How to Use AI as a Banker: What Works, What to Avoid, and the Rules

Banking is unusual among the jobs being reshaped by AI, because so much of the actual work is deciding who gets money and on what terms, and explaining why. A model can draft the explanation. It cannot be the reason a loan was declined, and it cannot be the one who answers for it if a regulator asks. That single fact should decide where these tools sit in a banker's day, far more than what they are technically capable of.

Start with the obligations, because they narrow the options

A customer's financial information is not general-purpose input for a consumer chat tool. It is data the bank has separate legal obligations to protect, and once it is pasted somewhere you have limited visibility over how it is retained or reused. Credit decisions carry their own, older constraint: fair-lending law requires that a bank be able to explain why an application was approved or declined, in terms a regulator can follow. A model producing a fluent-sounding reason is not the same as a reason the bank can actually stand behind, and model complexity has never been an accepted excuse for not being able to explain a decision.

None of this is unique to banking, and the general shape of the guidance is not new either. FINRA's reminder on generative AI and large language models is explicit that existing supervisory obligations are technology-neutral: a firm using generative AI still needs a reasonably designed supervisory system covering model risk, data privacy and the accuracy of what the tool produces, whether that firm is a broker-dealer or a bank running the same kind of tool internally. NIST's AI Risk Management Framework is the reference point many banks build that governance around, precisely because it is not specific to one regulator or one product line.

Practically, that gives you three tiers of work. Things fine to do with anonymised or generic inputs in any tool. Things that need an enterprise tool with no training on your inputs and a data-handling agreement your compliance team has actually reviewed. And judgement calls — a credit decision, a suspicious-activity disposition — that stay with a named person regardless of which tool produced the first draft.

What it is genuinely good at

The pattern across all of these: the model produces a draft or a first pass, and you supply the judgement that turns it into something you would put your name on.

1. 01Customer correspondence. Explaining a rate change, a documentation request, or why a wire is delayed, in plain language, without writing the same explanation from scratch for the fiftieth customer this week.
2. 02Turning a completed credit analysis into a memo narrative. You, or an analyst, already worked out which ratios moved and why. The model turns those conclusions into the structured prose a credit committee expects to read, which you then check line by line.
3. 03A first pass on an alert narrative. Anti-money-laundering monitoring generates a high volume of alerts that need a written rationale for why each was cleared or escalated. A draft rationale from the transaction details saves the writing; the disposition itself is still a decision a trained analyst makes.
4. 04Reading a loan agreement or credit facility against a specific question. "Where does this agreement say anything about a cross-default provision" is a search problem over a document already in front of the tool, and that is the category these tools handle most reliably.
5. 05Formula and spreadsheet work. Describing a covenant test or a cash-flow schedule in words and getting a formula back, which you then verify by running it against a figure you can check by hand.
6. 06Rehearsing a difficult conversation. Ask it to argue a customer's likely objection to a declined application or a changed term, so you have already heard the pushback once before the real call.

A worked example

A commercial loan officer has finished spreading the financials on a renewal request: debt service coverage slipped from 1.4x to 1.1x, driven by one large customer going slow-pay. The bad version is pasting the client's full financial statements into a general consumer tool and asking it to "write the credit memo" — client data in a tool nobody vetted, and a narrative that may confidently describe trends that are not actually there.

The version that works: strip identifying details, give the tool the two facts already established — the coverage ratio and the cause — and ask for a 150-word narrative in the bank's standard memo format. The officer edits two sentences, attaches the actual supporting schedules, and the credit committee reads a memo that took a third of the time to draft. The model wrote the paragraph. The officer did the underwriting, and still signs the recommendation.

Use it where you already know the answer and the work left is expressing it clearly. That holds for every task on the list above, and stops holding the moment the model is being asked to decide something instead.

Where it will quietly get you into trouble

• —Credit decisions dressed up as questions. "Should we approve this application" or "why did we decline this one" gets you an answer with the tone of authority and none of the accountability. Fair-lending obligations mean the bank has to be able to explain the actual decision, not a plausible-sounding one generated after the fact.
• —Accepting an AML alert disposition without opening the underlying activity. Drafting the narrative is fine. Treating the draft as the investigation is not — the review is the part regulators actually check for.
• —Regulatory citations produced from memory. A confident bulletin number or rule reference is not the same as the actual current text. Open the primary source every time, the same discipline how to use AI as an accountant covers for a neighbouring profession.
• —Customer financial data going into a general consumer tool without your bank's data policy covering it. The obligation does not pause because the tool is convenient.
• —Arithmetic across a large ledger or portfolio. These are language models, not calculators — a plausible total that is quietly wrong will not raise an error the way a broken spreadsheet formula would.

Getting started without a project plan

1. 01Write down your bank's rule on customer data before you touch a tool with anything real in it, even if the rule is one sentence borrowed from an existing policy.
2. 02Pick one recurring, checkable task first. Customer correspondence is usually the right answer, because you can judge a draft instantly against what you would have written. Find the repetitive part is a short test if the choice is not obvious.
3. 03Run it both ways for a couple of weeks and keep the time saved. You want evidence, not an impression.
4. 04Treat every generated number, citation or disposition as a lead you personally verify, never as the finished answer — checking an AI answer when you are not the expert is the general version of that habit.
5. 05Expand to a second task only once the first is routine. Rolling out several uses at once tends to end with the desk using none of them properly.

The part that actually protects the job

The tasks most exposed here are the ones furthest from judgement — drafting, first-pass summarising, matching a document to a question — which is what task-level exposure research finds across white-collar work generally, and Microsoft's occupational research makes the same distinction explicitly: a task being applicable to AI is not the same as a job being displaced by it. What is left in banking specifically is the decision and the name attached to it, which regulation will not let a tool hold regardless of how good the drafting gets. The same argument, built around a fiduciary's suitability obligation rather than a lender's fair-lending one, runs through how to use AI as a financial advisor, and the audit-evidence version of it is in how to use AI as an auditor.

That also lines up with the wider labour-market picture: PwC has measured a real, growing wage premium for workers who can use these tools well, and the World Economic Forum's Future of Jobs Report lists AI and big data skills among the fastest-growing employers are hiring for — which is a reason to get fluent with the drafting and flagging work above, not a reason to hand it the decision. If it is the risk to the job itself you are weighing rather than the day-to-day practice, will AI replace financial advisors covers the closest neighbouring role where that question has actually been worked through.

The short version

Use AI for correspondence, memo drafting, alert narratives and document search, on data your bank has actually cleared for it. Do not use it for the credit decision, the AML disposition, or any judgement call the bank has to be able to explain to a regulator. Verify every citation and every number against the primary source. Write down a data rule first, pick one checkable task, measure it, and expand only once it is dull.

Coursium teaches exactly that layer — practical use of the tools plus the checking habit that has to go with it. Stay ahead of AI by learning them on your phone.