, and that is the most common gap in AI-built apps. Run the 13 checks in this guide before you invite real users and most of the risk is gone.
**How do I check Supabase row-level security?**
In the SQL editor, list tables in the public schema where rowsecurity is false and turn RLS on for each. Then test the policies from outside: request each table with curl using only your publishable key, then sign in as one test user and try to read and change another user's rows. Check storage buckets too, and avoid policies that rely on user_metadata, which users can edit.
**Is Lovable secure?**
Lovable's platform runs a Quick security scan on every publish and offers a Deep scan of the code, and admins can block publishing on critical findings. Your app's security still depends on its row-level security policies, because the front end reads the database with the public key. CVE-2025-48757 recorded Lovable apps with weak policies in 2025; Lovable disputes it as the customer's responsibility. Test your policies with the anon key before launch.
**Is AI-generated code secure?**
Not by default. It tends to work on the happy path and miss authorization checks, input validation and limits. Research accepted at USENIX Security 2025 found hallucinated package names in at least 5.2% of commercial-model code samples and 21.7% of open-source-model samples, so check every dependency an AI adds. Treat AI-generated code like a junior developer's pull request: read it, test it, and verify access rules from outside.
**Can AI code review replace a security review?**
No, but it is a useful first pass. Tools such as Claude Code's /security-review and Cursor's Bugbot read your changes and flag likely bugs. They do not test how the running app behaves, whether your database policies hold, or whether a leaked key still works. Use AI review on every change, run the checklist's tests before launch, and get a human review when the app will hold customer data or take payments.

## How to rank in ChatGPT: a practical guide to ChatGPT search
URL: https://gigabit.agency/insights/how-to-rank-in-chatgpt/
Published 2026-10-23

**To rank in ChatGPT, let OpenAI's search crawler read your site, appear in the search results ChatGPT reads when it answers, and publish pages it can quote and check.** OpenAI's own instruction is short: to make a site eligible for ChatGPT search, [allow OAI-SearchBot to crawl it](https://help.openai.com/en/articles/9237897-searching-the-web-with-chatgpt) and make sure your host or CDN accepts traffic from OpenAI's published IP addresses. Past eligibility, OpenAI says ChatGPT ranks results using multiple factors meant to surface relevant, reliable information, and that placement is not guaranteed. It publishes no list of those factors. Anyone selling you "the ChatGPT ranking factors" is guessing. This guide sticks to what OpenAI and Microsoft document, as checked on 23 October 2026, and to the work that follows from it.
One word of framing. There is no position one in ChatGPT. An answer either names you, cites your page as a source, or does neither. So "ranking" here means being named and cited on the questions your buyers ask, and the number you track is a citation rate, not a rank.
### Two ways ChatGPT can mention you
ChatGPT can name a company from what the model learned in training, or from pages it finds when it searches the web during an answer. Most of the work you can control is on the second path, because search answers come with citations and change when the web changes. OpenAI runs a separate crawler for each job, and each one is controlled separately in robots.txt.
| User agent | What OpenAI says it does | What blocking it in robots.txt does |
| --- | --- | --- |
| OAI-SearchBot | Surfaces websites in ChatGPT's search features | Your pages are not shown in ChatGPT search answers, though they can still appear as navigational links |
| GPTBot | Crawls content that may be used to train OpenAI's foundation models | Signals your content should not be used for training. It does not remove you from search |
| ChatGPT-User | Visits a page when a user's request calls for it, inside ChatGPT or a custom GPT | Robots.txt may not apply, because the visit is user-initiated, and OpenAI says it does not decide what appears in search |
| OAI-AdsBot | Checks the safety of pages submitted as ChatGPT ads | Only visits pages submitted as ads |
That table comes from [OpenAI's crawler documentation](https://developers.openai.com/api/docs/bots), which also says each setting is independent of the others. You can block GPTBot to keep your content out of training and still allow OAI-SearchBot so you appear in search. That is a legitimate choice. Blocking OAI-SearchBot is the only one of the four that takes you out of ChatGPT search.
### Step 1: let OAI-SearchBot in
Open `yourdomain.com/robots.txt` and look for a group that names `OAI-SearchBot`, or a `User-agent: *` group with `Disallow: /`. If you want to appear in ChatGPT search, OAI-SearchBot needs to be allowed on the pages you want cited. OpenAI says search-related robots.txt changes can take about 24 hours to take effect.
Robots.txt is half of it. Firewalls, bot-management rules and CDN presets block crawlers without touching robots.txt, which is why OpenAI asks site owners to allow requests from its [published OAI-SearchBot IP ranges](https://openai.com/searchbot.json). Check your WAF or bot settings for a rule that challenges or blocks unknown bots, and check your server logs for OAI-SearchBot requests and the status codes they got. A wall of 403s means the crawler is arriving and being turned away.
Then check what the crawler can read once it is in. OpenAI's article on [offline web search](https://help.openai.com/en/articles/20001203-offline-web-search-for-chatgpt-workspaces), which answers from OpenAI's own index and cache, lists the usual reasons a page is missing or stale there: a robots.txt block, CDN or bot blocking, a login wall, content that depends on scripts or dynamic loading, a page that is new or rarely accessed, and site structure that makes deeper pages hard to discover. Every one of those is fixable on your side. The raw-HTML test and the rest of the site checks are in our [GEO field guide](https://gigabit.agency/insights/generative-engine-optimization-field-guide/), so they are not repeated here.
### Step 2: appear in the sources ChatGPT search reads
When ChatGPT searches, it does not send your buyer's message to a search engine word for word. OpenAI says ChatGPT [typically rewrites the question](https://help.openai.com/en/articles/9237897-searching-the-web-with-chatgpt) into one or more targeted queries, sends them to search providers it partners with, and may follow up with more specific queries after reading the first results. It can add the user's general location, inferred from their IP address, and, if memory is on, details from saved memories. OpenAI's article links Microsoft's and Shopify's privacy policies for the providers involved. It also keeps its own index, which its offline search mode uses.
Three things follow from that.
- **Be indexed where ChatGPT looks.** Microsoft is one of the providers OpenAI names, so being crawled and indexed by Bing is a sensible baseline. Verify the site in Bing Webmaster Tools, submit an XML sitemap of canonical URLs, and use IndexNow so new and changed pages are picked up quickly. OpenAI does not say how much of each answer comes from Bing, so treat this as one input, not the whole system.
- **Write for the query ChatGPT will generate.** A buyer types a long, chatty prompt. ChatGPT turns it into short search queries like "[category] pricing 2026" or "[category] for [industry]". You need a page that is a strong result for those short queries. A page per buying question beats one page that mentions everything.
- **Expect local and personal variation.** Location and memory change the rewritten query, so two people asking the same thing can get different sources. This matters when you test, below.
For the page-level rules, [Bing's Webmaster Guidelines](https://www.bing.com/webmasters/help/webmaster-guidelines-30fba23a) are the most detailed publisher guidance we found from a provider OpenAI names. They cover Copilot and grounding as well as search, and say SEO fundamentals still apply to AI answers. Several of the rules read like a ChatGPT checklist: put essential information near the top, make facts explicit and visible on the URL itself, keep each URL on one topic, use consistent names for people, organisations and products, and make sure structured data matches what the page shows. Bing also says markup does not guarantee grounding or citations. These are Microsoft's rules for Microsoft's products, but they line up with what OpenAI's own articles imply.
### Step 3: publish answer-first pages
Once a page is retrieved, the model has to lift something from it. The pages that get quoted state the answer in the first two sentences, use a heading phrased the way a buyer asks the question, and give facts a model can repeat without hedging: a price, a timeline, a named method, a dated figure. A paragraph of adjectives gives it nothing to quote.
The retrieval, extraction and verification steps behind this are in [how ChatGPT decides which companies to recommend](https://gigabit.agency/insights/how-ai-assistants-choose-recommendations/). The short version for a page you are writing today:
1. **Answer in the opening.** The first two sentences under the H1 should still make sense if quoted alone.
2. **One question per page.** Pricing, comparison, "is X right for Y" and "how does X work" are separate pages.
3. **Numbers with dates.** "Agents from $8,000 flat, as of October 2026" can be quoted. "Affordable" cannot.
4. **An FAQ written in buyer language.** Each answer complete in one paragraph, and the same facts as the body.
5. **Server-rendered.** If the answer only appears after JavaScript runs, assume the crawler never saw it.
### Step 4: keep your entity facts consistent
ChatGPT has to be confident it knows which company it is naming. If your site, your Organization schema, LinkedIn and directory listings use different names, addresses or descriptions, the safer move for a model is to leave you out. The same applies to your own pages: if the pricing page, an old blog post and your llms.txt disagree on a price, there is no figure it can defend.
Fix this in one sitting. Write down your legal name, trading name, one-sentence description, address and headline prices. Then make every surface match: site header and footer, About page, schema, LinkedIn, Google Business Profile, Bing Places and the two or three directories that matter in your category. The full list of conditions that remove a company from answers, with how to test each, is in [what gets a company excluded from AI answers](https://gigabit.agency/insights/what-gets-a-company-excluded-from-ai-answers/).
### Step 5: earn third-party mentions
This is the step most sites skip, and the one that follows directly from how ChatGPT searches. When a buyer asks for the best firm for a job, the rewritten query returns roundups, comparison pages, directories, review sites and forum threads. If those pages do not mention you, an answer built from them cannot either, however good your own site is. Your domain describing itself is also the weakest evidence a model has.
So the work looks like PR more than SEO. Get listed in the directories your buyers use. Contribute real data or expertise to the publications that write the roundups. Give a comparison site the facts it needs to include you accurately. Publish original numbers others want to cite. Bing's guidelines are explicit about the shortcuts: link buying, artificial promotion, keyword stuffing and content written to trigger AI responses, including attempts to manipulate language models, can reduce visibility or get a site removed. Do not hide instructions for AI in your pages.
An llms.txt file is not a substitute for any of this. It is cheap to add and no major assistant has said it reads it as a ranking input. We explain what it does and does not do in [llms.txt explained](https://gigabit.agency/insights/llms-txt-explained/).
### Step 6: measure with a fixed set of prompts
You cannot manage a citation rate you do not track. The method that works:
1. **Write 15 to 25 buyer questions** in the words buyers use, taken from sales calls and support tickets. Leave your brand name out. "What is [your company]" measures recognition, not whether you get recommended.
2. **Run each one in ChatGPT at least three times,** spread over days. OpenAI says people who are not signed in can use web search. Where that works in your country, test logged out. Otherwise use a clean account with memory off, because saved memories feed the rewritten query.
3. **Record four things per answer:** whether you were named, whether your page was cited as a source, which competitors were named, and which sources were cited. The source list tells you which third-party pages to get onto.
4. **Compute the share of answers that name you,** and the questions where you appear in every run. Repeat monthly on the same panel.
For a sense of scale, our [own benchmark](https://gigabit.agency/insights/what-is-a-good-ai-citation-rate/) put a fixed panel of buying questions about AI agencies through Google AI Mode and Bing Copilot Search; the most-named firm on Google appeared in 20% of answers. That study did not cover ChatGPT, and your category will differ, so judge your number against the leader on your own panel. What a good rate looks like, and why one run proves little, is in [what is a good AI citation rate](https://gigabit.agency/insights/what-is-a-good-ai-citation-rate/). If you want to do the first pass by hand this afternoon, follow [check your AI visibility in 30 minutes](https://gigabit.agency/insights/check-your-ai-visibility-in-30-minutes/).
Two more signals help. In your analytics, segment referral visits whose source contains chatgpt.com and watch which landing pages they reach. And expect clicks to understate visibility: Bing's guidelines make the same point, that content can appear as a citation without producing a click.
### ChatGPT SEO and LLM SEO: one job under two names
ChatGPT SEO and LLM SEO are the names people search for this work; GEO, generative engine optimization, is the name the research uses. They describe the same job: being named and cited when an AI assistant answers a buyer's question. It starts from ordinary SEO, because ChatGPT, Copilot and Google's AI answers all retrieve pages through search before they write, and then changes the target from a position on a results page to a share of answers. Everything in steps 1 to 6 applies whichever assistant you care about, with the crawler names and index sources changing per vendor.
### What not to spend money on
- **A promised placement.** OpenAI says placement is not guaranteed. A vendor guaranteeing a ChatGPT position is promising something OpenAI does not.
- **Mass-produced AI pages.** Bing's guidelines say large-scale generated content without editorial review may be excluded from indexing. It also gives a model nothing original to cite.
- **Blocking GPTBot to "protect" search visibility, or allowing it to "boost" it.** OpenAI says the settings are independent. Decide GPTBot on training grounds alone.
- **Rewriting everything at once.** Fix crawler access and entity facts first, then the five to ten pages that answer your most valuable questions, then go after third-party mentions.
### When to do it yourself, and when to get help
Steps 1 to 4 are a few days of work for someone who can edit robots.txt, read server logs and rewrite pages. Do them yourself first. Step 5 is slower, because it depends on other people, and step 6 only pays off if someone runs the same panel every month and acts on it.
Start with a baseline. Our [free AI Citation Audit](https://gigabit.agency/tools/geo-citation-audit/) runs your category's buying questions through the major assistants and shows where you are named, who is named instead, and where the answer is wrong. If you want the whole program run for you, [Ascent](https://gigabit.agency/services/geo/) is our 12-month growth program: site fixes, answer-first content, entity clean-up, third-party coverage and monthly citation-rate reporting, from $4,000 a month with no upfront fee. The exact monthly figure is set on a scoping call after the audit.

