It is 2:15 on a Wednesday afternoon and Marcus, a senior loan officer at a mid-size regional bank in the Midwest, pastes a 900-word prompt into an AI assistant window. He has attached the PDF of a self-employed borrower's two-year tax return and typed: "Analyze this and tell me what the income looks like." The model responds within four seconds with a clean, confident paragraph. It states that the borrower earns approximately $87,000 per year, that the business is profitable, and that debt coverage is comfortable. Marcus approves the file and moves on. Three weeks later, the underwriter pulls the tax return and finds that Schedule C shows a net loss both years after depreciation and business expenses. The $87,000 figure appears nowhere in the document. The model invented it, probably by averaging a line near the top of the return that captured gross receipts, not net self-employment income. The loan is now heading to a credit committee review that nobody wanted. Marcus has a new appreciation for a lesson he did not know he needed: how you structure a prompt determines whether the AI gives you a useful, grounded answer or a confident fabrication dressed up to look like analysis. (This scenario is a composite illustration drawn from common field patterns; identifying details are changed.)

Why Prompting Matters Differently in Lending

In most professional contexts, a poorly written prompt produces a disappointing but harmless result. In lending, a poorly written prompt can produce an invented income figure that ends up in a credit file, an AI-generated adverse-action reason that does not match the actual denial basis, or a regulatory summary that cites a rule that does not exist. The cost of a poor prompt in credit work is not a deleted email draft. It is a compliance exposure, a potential violation of the Equal Credit Opportunity Act (ECOA, the federal statute prohibiting credit discrimination, enacted in 1974 and amended in 1976) and its implementing regulation Regulation B (Reg B, 12 CFR Part 1002), and a credit decision that may not survive examination.

This distinction changes how you need to think about prompting. Generic prompting advice focuses on clarity, tone, and format. Banking prompts need three additional properties that have no equivalent in other professional contexts:

• Grounding: The answer must come from a specific document or data source, not from the model's general knowledge about what a typical borrower's file looks like.
• Citation: The model must tell you exactly where in the file it found the number or fact it is reporting, so you can verify it in the source document.
• Policy-anchoring: The analysis must be evaluated against your institution's specific credit policy, not against generic lending norms that may differ from your underwriting guidelines.

A prompt that achieves all three is not complicated to write. But it requires a different mental model of what you are asking the AI to do. You are not asking it to analyze a borrower. You are asking it to be a reading tool that extracts and applies specific information from specific documents, and to refuse to fill in gaps from its training data.

In credit work, the prompt is a contract: you tell the model exactly what documents to use, exactly what your policy requires, and exactly how to flag uncertainty. A prompt that omits those terms invites the model to substitute its own assumptions for the facts in the file.

The Three Habits That Separate Useful from Harmful

Nearly every prompt failure in lending can be traced to the absence of one or more of three habits. Understanding what each habit does and why it matters is how you build a reliable practice rather than occasional luck.

Habit One: Context First

Context means telling the model who is asking, what the purpose of the analysis is, and what constraints apply before you ask the actual question. A prompt that opens with "analyze this tax return" leaves the model to fill in the context from its training data: what kind of lender, what kind of borrower, what standards, what jurisdiction. The model will do this filling-in automatically and invisibly. The context it supplies will be generic and will not match your institution's specific requirements.

Compare that opening with this one: "You are a credit analyst at a federally regulated mortgage lender subject to ECOA and Reg B. I am reviewing a self-employed borrower's federal tax return for a conventional mortgage application. Our policy requires two years of filed tax returns and uses IRS Schedule C net income after depreciation add-back as the qualifying income basis. Do not infer, estimate, or calculate any figure that is not explicitly stated in the document I provide."

That context accomplishes four things before you ask a single question. It establishes the regulatory frame (ECOA, Reg B, federally regulated). It establishes the policy basis (Schedule C, depreciation add-back, two-year requirement). It establishes the document standard (filed returns, not estimates). And it establishes the refusal instruction (do not infer). Every subsequent question in that session inherits those constraints. The model is now oriented to behave like a document reader operating under credit policy, not like a general financial advisor making reasonable inferences.

