AI in fintech pays off first on the paperwork around money, not on the decision to move it. The work that suits an agent is reading onboarding documents, sorting monitoring alerts, assembling dispute evidence, matching bank lines to the ledger and drafting support replies. The release of funds, the credit decision and the suspicious-activity filing stay with written rules and named people. The industry already works this way: in a 2024 Bank of England and FCA survey of 118 UK financial firms, 75% were using AI, but only 2% of their AI use cases made fully autonomous decisions. This guide takes eight use cases, says what the agent does in each, where a person or a rule decides, and what the work needs first.
AI use cases in fintech at a glance
| Use case | What the agent does | Where a person or a rule decides | Typical systems |
|---|---|---|---|
| Onboarding documents (KYC and KYB) | Sorts documents, extracts fields, flags mismatches, assembles the case file | The onboarding decision, beneficial owners, any screening hit that is not clearly false | Identity verification provider, company registries, case management |
| Transaction-monitoring alerts | Groups related alerts, pulls account history, drafts the case narrative | Whether to clear, escalate or file a report | Monitoring engine, core ledger, case management |
| Fraud and disputes | Collects order, delivery and login evidence and drafts the response | Refunds above a limit, account closures | Processor dispute tools, order system, support desk |
| Payment-compliance documents | Reads waivers, certificates and contracts and drafts the reason a check failed | Rules release the payment; a person takes the exceptions | Rules engine, document store, payment processor |
| Reconciliation | Proposes matches for bank and processor lines the rules could not match | Write-offs and manual journal entries | Bank feeds, processor reports, ledger or ERP |
| Customer support | Answers status and how-to questions from account data, drafts replies | Disputes, complaints, hardship and anything regulated | Support desk, core APIs, help center |
| Credit and underwriting support | Extracts figures from statements and tax forms, summarizes the file | The credit decision and the reasons given to the applicant | Loan origination system, bank data, document store |
| Compliance knowledge and reporting | Answers staff questions from policy with a citation, drafts report sections | Policy changes and sign-off on anything filed | Policy library, data warehouse |
A note before we start
Gigabit has published two fintech builds: a B2B payments platform and a zero-downtime move from Heroku to AWS. Neither is an AI agent. We have not published a KYC case study or a fintech agent case study, and we do not claim one. Where we use our own numbers below, they come from those two builds, and we say what each one was. The realism rating on each use case is our engineering judgment: strong (works with the data most fintechs already have), conditional (works once a named gap is closed) or hard (the decision itself, which should not be handed to a language model).
What the public numbers say
- Most firms use AI, and few let it decide alone. In the Bank of England and FCA survey, published 21 November 2024, 75% of firms were already using AI and another 10% planned to within three years. Respondents said 55% of their AI use cases had some degree of automated decision-making, and only 2% were fully autonomous. A third of all use cases were third-party implementations. These are UK banks, insurers and other regulated firms, not startups.
- Customers already talk to bots. The CFPB reported in June 2023 that about 37% of the US population, over 98 million people, used a bank's chatbot in 2022, and that all of the top 10 commercial banks had deployed one. The same report lists the risks: breaking consumer financial law, weaker service when people cannot reach a person, and harm from wrong answers.
- One company's support numbers. Klarna said in February 2024 that its AI assistant handled 2.3 million conversations in its first month, two-thirds of its customer service chats, and resolved errands in under 2 minutes against 11 minutes before. That is a company describing its own product at very high volume. Read it as what is possible for simple questions, not as a forecast.
- Fraud volume keeps rising. The FTC told Congress in March 2026 that consumers filed 3 million fraud reports in 2025 and reported $15.9 billion in losses. That measures the workload, not what AI does about it.
- Regulators name the same risks. A GAO report from May 2025 found AI in use for automated trading, credit decisions and customer service, and listed potentially biased lending decisions, data quality issues, privacy concerns and new cybersecurity threats.
The eight use cases
1. Onboarding documents (KYC and KYB)
US rules set four core parts of customer due diligence: identify and verify the customer, identify and verify the beneficial owners of companies opening accounts, understand the nature and purpose of the relationship, and conduct ongoing monitoring. FinCEN's CDD rule page puts the ownership test at any individual who owns 25 percent or more of a legal entity, plus an individual who controls it. The page also lists exceptive relief issued in 2026, so check the current rulings with your compliance counsel.