### FAQ
**How do I rank in ChatGPT?**
Allow OpenAI's OAI-SearchBot in robots.txt and in your CDN or firewall, which OpenAI says makes a site eligible for ChatGPT search. Then make sure you appear in the search results ChatGPT reads, by being indexed (including in Bing, one of the providers OpenAI names), publishing one answer-first page per buying question, keeping your company facts identical everywhere, and earning mentions on the roundups and directories that answer category questions. OpenAI says placement is not guaranteed and publishes no list of ranking factors, so measure a citation rate on a fixed set of buyer questions and adjust.
**Does ChatGPT search use Bing?**
Partly. OpenAI says ChatGPT search sometimes partners with other search providers, rewrites the user's question into targeted queries for them, and links Microsoft's and Shopify's privacy policies for those providers. It also keeps its own index, built by OAI-SearchBot. OpenAI does not say how much of an answer comes from each source, so being indexed in Bing is a sensible baseline but not the whole picture. Checked on 23 October 2026.
**Does blocking GPTBot remove my site from ChatGPT search?**
No. OpenAI's crawler documentation says each setting is independent. GPTBot gathers content that may be used to train its models, and disallowing it signals you do not want that. OAI-SearchBot is the crawler for ChatGPT search, and blocking it is what keeps your pages out of search answers. ChatGPT-User makes visits on a user's behalf and OpenAI says it does not decide what appears in search.
**How long does it take to show up in ChatGPT search?**
OpenAI says search-related robots.txt changes take about 24 hours to take effect. Being crawled and cited takes longer and varies by page: OpenAI says there is no refresh guarantee for a specific URL in its index, and new or rarely visited pages are more likely to be missing or stale. Being named on competitive buying questions also depends on third-party mentions, which take months to build.
**What is LLM SEO, and how is it different from SEO?**
LLM SEO, or ChatGPT SEO, is the work of getting AI assistants to name and cite your company in their answers. It builds on SEO, because assistants like ChatGPT retrieve pages through search before they answer. The differences are what success means and what gets quoted: there is no list of ten results, so you track the share of buyer questions where you are named, and pages win when they state a checkable fact in the first two sentences.

## AI in logistics: where it pays off, and what it needs first
URL: https://gigabit.agency/insights/ai-in-logistics/
Published 2026-10-23

**AI in logistics pays off on the high-volume, rule-bound work that arrives as unstructured text: quote requests by email, orders and load tenders, bills of lading and proofs of delivery, status checks, missed pickups and "where is my freight" questions.** The clearest public evidence is from C.H. Robinson, which says it moves 37 million shipments a year and reported in April 2025 that its AI agents had [performed more than 3 million shipping tasks](https://www.chrobinson.com/en-us/about-us/newsroom/press-releases/2025/ai-performs-over-three-million-shipping-tasks/). AI pays off least where the data underneath is messy or the decision carries real money and judgment, such as contract pricing or tendering a load to a carrier you have never used. This guide covers where it works, with sourced figures only, what it needs before it works, the risks, and a table for deciding where to start.
### Where AI pays off in logistics operations
Most of the published numbers come from one company, so read them as what is possible at very high volume, reported by the company itself. A broker moving a few hundred loads a month will not see the same totals, but the kinds of work are the same.
### Quoting from email
Price requests arrive as free-text emails in every format a customer can think of. In May 2024 C.H. Robinson said it [received over 11,000 emails a day](https://www.chrobinson.com/about-us/newsroom/press-releases/2024/technology-breaks-a-decade-old-barrier-to-automation) asking for truckload pricing, and that its AI was answering 2,000 customer quote requests a day in an average of 2 minutes 13 seconds. The model reads the email, works out the mode (truckload, LTL, intermodal or air) and returns a price. By April 2025 the company said it had passed 1 million AI-delivered quotes. If you build the same thing, keep the model on the reading. The number should come from your own rate logic, never from the model.
### Order entry and load acceptance
Once a customer accepts a quote, someone has to key the order into the TMS. C.H. Robinson reported [1 million orders processed by AI](https://www.chrobinson.com/en-us/about-us/newsroom/press-releases/2025/ai-performs-over-three-million-shipping-tasks/) by March 2025, and more than 5,200 customers getting loads accepted in under 90 seconds, against waits of up to four hours for a person to reach the email. Its agents had also set more than 1 million pickup and delivery appointments. These are good first projects because the output is a structured record that can be checked against the email it came from.
### Exceptions such as missed pickups
Exception handling is where agents start to do more than extract data. In January 2026 C.H. Robinson described [agents that chase missed LTL pickups](https://www.chrobinson.com/en-us/about-us/newsroom/press-releases/2026/chrobinson-lean-ai-agents-for-ltl-pickup-efficiency/) and decide how to keep the freight moving; the company says they can make 100 calls and 100 decisions at once. It reported that 95% of missed-pickup checks were automated, saving more than 350 hours of manual work a day, with 42% fewer unnecessary return trips and freight moving up to a day sooner. The same release says a fleet of more than 30 agents also handles quotes, orders, freight classification, shipment tracking and proof of delivery.
### Documents: BOLs, PODs and rate confirmations
Bills of lading, proofs of delivery, rate confirmations and carrier invoices are the paperwork every broker reconciles by hand. A model can read a scanned or photographed document, pull out the load number, dates, signatures, piece counts and exception notes, match it to the load in the TMS, and flag anything that does not match for a person. C.H. Robinson lists proof of delivery among its agents' jobs but has not published a figure for it, so we will not quote one.
The closest Gigabit evidence is from an e-commerce operation, not a broker. In our [operations automation for a four-channel retailer](https://gigabit.agency/proof/ai-workflow-automation-ecommerce-ops/), a model parses wildly inconsistent vendor reply emails into structured ship dates and tracking numbers, and vendor PO follow-up dropped from 5 hours a week to 45 minutes. The same project processes 85% of returns with no manual handling and runs a nightly inventory reconciliation that lifted inventory accuracy from 94.2% to 99.1%. The pattern carries straight over to carrier paperwork: extract, match to the record, act on clean cases, route the rest to a person.
### Track-and-trace updates and customer questions
Shippers want to know where their freight is, and most of the answers already sit in carrier status messages, tracking APIs and the TMS. An agent can read those, write the update and answer "where is my load" without a person looking it up. In a support project for a direct-to-consumer brand, [order status was 23% of ticket volume](https://gigabit.agency/proof/ai-automation-ecommerce-support/); the AI agent we built resolves 64% of all tickets with no human and cut first response time from 4.2 hours to 6 minutes. That client sells home goods, not freight, but the question it answers most is the one logistics customers ask most. The rule that matters: the agent reports what the data says and never guesses an ETA.
### Carrier matching
Ranking carriers for a lane by history, price, on-time record and current capacity is a reasonable AI task. Booking them without a person is a different matter, covered under risks below. Treat carrier matching as a recommendation an agent prepares and a person approves, at least for any carrier you have not worked with before.
### The wider evidence
Outside freight brokerage, the most cited figure is McKinsey's: early adopters of AI-enabled supply-chain management [improved logistics costs by 15%, inventory levels by 35% and service levels by 65%](https://www.mckinsey.com/industries/metals-and-mining/our-insights/succeeding-in-the-ai-supply-chain-revolution) compared with slower-moving competitors. That is from April 2021, it covers supply chains broadly rather than brokerage, and early adopters tend to be better-run companies to begin with. Use it as direction, not a forecast for your business.
### What AI needs before it works
Every example above has the same thing underneath: a system of record with clean, structured data that an agent can read from and write to. Without that, an AI model is reading emails into nowhere.
- **A TMS or WMS with an API.** Loads, statuses, customers, carriers and documents need IDs the agent can look up and update. Spreadsheets and shared inboxes do not count.
- **Integrations to where the work arrives.** The quoting inbox, EDI load tenders and shipment status messages (X12 204 and 214 in the US), carrier tracking APIs or ELD feeds, load boards and accounting.
- **Reference data you trust.** Customer records, lanes, accessorials, rate tables and carrier profiles. If two people would price the same lane differently, an agent will too.
- **Written rules.** Margin floors, which customers get auto-quotes, when an exception goes to a person, and who approves what.
- **Logging and tests.** Every agent action recorded against the load, and a test set of real past emails and documents to measure accuracy before and after each change. Our [AI agent eval metrics](https://gigabit.agency/insights/ai-agent-eval-metrics/) guide covers what to measure.
This is where many brokers actually start. The founder behind our [logistics SaaS MVP](https://gigabit.agency/proof/saas-mvp-logistics-platform/), a 15-year freight veteran, saw mid-size brokers handling 50 to 500 shipments a month running on 2000s-era TMS software, Excel, email chains and phone calls. Gigabit built the platform in 11 weeks for $48,000: shipment creation, multi-stop routing, carrier assignment, document upload, real-time tracking on a map, automated notifications, a KPI dashboard and per-shipment billing. It reached 14 paying customers and $19,000 in monthly recurring revenue by month 6, and the founder closed a $2.1M Series A four months after launch. That build was a platform, not an AI agent. It is the structured data layer that agents need, and the right first step for a broker who does not have one.
### The risks, and how to contain them
- **A wrong quote commits margin.** Set a margin floor the agent cannot cross, send quotes above a value threshold or for new customers to a person, and log the inputs behind every price.
- **Misread documents.** A model can misread a smudged scan with full confidence. Match every extracted field to the load record, set a confidence threshold, and route mismatches and exception notes such as damage or shortage to a person.
- **Freight fraud.** An agent that books carriers or changes payment details automatically is a target. Check authority and contact details against FMCSA's records and your own carrier profiles before any tender, and never let an agent change remit-to details. The [FMCSA's fraud alerts page](https://www.fmcsa.dot.gov/registration/fraud-alerts) lists schemes that impersonate FMCSA itself, including its SAFER lookup, which shows how much of this is going on.
- **Made-up status updates.** Tracking messages must come from carrier data. If the data is missing, the agent says so and escalates.
- **Integrations that drift.** Carrier portals, APIs and email formats change. Monitor failure rates and alert a person, the way you would for any production system. See [why AI systems break after launch](https://gigabit.agency/insights/why-ai-systems-break-after-launch/).
- **Customer data in prompts.** Shipper names, addresses and rates go into the model. Check your model provider's data terms and keep the processing in accounts you control.
The general rule is human review where a mistake costs money or trust, and full automation where the output can be checked against a record. We cover how to place those checkpoints in [human-in-the-loop for AI agents](https://gigabit.agency/insights/ai-agent-human-in-the-loop/).
### Decision table: where to start
| Task | Good first AI project? | What it needs | Keep a person on |
| --- | --- | --- | --- |
| Quote requests by email | Yes, if you already have rate logic | Quoting inbox, rate tables, margin rules | New customers and quotes above a threshold |
| Order entry from email or PDF | Yes | TMS API, customer records | Low-confidence fields |
| BOL and POD processing | Yes | Document storage, load matching | Mismatches, damage and shortage notes |
| Track-and-trace updates | Yes | Status messages, tracking APIs or ELD feeds | Delays that break a delivery promise |
| Missed pickups and other exceptions | Yes, for triage and outreach | Status feed, carrier and shipper contacts | Rebooking that adds cost |
| "Where is my load" questions | Yes | Read access to the TMS | Claims and disputes |
| Carrier matching | As a recommendation | Lane history, carrier profiles, FMCSA checks | Every tender to a new carrier |
| Contract pricing and bids | No, analysis support only | Clean historical data | The decision |
Pick the row with the most volume and the cleanest data, prove it on real past emails or documents, then move to the next. If your TMS already ships one of these features, use it. Building your own is worth it when the vendor's version does not fit your process or your data lives across several systems.
### Where Gigabit fits
- **One workflow, such as email quoting or POD matching:** our [AI Quick Win Sprint](https://gigabit.agency/services/ai-agents/) is $8,000 fixed over 3 weeks, with integration to up to three tools, error handling and 30 days of light support.
- **A production agent, such as an exceptions or customer-status agent:** [Gigabit Agents](https://gigabit.agency/services/ai-agents/) from $8,000, flat per agent.
- **Not sure where to start:** the [AI Transformation Sprint](https://gigabit.agency/services/ai-transformation/) is $25,000 over 2 weeks. It maps your workflows, models the return and ends in a scoped, ready-to-build agent, and the fee credits toward a Gigabit Agents build.
- **No usable system of record yet:** start with the platform, as the [logistics MVP](https://gigabit.agency/proof/saas-mvp-logistics-platform/) did, through [custom software development](https://gigabit.agency/services/custom-software-development/).
If your volume is a few dozen loads a week and one person handles them comfortably, you probably do not need any of this yet. Clean up the data and the inbox first. The agents will be cheaper to build later.

### FAQ
**How is AI used in logistics?**
Mostly on high-volume text and document work: reading quote requests and orders from email, processing bills of lading and proofs of delivery, sending tracking updates, chasing exceptions such as missed pickups, and answering customer status questions. C.H. Robinson reported in April 2025 that its AI agents had performed over 3 million shipping tasks, including over 1 million price quotes. AI is weaker on decisions that need judgment and carry money, such as contract pricing.
**What does AI do in freight brokerage?**
It reads customer emails and returns quotes from the broker's pricing engine, enters orders into the TMS, books appointments, chases missed pickups, matches paperwork to loads and keeps shippers updated. In May 2024 C.H. Robinson said its AI answered 2,000 quote requests a day in an average of 2 minutes 13 seconds. Carrier selection is usually kept as a recommendation that a person approves.
**Do I need a TMS before using AI agents?**
You need a system of record with an API, whether a commercial TMS or your own platform. An agent has to look up loads, customers and carriers and write results back. If your operation runs on spreadsheets and email, building that layer comes first; agents built on top of a shared inbox have nowhere reliable to put their work.
**Can AI agents book carriers automatically?**
They can, but it is the riskiest place to remove people. FMCSA keeps a running list of fraud alerts, including schemes that impersonate FMCSA's own carrier lookup. A safer pattern is an agent that ranks carriers and checks their authority and contact details, with a person approving any tender to a carrier you have not used, and no agent permitted to change payment details.
**How much does an AI agent for logistics cost?**
With Gigabit, a single workflow such as email quoting costs $8,000 fixed over 3 weeks through the AI Quick Win Sprint, and a production agent starts at $8,000 flat per agent. The larger cost is often the data and integration work underneath: an agent needs a TMS with an API and clean customer, lane and rate data before it can work reliably.

