The credit memo arrived in the underwriting queue at 8:47 on a Tuesday morning. The commercial relationship manager had assembled the borrower's package over three days: two years of business tax returns, a personal financial statement, twelve months of bank statements, a rent roll for the collateral property, and a three-page borrower narrative. She handed the file to her AI drafting tool, entered a prompt, and received a five-page credit memo in eleven minutes. The memo read well. The debt service coverage ratio (DSCR, the ratio of the property's net operating income to its annual debt service, the core underwriting metric for income-producing real estate) was stated as 1.31. The loan-to-value ratio (LTV, the loan amount divided by the appraised or purchase value of the collateral) was 68%. The borrower's global cash flow, which the memo identified as sufficient for the requested loan size, traced to a specific schedule in the tax returns. The relationship manager reviewed the narrative, liked the tone, and submitted the file for underwriting review. The underwriter pulled the tax returns to check the DSCR. The actual figure, calculated from the Schedule E on the most recent return, was 1.09. The file was a borderline approval at 1.09, not a clean one at 1.31. The AI had produced a plausible number, not the actual number. The relationship manager, working fast, had not verified it before submission. That eleven-minute memo almost became a loan that failed its own underwriting standard, approved on a figure that existed only inside a language model's pattern-matching of what a 1.30-range DSCR looks like in a credit memo for a property of this type.

What AI Can and Cannot Do in a Credit Memo

A credit memo is a narrative document. It tells the story of the borrower's creditworthiness: who the borrower is, what the collateral is, what the financial data shows, how the request fits the institution's credit policy, and what risks the institution is taking on and why they are acceptable. For decades, producing this narrative was a purely manual task: an analyst would read every source document, calculate each ratio, and translate the numbers into prose. AI drafting tools change the economics of that task dramatically, but they do not change its content requirements, its accuracy obligations, or its legal consequences.

What AI does well in credit memo drafting is structural and narrative: it organizes sections, produces grammatically clean prose, identifies the categories of information that a credit memo of a given type should address, and drafts language that sounds like a senior analyst's work product. For a commercial real estate credit memo, an AI tool that has seen thousands of such memos can produce in minutes a document with all the right headings, the right analytical framework, and the right level of detail in the narrative sections. That structural and organizational capability is the real productivity gain. For an analyst who might spend four to six hours assembling a complete commercial credit memo, the AI draft can cut that to one to two hours of verification and refinement.

What AI does poorly in credit memo drafting, and what creates the specific risk illustrated in the opening story, is calculation and source verification. Generative AI models produce text based on patterns. When the model writes "DSCR of 1.31," it is not calculating 1.31 from the actual net operating income and debt service in the borrower's tax return. It is generating a number that fits the pattern of how DSCR figures appear in credit memos for borrowers with profiles like the one in the prompt. That number may be close to the actual figure, or it may be wrong by enough to change the credit decision. The model does not know the difference, and it does not signal uncertainty the way a cautious analyst would. It writes the number as if it knows.

This distinction is load-bearing for every person who uses an AI drafting tool in a lending context. The model is a narrative engine. It is not a calculator, not a document parser (unless a dedicated extraction system is built on top of it), and not an auditor. Treating its output as a first draft that requires verification is the only operationally safe posture.

The AI draft is the starting point; the verified figures are the product. Releasing a credit memo means verifying every number, ratio, and fact against the source document, not reading it for coherence.

Building the Drafting Prompt

The quality of the AI draft depends significantly on the quality of the prompt. A vague prompt produces a generic memo with invented specifics. A well-constructed prompt, grounded in the actual file data and the institution's credit policy, produces a draft that is closer to correct and requires less verification effort. Building the prompt well is the first skill in AI-assisted credit memo drafting.

An effective drafting prompt for a credit memo has five components.

Component one: the role and the policy context. Tell the model what institutional context it is operating in. For example: "You are drafting a commercial real estate credit memo for a community bank with the following credit policy parameters: maximum LTV of 75%, minimum DSCR of 1.20, maximum loan size of $5 million without committee approval, and a preference for borrowers with at least five years of operating history in the subject property type." This component prevents the model from defaulting to generic templates that may not match the institution's standards.

Component two: the specific financial data, entered by the analyst. Do not ask the AI to extract figures from documents. Provide the key figures explicitly, labeled with their source. For example: "Net operating income per Schedule E, 2024 federal return: $187,400. Annual debt service at proposed terms: $142,200. Appraised value: $2,150,000. Loan amount requested: $1,462,000. Borrower's personal liquid assets per personal financial statement: $340,000." When the analyst enters these figures directly, the model uses them rather than inventing approximations. This one practice eliminates the DSCR-hallucination failure mode from the opening story.

