AI in manufacturing: eight use cases, the data each needs, and how realistic they are

AI in manufacturing pays off soonest on the paperwork and knowledge work around production: quotes and RFQs, supplier follow-up, documents and certificates, maintenance know-how and the data moving between the ERP and everything else. The headline uses, vision inspection on the line and predictive maintenance, can work very well, but they need years of clean sensor or image data and engineering on the plant floor. Public data points the same way. When the US Census Bureau widened its business survey question in November 2025 from AI used in "producing goods or services" to AI used in "any of its business functions", reported adoption jumped most in manufacturing, by 159 percent (7.5 percentage points), according to a Federal Reserve note. That suggests much of the AI manufacturers already use is not on the production line. This guide takes eight use cases by function, says what data each one needs, and gives our judgement of how realistic each is.

A note before we start

Gigabit has not published a manufacturing case study, and we do not claim manufacturing clients. Where we use our own numbers below, they come from adjacent projects (an e-commerce operation, a support agent, a consulting firm) and we name each one for what it was. The pattern transfers; the figures are not a forecast for a factory. The realism rating on each use case is our engineering judgement, on a three-step scale: strong (works with the data most mid-sized manufacturers already have), conditional (works once a specific data gap is closed) and hard (specialist, plant-floor work with a longer path to payback).

What the public numbers say

• Adoption is still a minority. The Census Bureau's Business Trends and Outlook Survey put AI use at 19.8% of US firms as of 3 May 2026, and 37% among firms with 250 or more employees. A Federal Reserve note put it at about 18% at the end of 2025. Neither text gives a manufacturing-only level, so we do not quote one.
• Manufacturing workers are picking it up fastest. The same Fed note reports that year-on-year growth in work-related generative AI use, from the Real-Time Population Survey, was strongest in manufacturing, at about 58 percent.
• Maintenance is where the money is lost. A NIST survey-based study by Thomas and Weiss estimated $222.0 billion a year in US manufacturing maintenance costs and losses. Respondents that relied more on predictive and preventive maintenance reported 52.7% less unplanned downtime and 78.5% fewer defects than those that relied most on fixing machines after they break. That is a survey of maintenance strategy, not of AI, and better-run plants may simply do both. Read it as the size of the prize, not as what a model will deliver.

The eight use cases

1. Quality inspection

Cameras and a trained vision model check parts for visible defects: scratches, missing components, misaligned labels, bad welds. When it works, it inspects every part at line speed instead of a sample.

• Data it needs: labelled images of good parts and of each defect type, captured with fixed cameras and controlled lighting, plus a way to send a reject signal to the line's controls.
• What goes wrong: rare defects with too few examples, lighting that changes through the day, and new product variants the model has never seen.
• Realism: hard. It is proven technology but specialist work, with cameras, optics and line integration. Gigabit has not published a vision inspection project, so treat this as general guidance. One caution on tooling: AWS shut down its managed defect-detection service, Amazon Lookout for Vision, on 31 October 2025. If you build on a managed AI service, plan for the day it goes away.

2. Predictive maintenance

A model learns each machine's normal behaviour from sensor history and flags drift before a failure. AWS's own documentation for its Lookout for Equipment service is a useful description of what this takes: historical data from process historians or SCADA systems, up to 300 sensors per model, and labelled periods of past failures if you have them, for stationary equipment such as pumps, compressors, motors and CNC machines that runs under fairly steady conditions. That same page says AWS discontinued the service on 7 October 2026.

• Data it needs: sensor data (vibration, temperature, current, pressure) at a useful sampling rate, a historian that keeps it, and maintenance work orders from the CMMS that record what actually failed and when.
• What goes wrong: the sensors are not there, the work orders say "fixed" and nothing else, or failures are so rare there is nothing to learn from.
• Realism: conditional. Strong where you already have sensors and a historian on critical assets; a multi-year project where you do not. Many plants get most of the value from simple condition thresholds and good preventive schedules first, which is also what the NIST figures above measure.

3. Production planning and scheduling

Deciding what to run, on which line, in what order, given orders, capacity, changeovers and material.

• Data it needs: accurate routings, cycle and changeover times, machine capacity, bills of materials and the order book, all in the ERP or MES and kept current.
• What goes wrong: the routings in the ERP are years out of date, so any plan built on them is wrong.
• Realism: conditional. The scheduling itself is usually an optimisation problem, solved better by a planning tool than by a language model. Where a language model helps is around it: explaining why a plan changed, answering "when will order 4471 ship?" from the schedule, and turning a planner's request into a scenario to run.

4. Supply and inventory

Forecasting demand, setting reorder points, and the daily grind of purchase orders: confirmations, delivery dates and delays, mostly arriving as emails in every format a supplier can think of.

• Data it needs: order and usage history, supplier lead times, open POs in the ERP, and access to the purchasing inbox.
• Closest Gigabit evidence: in operations automation for a four-channel e-commerce business, a model parses inconsistent vendor reply emails into structured ship dates and tracking numbers and chases delays; vendor PO follow-up fell from 5 hours a week to 45 minutes. A nightly reconciliation across four systems lifted inventory accuracy from 94.2% to 99.1%. That client sells consumer electronics; it does not make them. The supplier side of the work is the same.
• Realism: strong for supplier follow-up and reconciliation; conditional for forecasting, which depends on how much clean history you have and how lumpy your demand is.

5. Maintenance and process knowledge assistants

Technicians and operators ask questions in plain language and get answers from the plant's own manuals, SOPs, past work orders and engineering notes, with a link to the source.