Habit Two: Cite the File

Citing the file means requiring the model, in every response, to name the specific line, page, field, or section of the document where it found the information it is reporting. This is not a nice-to-have. It is the mechanism that makes AI output verifiable.

Without a citation requirement, the model will answer your question about income with a number. With a citation requirement, the model will answer with a number and a pointer: "Schedule C, Line 31 (Net Profit or Loss), Year 1: $42,300; Year 2: $38,900. Note: depreciation add-back per Schedule C, Line 13 was $4,100 in Year 1 and $3,800 in Year 2, bringing adjusted qualifying income to $46,400 in Year 1 and $42,700 in Year 2, averaged to $44,550 over two years." Now you can open the tax return and check every number in under two minutes. The citation transforms the output from an answer you must trust into an answer you can verify.

The cite-the-file habit also forces the model to flag when it cannot cite. If the document does not contain the information needed to answer the question, a model operating under a citation requirement will say "I cannot find this information in the document provided" rather than supplying a plausible estimate. That refusal is exactly what you want. The alternative, a confident figure with no source, is the Marcus scenario from this lesson's opening.

Habit Three: Policy as Constraint

The third habit is stating your institution's credit policy as a constraint rather than asking the model to apply general lending standards. This matters because credit policy varies significantly between institutions. A model trained on general lending knowledge will apply norms that approximate the market average. Your institution's policy may require tighter debt-to-income limits, different income documentation standards, specific collateral rules, or geographic restrictions. Without your policy in the prompt, the model's analysis may be technically coherent and completely misaligned with your actual underwriting criteria.

The policy-as-constraint habit has an additional compliance dimension. The Community Reinvestment Act (CRA, the statute requiring banks to meet the credit needs of the communities they serve, particularly low-to-moderate income neighborhoods) and Unfair, Deceptive, or Abusive Acts or Practices (UDAAP, the broad consumer protection standard under the Dodd-Frank Act) both require that your credit criteria be applied consistently. If an AI tool is applying different standards to different applicants because different prompts supply different policy context, you have created a fair-lending exposure. The policy-as-constraint habit, applied consistently, is one of the mechanisms that keeps AI-assisted analysis from drifting into inconsistency.

Building a Credit-Work Prompt, Step by Step

The three habits translate into a specific prompt structure that works across most credit analysis tasks. Here is the structure, followed by a worked example.

A well-formed credit-work prompt has five components:

• Role and regulatory frame: What kind of analyst the model should behave as, and what regulatory regime governs the analysis.
• Task and document specification: Exactly what task you need performed, and exactly which document or documents are the source of truth.
• Policy anchor: The specific policy rules that govern the analysis (income method, DTI limits, documentation requirements).
• Citation requirement: An explicit instruction that every figure must be cited to its source location in the document.
• Refusal instruction: An explicit instruction that the model must not infer, estimate, or fill gaps from general knowledge.

Here is a worked example for a W-2 income verification task. The document attached is a single-year W-2 from a borrower applying for a home equity loan (HELOC).

The prompt reads: "You are a credit analyst at a federally regulated bank operating under ECOA, Reg B, and OCC model-risk requirements per OCC Bulletin 2026-13. I am verifying employment income for a HELOC application. The attached document is the borrower's W-2 for the most recent tax year. Our policy uses Box 1 (Wages, Tips, Other Compensation) as base income. If Box 1 is not visible or readable, report that fact and stop. For each figure you report, cite the Box number and the exact dollar amount shown. Do not average, project, or estimate any figure not explicitly shown in this document. If the document does not contain enough information to answer a question, say so rather than estimating."

That prompt will reliably produce a response along these lines: "W-2 Box 1 (Wages, Tips, Other Compensation): $74,250.00. W-2 Box 12 (Deferred Compensation Code D): $6,500.00. Note: Box 12 Code D represents pre-tax retirement deferrals. Depending on your policy, this may or may not be addable to base income. I cannot determine from this document alone whether your policy permits the add-back."