Identity verification itself, the document and liveness check, is bought from a specialist provider. The agent's job is the work around it: sorting what arrived, rejecting unreadable files, extracting fields, flagging where the form and the document disagree, and assembling one case file for the analyst.
- Data it needs: the provider's results by API, registry data, your written onboarding policy and a case-management queue.
- What goes wrong: a model accepts a low-quality scan with confidence, or two records that look alike are merged.
- Realism: strong for preparation; hard for the decision, which stays with a person. Our page on AI for KYC and KYB goes step by step.
2. Transaction-monitoring alert triage
Monitoring engines produce more alerts than analysts can read closely. An agent can group alerts that belong to one pattern, pull the account history and prior cases, and draft the narrative an analyst would otherwise type. It does not clear an alert and it does not file anything.
- Data it needs: alerts and case history by API, and past cases with their outcomes to test against.
- What goes wrong: the draft reads well and is wrong. Every sentence in the narrative should point to the transaction it came from.
- Realism: conditional. It depends on a case-management system the agent can read and write, and on enough labeled history to measure it. We have no published case here.
3. Fraud and disputes
Fraud scoring at the moment of payment is the processor's job and should stay there. The slow part is afterwards: a dispute arrives, and someone collects the order, the delivery proof, the login history and the customer messages into a response before the deadline. An agent can assemble that file and draft the response for review.
- Data it needs: the processor's dispute feed, order and delivery records, and your refund policy in writing.
- Keep a person on: refunds above a set value, account closures and anything that looks like account takeover.
- Realism: strong for evidence assembly.
4. Payment-compliance documents
This is where our own evidence sits. In the construction payments platform we built, a general contractor cannot release money to a subcontractor until the paperwork clears. A rules engine checks the lien waiver (signed, correct amount and period), the insurance certificate (current, meets contract minimums), the contract terms (the draw matches the schedule, retainage is correct) and that the prior payment settled. Every rule returns pass or fail with a specific reason.
In its first year the platform processed $41.2M across 3,847 transactions with a 0.02% processing error rate. Compliance documents were verified automatically 94.6% of the time, and the other 5.4% were flagged for manual review. Subcontractors' days sales outstanding fell from a 67-day average to 12. The MVP cost $118,000 over 4 months.
The engine is rules, not AI, and that is deliberate. A check that releases money has to give the same answer every time and explain itself. The agent belongs on the 5.4%: read the document that failed, classify it, draft the reason and route it to a person. Our guide to rules-based automation versus AI agents covers where that line sits.
- Realism: strong, if the rules are written down first.
5. Reconciliation
Exact matches between bank lines, processor payouts and ledger entries are a job for rules. What is left is the pile with a missing reference, a split payment or a fee netted out. An agent can propose a match with its reasoning and leave the posting to a person.
- Data it needs: bank and processor feeds, the ledger by API, and tolerance rules.
- Keep a person on: write-offs, manual journals and any change to a customer balance.
- Realism: strong where the ledger is sound. If the ledger does not balance to the cent today, fix that first.
6. Customer support
Most questions to a fintech are about status: where a payout is, why a payment failed, how to change a detail. An agent can answer those from account data and hand the rest to a person with the context attached. Two rules matter. The agent reports what the system says and never invents a balance, a date or a policy. And a customer can always reach a person, which is the failure the CFPB report describes when bots stand between people and help.
- Data it needs: read access to accounts and payments, a current help center and a support desk with handoff.
- Keep a person on: disputes, complaints, hardship, fees and anything a regulation covers.
- Realism: strong for status and how-to; conditional beyond that.
7. Credit and underwriting support
A lender that turns an applicant down in the US has to say why. Regulation B requires a statement of specific reasons that indicates the principal reasons for the action, and says that pointing to internal standards or policies, or to a failed credit score, is insufficient. That shapes where AI fits. An agent can extract income and cash-flow figures from statements and tax forms and summarize the file. The decision should come from a documented scorecard or an underwriter whose reasons can be stated.
- Data it needs: the application, bank data and documents, and the credit policy in writing.
- What goes wrong: a model's summary leaves out the one fact that mattered. Show the source line next to every extracted figure.
