AI for Fintech and Payments: Where It Pays and Where the Rules Bite

Start where the answer already exists. In a fintech or payments firm, that means three places.

First, the support queue. When will my payout land? Why is this transfer on hold? What does this fee mean? Second, the failed card payment that needs a reminder at the right hour. Third, the onboarding form someone has to read.

AI does those well. And the rules that apply are ones a good support team follows now.

Each step toward acting on an account pays less. It also costs more to run safely. The same goes for moving money or deciding who gets credit. So build from the bottom. First the AI answers. Then it reads the user's own account. Then it acts, with a yes from the user. Leave moving money and credit decisions with people. Wait until the lower rungs have earned their keep.

The Autonomy Ladder: what AI should do in fintech, rung by rung

We sort each fintech use onto four rungs, drawn at the top of this page. Each rung adds a skill, and each one adds rules.

The ladder is the plan as well as the map. Start on rung one and a team learns three things. It learns where its help docs are wrong. It learns how users really phrase things. And it learns which questions a machine should never answer. All three are needed before rung three is safe.

A team that starts on rung four learns them too. It just learns them in front of a regulator.

Where AI pays in fintech, ranked

This ranking is for a fintech or payments firm of 5 to 50 people. It goes by value. It also goes by how soon each use pays back.

1. Support on the repeat questions (rung one). The same questions come in through Slack, WhatsApp, email and the app. The founder answers them. No one else knows how. A support agent takes 40 to 60% of those questions off founders and staff. For a team that size, we estimate it is worth $100,000 to $400,000 or more in year one.
2. Answers that use the account (rung two). "Where is my payout?" needs the user's own record. That means identity checks and access rules. It means the Safeguards Rule in full. It's worth it once rung one is stable.
3. Billing and failed-payment recovery (rung three). Trials that turn into paid plans. A reminder at the right time after a card fails. A self-serve portal, so the user fixes it with no ticket. The card number never touches the AI. We estimate a billing layer like this is worth $100,000 to $1 million or more. That is per year, and it depends on scale. Much of that comes from three things. Billing runs itself. A newsletter writer is built to cut content time by 60 to 80%. And a web app keeps subscriptions out of the app stores. That keeps them out of the stores' 15 to 30% cut. On one fintech product, we built that layer. It has self-serve signup. Trials run 7, 14, 15 or 30 days. Staff can override the length for any one user. It has Stripe subscriptions. Its reminders step up after each failed payment. There is a web app that works on phones. And there is an admin dashboard where staff can review each answer.
4. Dispute intake (rung three). The agent spots the dispute and logs it with the time. It opens the case and hands it to a person. The value is the fine you never pay.
5. Onboarding and document checks (rung three). It reads the ID and the business papers. It pulls out the fields and flags what doesn't match. A person approves.
6. Fraud scoring and credit decisions (rung four). These are the biggest prizes in the field. They carry the heaviest rules. Most teams this size buy fraud scoring. They get it from their processor or an outside expert. Credit decisions stay with an underwriter, the person who judges credit risk. The model they use is one they can explain.

How a fintech support agent should be built

The pattern that holds up starts with a knowledge base. That is the set of written answers the agent draws on. Your team wrote it, across chats and docs. It is cleaned, then indexed so it can be searched. The agent answers only with a citation. And it hands off when it isn't sure.

• One source of truth. Pull the answers out of where they live, into one index. That means Slack threads, chat groups, help docs and policy PDFs. Note where each one came from.
• Clean it first. Before indexing, run a step that finds personal data in the history. That way, an account number from a 2024 thread stays out. It never ends up in someone else's answer.
• Answer with the source. Each answer carries the snippet it came from. That is how staff catch an old fee table. The design is in RAG with citations.
• A confidence line, not a guess. The agent scores how sure it is. On an approved topic, that score may fall below a set line. Then it says so and hands off. On our build, the line was set at 95%.
• Watch the dashboard. It shows the top questions. It shows the deflection rate and the spread of scores. Deflection rate is the share of questions the agent closes with no person. The three tell you which docs to fix next.