Component three: the loan purpose and collateral description. Describe the transaction in concrete terms: the property type, location, use, occupancy status, and any relevant history. The more specific the collateral description, the more accurate the memo's risk narrative will be.

Component four: the risk factors to address. If the analyst has identified specific concerns (a single-tenant occupancy concentration, a lease expiring within 24 months, a recent environmental inquiry, a borrower who has had a prior bankruptcy), name them explicitly and instruct the model to address them in the risk section. The model will not reliably surface risks it cannot infer from general context, and a credit memo that ignores a known risk is a more serious problem than one that overstates it.

Component five: the "cite the file" instruction. End every drafting prompt with a direct instruction: "For every numerical figure or factual claim in the memo, indicate the source document and line item in parentheses so I can verify it during review." This instruction does two things: it forces the model to signal where it is working from the analyst-supplied data (which you verified entering) versus where it is generating from context (which you need to verify), and it creates a verification map that makes the review step faster. When the model cites "(Schedule E, 2024, Line 28)" next to a figure, you know exactly where to look. When it cannot cite a source, you know to be skeptical.

A drafting prompt built this way takes five to seven minutes to construct for a standard commercial real estate deal. It saves significantly more time in the verification step than it costs in the prompt-building step, and it substantially reduces the risk that the verification step will find material errors.

The Verify-Every-Figure Step

Verification is not optional and is not a quick scan. It is the step that transforms an AI draft into an institution's credit analysis, and it is the step that the relationship manager in the opening story skipped. The verify-every-figure step requires a specific method, not just a general reading.

The method is systematic and source-document-first. Before reading the AI memo for quality or narrative coherence, create a separate verification checklist. The checklist has three columns: the figure or claim in the memo, the source document and specific location where it should appear, and the actual value found in the source document. Work through every quantitative statement in the memo against this checklist before evaluating the narrative.

For a standard commercial real estate credit memo, the verification checklist covers these categories:

Income figures. Gross rental income, vacancy allowance, operating expenses, and net operating income all trace to the Schedule E or the borrower-prepared operating statement. The analyst verifies each line against the actual return. Common AI errors in this category include using gross income where net is appropriate, applying industry-average vacancy rates instead of the borrower's actual vacancy history, and treating non-recurring income as recurring.

Debt service and coverage. The DSCR calculation requires the correct net operating income figure (from the verified income section) and the correct annual debt service (from the proposed loan terms: principal amortization at the proposed rate over the proposed term, calculated or quoted from the institution's rate sheet). The model will often approximate both. Verify both. The DSCR is the most consequential single figure in a commercial real estate credit memo, and it is the one most commonly wrong in AI drafts.

Collateral values and LTV. The appraised value or purchase price, and the resulting LTV, trace to the appraisal report or the purchase and sale agreement. The model should not be generating its own collateral value estimates; if it is, the prompt construction was insufficient. Verify the LTV against the actual loan amount divided by the actual collateral value.

Borrower financial data. Income from the personal financial statement and personal tax returns, including self-employment income, Schedule C or Schedule K-1 income, and any applicable add-backs for non-cash expenses, all trace to specific documents and lines. Global debt-to-income (DTI, the ratio of total monthly debt payments to gross monthly income, the primary consumer underwriting metric also used in commercial analysis for personally guaranteed loans) requires verifying both the income figure and the complete debt inventory.

Credit history assertions. Claims about the borrower's credit history, prior derogatory items, or payment history trace to the credit report, not to the borrower's narrative or to the AI's interpolation. If the memo states "no derogatory items in the past 36 months," verify that against the credit report.

Policy compliance statements. If the memo asserts that the loan meets specific credit policy thresholds (maximum LTV, minimum DSCR, maximum term), verify each assertion against the institution's current credit policy document. AI tools do not have access to the institution's current credit policy and may apply generic or outdated thresholds.

After the quantitative verification is complete, review the narrative for qualitative accuracy: are risk factors presented accurately, are mitigants correctly described, and are any conditions or covenants correctly stated? The narrative review is faster and less mechanical than the quantitative verification, but it requires credit judgment, not just source-checking.

The total verification time for a standard commercial real estate memo, using the checklist method, is 45 minutes to 90 minutes depending on the complexity of the borrower's financial structure. Compared to the four to six hours of original drafting it replaces, the total process (prompt construction plus AI drafting plus verification) is materially faster. The time savings are real. The verification step is not negotiable.

Consumer and Small-Business Credit Memos

The verification framework applies to consumer and small-business credit memos as well, with adjustments for the different data sources and metrics involved.

