The spreading request came in on a Friday afternoon with a note from the commercial lender: "Three years of financials, just need the spread and the DSCR story by Monday morning." In the old world, the credit analyst would have spent most of Saturday extracting income statement and balance sheet figures from three sets of tax returns into a spreading template, calculating the ratios by hand, and then writing the commentary that explained what the trends meant. The total time was roughly six hours, spread across an evening and a Saturday morning. In the new world, the analyst loaded the tax returns into the institution's AI-assisted spreading tool, received a populated template in under thirty minutes, and then used a generative AI tool to draft the ratio commentary. By noon on Saturday, the spread and the narrative were both done. The analyst spent Sunday doing something other than banking. The Monday morning deliverable was on time, it was clean, and it was defensible, because the analyst had verified every figure the AI extracted before accepting the spread, and had confirmed the commentary's analytical claims before submitting the narrative. That is the right story. The wrong story is the version where the analyst accepted both the spread and the commentary without verification, submitted on Friday at 4:00 PM to clear the queue, and the commercial lender discovered on Monday that the Year 2 depreciation add-back was calculated on the wrong line of the return, producing a DSCR of 1.28 instead of the correct 1.07, and the loan they had verbally approved at 1.28 was actually below the institution's 1.20 floor.

What Spreading Is and Why AI Changes It

Financial statement spreading (spreading, the process of extracting income statement and balance sheet data from borrower-provided financial statements into a standardized analytical template, recasting the figures to remove non-recurring items and normalize for comparison, and calculating the key credit ratios the institution uses to evaluate creditworthiness) has been a fundamental credit analysis task for as long as commercial lending has existed. The spread is the bridge between the borrower's accountant-prepared or IRS-formatted financial data and the standardized analytical language the institution's credit policy speaks.

For decades, spreading was done entirely by hand. An analyst worked through each document line by line, entering figures into a template, and the template's formulas calculated the ratios. The quality of the spread depended entirely on the analyst's attention to detail, understanding of the tax code and accounting conventions, and familiarity with the institution's recast methodology. Common errors included using the wrong year's figures, applying the wrong depreciation add-back, conflating business and personal income for pass-through entities, and misclassifying one-time items as recurring.

AI-assisted spreading changes the economics of the extraction step significantly. A well-designed AI spreading tool (one built specifically for financial statement extraction, not a general-purpose language model) can read a tax return, identify the relevant line items, extract the figures, and populate a spreading template in minutes rather than hours. The accuracy of extraction-focused tools has improved substantially; modern systems using optical character recognition (the process of converting scanned document images into machine-readable text) combined with document-specific training data can achieve extraction accuracy rates in the high nineties on well-formatted tax returns, according to vendor-reported benchmarks. The accuracy drops on handwritten amendments, unusual entity structures, and documents with complex recast requirements.

The important distinction is between extraction-focused AI tools (purpose-built for financial statement parsing, with specific training on tax returns and financial statements) and general-purpose generative AI tools (large language models used to draft text). Extraction tools extract from actual source documents. Generative tools generate plausible text based on learned patterns. Both are useful in the spreading process, but they perform different functions and carry different risk profiles. Treating a generative tool as an extraction tool is the failure mode that produces the wrong-DSCR scenario in the opening paragraph.

AI speeds the extraction; the analyst owns the conclusion. The ratios are the model's job to calculate from the data the analyst has verified. The judgment about what those ratios mean for the credit decision belongs to the human.

The Spreading Verification Workflow

Whether the spread is produced by an extraction tool, a generative drafting tool, or a combination, the verification workflow is non-negotiable. The workflow has a specific structure that makes verification efficient rather than exhausting.

The anchor-and-trace method. Before accepting any figure in the spread, identify the single authoritative source for that figure: the specific document, the specific year, and the specific line or schedule. For a federal business tax return, the authoritative source for each income category is the specific form and line number: net income from Form 1120 (the standard corporate income tax return), line 28; depreciation from Form 4562; interest expense from Schedule K for pass-through entities. For a personal tax return analyzed in a global cash flow, the authoritative sources include Schedule E for rental income, Schedule C for self-employment income, and the relevant K-1 forms for partnership or S-corporation income. The anchor-and-trace method requires the analyst to locate each anchor before accepting the spread's figure for that line.

The recast verification. Spreading is not simple transcription. It requires recast: the process of adjusting reported figures to reflect the borrower's true recurring economic cash flow by adding back non-cash expenses (depreciation and amortization are always added back), removing non-recurring items (a one-time lawsuit settlement, a casualty loss, or a gain on sale of equipment are excluded from the recurring cash flow analysis), and adjusting for owner compensation at market rates for the analysis of closely-held businesses. AI tools may apply recast rules mechanically and correctly for standard items. They may apply them incorrectly for unusual items or for entity structures with complex income flows. The verification workflow must include a specific check of each recast adjustment: what was added back, why, and from which line in the source document.