The last sentence is exactly what you want. The model is not guessing whether your policy permits the add-back. It is flagging a decision point and returning it to you. That is the behavior of a useful tool operating under the right constraints.

Common Prompt Failures and What They Produce

Cataloging the failure modes helps you recognize them in practice. Each failure has a characteristic symptom in the output.

The open-ended income prompt. Prompt: "What is this borrower's income?" Failure: The model identifies income from multiple sources, applies its own averaging logic, and produces a single qualifying income figure based on norms it learned in training. The figure may be plausible but is based on assumptions the analyst never authorized. The symptom is a specific, confident number with no cited source.

The context-free policy prompt. Prompt: "Does this borrower meet our debt-to-income requirements?" Failure: The model has no idea what your requirements are. It applies an approximate market norm (often 43 percent back-end DTI, the qualified mortgage standard under the Ability-to-Repay rule) and tells you whether the borrower meets it. Your policy may be 45 percent or 38 percent. The symptom is a clear pass or fail with a DTI threshold that is not your institution's threshold.

The implicit-estimation prompt. Prompt: "Based on the tax return, what will this borrower's income be next year?" Failure: The model estimates future income based on trends it sees in the provided document, using projection logic it learned in training. The symptom is a future income figure with no disclaimer that it is a projection, not a document-sourced fact.

The missing-refusal prompt. Prompt: "What is the property value?" (attached document is an application, not an appraisal). Failure: The model reports the borrower's self-stated property value from the application as though it were an appraised value, or estimates a value based on location data. The symptom is a dollar figure where no appraisal document was provided, sometimes with a hedge buried in a subordinate clause that the reader misses.

Each of these failures can be prevented by the three-habit structure described above. The failures are not random. They are predictable consequences of prompts that leave the model without the constraints it needs to behave like a reliable document reader rather than a general knowledge synthesizer.

The Human Accountability Line

Better prompting substantially reduces the rate of AI error in credit work, but it does not eliminate the need for human review and sign-off. This distinction is important, and it has a specific legal dimension under ECOA and OCC Bulletin 2026-13 (the April 2026 interagency model-risk guidance that superseded OCC 2011-12 and explicitly brought AI and generative AI under model-risk, fair-lending, third-party, and board-governance expectations).

OCC Bulletin 2026-13 establishes that the institution, not the AI tool, retains accountability for every AI-assisted decision. The bulletin's model-risk framework requires that AI outputs be validated before use in credit decisions, that the institution understand how the AI produces its outputs, and that human oversight be maintained at defined decision points. A well-prompted AI that produces a well-cited income analysis still requires a human credit analyst to open the source document, confirm that each cited figure is correctly read, and sign off on the analysis before it enters the credit record. The prompt structure makes that verification efficient, sometimes reducing a manual two-hour income analysis to a ten-minute document check. It does not replace the check.

In practice, the accountability line sits at the point where the AI output becomes part of the credit record. Before that point, the AI can draft, analyze, flag, and calculate. After that point, a named human analyst or officer has reviewed the output, confirmed the citations, and accepted responsibility for the accuracy of the information entering the decision. The Bank Secrecy Act/Anti-Money Laundering (BSA/AML, the combined regulatory regime governing financial crime compliance) context adds a parallel accountability requirement: AI can help triage alerts and summarize suspicious activity, but Suspicious Activity Reports (SARs, the federal filings required when a bank identifies potential money laundering or financial crime) must be filed based on human review and human judgment. No prompt structure changes that.

The 38 percent of mortgage lenders who were using AI in credit work by 2024 (up from 15 percent in 2023) learned this distinction through experience, often expensively. The institutions that built verification disciplines early, where AI drafts and a human confirms before the output reaches the file, have generally fared better in examinations and in credit quality than those that treated AI output as a finished product. The performance gap is not about model quality. It is about whether a human is closing the loop before the output matters legally.