At one payments firm, the founders kept answering the same questions. The questions came in on Slack, WhatsApp and Telegram. We built one knowledge base. It drew on those channels and on the docs the team uploaded. It has a dashboard of what is asked and what is deflected. It has a path to a person. It was built to cut repeat questions by 40 to 60%. And it was built to send a first reply in under 10 seconds.

Most of the value is kept or lost at the handoff. Hand a person the whole thread and the agent's draft. They can close it in a minute. The longer case for that is in how to reduce churn with AI customer support.

The sentence your agent must not miss

The CFPB looked at chatbots in consumer finance. Its report came out in June 2023. It is plain about where they fail. Scripted bots may miss a dispute. They log it only if the user says certain words. And they may not notice a user claiming their rights under US law.

A large language model, or LLM, is the AI behind today's chat tools. It is better at hearing "that wasn't me," however it's said. It still has to act on it.

Reg E sets the clock. A notice of error can be spoken or written. It counts if it reaches the firm in time. That means no later than 60 days after the firm sent the statement that shows the error. Then the firm has 10 business days. In that time it looks into it and decides. It has 3 more to report the result. And it has 1 to fix an error it finds.

It can take up to 45 days instead. But it must give provisional credit within those 10 business days. That means it puts the money back for now, while it looks.

So the design rule is short. Say a message reads as a dispute of an electronic transfer. The agent stops answering and opens a case. The case holds the time, account, amount and date. It also holds the user's own words. Then it goes to the team that owns disputes.

The agent doesn't argue. It doesn't troubleshoot. It doesn't ask the user to come back later. Card disputes and credit card billing errors have their own rules. The same reflex applies.

The rules that apply to AI in fintech

The FTC is the US trade watchdog. PCI DSS is the card brands' rule book for keeping card data safe. SOC 2 is an outside report on how well a vendor guards data.

This is not legal advice. It is the list your lawyer will start from. The design should answer each line of it. Do that before the first user types a question.

What it costs to run

We scoped the support agent above at $50 to $500 a month in tools and model costs. Clients at 1,000+ daily AI interactions typically see all-in model costs of $200 to $1,200 a month. Those bills sit on their own accounts. So no one marks up the tokens.

Most fintech questions need only a small, cheap model reading the right doc. The costly model is for the few that need reasoning. And a handoff costs a person's time, not tokens. The full breakdown is in what it costs to run an AI system.

The model isn't what takes the time. The time goes into the links to your ledger, your processor and your support desk. It goes into the access rules for rung two. And it goes into the test set. That is a few dozen real questions. Each has the right answer and source. You run it again after each change.

What goes wrong with AI in fintech

What to ask an AI vendor

• Which of our systems will the agent read? Which will it write to? What does it need a person to approve? Put the three lists in the contract.
• How does it spot a dispute or a notice of error? What exactly happens next? Show us a test.
• How is personal data found and masked before indexing? How are card numbers kept out of transcripts and logs?
• Who are your subprocessors, the model provider among them? How do you meet our Safeguards Rule duty to watch over them?
• Is each answer cited to a source doc we can open?
• What is the confidence line for a handoff? Can we see the spread on our own questions?
• How many asks does it take a user to reach a person?
• Whose accounts do the model, hosting and logs run on? If we leave, what do we keep?

A vendor who answers the first two with a test, not a slide, has built this before.

When not to use AI in a fintech

Light volume. Fewer than 30 to 50 questions a week? A well-kept FAQ page does the job.

Nothing written down. The agent answers from your docs. Does the know-how live only in the founder's head? Write it down first.

You want it to move money on its own. An agent that answers and routes is a support system. One that issues refunds, credits or approvals with no person is another kind of system. It needs other controls. It belongs at the top of the ladder. It comes after each rung below it has run clean.