The ratio calculation verification. After the spread's figures are verified, the ratios must be checked. The debt service coverage ratio (DSCR, the ratio of the property's net operating income or the borrower's adjusted cash flow to the annual debt service on the proposed loan and all existing senior obligations) is the most critical ratio in commercial credit analysis and the one most commonly stated incorrectly in AI-generated commentary. The verification requires confirming: the correct numerator (the institution's credit policy specifies whether DSCR is calculated on net operating income, global cash flow, or another measure), the correct denominator (the annual debt service on the proposed loan at the proposed rate and term, plus all existing obligations that remain after the transaction), and the correct treatment of existing obligations that will be paid off at or before closing. A DSCR error of the magnitude in the opening story, 1.28 versus 1.07, is the difference between an approvable credit and one that is 0.13x below the policy floor. That gap matters.

Trend analysis verification. Spreading three years of financials is not just about the most recent year; it is about the trend. AI-generated commentary often presents trend observations that are directionally correct but numerically imprecise, or that interpret trends through a default optimistic framing rather than the balanced analytical framing a credit officer needs. The verification for trend analysis is qualitative as well as quantitative: confirm the direction of revenue, margin, and cash flow trends, confirm the numerical magnitude of the changes, and form an independent judgment about whether the trend supports the creditworthiness conclusion the commentary draws.

The total verification time for a three-year commercial spread depends on entity complexity, but a typical small-business spread (Schedule C or K-1 based, single entity, standard income categories) requires 30 to 45 minutes of verification after an AI extraction. A more complex commercial spread (multiple entities, international operations, significant one-time items, or a reorganization in one of the analysis years) may require 90 to 120 minutes. These times are still materially shorter than doing the spread entirely by hand.

Ratio Commentary: Prompts and What to Verify

Ratio commentary is the analytical narrative that accompanies the spread: what the numbers mean, what the trends show, where the risks are, and why the figures support or challenge the credit decision. It is the most judgment-intensive part of the spreading deliverable, and it is also the part where AI assistance is most useful and most risky.

Useful because: ratio commentary is a structured narrative task. It follows patterns: context the ratio against industry norms, describe the trend over the analysis period, identify the drivers of significant changes, note the risks, and state the credit implication. An AI tool that has seen thousands of commercial credit analyses can produce a well-structured commentary draft quickly, and the draft will often capture the right categories of analysis even if the specific claims require verification and adjustment.

Risky because: AI-generated ratio commentary tends toward optimism and coherence. The model has learned what credit memos that result in approvals look like, and it produces text that sounds like an approval-track analysis. Risk factors are noted, but they are noted in language that is slightly more mitigated than the analyst might choose. Trend reversals are contextualized charitably. Margin compression is described as "stabilizing" when it might be more accurately described as "continuing but at a decelerating pace." The analyst who reads the commentary looking for errors will miss these systematic framings if they are reading for factual accuracy only. The qualitative judgment review is as important as the quantitative figure check.

A well-constructed commentary prompt has the same five-component structure as the credit memo drafting prompt: role and policy context, specific financial data from the verified spread, entity and borrower description, explicit identification of analytical concerns the analyst has already identified, and the "cite the file" instruction. In the commentary context, the "cite the file" instruction is slightly different: it asks the model to cite the specific year and ratio when making trend claims ("Revenue declined 8.3% from Year 2 to Year 3, per the verified spread") rather than making trend claims as general observations. This approach makes the verification step straightforward: every trend claim has a number attached, and the analyst confirms the number against the verified spread.

The commentary verification checklist covers: every numerical trend claim (confirmed against the spread), every assertion about industry norms (confirmed against the institution's credit policy benchmarks or, where relevant, external industry data the institution uses), every risk identification (confirmed against the analyst's own read of the financial data and the file), and every credit implication statement (confirmed against the institution's credit policy: does a DSCR at this level, with this trend, in this property type, actually support the approval recommendation the commentary suggests, given the policy parameters?).

One specific verification target: the global cash flow statement. For small-business loans where the owner's personal guarantee is a primary credit support, the spreading deliverable often includes a global cash flow analysis that combines the business's cash flow (from the business returns) with the owner's personal cash flow (from the personal return). AI tools frequently make errors in this combination: double-counting income that flows from the business to the personal return, failing to apply the correct personal living expense deduction, or including personal investment income that the institution's credit policy excludes from qualifying income. The global cash flow is the most complex single calculation in small-business credit analysis and requires the most careful verification.