## n8n templates for business: which to start from, and what to fix before production
URL: https://gigabit.agency/insights/n8n-templates-for-business/
Published 2026-10-23

**n8n templates are ready-made workflows you import into your own n8n instance and adapt, and the library is large: on 23 October 2026 [n8n's template page](https://n8n.io/workflows/) listed 13,652 of them, grouped under AI, Sales, IT Ops, Marketing and Document Ops.** n8n describes them as pre-built workflows designed by n8n and community members. That is their strength and their limit. A template shows you a working shape in minutes, but it was written for someone else's tools, data and tolerance for failure. This guide covers which kinds of template are worth starting from for a business, how to bring one in, and what has to change before it touches real orders, customers or money.
### What an n8n template is, and is not
A template is a workflow file: the nodes, how they connect, and their settings, with your credentials left blank. In the editor you open the library from the Templates icon, pick one and import it onto the canvas. [n8n's documentation](https://docs.n8n.io/workflows/templates/) is plain about what happens next: you may need to fill in credentials and adjust the configuration to suit your needs.
Most of the library is written by the community. n8n lists [verified creators](https://n8n.io/creators/) with how many templates each has published, and invites people to submit templates and earn through its affiliate programme. n8n does not publish how templates are reviewed before they are listed, so treat a template as you would code from a public repository: useful, often clever, and yours to check.
### The template families worth starting from
For a business, the useful templates fall into a handful of families. Search the library for these patterns rather than for a specific template name; names and versions change often.
| Family | What it does | What usually needs changing |
| --- | --- | --- |
| Lead capture to CRM | Form or inbox to HubSpot, Salesforce or Pipedrive, with enrichment and routing | Your field mapping, duplicate rules and who gets notified |
| Order and returns operations | Store events to fulfilment, refunds, stock and customer emails | Idempotency, so a retried webhook does not refund twice |
| Document and invoice extraction | Email attachments or uploads to structured fields in the ERP or a sheet | Confidence thresholds and a review queue for low-confidence fields |
| Support triage | Tickets classified, tagged and routed, sometimes with a drafted reply | Your categories, escalation rules and a human approval on replies |
| Scheduled reporting | Data pulled from several systems into a daily or weekly summary | Time zones, missing-data handling and who receives it |
| AI agent starters | An AI Agent node with a model, memory and a few tools | Fewer tools, tighter instructions and a test set of real cases |
| Internal alerts | Failures, thresholds or new records posted to Slack, Teams or email | Alert fatigue: who needs it, and how often |
Templates in the first five families tend to survive contact with a business with modest changes, because the logic is simple and deterministic. AI agent templates need the most work. They demonstrate the AI Agent node well, but they often give the agent more tools than it needs and no way to tell whether its answers are right.
### How to import a template without surprises
1. **Read it before you import it.** Open the template page and look at every node. Note which apps it connects to and which credentials it will ask for.
2. **Check for community nodes.** Some templates use nodes published by the community rather than n8n. On self-hosted n8n, whether those can be installed at all is controlled by settings such as `N8N_COMMUNITY_PACKAGES_ENABLED` and `N8N_UNVERIFIED_PACKAGES_ENABLED`. Decide your policy before a template decides it for you.
3. **Import into a test instance or project first,** with test credentials that cannot touch production data.
4. **Run it on real past cases,** not the sample data. Ten real orders, tickets or invoices will show more than any description.
5. **Only then connect production credentials,** with the narrowest permissions the workflow needs.
If you run n8n yourself, two settings are worth knowing. `N8N_TEMPLATES_ENABLED` turns the template library on or off for your users and is on by default. `N8N_TEMPLATES_HOST` points the library at your own template API instead of n8n's, which some companies use to offer only approved, internal templates; n8n's [deployment settings](https://docs.n8n.io/hosting/configuration/environment-variables/deployment/) note that your API must provide the same endpoints and response structure as n8n's.
### Eight fixes a template needs before production
Nearly every template we have taken over or reviewed was missing some of these. They are the difference between a demo and a system that runs thousands of times a week.
1. **Error handling.** An error workflow that alerts a named person, and retries with backoff on steps that call external APIs.
2. **Idempotency.** A check that a record has not already been processed, so a retried webhook or a re-run does not create a second invoice or refund.
3. **A dead-letter path.** Records that cannot be processed go to a queue to be fixed and replayed, instead of disappearing.
4. **Validation between steps.** Required fields checked, and AI output parsed into a fixed structure before anything is written.
5. **Human approval on costly actions.** Refunds, stock changes and outbound messages to customers wait for a person until the workflow has earned trust.
6. **Rate limits and batching.** Templates often loop over records one call at a time; production volumes hit API limits.
7. **Credentials with least privilege.** A dedicated service account per workflow, not an admin's personal login.
8. **Version control and a test set.** The workflow exported to Git, and a set of past cases it must pass after every change.
### Template, adapt or build?
Use a template as-is for internal, low-stakes jobs: alerts, a weekly report, saving attachments to a folder. Adapt a template when the shape fits and the eight fixes above are a day or two of work. Build from scratch when the process touches money, stock or customers at volume, or when the logic would need so many changes that the template is no longer saving time. We often use a template as the sketch and rebuild the parts that matter in tested code alongside n8n.
For scale, five automations we built on self-hosted n8n for a $28M multi-channel e-commerce operation cut its team's manual work from 32 hours a week to 4.5 and now process 85% of returns with nobody involved. None of them started as a template, but each one has every item in the list above, including a human approval kept on order exceptions on purpose. The full write-up is in the [case study](https://gigabit.agency/proof/ai-workflow-automation-ecommerce-ops/).
If you would rather hand over the workflow, our [n8n team](https://gigabit.agency/platforms/n8n/) builds and runs it on your own instance, and our [business process automation](https://gigabit.agency/services/business-process-automation/) page covers the same work across Make, Zapier and code. One workflow, scoped, built and deployed with up to three tool integrations, is the AI Quick Win Sprint at $8,000 fixed over three weeks. What n8n itself costs at different volumes is in our [n8n pricing guide](https://gigabit.agency/insights/n8n-pricing/).

### FAQ
**How many n8n templates are there?**
On 23 October 2026, n8n's workflow template page listed 13,652 templates, grouped under categories including AI, Sales, IT Ops, Marketing and Document Ops. The number grows every week, because most templates are submitted by community creators.
**Are n8n templates free?**
The templates in n8n's own library can be browsed and imported into your n8n instance without paying for the template. You still pay for running n8n (n8n Cloud, or your own server for the self-hosted Community edition) and for any paid apps or AI models the workflow calls.
**Are n8n templates safe to use?**
Treat them like code from a public repository. Most are written by community members and n8n does not publish how they are reviewed. Read every node before importing, check whether it needs community nodes, import into a test environment first, and give it credentials with the narrowest permissions it needs.
**Can I use an n8n template in production?**
Yes, after adapting it. Production workflows need error alerts, retries, idempotency checks, a queue for records that fail, validation between steps, human approval on costly actions, rate-limit handling and version control. Most templates have some of these, few have all of them.
**Can I hide the template library from my team?**
On self-hosted n8n, yes. The N8N_TEMPLATES_ENABLED setting turns the library on or off, and N8N_TEMPLATES_HOST points it at your own template API, so you can offer only approved internal templates. Your API has to follow the same endpoints and response structure as n8n's.

## What are agentic workflows? A plain definition, six patterns and when to use one
URL: https://gigabit.agency/insights/what-are-agentic-workflows/
Published 2026-10-23

**Agentic workflows** are processes in which a large language model does more than answer once. It works through a task in steps: it plans, calls tools such as a CRM or a search index, checks its own output and revises it, and in the more autonomous versions decides for itself what to do next. That puts them on a spectrum. At one end is fixed automation, where every step is written in advance. At the other is an autonomous agent that chooses its own steps until the job is done. Most systems that work reliably in a business sit nearer the fixed end than the marketing suggests. This guide defines the terms using two primary sources, walks through the six common patterns with a business example for each, and covers when not to build one at all.
### Workflow vs agent: the distinction that matters
Anthropic's engineering guide [Building effective agents](https://www.anthropic.com/engineering/building-effective-agents) (December 2024) gives the clearest split. A **workflow** is a system where models and tools are orchestrated through "predefined code paths": a developer decided the order of steps, and the model fills in the judgement at each one. An **agent** is a system where the model directs its own process and tool use, deciding which step comes next and when it is finished. Anthropic groups both under the name *agentic systems*.
Andrew Ng popularised the term from a different angle. In a [March 2024 letter in The Batch](https://www.deeplearning.ai/the-batch/how-agents-can-improve-llm-performance/) he contrasted asking a model for an answer in one pass with letting it iterate: outline, research, draft, review, revise. He named four design patterns for that iteration (reflection, tool use, planning and multi-agent collaboration) and reported that on the HumanEval coding benchmark, GPT-3.5 scored 48.1% in one pass and GPT-4 67.0%, while GPT-3.5 wrapped in an agent loop reached up to 95.1%. The point was that the loop around a model can matter more than which model you pick. Those numbers are from early 2024 models, so read them as a demonstration of the effect, not a current benchmark.
Put the two together and you get a usable definition. A workflow becomes agentic when a model's output decides something about the flow, whether that is which branch to take, whether a draft is good enough, or which tool to call next. The more of those decisions the model makes, the more agentic it is.
### The spectrum, from fixed automation to autonomous agent
| Level | Who decides the next step | Example | Predictability | Cost per run |
| --- | --- | --- | --- | --- |
| Fixed automation | Your rules | New Shopify order creates a row in the warehouse sheet | Identical every time | Almost nothing |
| One model call in a fixed flow | Your rules; the model fills one field | Classify a support email's language, then route by rule | High | Low |
| Chained or routed workflow | Your code, using the model's outputs | Extract invoice fields, validate in code, draft a payables note | High, if each step is checked | Low to medium |
| Orchestrated workflow | A model splits the work; code assembles it | Answer a 60-question security questionnaire from your policy library | Medium | Medium |
| Tool-using agent with limits | The model, within a fixed tool set and step cap | Resolve an order exception by checking order, carrier and stock systems | Lower; needs evals | Medium to high |
| Open-ended autonomous agent | The model, with broad tools | A coding agent working on a repository | Lowest | Highest |
Moving down the table buys flexibility and pays for it in predictability and money. Anthropic says this directly: agentic systems trade latency and cost for better task performance, and you should only make that trade when it pays.
### The six patterns, with a business example for each
The first five are the workflow patterns Anthropic describes; the sixth is its agent pattern. The examples are ours, chosen from the kinds of jobs mid-sized companies bring us.
### 1. Prompt chaining
**How it works:** the task is split into fixed steps, each model call working on the output of the last, with checks in code between them. **Business example:** a supplier invoice arrives as a PDF. Step one extracts the fields as structured data. Code then checks that line items add up to the total and the supplier exists. Step two drafts the approval note for accounts payable only if the checks pass. **Use it when** a task splits cleanly into stages and you want to check each one. **Watch for** errors carrying forward: without the check between steps, a misread total becomes a confident wrong note.
### 2. Routing
**How it works:** a first step classifies the input and sends it to a specialised path, which may be a different prompt, a different model, or a person. **Business example:** inbound support email is sorted into order status, returns, product questions and complaints. Order status goes to a path that looks the order up; complaints go straight to a human. Anthropic also suggests routing easy questions to a smaller, cheaper model. **Watch for** misclassification. In a support agent we built for a Shopify Plus brand, the agent replies on its own only when classification and retrieval both clear confidence thresholds, and everything else goes to a person; it now resolves 64% of tickets that way ([case study](https://gigabit.agency/proof/ai-automation-ecommerce-support/)).
### 3. Parallelisation
**How it works:** several model calls run at once and their results are combined. Anthropic distinguishes *sectioning*, where independent parts of a task run side by side, from *voting*, where the same task runs several times and the answers are compared. **Business example of sectioning:** a contract review where separate calls check payment terms, liability caps and termination clauses against your playbook, while another screens the document for anything that should stop the process. **Business example of voting:** three independent checks on whether a product listing breaks marketplace rules, with a flag raised if any one says yes. **Watch for** cost, which multiplies with each parallel call.
### 4. Orchestrator-workers
**How it works:** a central model reads the task, decides how to break it up, hands the pieces to worker calls and assembles the result. Unlike parallelisation, the subtasks are not known in advance. **Business example:** answering a customer's security questionnaire. The orchestrator groups the questions, each worker drafts answers from the relevant policy documents, and the orchestrator assembles them and lists the questions it could not answer for a person to complete. **Watch for** the orchestrator inventing work. Cap the number of workers and log the plan it made so you can see why.
### 5. Evaluator-optimiser
**How it works:** one call produces a draft, a second call judges it against explicit criteria and sends feedback, and the loop repeats until the draft passes or a limit is hit. This is the reflection pattern in Ng's list. **Business example:** product descriptions written against a brand guide and a list of banned claims, where the evaluator checks tone, length and compliance and returns specific fixes. **Use it when** you can write down what good looks like. **Watch for** loops that never converge. Set a maximum number of rounds, and check that the evaluator agrees with your human reviewers before you trust it; our guide to [calibrating an LLM judge](https://gigabit.agency/insights/calibrate-llm-as-judge-for-ai-agents/) covers how.
### 6. The tool-using agent loop
**How it works:** the model is given a goal and a set of tools, then repeats a cycle of choosing an action, calling a tool, reading the result and deciding what to do next, until it finishes or hits a stop condition. **Business example:** an operations agent handling an order exception. It looks the order up, checks the carrier's tracking, checks stock at a second warehouse, and proposes a fix (reship, refund or wait) for a person to approve with one click. **Use it when** the steps genuinely cannot be known in advance. **Watch for** compounding error: small per-step failure rates multiply over a long loop, which is why we split or gate long agents, as explained in [when a multi-step AI agent fails in production](https://gigabit.agency/insights/when-a-multi-step-ai-agent-fails-in-production/).
### When not to use an agentic workflow
Anthropic's first recommendation is to find the simplest solution that works and add complexity only when it clearly improves results. For many tasks, it notes, a single well-prompted model call with retrieval and a few examples is enough. We would go a step further: for a lot of business work, no model is needed at all.
- **The input is already structured.** If the data arrives in fields, rules route it faster, cheaper and identically every time.
- **Nobody can define a correct answer.** If you cannot write down what good output looks like, you cannot test an agent, and you will not know when it gets worse.
- **A mistake is expensive and cannot be checked.** Payments, legal commitments and medical decisions need a person in the path, at which point a simpler draft-and-approve workflow often does the job.
- **The volume is tiny.** Ten cases a week rarely justify the build and the monitoring.
- **Your exception rate is low.** If rules already handle nearly everything, keep them. Our guide to [rules-based automation vs AI agents](https://gigabit.agency/insights/rules-based-automation-vs-ai-agents/) gives exception-rate thresholds for when the upgrade is worth it.
The best designs we see are hybrids: rules handle the predictable majority, and an agentic step takes only the cases the rules cannot.
### What makes an agentic workflow safe to run in production
None of this is specific to one pattern. Each point has its own guide on this site, so here is the short version and where to read more.
- **Evals before and after every change.** A fixed set of real past cases with known right answers, run automatically, with a pass mark that blocks a release. What to measure is in [AI agent eval metrics](https://gigabit.agency/insights/ai-agent-eval-metrics/), and how to build the set from your own history is in [building an eval set from support tickets](https://gigabit.agency/insights/build-ai-agent-eval-set-from-tickets/).
- **A person where it counts.** Approval on anything hard to undo, and a path to a human when confidence is low. The four supervision patterns are in [when an AI agent should stop and ask a human](https://gigabit.agency/insights/ai-agent-human-in-the-loop/).
- **Observability.** A trace of every step: the prompt, the tool calls, the outputs and the cost, so you can explain any single decision after the fact. Anthropic lists showing the agent's planning steps as one of its three core principles. The rest of the stack is in [the production AI agent stack](https://gigabit.agency/insights/production-ai-agent-stack/).
- **Cost caps.** A maximum number of steps or rounds per run, token limits per call and a spend limit at the provider. Where the money really goes is in [what an AI agent costs](https://gigabit.agency/insights/ai-agent-cost/).
- **Narrow, well-described tools.** Give the agent the fewest tools that do the job, with least-privilege access, and document each one as carefully as a public API; Anthropic calls this the agent-computer interface. Treat anything the agent reads from outside as untrusted, as covered in our [AI agent security framework](https://gigabit.agency/insights/ai-agent-security-framework/).
### Which pattern fits which job
A starting point, not a rule. Pick the lowest row on the spectrum that does the job.
| Job | Pattern | Why |
| --- | --- | --- |
| Extract fields from invoices or purchase orders | Prompt chaining | Fixed stages, each checkable in code |
| Sort inbound email or tickets by intent | Routing | One classification decides the path |
| Draft replies to common support questions | Routing plus a draft for approval | Easy cases automated, hard ones to people |
| Review contracts against a playbook | Parallelisation (sectioning) | Independent clauses checked side by side |
| Flag risky listings or content | Parallelisation (voting) | Several checks lower the miss rate |
| Answer RFPs or security questionnaires | Orchestrator-workers | The breakdown depends on the document |
| Write on-brand product copy | Evaluator-optimiser | Clear criteria to check against |
| Resolve order or shipping exceptions | Tool-using agent, with approval | Next step depends on what each system returns |
| Research a prospect before a sales call | Tool-using agent, read-only | Open-ended search, low risk |
| Move an order into the ERP | No model; rules | Structured input, must be exact |
### How agentic workflows get built
You can build the workflow patterns in an automation tool such as n8n, which has an AI Agent node and handles the plumbing; our guide to [n8n templates](https://gigabit.agency/insights/n8n-templates-for-business/) covers which ones suit a business and what to fix before they go live. You can also write them directly against a model provider's API in your own code. Anthropic recommends starting with the API directly, and warns that frameworks can hide the prompts and responses you need to see when debugging. Whichever you choose, the patterns are the same, and so is the work of testing and monitoring them.
### Getting one built
If you have a workflow in mind and want it in production, [Gigabit Agents](https://gigabit.agency/services/ai-agents/) builds and deploys a production agent for a flat fee from $8,000 per agent; the fee covers build, deployment and 30 days of tuning. If the job is a single automation with an AI step, the AI Quick Win Sprint is $8,000 fixed over 3 weeks. If you are not sure which workflow should come first, the AI Transformation Sprint ($25,000, 2 weeks) maps your workflows, models the return and ends with one scoped, ready-to-build agent. And if the honest answer is that rules will do, we will tell you that too.

