Your First AI Prompts in Banking (and the Privacy Rules)
Knowing that AI can help is one thing. Getting a useful, compliant answer out of it is another. The difference between a vague reply and a deal-ready draft is almost always the prompt. This lesson teaches you the prompt structure that works for lending tasks — and, just as important, how to use these tools without ever putting customer information at risk.
What You'll Learn
- The four-part prompt structure that produces banker-grade output
- How to redact customer data so you can use AI safely and compliantly
- Five copy-paste starter prompts for common lending tasks
- How to iterate when the first answer isn't right
Privacy First: The Redaction Habit
Before any prompt, build one reflex: strip the non-public personal information (NPI) before it leaves your screen. Under GLBA and your institution's data policy, names, Social Security numbers, account numbers, dates of birth, and full addresses do not belong in a public AI tool.
The good news: AI doesn't need real identifiers to do the work. It needs the shape of the data, not the identity. Replace specifics with placeholders:
| Real data | What you paste instead |
|---|---|
| Jane Q. Borrower | "the applicant" or [BORROWER] |
| SSN 123-45-6789 | (delete entirely — never needed) |
| Acct #4471002938 | [ACCOUNT] |
| 412 Maple Street, Apt 3 | "the subject property" |
| Acme Welding LLC | [BUSINESS] |
A credit memo, a decline letter, or a financial-statement analysis works perfectly well with placeholders. You drop the real names back in afterward, inside your secure system. If your institution has an enterprise AI agreement (ChatGPT Enterprise, Microsoft Copilot with data protection, a private Claude deployment), follow your policy — but when in doubt, redact.
The Four-Part Prompt Structure
Every strong banking prompt has four parts:
- Role — who you want the AI to be.
- Task — exactly what to produce.
- Context — the situation, numbers (redacted), and constraints.
- Format — length, structure, and tone.
Weak prompt:
Explain a HELOC.
Strong prompt:
ROLE: You are an experienced consumer loan officer.
TASK: Write a short explanation of a home equity line of credit (HELOC)
for a borrower.
CONTEXT: The borrower is a homeowner with strong equity who wants
flexibility to draw funds over time for a remodel. They have never
borrowed against their home before and are nervous about variable rates.
FORMAT: Plain language, under 200 words, with one sentence on the main
risk (variable rate and using the home as collateral). Warm, reassuring
tone. No jargon.
The second prompt produces something you could send today. The structure forces you to supply what the AI cannot guess.
Five Starter Prompts for Lending Tasks
1. Explain a product in plain language
Explain the difference between APR and interest rate to a first-time
auto-loan borrower. Under 120 words. Use one simple example with round
numbers. Friendly, non-technical tone.
2. Draft a stipulation (document request) email
Draft a polite email to a mortgage applicant requesting three missing
documents: their two most recent pay stubs, last year's W-2, and a
letter explaining a recent large deposit. Keep it warm, set a 5-business-day
deadline, and reassure them this is a routine part of the process.
3. Summarize a document (paste redacted text)
Summarize the key terms of this commercial lease for a credit file:
the tenant's monthly rent, term length, renewal options, escalation
clauses, and any personal guaranty. Use bullet points. Flag anything
unusual. [PASTE REDACTED LEASE TEXT]
4. Turn analysis into a narrative
Turn these underwriting notes into a two-paragraph credit memo narrative
for a small-business term loan: [PASTE REDACTED BULLET NOTES]. Professional,
objective tone suitable for a loan committee.
5. Prepare for a borrower conversation
I'm meeting a small-business owner who was pre-qualified for a $150,000
equipment loan but is worried about the monthly payment. Give me five
talking points to walk them through affordability, plus two questions
I should ask to understand their cash flow.
Iterating When the Answer Isn't Right
The first response is a draft, not a verdict. Refine with follow-ups in the same conversation:
- Too long: "Cut this to 100 words and keep only the affordability point."
- Too formal: "Rewrite this for a borrower with no finance background."
- Missing something: "Add a sentence about the prepayment penalty."
- Wrong tone: "This sounds like a rejection. Make it encouraging — the loan is approved, we just need one more document."
Treat the AI like a sharp junior colleague: give clear direction, then edit. Two or three rounds of refinement usually gets you to a final draft.
A Word on Verification
AI will sometimes invent a regulation, miscalculate a ratio, or quote a rate that doesn't exist. For any factual claim — a guaranty fee, a TILA disclosure requirement, a debt-to-income threshold, a payment amount — verify against your authoritative source before relying on it. The prompt gets you 90% of the way; your verification is the last 10% that protects you and the borrower.
Key Takeaways
- Redact all customer NPI before using public AI — replace names, SSNs, and accounts with placeholders.
- Use the four-part structure: Role, Task, Context, Format.
- Specific context and format instructions are what turn a vague answer into a usable draft.
- Iterate with follow-up instructions; the first draft is a starting point.
- Always verify rates, ratios, and regulations against authoritative sources before relying on AI output.