Consumer credit memos (mortgage, auto, personal loan, home equity) are typically shorter than commercial memos but involve their own specific calculations. The debt-to-income ratio is the primary underwriting metric for most consumer credit products, and it requires verifying both the income figure (from W-2s, pay stubs, tax returns, or Social Security award letters, depending on the income type) and the complete debt inventory (from the credit report, including all installment and revolving obligations, plus the proposed new obligation). AI drafts of consumer credit memos commonly err in the income verification direction: using gross income rather than qualifying income, failing to apply the correct income calculation for variable or self-employment income, or omitting a debt obligation that was disclosed on the application but not prominently featured in the prompt.

Small-business credit memos (SBA loans, business lines of credit, commercial vehicle financing, equipment financing) combine elements of consumer and commercial analysis. The business entity's income is typically analyzed through the Schedule C or Schedule K-1, and the owner's personal guarantee brings in a personal financial statement and personal tax return analysis. AI tools frequently confuse the business and personal income streams, particularly when the business is a pass-through entity whose income flows to the owner's personal return. The verification step for a small-business memo must trace the global cash flow calculation explicitly, confirming that the business income is correctly isolated, the personal draw is correctly added back, and the total qualifying income is correctly calculated before debt service is applied.

For any credit type, the ECOA (Equal Credit Opportunity Act, the federal statute prohibiting credit discrimination on the basis of race, color, religion, national origin, sex, marital status, age, or the receipt of public assistance) and Regulation B (Reg B, the CFPB regulation implementing ECOA, 12 CFR Part 1002) implications of the credit memo are worth noting. The credit memo is not itself the adverse-action notice, but its content forms the basis for the adverse-action reasons if the loan is declined. If the memo contains errors in the financial analysis that contribute to the denial, and the denial triggers an adverse-action notice, the adverse-action reasons drawn from the erroneous memo are inaccurate. That inaccuracy is an ECOA problem. Verifying the memo's figures is not just a credit quality step; it is a fair-lending prerequisite.

The Human Owns the Conclusion

OCC Bulletin 2026-13, the April 2026 interagency model-risk guidance that superseded OCC 2011-12 and explicitly pulled generative AI under model-risk, fair-lending, third-party, and board-governance expectations, is clear on one point that every AI-assisted credit memo author needs to internalize: the institution's human officer, not the AI tool, is accountable for the credit decision and for the analysis that supports it.

This accountability principle has two dimensions in the credit memo context.

The first dimension is analytical accountability: the analyst who submits the credit memo under their name is representing to the institution that the analysis is accurate, that the figures are verified, and that the risk assessment reflects their professional judgment. An AI draft that the analyst submits without verification is not the analyst's memo in any meaningful sense; it is the AI's output with the analyst's name attached. When an examiner reviews the memo and finds the DSCR was stated incorrectly by a margin that would have changed the credit decision, the analyst cannot respond with "the AI generated it." The memo has the analyst's name. The institution approved the loan on the basis of a figure the analyst did not verify. The accountability is not distributed; it is the analyst's.

The second dimension is judgment accountability: the conclusion the credit memo reaches, whether approval or conditional approval or decline recommendation, is a professional credit judgment that belongs to the human analyst, not to the AI. The AI may draft the conclusion section. The analyst who submits the memo has determined whether that conclusion is correct. If the analyst believes the conclusion is wrong (perhaps the AI draft recommended approval but the analyst's full review of the file, including risk factors the AI underweighted, leads the analyst to a conditional approval recommendation with additional collateral or covenant requirements), the analyst's judgment supersedes the AI draft. The AI is a drafting tool, not a credit officer.

In practical terms, this means two things. First, the analyst's signature on the credit memo means the analyst has verified the figures, agrees with the analysis, and stands behind the conclusion. Second, if the analyst's read of the file differs from the AI draft's narrative or conclusion, the analyst corrects the draft. The correction is not a failure of the AI tool. It is the value the analyst adds to a process the AI started.

Under OCC 2026-13, institutions must also maintain model-risk documentation for AI tools used in credit analysis. For an AI drafting tool used to produce credit memos, this documentation includes the tool's intended use (drafting assistance, not credit decisioning), the verification workflow required before the draft is submitted, and evidence that the workflow was followed for individual credits. The documentation requirement reinforces the accountability principle: the institution is expected to know exactly what role the AI played in the credit process and to demonstrate that human judgment was applied at the decision points. A loan file where the AI-generated memo was submitted without visible human verification creates a model-risk documentation problem on top of the credit quality problem.

The Bank Secrecy Act and Anti-Money Laundering (BSA/AML, the regulatory regime requiring financial institutions to maintain programs to detect, report, and prevent money laundering and other financial crimes) obligation adds one more dimension to credit memo accuracy. If the borrower's financial data reveals any patterns that could indicate suspicious activity (income inconsistent with stated business activity, transactions that do not fit the borrower's profile, or unexplained large deposits in the bank statements), the credit memo is not the right vehicle for that analysis. But the credit analyst who reviews the file and observes such patterns has an obligation to route the file appropriately within the institution's BSA program. An AI tool will not make that referral. The human analyst must recognize the pattern and act on it.