### FAQ
**What is an agentic workflow?**
A process in which a language model works through a task in steps, planning, calling tools, checking and revising its work, rather than answering once. It is agentic to the degree that the model's output decides the flow: which branch to take, whether a draft is good enough, or which tool to call next.
**What is the difference between an AI workflow and an AI agent?**
In Anthropic's definitions, a workflow orchestrates models and tools through code paths a developer wrote in advance, while an agent lets the model direct its own process and tool use. Workflows are more predictable and cheaper; agents handle tasks whose steps cannot be known in advance.
**What are the main agentic workflow patterns?**
Anthropic describes prompt chaining, routing, parallelisation, orchestrator-workers and evaluator-optimiser as workflow patterns, plus the autonomous agent that uses tools in a loop. Andrew Ng's four design patterns, reflection, tool use, planning and multi-agent collaboration, cover similar ground from a different angle.
**What is an example of an agentic workflow in business?**
Invoice processing is a common one: a model extracts the fields, code checks the totals and supplier, and a second step drafts the approval note only if the checks pass. Support routing is another: a model classifies each email, answers routine ones when confidence is high and sends the rest to a person.
**When should you not use an agentic workflow?**
When the input is already structured, when nobody can define a correct answer to test against, when a mistake is expensive and cannot be checked, or when volume is tiny. Rules-based automation is cheaper and identical every time for work that fits it.
**Are agentic workflows the same as agentic AI?**
They overlap. Agentic AI is the broad label for systems that act with some autonomy. Agentic workflows are the concrete designs used to build them, from fixed chains with a model at each step to agents that choose their own steps.

## AI agent use cases: 20 jobs agents can do, rated for how well they work
URL: https://gigabit.agency/insights/ai-agent-use-cases/
Published 2026-10-23

The **AI agent use cases** that hold up in production share a shape. There is a repeated job with a knowable right answer, the systems involved can be reached through an API, and there is one point where a person checks the work before a mistake costs money or trust. Below are 20 of them, grouped by function. For each one we list the job, what the agent reads and writes, the human checkpoint, the metric that proves it is working, and a realism rating. Where Gigabit has a published case study for a use case, we link it and quote its numbers exactly as published. Where we do not, the row says nothing about results, because we have none of our own to show.
### How to read the ratings
- **Proven.** The pattern is well understood, the failure modes are known, and a narrow version can go live in a normal build with confidence gates in place.
- **Works with care.** It works when the scope stays narrow and the human checkpoint is real. Widen either one and quality drops quickly.
- **Still experimental.** Run it in shadow mode, where the agent drafts and a person does the job as before, until your own numbers say it is ready.
The ratings are our judgement from building agents, not an industry survey. A chatbot answers questions, while an agent reads from and writes to your systems; [AI agent vs chatbot](https://gigabit.agency/insights/ai-agent-vs-chatbot/) covers the difference. Several jobs below are better done by plain rules with a model only on the messy step, and [rules-based automation vs AI agents](https://gigabit.agency/insights/rules-based-automation-vs-ai-agents/) has the test for telling which.
### Customer support
| Use case | Reads and writes | Human checkpoint | Metric that proves it | Realism |
| --- | --- | --- | --- | --- |
| 1. Tier-1 ticket resolution | Reads tickets, order and shipping data, policies and product docs. Writes replies; can start returns, apply codes or change addresses | Below a confidence threshold the ticket goes to a person with the draft attached. Order changes ask the customer to confirm | Share of tickets resolved with no human touch, first response time, CSAT, reopen rate | Proven |
| 2. Agent assist: triage and drafts | Reads the ticket and customer history. Writes category, priority and a draft reply into the help desk | A person sends every reply | Handle time per ticket, share of drafts sent with light edits | Proven |
**Published proof for row 1.** In our [D2C support agent case study](https://gigabit.agency/proof/ai-automation-ecommerce-support/), over the first 90 days of full deployment the agent resolved **3,149 tickets a month (64%)** with no human involvement. First response time fell from **4.2 hours to 6 minutes** for AI replies, CSAT went from **3.9 to 4.5**, annual savings were **$217,200** and payback took **5.1 months**. The checkpoint was a pair of gates: the agent replied on its own only when intent confidence exceeded 0.85 and retrieval relevance exceeded 0.80. Below either, the ticket went to a person with the draft and context pre-loaded. Three of the human agents were moved to proactive outreach, VIP management and post-purchase work.
Row 2 is the safer first step when your help desk data is messy or your policies change weekly. It earns trust and produces the labelled data that row 1 needs. If you are weighing a custom build against a help-desk add-on, our [custom support agent vs Intercom Fin](https://gigabit.agency/compare/custom-support-agent-vs-intercom-fin/) comparison covers the trade.
### Sales and revenue
| Use case | Reads and writes | Human checkpoint | Metric that proves it | Realism |
| --- | --- | --- | --- | --- |
| 3. Inbound lead qualification and routing | Reads form fills, enrichment data and the CRM. Writes a score, an owner and a booked meeting | Reps see the reasoning behind each score. Disqualified leads are reviewed weekly | Time to first response, share of booked meetings that sales accepts | Proven |
| 4. Proposal and RFP first drafts | Reads past proposals, case material and the brief. Writes a draft document | A person owns every word that goes out | Hours per proposal, with win rate held steady | Proven |
| 5. Outbound prospecting | Reads target lists, enrichment and company sites. Writes personalised emails, handles replies, books meetings | Sequences and a sample of messages approved before sending. Hard caps on daily volume | Positive reply rate, meetings booked, bounce and spam-complaint rates | Works with care |
**Published proof for row 4.** In our [consulting firm case study](https://gigabit.agency/proof/ai-transformation-professional-services/), a proposal assistant built on the firm's knowledge base cut proposal creation from **4–8 hours to 45 minutes** of review. Row 5 is rated lower because the risk sits outside your systems: a badly tuned outbound agent can damage your sending domain faster than it books meetings, so deliverability limits matter as much as the writing.
### Finance and back office
| Use case | Reads and writes | Human checkpoint | Metric that proves it | Realism |
| --- | --- | --- | --- | --- |
| 6. Time entry suggestions | Reads calendar, email, documents and chat activity. Writes suggested time entries | Each person approves their own entries | Time capture rate | Proven |
| 7. Management reports with a written summary | Reads the warehouse or source systems. Writes scheduled reports with a short narrative | The numbers come from queries, never from the model. Someone reads the narrative before it goes to leadership | Report delivery time, errors found in the narrative | Proven |
| 8. Invoice capture and AP coding | Reads emailed invoices and purchase orders. Writes draft bills with GL codes in the accounting system | A person approves every bill before payment. Failed PO matches go to finance | Share of matched invoices needing no edits, exception rate, posting errors found at close | Works with care |
**Published proof.** In the same [consulting firm case study](https://gigabit.agency/proof/ai-transformation-professional-services/), a time entry assistant that suggests entries each morning for one-tap approval lifted time capture from **80–85% to 94%** (row 6). In our [e-commerce operations case study](https://gigabit.agency/proof/ai-workflow-automation-ecommerce-ops/), automated reports with AI-written narrative moved daily delivery from **10:30 AM to 7:00 AM** (row 7). We have no published case for row 8. It is rated lower because a wrong GL code is quiet: nobody notices until close, so the checkpoint has to be a real approval step rather than a glance.
### Operations and logistics
| Use case | Reads and writes | Human checkpoint | Metric that proves it | Realism |
| --- | --- | --- | --- | --- |
| 9. Returns processing | Reads the order, return policy and the customer's free-text reason. Writes the return record, label, confirmation and refund on delivery | Hard stop and human review when a policy or confidence check fails. Likely defects flagged to QA | Share processed with no touch, minutes per return, refund errors | Proven |
| 10. Inventory reconciliation | Reads stock levels across channels and the system of record. Writes corrections for known discrepancy causes | Only true exceptions reach a person | Inventory accuracy, hours per week | Proven |
| 11. Supplier PO follow-up | Reads purchase orders and vendor reply emails. Writes structured ship dates, tracking numbers and chasers | Escalates when a date slips past the threshold you set | Hours per week on follow-up, late POs caught before customers notice | Proven |
| 12. Order exception handling | Reads failed orders: address, payment, split shipments. Writes a classification and a prepared fix | A person clicks the fix | Minutes per exception | Proven |
| 13. Freight quotes and orders from email | Reads quote requests and rate data. Writes quotes and orders into the TMS | Pricing rules and margin floors in code. A person checks unusual lanes | Quote turnaround, quote-to-book rate | Works with care |
**Published proof for rows 9 to 12.** In our [e-commerce operations case study](https://gigabit.agency/proof/ai-workflow-automation-ecommerce-ops/), **85% of returns** processed with no human touch and per-return handling fell from **12–18 minutes to zero** on that slice. Inventory accuracy rose from **94.2% to 99.1%**, vendor PO follow-up dropped from **5 hours a week to 45 minutes**, and exception resolution fell from **25 minutes to 6 minutes** with a human approval deliberately kept on the final action. Those were n8n automations with GPT-4o-mini on the classification and email-parsing steps, which shows how much agent work is ordinary workflow with a model in the right place. The full write-up is in [where AI agents pay off first in ecommerce ops](https://gigabit.agency/insights/where-ai-agents-pay-off-first-in-ecommerce-ops/). For row 13 we have no case of our own; [AI in logistics](https://gigabit.agency/insights/ai-in-logistics/) covers the published freight-broker results.
### HR and IT
| Use case | Reads and writes | Human checkpoint | Metric that proves it | Realism |
| --- | --- | --- | --- | --- |
| 14. Internal knowledge assistant | Reads wikis, drives and past work, filtered by each user's access. Writes cited answers in Slack or Teams | Every answer cites its source. Low-confidence answers say so and name an owner | Repeat questions reaching HR, IT and ops, search time, accuracy on a fixed test set | Proven |
| 15. IT helpdesk triage and routine requests | Reads tickets, device and identity systems. Writes category and routing, and performs a short list of safe actions | Anything that grants access goes through your existing approval | Time to resolution, share resolved at first touch | Works with care |
| 16. Candidate screening | Reads CVs and applications. Writes summaries and a suggested shortlist | A person makes every decision, with regular checks for adverse impact | Recruiter hours, agreement between reviewers | Still experimental |
**Published proof for row 14.** In the [consulting firm case study](https://gigabit.agency/proof/ai-transformation-professional-services/), a consultant asks a question in Slack and gets a sourced answer in 4 seconds, and the knowledge base recovered **280 hours** in the first full quarter with all three phases live. Row 16 is experimental for legal as well as technical reasons. The [EU AI Act](https://eur-lex.europa.eu/eli/reg/2024/1689/oj) lists AI used for recruitment and candidate selection among high-risk uses in Annex III, which brings obligations for whoever deploys it. If you hire in the EU, involve employment counsel before you pilot. Our [AI governance service](https://gigabit.agency/services/ai-governance/) helps with the controls; it does not give legal advice.
### Healthcare and regulated work
| Use case | Reads and writes | Human checkpoint | Metric that proves it | Realism |
| --- | --- | --- | --- | --- |
| 17. Patient admin questions and prep instructions | Reads approved SOPs, prep guides and insurance FAQs. Writes cited answers on web, SMS and after hours | Scope limited to admin topics with no clinical advice, escalation triggers, staff handoff with a summary | Repetitive inquiries reaching the front desk, time to answer, after-hours self-service | Works with care |
| 18. Prior authorisation and claims preparation | Reads chart notes and payer rules. Writes a draft submission packet | A trained person checks and submits every packet | Packet completeness, denials for missing information, staff minutes per case | Still experimental |
**Published proof for row 17.** In our [specialty clinic case study](https://gigabit.agency/proof/clinic-llm-assistant-front-desk/), after 8 weeks repetitive inquiries handled by the front desk fell **62%**, median time-to-answer on common questions dropped from **4:10 to 1:20** (68%), and **40%** of after-hours interactions resolved self-serve. It is still rated "works with care" because the narrow scope did the work: administrative and prep topics only, refusal rules, PHI redaction and human handling for edge cases. Anything touching PHI needs the controls in [HIPAA-compliant AI agents](https://gigabit.agency/insights/hipaa-compliant-ai-agents/), and intake that writes to the EHR is covered in [AI patient intake and EHR integration](https://gigabit.agency/insights/ai-patient-intake-ehr-integration/).
### E-commerce storefront
Rows 1 and 9 came from e-commerce clients. Two more storefront jobs come up often:
| Use case | Reads and writes | Human checkpoint | Metric that proves it | Realism |
| --- | --- | --- | --- | --- |
| 19. Product questions and sizing help | Reads the catalogue, size guides, stock and reviews. Writes answers and add-to-cart links | It never states a spec it cannot find. Unknowns go to support | Conversion of chat sessions, returns citing wrong size | Works with care |
| 20. Catalogue enrichment | Reads supplier data sheets and images. Writes titles, attributes and descriptions as drafts in the store admin or PIM | A merchandiser approves before anything publishes | Attribute completeness, time to list a new SKU, error rate in sample audits | Works with care |
We have no published case for either. The common failure in both is the agent filling a gap with something plausible, which is why the checkpoint is about missing data rather than tone.
### How to pick your first one
Pick a job where the hours are already measured, the right answer is knowable, the systems can be reached, one person owns the number, and the worst case is survivable. Those are the five conditions in [five signs your business is ready for an AI agent](https://gigabit.agency/insights/signs-your-business-is-ready-for-an-ai-agent/), so score yourself there rather than here. Our free [AI readiness assessment](https://gigabit.agency/tools/ai-readiness-assessment/) asks six questions, one each on data, process, tooling, team, alignment and the problem itself, and returns a score with a recommended first move.
Three filters from the catalogue above narrow the list further:
1. **Start with a Proven row.** Your first agent should teach your team how to run one, not test whether the idea works at all.
2. **Prefer a job where mistakes are caught before a customer sees them.** Drafts, prepared fixes and internal answers forgive more than outbound email or refunds.
3. **Measure the baseline for 30 days before you build.** Without the before number, the metric column above proves nothing.
Then write a one-page brief. Copy this and fill every line; a line you cannot fill is the first thing to go and find out.
| Field | Example (row 9, returns) |
| --- | --- |
| Job, in one sentence | Process eligible returns from request to refund |
| Volume | Returns per month, and the share that are routine |
| Hours per week today | Staff time on returns, from a two-week log |
| Systems it reads | Store, OMS or ERP, carrier tracking, return policy |
| Systems it writes | Return record, label, customer email, refund |
| Human checkpoint | Which cases stop for review, and who reviews them |
| Metric and target | Share processed with no touch; minutes per return |
| Worst case | A refund issued in error: how much, how often, how it is caught |
For the checkpoint line, [when should an AI agent stop and ask a human?](https://gigabit.agency/insights/ai-agent-human-in-the-loop/) has an escalation matrix you can copy. For the metric line, [what to measure in an AI agent eval suite](https://gigabit.agency/insights/ai-agent-eval-metrics/) covers which numbers should gate a deploy. To check the payback before you commit, put your volume and hours into the [AI ROI calculator](https://gigabit.agency/tools/ai-roi-calculator/).
### What it costs to put one live
A [Gigabit Agents](https://gigabit.agency/services/ai-agents/) build is one flat fee per agent, from $8,000. The service page puts the Standard build at chat, up to two integrations and your existing knowledge base. Voice, three or more integrations, custom back-end actions and HIPAA work are scoped higher. The fee covers the scoping workshop, design and build, integration, testing and QA, deployment, handoff and 30 days of tuning. After that, [Managed AI Operations](https://gigabit.agency/services/managed-ai-operations/) from $3,000 a month keeps it monitored, re-tested and updated. For running costs such as model tokens and hosting, see [what an AI agent costs](https://gigabit.agency/insights/ai-agent-cost/).
If the job turns out to be a fixed sequence of steps with little judgement, it may not need an agent at all. The [AI Quick Win Sprint](https://gigabit.agency/pricing/) builds one workflow with up to three tools for $8,000 fixed in 3 weeks. If you know you want AI but not where to start, the [AI Transformation Sprint](https://gigabit.agency/services/ai-transformation/) is $25,000 for 2 weeks and ends in a scoped agent. And if your best candidate is a Still experimental row, or the job takes less than about 15 hours a week of paid time, the honest advice is to wait or run it in shadow mode first.