Prompting for Different Lending Tasks

The three-habit structure adapts to different credit-work tasks by varying what goes into the policy anchor and the document specification. Here is how the structure applies across the most common lending AI use cases.

Income and Asset Extraction

Policy anchor: income method (Schedule C with depreciation add-back, W-2 Box 1, 1099 average over 24 months), documentation requirements (how many years of returns, whether paystubs are required), and asset documentation standard (two months of statements, retirement account eligibility, gift letter requirements). Document specification: the specific filed tax form, statement, or paystub attached. Citation requirement: form, line number, and exact dollar amount for every figure.

Adverse-Action Reason Drafting

Policy anchor: your institution's Reg B reason code library (the specific codes your institution uses, mapped to the credit policy criteria they represent), the denial basis from the credit decision memo, and any policy-level factors that contributed (geography, product-type restrictions, exception requirements). Document specification: the credit decision record and the file data supporting each reason. Citation requirement: for each proposed reason code, the specific data point from the file that supports it. Refusal instruction: do not propose any reason code that is not directly supported by a data point in the file provided; if a reason cannot be cited to the file, flag it as unverifiable rather than including it.

The adverse-action use case has a specific risk that the income-extraction case does not: the Loan Origination System (LOS, the software platform that manages loan applications from intake through funding) may auto-generate reason codes that were not explicitly reviewed against the prompt-structured analysis. Any LOS-generated reason code that bypasses the prompt-structured citation check is a potential Reg B exposure. A well-structured prompting workflow for adverse action should feed into a review step where the credit analyst compares the AI-proposed reasons against the LOS-generated codes and reconciles differences before the notice is finalized.

Credit Memo Narrative

Policy anchor: your institution's credit memo template (the sections required, the analytical standards for each section, the format for risk ratings and recommendations), the credit decision being supported, and the risk rating criteria. Document specification: the financial statements, credit reports, and collateral documentation being summarized. Citation requirement: every financial figure in the narrative must be traceable to a specific statement, line, or report. Refusal instruction: do not characterize the borrower's risk favorably or unfavorably beyond what the cited documents support; flag any area where the documentation is insufficient to reach a conclusion.

Key Takeaways

• The three habits that define a useful banking prompt are context first (establishing the regulatory frame, the institution type, and the policy basis before asking the question), cite the file (requiring the model to name the specific line, page, or field where every reported figure comes from), and policy as constraint (stating your institution's actual underwriting criteria rather than letting the model apply generic market norms).
• An open-ended prompt in credit work does not produce a neutral result. It produces a response based on the model's training data about what typical files look like, which may bear no relationship to the specific document in front of you or to your institution's specific policy requirements.
• The citation requirement is the mechanism that makes AI output verifiable. A model that cites its sources produces output you can check in under ten minutes. A model that does not cite produces output you must either trust or re-do from scratch.
• The refusal instruction is as important as the question. Telling the model not to estimate, infer, or fill gaps from general knowledge is what prevents a confident fabrication when the document does not contain the information needed to answer the question.
• OCC Bulletin 2026-13 establishes that the institution retains full accountability for AI-assisted credit decisions. Better prompting reduces error rates; it does not transfer accountability from the human reviewer to the AI tool. The human credit analyst who signs off on an AI-assisted analysis is responsible for its accuracy.
• The human accountability line in credit work sits at the point where AI output enters the credit record. Before that line, AI drafts and analyzes. After that line, a named human has reviewed, verified the citations, and accepted responsibility.
• ECOA and Reg B compliance requirements apply to AI-assisted adverse-action drafting exactly as they apply to manual drafting. A prompt that produces reason codes not grounded in the specific file creates the same Reg B exposure as a manually written notice with inaccurate reasons.
• Consistent prompt structure, applied uniformly across analysts and files, is also a fair-lending control. Variable prompting that applies different policy constraints to different applicants creates the inconsistency that fair-lending examiners look for and that CRA examination scrutinizes in geographic distribution of credit.