- Realism: conditional for document work; hard for the decision.
8. Compliance knowledge and reporting
Staff ask the same policy questions all week: which documents a customer type needs, what the limit is, who approves an exception. A search assistant over the policy library answers with a link to the paragraph, and says not found when the policy is silent. The same pattern drafts sections of board and partner-bank reports from the data warehouse for a person to check and sign.
- Realism: strong, with a named owner who keeps the documents current.
Rules, models and agents under bank guidance
If you work with a partner bank, its model risk team will ask how your AI is governed. The reference point changed this year. On 17 April 2026 the Federal Reserve, OCC and FDIC issued revised guidance on model risk management (SR 26-2), which replaced SR 11-7 from 2011. Three points in the guidance text matter for a fintech:
- Rules are not models. The term model excludes deterministic rule-based processes with no statistical, economic or financial theory behind them. A compliance gate built from explicit rules sits outside it.
- Generative and agentic AI are out of scope. A footnote says they are novel and rapidly evolving and not within the scope of the guidance, and that a bank's own risk management and governance should decide the controls for tools it does not cover.
- It is aimed at large banks. The agencies expect it to be most relevant to banking organizations with over $30 billion in total assets.
The practical result is that there is no federal template for governing an agent. Expect each bank to apply its own. Bring the evidence yourself: an inventory of what each agent can read and do, a test set with results, logs of every action, and the list of steps where a person approves. This is our reading of a published document, not legal or compliance advice.
What AI needs before it works
- Systems with APIs. A ledger, a case-management tool and a support desk the agent can read from and write to. A shared inbox does not count.
- Written rules. Onboarding policy, refund limits, tolerance levels and who approves what. If two analysts would decide a case differently, an agent will too.
- A queue with a person on it. Every agent needs somewhere to send what it cannot finish. Our guide to human-in-the-loop for AI agents covers how to place the checkpoints.
- A test set. Real past cases with known outcomes, to measure accuracy before launch and after every change. See AI agent eval metrics.
- An audit log. Every read, draft and action recorded against the case, so a reviewer can replay it.
- A platform that can change safely. In our migration project, a payments platform handling about $14M a month moved from Heroku to AWS in 14 weeks for $74,000 with zero downtime. The payment core moved last, after 48 hours of shadow traffic across 12,000 transactions showed no discrepancy, and traffic shifted from 1% to 100% over 10 days.
The risks, and how to contain them
- Wrong answers about money. Numbers come from the ledger, never from the model. If the data is missing, the agent says so and escalates.
- Decisions nobody can explain. Keep explicit rules on anything that releases funds, sets a price or declines a customer.
- Instructions hidden in documents. An uploaded file can carry text aimed at the model. Treat every document as data, limit what the agent can do, and test for it. An AI Security Assessment covers prompt injection and data leakage.
- Customer data in prompts. Names, account numbers and documents go to the model. Check the provider's data terms and keep processing in accounts you control.
- Bias in lending. GAO lists it first among the risks. Keep the model out of the decision and test the inputs.
- Drift. Document formats, processors and policies change, and accuracy decays quietly. Monitor it like any production system.
Where to start
List every check and queue that can hold money back, and put the dollars waiting next to each one. Start with the largest that has a written rule and a clean system of record. For most payment and lending companies that is a document queue, not a chatbot. If your provider already ships the feature, switch that on before building.
Where Gigabit fits
- One workflow, such as dispute evidence or exception drafting: the AI Quick Win Sprint 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 on one queue: Gigabit Agents start at $8,000, flat per agent, with the evaluation set and audit log included.
- Not sure which queue: the AI Transformation Sprint is $25,000 over 2 weeks and ends in a scoped pilot. The fee credits toward an agent build.
- An AI feature that has to pass a partner's review: the AI Security Assessment is from $12,000 fixed over 2 to 4 weeks.
- The platform underneath: fintech app development for payment platforms, ledgers and compliance checks, quoted as a fixed price in writing.
- After launch: Managed AI Operations runs $3,000–$20,000 a month, by the number of systems.
We do not sell identity verification or fraud scoring, and we do not let a model decide what money moves. Buy the first two from specialists. Our fintech page shows what we build around them.