### FAQ
**What are the most common AI agent use cases?**
The most common production use cases are tier-1 support ticket resolution, inbound lead qualification, internal knowledge assistants in Slack or Teams, proposal drafting, returns processing, inventory reconciliation, supplier follow-up and order exception handling. They share a repeated job, a knowable right answer, systems with APIs and a clear human checkpoint.
**Which AI agent use cases are proven in production?**
Support ticket resolution with confidence gates, agent assist, lead routing, proposal first drafts, time entry suggestions, internal knowledge assistants, returns processing, inventory reconciliation, supplier PO follow-up and order exception handling are the ones we rate as proven. In a published Gigabit case, a support agent resolved 64% of tickets (3,149 a month) with no human involvement.
**Which AI agent use cases are still experimental?**
We rate candidate screening and prior authorisation preparation as still experimental. Both carry high stakes and, for recruitment, legal obligations: the EU AI Act lists AI used for recruitment and candidate selection as high-risk. Run them in shadow mode, where the agent drafts and a person does the job as before, until your own data says otherwise.
**How do I choose my first AI agent use case?**
Choose a job with measured hours, a knowable right answer, reachable systems, one owner for the metric and a survivable worst case. Start with a proven pattern, prefer jobs where mistakes are caught before a customer sees them, and measure a 30-day baseline before building. The free AI readiness assessment on our site gives a score and a recommended first move.
**How much does it cost to build an AI agent for one of these use cases?**
Gigabit builds production agents for one flat fee per agent, from $8,000 for a Standard build with chat, up to two integrations and your existing knowledge base. Voice, more integrations, custom actions and HIPAA work are scoped higher. Ongoing monitoring and updates through Managed AI Operations start at $3,000 a month.
**Do AI agents replace staff?**
In our published cases the agents took over repeated work and people moved to other work. After the D2C support agent went live, three human agents were reassigned to proactive outreach, VIP management and post-purchase experience. In the e-commerce operations case, the remaining 4.5 weekly hours were exception review, allocation decisions and vendor negotiation.

## AI agent observability tools: what to capture, and which tool fits your team
URL: https://gigabit.agency/insights/ai-agent-observability-tools/
Published 2026-10-23