Putting It Together: A Sample Workflow

The following workflow integrates AI drafting assistance into a commercial real estate credit memo process while preserving the verification controls and human accountability that compliance requires. It is designed for a loan origination system (LOS, the software platform used to manage the loan application, document collection, underwriting, and approval process from application through closing) environment where the credit memo is a required deliverable.

Step one: document review and data extraction (15 to 20 minutes). The analyst reviews every source document in the file and manually extracts the key quantitative data points onto a data sheet. The data sheet captures: income figures by category and source (W-2 income, Schedule E income, Schedule C income, K-1 income, each with the document and tax year), debt obligations (from the credit report, identified by creditor, balance, and monthly payment), collateral data (appraised value, property type, occupancy, lease terms), and proposed loan terms (amount, rate, amortization, term). This step is done by the analyst, not delegated to the AI. The data sheet is the foundation of the verification step.

Step two: prompt construction (5 to 7 minutes). Using the data sheet, the analyst constructs the drafting prompt per the five-component structure described earlier, entering the key quantitative figures from the data sheet directly into the prompt. The "cite the file" instruction is included. The credit policy parameters are stated explicitly.

Step three: AI drafting (8 to 12 minutes). The analyst submits the prompt and receives the AI draft. A brief initial review confirms that the draft addresses all required sections and is internally coherent. The analyst does not verify figures at this stage. This is a structure review only.

Step four: quantitative verification (45 to 75 minutes). Using the verification checklist method, the analyst traces every figure in the memo to its source document. Discrepancies are corrected in the draft. Sources are noted in the LOS or the document management system. This step produces the verified credit memo.

Step five: narrative and judgment review (15 to 30 minutes). The analyst reads the verified memo for analytical accuracy: are risk factors correctly characterized, are the mitigants reasonable, is the conclusion supportable given the verified financial data and the institution's credit policy? Adjustments to the narrative and conclusion are made as needed. The analyst's conclusion supersedes the AI draft's conclusion if they differ.

Step six: LOS submission and documentation (5 to 10 minutes). The verified, analyst-reviewed memo is entered into the LOS. A notation in the file records that AI drafting assistance was used and that the figures were verified against source documents per the institution's verification workflow. This notation satisfies the OCC 2026-13 model-risk documentation requirement for AI tool usage in credit analysis.

Total elapsed time for a typical commercial real estate credit memo using this workflow: approximately 100 to 140 minutes. Total elapsed time for the same memo produced entirely by hand by an experienced analyst: approximately four to six hours. The productivity gain is real, substantial, and achieved within a framework that satisfies credit quality, fair-lending, and model-risk requirements.

Key Takeaways

• AI drafting tools produce narrative structure efficiently, but they do not calculate from source documents: they generate plausible numbers based on learned patterns. Treating AI output as a first draft that requires systematic quantitative verification is the only operationally safe posture.
• The most dangerous AI error in a credit memo is a convincingly stated figure that is wrong by enough to change the credit decision. DSCR is the highest-risk figure in commercial real estate memos; DTI is the highest-risk figure in consumer memos. Verify both independently from source documents every time.
• A well-constructed drafting prompt, with analyst-supplied figures and a "cite the file" instruction, substantially reduces AI error rates and makes the verification step faster. The five minutes spent building the prompt pays off in the verification step.
• The verify-every-figure step covers income figures, debt service and coverage ratios, collateral values, borrower financial data, credit history assertions, and credit policy compliance statements. It requires a checklist and source-document confirmation, not a general reading of the memo for coherence.
• The analyst's signature on the credit memo means the analyst has verified the figures and stands behind the conclusion. "The AI generated it" is not a defense when an examiner finds an error. The accountability belongs to the human, regardless of what tool produced the draft.
• Credit memo accuracy is a fair-lending prerequisite: if the memo contains errors that contribute to a denial, the adverse-action reasons drawn from the erroneous memo are inaccurate, creating an ECOA and Reg B problem on top of the credit quality problem.
• OCC Bulletin 2026-13 requires institutions to document AI tool use in credit analysis, including the intended use of the tool, the verification workflow required, and evidence the workflow was followed. A file where the AI draft was submitted without documented human verification creates a model-risk compliance problem alongside the credit quality issue.
• The productivity gain from AI-assisted credit memo drafting is real and material: a 100-to-140-minute total process versus a four-to-six-hour manual process. That gain is sustainable only if the verification workflow is consistent. Institutions that skip verification to capture more time savings are trading a known efficiency against a contingent but severe compliance liability.