• Data it needs: the documents, in searchable form, with access rules; work order history; a named owner who keeps documents current.
• Closest Gigabit evidence: for a 140-person management consulting firm, we put 12,000 of its most relevant documents behind a search assistant with access control by practice area; a question in Slack returns a sourced answer in 4 seconds, and the knowledge base recovered 280 hours in its first full quarter. The documents were proposals and deliverables, not equipment manuals, but the build is the same.
• Realism: strong, with one firm rule. The assistant points to the procedure; it never replaces it. Any answer touching lockout, guarding or a safety-critical setting must quote the controlled document and send the user to it, and must say "not found" rather than guess.

6. Quoting and RFQs

Job shops and contract manufacturers live on quotes. An RFQ arrives by email with a drawing, a material, quantities and a due date; someone reads it, estimates the routing and cost, and writes the quote.

• Data it needs: the RFQ inbox, your costing rules (machine rates, material prices, setup times, margins), past quotes and their win or loss, and the ERP for current material costs.
• The rule that matters: the model reads the RFQ and drafts the quote; the price comes from your costing logic, never from the model. We make the same point for freight quotes in our AI in logistics guide.
• Closest Gigabit evidence: the consulting firm above uses a proposal assistant that drafts full proposals from a structured brief, cutting creation from 4 to 8 hours to 45 minutes of review.
• Realism: strong for intake, extraction and the draft; conditional for anything that has to read geometry from drawings, which is harder and less reliable than reading text.

7. Documents and compliance

Certificates of conformance, material test reports, supplier documents, packing lists, nonconformance reports and corrective-action write-ups. The work is reading, matching and filing, much like invoices, which our invoice processing automation guide covers in detail.

• Data it needs: the documents, the records they must match (POs, lot numbers, specifications), and written rules for what counts as a pass.
• Regulated records: if you make medical devices or drugs, records the AI creates or changes may fall under FDA's electronic records rule. FDA's Part 11 scope guidance is the place to start, with your quality team, before anything is automated.
• Realism: strong for extraction and matching with a person reviewing mismatches; the AI drafts, a person signs.

8. ERP integration and customer questions

Every use case above reads from or writes to the ERP. If orders, stock, routings and costs live in spreadsheets and email, AI has nowhere reliable to put its work, and integration comes first. Once it is in place, the easy win is often customer service: "where is my order?" answered from the ERP. In a support agent for a direct-to-consumer brand, order status was 23% of tickets; the agent resolves 64% of all tickets with no human and cut first response from 4.2 hours to 6 minutes. That brand sells home goods online, not to industrial buyers, but the question is the same one manufacturers' customer service desks answer all day.

• Data it needs: an ERP or MES with an API, and a decision on which system owns each record.
• Realism: strong, and a prerequisite for the rest.

The use cases side by side

What AI needs before it works

• A system of record with an API. An ERP or MES where orders, items, routings and stock have IDs an AI system can look up and update.
• The plant data, where the use case needs it. A historian for sensor data, a CMMS with real failure notes, image capture for inspection. If none of these exist, start with the office-side use cases.
• Documents in one place. Manuals, SOPs, specs and certificates in a store with access rules, not on personal drives.
• Written rules. Costing rules, approval limits, what counts as a pass, and when a case goes to a person.
• A test set. Real past RFQs, documents or questions with known right answers, to measure accuracy before launch and after every change. Our AI agent eval metrics guide covers what to measure.

The risks, and how to contain them

• Plant networks. Connecting operational technology (machines, PLCs, SCADA) to cloud AI services opens a path into systems that move physical things. Keep AI read-only towards the plant floor unless there is a strong reason, route data out through a controlled point, and follow NIST's Guide to Operational Technology Security (SP 800-82 Rev. 3).
• Safety. No AI output should change a safety-relevant setting or replace a controlled procedure. Assistants cite the procedure; people follow it.
• Drawings and IP. Customer drawings and your process know-how go into prompts. Check the model provider's data terms, keep processing in accounts you control, and check what your customer contracts allow.
• Wrong quotes. Prices come from costing rules, with a margin floor and a person approving quotes above a set value.
• Vendor churn. AWS has retired Lookout for Vision, AWS Panorama, IoT Analytics and IoT Events, as well as Lookout for Equipment. Keep your data and models portable, and own the integration code.
• Drift. Product mix, suppliers and documents change, and accuracy decays quietly. Monitor it like any production system; see why AI systems break after launch.

Where to start

Pick the use case with the most volume and the cleanest data, prove it on real past cases, then move to the next. For most mid-sized manufacturers that is RFQ intake, supplier follow-up, document matching or a knowledge assistant, not the line. If your ERP already ships one of these features, switch it on first. Build your own when the built-in version does not fit your process or the data spans several systems.

Where Gigabit fits

• One office-side workflow, such as RFQ intake, supplier follow-up or certificate matching: AI integration through the AI Quick Win Sprint, $8,000 fixed over 3 weeks, with integration to up to three tools, error handling and 30 days of light support.
• Systems that do not talk to each other: ERP and API integration, one integration at the same $8,000 fixed price.
• No ERP yet, or one that has to be replaced: Gigabit is an Odoo Partner, and our Odoo implementation starts at $15,000 fixed. Manufacturing set-ups are quoted above that from-price, as a fixed price before work starts; Odoo's licence is separate.
• A production agent, such as a quoting or customer-status agent: Gigabit Agents from $8,000, flat per agent.
• Not sure where to start: the AI Transformation Sprint is $25,000 over 2 weeks, maps your workflows and ends in a scoped, ready-to-build agent.

Vision inspection and predictive maintenance on the line are better handled by specialists in machine vision and industrial controls. We can connect what they build to your ERP and the rest of your systems.