**AI agent observability tools** record what an agent did on every run: each model call, tool call and retrieval step, with its inputs, outputs, latency, token cost and a quality score. That lets you find the bad runs and see why they went wrong. The main options are Langfuse, LangSmith, Arize Phoenix, Helicone, Braintrust, W&B Weave and Datadog LLM Observability, plus the OpenTelemetry GenAI semantic conventions as the open standard for the data itself. Below is what a setup must capture, a neutral comparison with prices read on each vendor's pricing page on 23 October 2026, and how to choose by team situation. Gigabit is not a partner of any of these vendors and earns nothing from your choice.
### Why agents need more than ordinary monitoring
Uptime and error-rate dashboards stay green while an agent gets worse, because a wrong answer returns HTTP 200 like a right one. Quality drifts when a vendor changes a model, a policy document goes stale or traffic shifts, and nobody notices for weeks; [why AI systems break after launch](https://gigabit.agency/insights/why-ai-systems-break-after-launch/) covers the five ways it happens. Agents add a second problem: a run is a chain of steps, and small per-step error rates multiply. [When a multi-step AI agent fails in production](https://gigabit.agency/insights/when-a-multi-step-ai-agent-fails-in-production/) does the arithmetic. Observability exists so you can see inside the chain.
### What agent observability must capture
Use this as a requirements list when you evaluate any tool, including one you build yourself.
| Signal | What to record | Why it matters |
| --- | --- | --- |
| Traces | One trace per request, with a nested span for each model call, tool call, retrieval and sub-agent, in order | You can replay what the agent did, step by step |
| Inputs and outputs | Prompt and its version, retrieved context, tool arguments and results, final answer | You cannot debug what you did not keep |
| Latency | Per span and end to end, watched at the median and the slow tail | A slow tool looks like a slow agent until you split it out |
| Token usage and cost | Input and output tokens per model call; cost per run and per resolved case | Retry loops burn money quietly, and cost per outcome is the number finance asks for |
| Eval scores | Offline scores on a fixed test set for every release; online scores on a sample of live runs | Turns "it feels worse" into a number with a date |
| User feedback | Ratings, CSAT, escalations and human edits to drafts, joined to the trace ID | Real users find failures your test set missed |
| Versions | Model, prompt, tool and retrieval index version on every trace | Links a quality drop to the change that caused it |
| PII handling | Redaction before storage, retention limits, access control, data location | Traces hold customer messages word for word |
| Alerts | Cost spikes, latency, tool errors, eval score drops, escalation rate | Someone hears about a problem before a customer does |
A quick test for any setup: pick one run from yesterday and answer these five questions in under five minutes. What did the user ask? What did the agent retrieve, and which tools did it call, in what order, with what arguments? What did each step return? What did the run cost and how long did it take? Which model and prompt version produced it? If any answer needs a developer and a log search, the setup is not finished. Our [AI operations runbook](https://gigabit.agency/insights/ai-operations-runbook/) explains how to write the trace-reading section around that test.
On scores, two posts go deeper than this one: [what to measure in an AI agent eval suite](https://gigabit.agency/insights/ai-agent-eval-metrics/) for which metrics should block a deploy, and [calibrating LLM-as-judge](https://gigabit.agency/insights/calibrate-llm-as-judge-for-ai-agents/) for automated scoring of live runs you can trust.
### The open standard: OpenTelemetry GenAI conventions
OpenTelemetry is the vendor-neutral standard most infrastructure monitoring already uses. Its [GenAI semantic conventions](https://github.com/open-telemetry/semantic-conventions-genai) define spans, metrics and events for generative AI: model call spans, token usage metrics and agent spans, with operation names such as `invoke_agent` and `execute_tool`. Their status, checked on 23 October 2026, is "Development", so attribute names can still change. Some content fields are opt-in; system instructions, for example, are not recorded unless you choose to. The conventions are a format, not a product: you still need a backend to store, search and score the traces. Instrumenting against them is the cheapest insurance against being locked into one vendor.
### The tools compared
Every vendor bills on a different unit: units, traces, spans, requests or gigabytes. Entry prices below are the lowest paid tier each vendor lists, read on 23 October 2026, in US dollars per month.
| Tool | Open source | Self-host option | Pricing model and entry price | Best at |
| --- | --- | --- | --- | --- |
| [Langfuse](https://langfuse.com/pricing) | Yes, [MIT licence](https://langfuse.com/pricing-self-host) | Yes, all core features free; Enterprise self-host at custom price | Per unit, where traces, observations and scores each count. Hobby free with 50k units; Core $29 with 100k units, then $8 per extra 100k | Tracing, prompt management, evals and datasets in one tool you can run yourself at no licence cost |
| [LangSmith](https://www.langchain.com/pricing) | No | Enterprise plan only, hybrid or fully self-hosted | Per seat plus traces. Developer free for one seat and 5k traces; Plus $39 per seat with 10k traces, then pay as you go | Teams building on LangChain or LangGraph, which come from the same company |
| [Arize Phoenix](https://github.com/Arize-ai/phoenix) and [Arize AX](https://arize.com/pricing/) | Phoenix is source-available under the Elastic License 2.0 | Phoenix with Docker or Kubernetes; AX self-hosted on Enterprise | Phoenix free to run yourself. AX Free with 25k spans; AX Pro $50 with 50k spans | Tracing and evals built on OpenTelemetry that you can start locally and move to a managed service later |
| [Helicone](https://www.helicone.ai/pricing) | Yes, [Apache 2.0](https://github.com/Helicone/helicone) | Yes, with Docker or Helm | Per request. Hobby free with 10,000 requests; Pro $79 plus usage; Team $799 | The fastest start: route model calls through its gateway by changing the base URL, then see cost and latency per request |
| [Braintrust](https://www.braintrust.dev/pricing) | No | Enterprise plan, with the [data plane in your cloud](https://www.braintrust.dev/docs/admin/self-hosting) | Per GB of processed data plus scores. Starter free with 1 GB and 10k scores; Pro $249 with 5 GB and 50k scores | Eval-led teams: experiments, scoring and regression tests next to production logs |
| [W&B Weave](https://coreweave.com/forge-pricing) | SDK under Apache 2.0 | [Self-managed](https://docs.wandb.ai/weave/guides/platform/weave-self-managed) needs a W&B platform install and a Weave-enabled licence | Per GB ingested. Free with 1 GB; Pro from $60 with 1.5 GB, then $0.10 per MB | Teams already training models in Weights & Biases |
| [Datadog LLM Observability](https://www.datadoghq.com/pricing/?product=llm-observability) | No | No, it runs in Datadog's cloud | Per LLM span; tool, workflow and retrieval spans are not billed. Free up to 40K LLM spans; Pro from $160 billed annually ($200 month to month) for 100K | Teams already on Datadog who want agent traces next to infrastructure and APM data |
| [OpenTelemetry GenAI](https://github.com/open-telemetry/semantic-conventions-genai) | Yes, an open standard | Not applicable; send data to any backend that accepts it | Free | Instrumenting once and keeping the freedom to switch backends |
Retention differs as much as price. On the free tiers whose retention is published, it ranges from 7 days (Helicone Hobby) to 30 days of data access (Langfuse Hobby); Langfuse Pro gives 3 years of data access, LangSmith keeps base traces 14 days and extended traces 180 days, and Datadog sells 30, 60 and 90-day retention as add-ons. If you need a year of traces for audits, price that in before anything else.
### Price on your own volume
Because the units differ, count your own. Take a typical run and list its steps. An agent that makes three model calls, two tool calls and one retrieval produces one trace and six spans per run. At 20,000 runs a month, that is 140,000 units on Langfuse, which counts the trace and every observation. On Datadog it is 60,000 LLM spans, because only the model calls are billed. Request-based and data-based pricing need a different count again: model requests per run for Helicone, gigabytes of trace data for Braintrust and Weave. Run that arithmetic for your top two candidates before you compare headline prices.
### How to choose
Pick by situation, not by feature list. The products above overlap heavily on core tracing, and the differences that decide most choices are hosting, the pricing unit and what each was built around.
- **One agent, small team, want traces this week.** Helicone's gateway is a base-URL change. Langfuse's free Hobby tier or $29 Core tier is the other easy start.
- **Data must stay on your own infrastructure** (health, finance, strict EU customers). Self-host Langfuse (MIT) or Helicone (Apache 2.0) at no licence cost, or Phoenix under the Elastic License. If you want a vendor-supported private deployment, LangSmith, Braintrust, Arize AX and Weave offer one on enterprise terms. Self-hosting means you run the database and its upgrades. For agents handling PHI, read [HIPAA-compliant AI agents](https://gigabit.agency/insights/hipaa-compliant-ai-agents/) first.
- **Built on LangChain or LangGraph.** LangSmith fits most naturally, since one company makes both.
- **Quality is the main worry and evals gate every release.** Braintrust is built around experiments and scores; Langfuse and Phoenix also run evals and cost less to start. Build the test set first: [how to build an eval set from support tickets](https://gigabit.agency/insights/build-ai-agent-eval-set-from-tickets/).
- **Already paying for Datadog.** Its LLM Observability keeps agent traces beside the rest of your monitoring, and because it bills per LLM span, an agent's tool and retrieval steps do not add to the bill.
- **Already on Weights & Biases for model training.** Weave keeps experiments and production traces in one place.
- **Want to keep options open.** Instrument against the OpenTelemetry GenAI conventions and check, in each vendor's own docs, that it accepts them before you commit.
One honest note: a single low-volume agent can start with structured logs in your own database, keyed by a trace ID, and a weekly review of a sample. Buy a tool when you find yourself reading traces every week and the logs slow you down.
### Questions to ask any vendor
Copy these into your evaluation sheet and get the answers in writing:
1. What exactly is a billable unit, and how many will one of our runs produce?
2. Where is our data stored, and can we choose the region?
3. Can we redact or mask PII before data leaves our servers?
4. What is the retention on the plan we would buy, and what does a longer one cost?
5. Can we export raw traces through an API or in OpenTelemetry format?
6. Can we attach eval scores and user feedback to a trace after the run?
7. Does a multi-step run show tool calls and sub-agents nested in order?
8. Who at the vendor can see our traces, and is that logged?
### What Gigabit sets up in production agents
Observability is one of six layers in every agent we ship, alongside a model tier, orchestration, retrieval, evals and guardrails, as our [production AI agent stack](https://gigabit.agency/insights/production-ai-agent-stack/) describes. In that stack every action is logged with full context, cost and latency are tracked per run, and drift is surfaced on dashboards. Permissions are bounded per step, low-confidence states escalate to a person, and rollback paths are rehearsed. Before an agent ships we build a scored test set from real cases, including the awkward ones, and every later change to prompt, model or retrieval runs against it. No green run, no deploy.
A [Gigabit Agents](https://gigabit.agency/services/ai-agents/) build is a flat fee from $8,000 per agent, and that fee includes evals and observability rather than treating them as change orders. It covers testing against real cases, guardrails and escalation paths, deployment and 30 days of tuning. After that, [Managed AI Operations](https://gigabit.agency/services/managed-ai-operations/) from $3,000 a month handles monitoring, eval regressions, model updates and new intents.

### FAQ
**What is AI agent observability?**
AI agent observability is recording each run of an agent as a trace: every model call, tool call and retrieval step, with inputs, outputs, latency, token cost, eval scores and user feedback. It lets you find the runs that went wrong and see which step caused it, which uptime and error-rate monitoring cannot do because a wrong answer still returns success.
**What are the best AI agent observability tools?**
The main options are Langfuse, LangSmith, Arize Phoenix, Helicone, Braintrust, W&B Weave and Datadog LLM Observability, with OpenTelemetry's GenAI conventions as the open standard. The best one depends on your situation: self-hosting needs point to Langfuse, Helicone or Phoenix, LangChain teams to LangSmith, eval-led teams to Braintrust, and existing Datadog or W&B customers to those vendors.
**Is Langfuse free?**
Self-hosted Langfuse is free: it is open source under the MIT licence, and its pricing page says all core features can be self-hosted without usage limits. Langfuse Cloud has a free Hobby tier with 50k units a month and 30 days of data access, and a Core tier at $29 a month with 100k units, checked on 23 October 2026.
**What is the difference between Langfuse and LangSmith?**
Langfuse is open source under MIT and can be self-hosted for free; its cloud bills per unit, from $29 a month on Core. LangSmith is closed source, comes from the company behind LangChain and LangGraph, bills per seat plus traces from $39 per seat a month on Plus, and offers self-hosting only on its Enterprise plan. Prices checked on 23 October 2026.
**What are the OpenTelemetry GenAI semantic conventions?**
They are OpenTelemetry's standard names for generative AI telemetry: spans for model calls and agent operations such as invoke_agent and execute_tool, token usage metrics and events. On 23 October 2026 their status was Development, so names may change. Instrumenting against them lets you switch observability backends without rewriting your tracing code.
**How should agent traces handle personal data?**
Treat traces as customer data, because they store messages word for word. Redact or mask personal data before it leaves your servers where the tool allows it, set the shortest retention that still covers your audits, restrict who can read traces, and choose the storage region or self-host when regulation requires it. Ask each vendor these questions in writing before you buy.

## AI policy template: a company AI acceptable use policy you can copy
URL: https://gigabit.agency/insights/ai-policy-template/
Published 2026-10-23

An **AI policy template** gives you the rules your staff follow when they use AI tools at work and when your company builds AI into its own products. Below is a complete AI acceptable use policy in twelve sections, from an approved tools list and data rules to human review, customer disclosure, building and buying AI, incidents and training. Fill in every field in [square brackets] and delete what does not apply. After the template, a table maps each section to the NIST AI Risk Management Framework, ISO/IEC 42001 and the EU AI Act, with dates checked on 23 October 2026. It is a starting draft written by engineers, not legal advice: have counsel review it before you adopt it.
### Before you fill it in
Three things decide whether people follow it.
1. **Make a list of the AI already in use.** Ask each team which assistants, copilots, meeting recorders and AI features in existing software they use, and on which accounts. Section 3 is only useful if it starts from what people already do.
2. **Name one owner.** Pick one person, usually in security, IT, legal or operations, who approves tools, answers questions and runs the review.
3. **Write for the person at the keyboard.** Most of the value is in sections 3 and 4: which tools, which accounts, which data. If those two sections are clear, most mistakes never happen.
If you only use off-the-shelf tools and keep customer data out of them, you can drop sections 7 and 8 until you build or buy AI for customer-facing work.
### The AI acceptable use policy template
Copy from the heading below to the end of section 12.
### [Company name] AI acceptable use policy
Version [1.0]. Effective [date]. Policy owner: [name, role]. Approved by [name, role]. Next review: [date, no more than 12 months from the effective date].
### 1. Purpose and scope
This policy sets out how people at [Company name] may use artificial intelligence tools, and how we build and buy AI systems. Its purpose is to let us use AI productively while protecting our customers, our data, our staff and our reputation.
It applies to all employees, contractors, interns and anyone else working on our behalf ([list any agencies or partners]), on any device, whenever they use an AI tool for [Company name] work or with [Company name] data. It covers AI tools we approve, AI features inside software we already use, and AI systems we build. Where this policy conflicts with a law, regulation or customer contract, the stricter rule applies.
### 2. Definitions
- **AI tool:** any software that generates text, images, audio, video or code, or that makes predictions, recommendations or decisions, using a machine-learning model. This includes chat assistants, coding assistants, meeting recorders, AI features in existing software and AI agents.
- **AI agent:** an AI system that can take actions, such as sending messages, changing records, calling other software or spending money, not only answer questions.
- **Approved tool:** an AI tool listed in section 3, used through the account type listed there.
- **Output:** anything an AI tool produces, including actions.
- **Customer-facing use:** any use where output reaches a customer, prospect, patient, candidate or member of the public without being rewritten by a person.
- **Policy owner:** the person named at the top of this policy.
### 3. Approved tools
Only the tools below may be used for [Company name] work. Use them only through the account type listed, signed in with your work identity. Personal accounts and free consumer versions are not approved, even for the same product, unless the table says so.
| Tool | Approved for | Account type | Highest data class allowed (see section 4) | Tool owner | Last reviewed |
| --- | --- | --- | --- | --- | --- |
| [Chat assistant, e.g. vendor and plan] | [Drafting, research, summarising internal documents] | [Company enterprise workspace, SSO] | [Confidential] | [Name] | [Date] |
| [Coding assistant] | [Writing and reviewing code in approved repositories] | [Company business licence] | [Confidential, no secrets] | [Name] | [Date] |
| [Meeting recorder] | [Internal meetings only, with notice to attendees] | [Company account] | [Internal] | [Name] | [Date] |
| [AI feature in existing software, e.g. CRM or helpdesk] | [Drafting replies that a person reviews] | [Existing company licence] | [Confidential] | [Name] | [Date] |
To request a new tool, send [request form or channel] the tool's name, what you want it for, the data it would touch, and a link to its terms. The policy owner will reply within [10 working days]. Until a tool is approved, do not put any [Company name] data into it.
### 4. Data rules
Every piece of information belongs to one class. When in doubt, treat it as the higher class.
| Class | Examples | Approved tools | Unapproved tools |
| --- | --- | --- | --- |
| Public | Published website copy, press releases, public documentation | Allowed | Allowed |
| Internal | Internal process documents, non-sensitive meeting notes, draft marketing copy | Allowed | Not allowed |
| Confidential | Customer names and contact details, contracts, pricing, financials, unreleased product plans, source code | Allowed only where section 3 permits Confidential | Not allowed |
| Restricted | See the list below | Not allowed, unless the policy owner approves a specific tool and use in writing | Not allowed |
Never paste, upload or dictate the following into any AI tool unless the policy owner has approved that exact use in writing:
- Passwords, API keys, access tokens, private keys or other credentials.
- Payment card numbers, bank account details or government identifiers such as passport, national insurance or social security numbers.
- Health, medical or biometric information about any person.
- Special category data, such as ethnicity, religion, sexual orientation or criminal records.
- Personal data about customers or staff beyond what the approved task needs.
- Information held under a confidentiality agreement that forbids it.
- Legal advice or material subject to legal privilege.
- Material non-public financial information.
- [Any other category specific to your business or your customers' contracts.]
### 5. Human review
You are responsible for any output you use, in the same way you are responsible for your own work. Before you rely on an output, check it.
- **Facts, figures, quotes and citations:** check each against the original source.
- **Code:** review and test it like a colleague's pull request, including any packages it adds.
- **Anything sent outside the company:** a person reads it before it is sent, unless the use is listed in section 6 as an approved automated reply.
- **Decisions about people:** AI output may inform, but a person makes and owns any decision about hiring, performance, pay, credit, pricing for an individual customer, access to services or [other decisions you name]. That person must be able to explain the decision without referring to the AI.
- **Actions by AI agents:** any agent action that sends money, deletes data, changes a customer record or contacts a customer needs a person's approval first, unless the system owner has documented a lower-risk exception under section 7.
### 6. Telling customers when they deal with AI
- Any chat, voice or messaging system that talks to customers or the public must tell them they are interacting with an AI system at the start of the conversation, and offer a way to reach a person.
- Images, audio or video generated or altered by AI that show real people, places or events must be labelled as AI-generated.
- Text generated by AI and published to inform the public must be labelled as such unless a person has reviewed and edited it and [Company name] takes editorial responsibility for it.
- Approved automated replies: [list any, with the system that sends them and who reviews a sample, or write None].
- Do not present AI output as a named person's personal work or opinion without that person's agreement.
### 7. Building AI into our products and processes
This section applies when we build, configure or deploy an AI system, including agents, chatbots, scoring models and AI features in our products.
- **Register it.** Every AI system has an entry in the AI register kept by [team]: owner, purpose, users, data used, decisions it influences, model, vendor and risk tier.
- **Risk review before build.** Who could be harmed, what data is used, what the system can do on its own and what happens when it is wrong. Systems rated [high] need approval from [role], and counsel checks whether any law classifies the use as high-risk.
- **Evaluations.** A written set of test cases with pass thresholds. A release that fails them does not ship. Re-run them after any change to the model, prompt, tools or data.
- **Human oversight.** Design where a person reviews, approves or can stop the system, with the context and authority to do it.
- **Logging.** Record inputs, outputs, tool calls and model version for [12 months, or longer if law or contract requires].
- **Least privilege.** Agents get their own credentials, the minimum access the task needs, an allowlist of actions, and spending and rate limits.
- **Security testing.** Customer-facing systems and agents with write access are tested for prompt injection and data leakage before launch.
- **Change control.** Model, prompt and tool changes go through the same review and release process as code.
### 8. Buying AI tools and AI features
Before we buy or switch on an AI tool or AI feature, [procurement or the policy owner] confirms in writing:
- Whether the vendor uses our inputs or outputs to train its models, and that this is switched off or contractually excluded.
- How long inputs and outputs are kept, where, and how to delete them, and which sub-processors and model providers are involved.
- A data processing agreement is signed where personal data is involved.
- Security evidence, such as a current SOC 2 Type II report or ISO/IEC 27001 certificate, and whether the vendor holds ISO/IEC 42001 certification.
- Admin controls: single sign-on, user management, audit logs and feature restrictions.
- How the vendor announces model changes and incidents, and how we export our data and leave.
### 9. Reporting AI incidents
Report to [channel, email or form] within [24 hours] if you see or cause any of the following. Report even if you are not sure. Nobody will be penalised for a prompt, honest report.
- Restricted or Confidential data put into a tool that is not approved for it.
- Harmful, false, offensive or biased output that reached a customer or the public.
- An AI agent taking an action it should not have.
- Signs that a tool was manipulated, for example by instructions hidden in a document, email or web page.
- A customer complaint that involves an AI system.
The policy owner sets a severity of [low, medium, high], contains the issue, informs affected people and authorities where required, and records cause and fix in the AI register.
### 10. Training
Everyone covered completes [Company name]'s AI training before using an approved tool, and a refresher every [12 months]. People who build or run AI systems complete additional training on [evaluations, oversight, security and the rules that apply to their systems]. [Team] keeps a record of who completed which training and when.
### 11. Ownership, exceptions and review
- The policy owner maintains this policy, the approved tools list and the AI register.
- Exceptions must be requested in writing, approved by [role], time-limited and recorded.
- This policy is reviewed at least every [12 months], and sooner when we adopt a significant new AI system, a law or contract changes what we must do, or an incident shows a gap.
- Breaches of this policy are handled under [disciplinary policy or contractor agreement].
- Questions go to [policy owner's contact].
### 12. Acknowledgement
I have read the [Company name] AI acceptable use policy, version [1.0], and I agree to follow it. Name: [ ]. Role: [ ]. Date: [ ].
### How the template maps to NIST AI RMF, ISO/IEC 42001 and the EU AI Act
Customers and auditors tend to ask about three frameworks. The [NIST AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework) (AI RMF 1.0, released 26 January 2023) is voluntary and organises the work into four functions: Govern, Map, Measure and Manage. NIST says it is being revised; its Generative AI Profile (NIST AI 600-1, July 2024) covers generative models. [ISO/IEC 42001:2023](https://webstore.iec.ch/en/publication/90574) specifies requirements for establishing, implementing, maintaining and continually improving an AI management system, and is the standard an accredited body can certify you against. The [EU AI Act](https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai), Regulation (EU) 2024/1689, is law.
| Template section | NIST AI RMF function | EU AI Act link (ask counsel whether it applies to you) |
| --- | --- | --- |
| 1, 2, 11: scope, definitions, ownership | Govern | Defines who is a provider or deployer of each system |
| 3: approved tools | Govern, Map | Gives you the list of AI systems you deploy |
| 4: data rules | Govern, Manage | Supports data governance; personal data stays under data protection law |
| 5: human review | Manage | Article 14 human oversight for high-risk systems |
| 6: customer disclosure | Govern | Article 50 transparency duties |
| 7: building AI | Map, Measure, Manage | Risk management, logging and oversight duties for high-risk systems |
| 8: buying AI | Govern, Map | Deployer duties depend on what the vendor's system is and does |
| 9: incidents | Manage | Serious-incident reporting for high-risk systems |
| 10: training | Govern | Article 4 AI literacy |
For ISO/IEC 42001, a policy is a starting document, not a management system. Certification needs evidence that controls run, such as risk assessments, records and reviews; the register in section 7 and the incident log in section 9 are where that evidence starts.
### What the EU AI Act asks of most companies today
**AI literacy (Article 4)** has applied since 2 February 2025, according to the [Commission's AI Act page](https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai), and it covers every provider and deployer, whatever the risk level of the system. The Digital Omnibus on AI, in force since 27 July 2026, rewrote the article. The [current text](https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-4) says providers and deployers shall take measures to support the development of AI literacy of their staff and others using AI systems on their behalf, taking into account their knowledge, experience and the context of use, and that this does not require them to guarantee any specific level of literacy for any individual. Section 10 and its training records are how you show those measures.
**Transparency (Article 50)** rules come into effect in August 2026, according to the Commission. [Article 50](https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-50) requires providers to tell people when they are interacting with an AI system, unless it is obvious, and deployers to disclose deep fakes and AI-generated text published to inform the public, unless a person has reviewed it and someone holds editorial responsibility. Section 6 follows that shape.
**High-risk systems**, such as AI used in recruitment, credit scoring or access to essential services under Annex III, carry far heavier duties. After the July 2026 amendment those rules apply from 2 December 2027 for Annex III uses and 2 August 2028 for AI in products covered by Annex I (dates checked on 23 October 2026). The Act also reaches companies outside the EU when an AI system's output is used in the EU. Whether a system of yours is in scope, and in which category, is a question for counsel; section 7 makes sure it gets asked before the build.
### What this template does not do
It does not tell you which laws apply. Data protection, employment and consumer law and sector rules may add duties or forbid uses this template allows. If you handle patient data, read our guide to [HIPAA-compliant AI agents](https://gigabit.agency/insights/hipaa-compliant-ai-agents/) and talk to counsel.
Signing it does not make it true. The tools list needs someone keeping it current, and human review only works if your systems put a person in the loop at the right step. Our guides to [human-in-the-loop design](https://gigabit.agency/insights/ai-agent-human-in-the-loop/), [AI agent evaluation metrics](https://gigabit.agency/insights/ai-agent-eval-metrics/) and the [AI agent security framework](https://gigabit.agency/insights/ai-agent-security-framework/) cover how to build sections 5 and 7 into real systems.
And it is not legal advice. Gigabit is an engineering firm. Have your counsel review the policy, decide which systems fall into which legal category, and adjust the wording for your jurisdiction.
### When you need more than a policy
For a company that uses a handful of approved tools and builds no AI into its products, a well-kept version of this template is enough.
It stops being enough when teams have shipped AI agents that take actions, when a customer questionnaire or an auditor asks for evidence, or when you sell into the EU or regulated buyers. Our [AI governance service](https://gigabit.agency/services/ai-governance/) inventories every AI system you build and buy, tiers each one against NIST AI RMF, ISO/IEC 42001 and, where counsel says it applies, the EU AI Act, then builds the controls the policy promises into the systems: human review steps, eval gates, logging, vendor checks and an incident runbook. The diagnostic is the [AI Transformation Sprint](https://gigabit.agency/services/ai-transformation/), $25,000 fixed over 2 weeks, which ends in a working pilot of the first control. The build that follows is a fixed price in writing after scoping, and ongoing monitoring runs on Managed AI Operations at $3,000–$20,000 a month. If you need an agent or LLM feature tested for prompt injection and data leakage, the [AI Security Assessment](https://gigabit.agency/services/ai-security/) starts at $12,000 fixed. We build and document controls; we do not certify, audit for compliance or give legal opinions.

### FAQ
**What should a company AI policy include?**
At minimum: who and what it covers, a list of approved AI tools and the account type for each, data rules that say what may never be put into a tool, a human review requirement before output is relied on or sent out, rules for telling customers when they deal with AI, a way to report incidents, training, and a named owner with a review date. If you build AI into products, add risk review, evaluations, logging and vendor checks.
**Is an AI acceptable use policy legally required?**
No law we know of requires a document with that name. The EU AI Act's Article 4, which has applied since 2 February 2025, requires providers and deployers of AI systems to take measures to support AI literacy among their staff, and a policy with training is a common way to show that. Sector rules and contracts may require more. Ask your counsel what applies to you.
**What does the EU AI Act say about AI literacy?**
Article 4, as amended by the Digital Omnibus on AI in force since 27 July 2026, says providers and deployers shall take measures to support the development of AI literacy of their staff and others using AI on their behalf, considering their knowledge, experience and the context of use. It does not require them to guarantee a specific level for any individual. The obligation has applied since 2 February 2025 (checked on 23 October 2026).
**Should employees be allowed to use ChatGPT or other public AI tools at work?**
Usually yes, through a business or enterprise account the company controls, with data rules and training. The risk sits mostly in personal accounts and free consumer versions, where you have no admin control and may not have agreed terms on training and retention. List the approved tools and account types in the policy, and ban putting company data into anything else.
**How does an AI policy relate to ISO 42001 and the NIST AI RMF?**
Both expect one. The NIST AI RMF is voluntary and puts policies, roles and accountability under its Govern function. ISO/IEC 42001:2023 is a certifiable management-system standard, and a policy is a starting document for it. Certification also needs evidence that controls run, such as risk assessments, an AI register, incident records and reviews. A policy alone does not get you certified.
**How often should we review our AI policy?**
At least once a year, and sooner when you adopt a significant new AI system, a law or customer contract changes what you must do, or an incident shows a gap. AI tools change their terms and features often, so review the approved tools list more often than the policy text, for example every quarter.

## WordPress to Webflow migration: a step-by-step guide with a redirect map
URL: https://gigabit.agency/insights/wordpress-to-webflow-migration/
Published 2026-10-23

A **WordPress to Webflow migration** is a rebuild with a content import, not a transfer. Webflow's own help centre says [there is no automated way](https://help.webflow.com/hc/en-us/articles/51622865825171-Migrate-your-site-to-Webflow-overview) to convert a site from another platform: you recreate the layouts, styles and CMS structure in Webflow, then bring the content in, usually as CSV files. The steps that decide whether you keep your search traffic are the dull ones: an inventory of every URL, collections designed before import, a 301 redirect map tested before the DNS changes, and the SEO fields carried across. Below are nine steps in the order we run them, a redirect map you can copy, a cutover checklist, and an honest list of what does not migrate cleanly. Webflow facts were checked on its help centre and pricing page on 23 October 2026.
### Step 1: Decide that Webflow is the right destination
Do this before anything else, because some sites should not move. Our [Webflow vs WordPress comparison](https://gigabit.agency/compare/webflow-vs-wordpress/) covers licence, hosting, editing, custom code and price side by side, so this guide does not repeat it. The short version: Webflow suits a marketing site that a design or marketing team wants to edit visually without plugin upkeep. It is a poor fit for sites whose value lives in plugins.
Three cases deserve a hard look. **Memberships:** Webflow [retired its User Accounts feature](https://help.webflow.com/hc/en-us/articles/36046006227731-User-Accounts-sunset) on 29 January 2026, so gated content now needs a third-party membership tool. **Stores:** a WooCommerce shop with a large catalogue, subscriptions or wholesale rules usually belongs on a commerce platform, not on Webflow Ecommerce. **Very large or complex content:** thousands of posts across many interlinked post types can hit Webflow's CMS limits (step 3).
### Step 2: Inventory the WordPress site
You cannot redirect or rebuild what you have not listed. Build one spreadsheet with a tab for each of these:
- **URLs.** Combine the XML sitemap, a full crawl with a tool such as Screaming Frog or Sitebulb, and the Pages report from Google Search Console for the last 12 months. The crawl finds orphaned pages; Search Console finds URLs that still get traffic but are no longer linked. Include PDFs and images under /wp-content/uploads/ that other sites link to.
- **Post types and taxonomies.** Posts, pages, every custom post type, categories, tags and custom taxonomies, with item counts.
- **Custom fields.** Every ACF, Meta Box or Pods field group, the field type and which post types use it.
- **Plugins.** Every active plugin, what it does on the front end, and what replaces it.
- **Traffic and links.** Clicks per URL from Search Console and referring domains per URL from your backlink tool. This ranks the redirect work.
The plugin tab is where migrations get underestimated. A typical mapping:
| WordPress plugin type | What it does | Webflow replacement |
| --- | --- | --- |
| SEO plugin (Yoast, Rank Math) | Titles, meta descriptions, canonicals, schema, sitemap | Native SEO settings per page and per CMS template, auto sitemap, custom code for schema |
| Form plugin (Gravity Forms, WPForms) | Forms, conditional logic, notifications, CRM sync | Native Webflow forms, plus Zapier, Make or n8n for routing and logic |
| Page builder (Elementor, Divi) | Layouts | Rebuilt in the Webflow Designer; nothing imports |
| Redirect plugin (Redirection) | 301s | Webflow 301 redirect settings (step 6) |
| Multilingual (WPML, Polylang) | Translations, hreflang | Webflow Localize, a paid add-on |
| Membership (MemberPress, Restrict Content) | Gated content, logins | Third-party membership tool |
| WooCommerce | Store | Webflow Ecommerce or a commerce platform |
| Caching, security, backups | Speed and safety on your own hosting | Handled by Webflow hosting; drop them |
| Search (Relevanssi, SearchWP) | Site search | Webflow site search on Premium, or a third-party search |
Webflow switched off its own automation tool, [Logic](https://help.webflow.com/hc/en-us/articles/36045959707667-Logic-sunset), in June 2025, so form routing and conditional workflows now run in an outside tool. Our [n8n vs Zapier vs Make comparison](https://gigabit.agency/compare/n8n-vs-zapier-vs-make/) helps you pick one.
### Step 3: Design the CMS collections
Design collections on paper before importing anything. A WordPress post type usually becomes a Webflow Collection; categories, tags and authors become their own Collections linked by reference fields; ACF fields become Collection fields. Then check the result against Webflow's limits.
| Webflow limit (checked 23 October 2026) | Starter | Basic | Premium | Team |
| --- | --- | --- | --- | --- |
| CMS items | 50 | No CMS | 20,000 | 20,000, more for a fee |
| CMS Collections | 20 | No CMS | 40 | 100 |
| CMS API requests per minute | 60 | None | 120 | Not listed |
Figures are from [Webflow's pricing page](https://webflow.com/pricing); Team also lists 100 fields and 20 reference fields per Collection. Webflow's help centre says non-Enterprise plans get [up to 10 reference fields](https://help.webflow.com/hc/en-us/articles/33961317363091-Reference-Collection-field) per Collection depending on the site plan, and Enterprise up to 20. On the page itself, a Collection list shows [up to 100 items](https://help.webflow.com/hc/en-us/articles/33961370432275-Dynamic-content-limits) unless you paginate, and a page can hold up to 40 Collection lists with nesting up to three levels deep. Plan prices are in our [Webflow pricing guide](https://gigabit.agency/insights/webflow-pricing/); any site with a blog needs Premium at least, because Basic has no CMS.
Where a WordPress site uses many relationships between post types (products to features to industries to case studies), count the reference fields each Collection needs. That is the limit that most often forces a redesign of the content model.
### Step 4: Export from WordPress and import into Webflow
WordPress's [Tools > Export screen](https://wordpress.org/documentation/article/tools-export-screen/) produces a WXR file, an XML file holding posts, pages, custom post types, comments, custom fields, categories, tags, custom taxonomies and users. Webflow's [WordPress migration guide](https://help.webflow.com/hc/en-us/articles/33961193662739-Migrate-your-site-from-WordPress-to-Webflow) suggests converting that XML to CSV and importing it Collection by Collection. An export plugin that writes CSV directly saves the conversion step and lets you pick columns.
Webflow's [CSV import](https://help.webflow.com/hc/en-us/articles/33961290794771-How-do-I-import-content-into-the-Webflow-CMS) has rules worth knowing before you build the files:
- Each CSV can be up to **4 MB**, so split large post types into batches.
- Name and slug values must be under 256 characters. Keep the WordPress slug in the slug column so URLs can stay the same.
- Reference fields match on slug. A plain-text value such as "Jane Austen" is turned into jane-austen and matched; if the item has a custom slug, put the slug in the CSV. Multi-reference values are separated by semicolons.
- Option fields take up to 100 distinct values. Number fields reject values with commas or currency symbols.
- Image fields need a direct URL to an image file. Images imported this way do not get responsive variants, so resize them first. File fields import nothing; upload files by hand.
- Webflow offers to back up the site before each import. Accept it.
Import in dependency order: authors, categories and tags first, then the posts that reference them. Run a test import of ten items, check every field on the published template, and only then import the rest.
For simple blogs there is a shortcut. The [WP to Webflow app](https://webflow.com/apps/detail/wp-to-webflow) on Webflow's marketplace, built by Eclipse SRL, is free, reads a live WordPress site and imports posts, images, tags and categories into Collections. Its listing does not mention pages, custom fields or custom post types, so treat it as a blog importer. For large sites, the CMS API at 120 requests a minute on Premium is the scripted route.
### Step 5: Clean the rich text and move the images
Post bodies are where imports go wrong quietly. WordPress content carries shortcodes, block-editor comments, page-builder markup, inline styles and embeds. Webflow accepts HTML in rich text fields but may strip elements, classes and attributes it does not support, so a shortcode such as a gallery or a call-to-action button comes through as literal text or disappears. Search the export for square-bracket shortcodes and for page-builder class names, and decide for each whether to convert it to plain HTML, rebuild it as a Webflow component, or drop it.
Images inside the body need the same care. Webflow's import guide covers image fields, but it does not say whether images referenced inside rich text HTML are copied to Webflow. Import a few posts and inspect the image URLs. If they still point at /wp-content/uploads/ on the old server, they will break the day that server is switched off, so move them to Webflow's assets first or keep the old media host running.
Rewrite internal links inside the body to the new URLs at the same time. Redirects will catch old links, but every internal link that goes through a redirect costs a hop.
### Step 6: Build and test the redirect map
Every old URL that had traffic, backlinks or an index entry needs a 301 to its closest new equivalent. Google's [site-move guidance](https://developers.google.com/search/docs/crawling-indexing/site-move-with-url-changes) asks for permanent redirects straight to the final URL, without chains, kept for at least a year. Use this format for the working map:
| Old URL | New URL | Rule type | Clicks, last 12 months | Referring domains | Checked by | Tested on staging |
| --- | --- | --- | --- | --- | --- | --- |
| /2023/04/pricing-update/ | /blog/pricing-update | Pattern | 412 | 6 | [Name] | [Date] |
| /services/seo-audits/ | /services/seo | One-to-one | 1,180 | 14 | [Name] | [Date] |
| /category/news/ | /blog | One-to-one | 35 | 1 | [Name] | [Date] |
| /?p=1234 | /blog/launch-notes | One-to-one | 0 | 2 | [Name] | [Date] |
| /wp-content/uploads/2022/05/brochure.pdf | /resources/brochure | One-to-one | 90 | 9 | [Name] | [Date] |
| /author/sam/ | /about | One-to-one | 4 | 0 | [Name] | [Date] |
The rows above are examples. Your real rows come from the inventory in step 2. Sort by clicks and referring domains so the URLs that matter most are checked by a person.
In Webflow, redirects live in **Site settings > Publishing > 301 redirects**, which needs a paid plan. You can add them one by one or [import a CSV](https://help.webflow.com/hc/en-us/articles/33961211526291-Import-export-301-redirects) of up to 16 MB with two columns, old path and new path. Four details from [Webflow's redirect guide](https://help.webflow.com/hc/en-us/articles/33961294898835-How-do-I-set-up-redirects-in-Webflow) catch people out:
1. **An import replaces every existing redirect.** Export the current list first, add your rows to it, then import.
2. **Patterns use capture groups.** Webflow's own example sends `/blogs/(.*)/(.*)` to `/articles/%1/%2`. A rule such as `/2023/(.*)/(.*)` to `/blog/%2` can cover a year of date-based WordPress permalinks; test how trailing slashes resolve before you rely on it. Webflow recommends no more than about 1,000 redirect rules in total, so use patterns where you can.
3. **Order matters.** Webflow runs redirects oldest first, so add specific rules before broad wildcards that would catch them.
4. **Some characters must be escaped** with % in old paths, including ?, =, & and the hyphen, and a path with a trailing slash before a query string will not match. Localised paths such as /es/old-url need their own rules.
Test before launch. Point a staging copy of the site at the redirect list, run every old URL through a crawler in list mode, and confirm each returns a single 301 to a live 200 page. Fix chains and loops on staging, not in production.
### Step 7: Carry the SEO over
- **Titles and meta descriptions.** Export them from your SEO plugin's fields (they travel in the WXR file as custom fields), import them into plain-text fields, and bind those fields to the SEO title and description on each Collection template. Set static page titles by hand.
- **Slugs.** Keep them the same wherever you can. An unchanged slug under a new folder still needs a redirect, but it is easy to pattern.
- **Canonicals.** Set the global canonical URL in Webflow and check any page-level canonicals your SEO plugin set.
- **Schema.** Rebuild Article, Organization, FAQ or Product markup in custom code on the templates, filled from CMS fields.
- **Sitemap and robots.txt.** Webflow generates the sitemap automatically; check it lists only pages you want indexed, and copy over any robots.txt rules you rely on.
- **Alt text, Open Graph images and hreflang.** Import alt text with images where you can, and set Open Graph images on templates. If the site is multilingual, set up Localize and check hreflang before launch.
- **A useful 404 page**, with search or links to the main sections, for the URLs your map missed.
### Step 8: Rebuild forms and integrations
List every form, where its submissions go and what happens next. Native Webflow forms collect submissions and send notification emails; routing to a CRM, a Slack channel or a second system needs an integration or an automation tool. Re-add analytics and tag manager, chat widgets, consent banners, CRM tracking scripts and any third-party embeds, then submit every form on staging and confirm each submission arrives where it should. Comments have no import path in Webflow's CSV field types, so decide whether to move them to a third-party comment tool or archive them.
### What does not migrate cleanly
- **Plugin functionality.** Nothing a plugin does on the front end moves on its own. Each one is rebuilt, replaced with a third-party tool or dropped.
- **Complex custom post types.** Deep relationships between post types run into reference field limits, and repeater or flexible-content fields in ACF have no direct equivalent in a Webflow Collection.
- **Memberships and user accounts.** These need a third-party tool since Webflow retired User Accounts.
- **E-commerce.** WooCommerce orders, customers and subscriptions do not import into Webflow. A store with real volume should be planned as its own migration.
- **Files and media.** File fields do not import, and media inside post bodies needs checking (step 5).
- **Design.** Themes and page-builder layouts are rebuilt by hand in the Designer.
### Cutover checklist
| When | Task | Done |
| --- | --- | --- |
| 2 weeks before | Content freeze date agreed; final URL inventory exported; redirect map signed off | [ ] |
| 1 week before | Full import on staging; redirects tested in list mode; forms submitted end to end; analytics firing | [ ] |
| 1 week before | DNS TTL lowered; access to the domain registrar and DNS confirmed; old site backed up in full, including the database and uploads | [ ] |
| Launch day | Final delta import of anything published since the freeze; redirects imported; site published; DNS switched | [ ] |
| Launch day | Spot-check the top 50 old URLs by traffic; submit the new sitemap in Search Console; check robots.txt is not blocking the site | [ ] |
| Day 1 to 7 | Watch 404s and server errors daily; fix missed redirects; keep the old server reachable for media until step 5 is confirmed | [ ] |
| Day 30 | Compare clicks and impressions with the 30 days before launch; review indexing; remove temporary workarounds | [ ] |
### Step 9: Monitor Search Console after launch
Google says that for a medium-sized site it [can take a few weeks or more](https://developers.google.com/search/docs/crawling-indexing/site-move-with-url-changes) for the new URLs to replace the old ones, and that rankings may fluctuate while it happens. In the first month, check the Pages report for new "Not found (404)" and redirect errors, the Sitemaps report for how many new URLs are indexed, and the Performance report filtered to your top pages, compared with the same period before launch. Keep the old sitemap for a while so you can watch the old URLs drop out. Use the Change of Address tool only if the domain itself changed; a migration on the same domain does not need it. Keep the redirects for at least a year, and in practice for as long as anyone links to the old URLs.
### Doing it yourself, or bringing in help
A small brochure site or a blog with a few hundred posts and few plugins is a reasonable do-it-yourself project with this guide and a test import. The WordPress site keeps running until DNS switches, so the risk is mostly time.
Bring in help when the site has thousands of URLs, several custom post types, forms that must land in a CRM, a membership or store component, or search traffic the business depends on. Our [Webflow work](https://gigabit.agency/platforms/webflow/) includes migrations from WordPress, and a [Platform Migration](https://gigabit.agency/pricing/) starts at $7,000 fixed, from 4 weeks: data migrated and reconciled, a full redirect map, a rehearsed cutover with a rollback plan and 30 days of post-launch monitoring. If the move is also a redesign, a Webflow [Website Build](https://gigabit.agency/pricing/) starts at $6,000 fixed. A store that should leave WooCommerce for a commerce platform is a different project; a store build starts at $9,000. Custom code components and integrations after launch can be handled by an embedded engineer at $3,000–$7,500 per engineer per month. We build on [WordPress](https://gigabit.agency/platforms/wordpress/) too, so if the inventory shows the site belongs where it is, we will say so.

### FAQ
**Can I import a WordPress site into Webflow automatically?**
Not the whole site. Webflow says there is no automated way to convert a site from another platform; the design is rebuilt in the Designer. Content can be imported: export WordPress to XML, convert it to CSV and use Webflow's CSV import per Collection, or use a third-party marketplace app such as WP to Webflow, which imports posts, images, tags and categories from a live site.
**How long does a WordPress to Webflow migration take?**
It depends on page templates, post types and how much plugin functionality must be replaced. A small blog can move in days. A marketing site with several custom post types, forms into a CRM and a few thousand URLs usually takes weeks, most of it on content modelling, cleaning rich text and testing redirects. Plan for a content freeze of a few days around the cutover.
**Will I lose SEO rankings moving from WordPress to Webflow?**
You can keep most of them if every old URL with traffic or backlinks gets a single 301 to its closest new page, titles and meta descriptions carry over, and internal links are updated. Google says rankings may fluctuate during a move and that a medium-sized site can take a few weeks or more to settle. Monitor 404s and indexing in Search Console daily for the first week.
**How do I set up 301 redirects in Webflow?**
Go to Site settings, Publishing, 301 redirects. Add rules one by one or import a CSV of up to 16 MB with old and new paths; an import replaces all existing redirects, so export first. Use capture groups such as /blog/(.*) to /articles/%1 for patterns. Webflow runs redirects oldest first, recommends about 1,000 rules at most, and needs a paid plan. Publish the site for redirects to take effect.
**Which Webflow plan do I need after migrating from WordPress?**
Any site with a blog or other CMS content needs Premium or above, because Basic has no CMS. Premium includes 20,000 CMS items in 40 Collections; Team raises Collections to 100. Starter allows only 50 items and cannot use a custom domain. Limits checked on webflow.com/pricing on 23 October 2026; see our Webflow pricing guide for plan prices.
**What doesn't transfer from WordPress to Webflow?**
Plugin functionality, themes and page-builder layouts, memberships (Webflow retired User Accounts in January 2026), WooCommerce orders and customers, file fields, and comments. Complex custom post types may need redesigning to fit Webflow's reference field limits. Shortcodes inside post content come through as text or disappear unless you convert them first.