Commercial Real Estate Spreads and the DSCR Story

Commercial real estate (CRE) spreads have a specific structure that differs from operating business spreads, and the AI-assisted verification workflow for CRE is organized around the DSCR narrative rather than a global cash flow.

For an income-producing property, the spread begins with the property's operating income: gross potential rent (the total rent if every unit is occupied at the stated lease rate), less vacancy and credit loss (the allowance for economic vacancy and non-payment, typically 5% to 10% for stabilized properties in most markets, with higher allowances for transitional or single-tenant assets), less operating expenses (taxes, insurance, management fees, maintenance reserves, utilities, and any other recurring property-level costs), yielding net operating income (NOI, the core metric for income-property credit analysis). NOI divided by annual debt service is the DSCR.

AI tools make two characteristic errors in CRE spreads that are worth specifically checking. The first is vacancy: using stabilized market vacancy assumptions rather than the property's actual historical vacancy (from the rent roll and the borrower's operating history), which may be higher or lower than market. The second is the management fee: either omitting it (producing an overstated NOI and an overstated DSCR) or applying a generic percentage when the actual management arrangement has a different fee structure. Both errors systematically bias the DSCR upward, which means AI-generated DSCR figures for income properties should be reviewed with the prior expectation that they may be overstated, not just checked for accuracy.

The debt service denominator is also frequently wrong in AI-generated CRE commentary because the model may not know the institution's proposed loan terms with precision. The analyst must calculate the correct annual debt service using the actual proposed rate, amortization period, and loan amount, verify the AI's debt service figure against that calculation, and confirm that the denominator includes all existing senior debt on the collateral (a second mortgage, a subordinate SBA loan) that remains in place after the transaction. A DSCR calculated on the proposed loan's debt service alone, when the collateral also carries senior debt from another lender, overstates coverage substantially.

The "DSCR story" in the commentary is the narrative that contextualizes the DSCR figure: how it has changed over the analysis period (for refinances where operating history is available), how it compares to the institution's policy minimum, what the interest rate stress test shows at a higher rate environment, and what the DSCR would be at a stressed vacancy level. An AI tool can draft this narrative from the figures the analyst provides, and the draft is often structurally good. The analyst's role is to confirm that the story the AI tells is the story the verified figures actually support, and to correct the narrative where the AI's framing is more optimistic than the figures warrant.

Regulatory Anchors for Spreading Work

Spreading and ratio commentary exist in a regulatory context that the AI tool does not understand and that the analyst must. Several specific regulatory frameworks touch the spreading deliverable directly.

ECOA and Reg B. The Equal Credit Opportunity Act and Regulation B (12 CFR Part 1002) apply to every credit decision that uses spread data. If the spread produces a DSCR or a global cash flow figure that, combined with other factors, leads to a denial, that denial requires specific, accurate adverse-action reasons. If the spread contained errors that the analyst did not verify and correct, the adverse-action reasons drawn from the spread may be inaccurate. Inaccurate adverse-action reasons are an ECOA violation. The verification workflow is a fair-lending prerequisite, not just a credit quality control.

OCC Bulletin 2026-13. The April 2026 interagency model-risk guidance explicitly covers AI tools used in credit analysis, including AI-assisted spreading tools. Under 2026-13, the institution must maintain documentation of the AI tool's intended use, performance characteristics, known limitations, and the human oversight controls in place. For a spreading tool, this means the institution should document: the tool's accuracy rates on the document types it processes, the specific document types and entity structures where the tool's accuracy is lower, the verification workflow required before a spread produced by the tool can be submitted as an institution deliverable, and the evidence that the workflow was followed for individual credits. The 2026 update treats AI assistance in credit analysis as a model-risk matter, not just a productivity feature, and requires the governance infrastructure to match.

Unfair, Deceptive, or Abusive Acts or Practices (UDAAP, the broad consumer protection standard under the Dodd-Frank Act, administered by the CFPB and applicable to any practice that harms consumers). UDAAP applies to small-business credit in some contexts and to consumer credit broadly. If the spreading process produces a global cash flow figure for a consumer loan that is materially incorrect because the AI tool misread the borrower's income documents, and the institution relies on that figure without verification, and the borrower is denied based on the incorrect figure, the institution may have a UDAAP exposure alongside the ECOA and Reg B issue. The harm to the borrower (a creditworthy individual denied credit on the basis of an error) is the UDAAP concern, and the institution's failure to maintain adequate verification controls is the practice that caused the harm.

The Community Reinvestment Act (CRA, the statute requiring banks to meet the credit needs of their communities, including low-and-moderate income neighborhoods and borrowers). CRA examiners review the institution's credit underwriting practices for small businesses and community development purposes. If AI-assisted spreading tools are applied inconsistently (used for some borrower profiles and not others, or used with different verification rigor for different segments), the inconsistency may appear as a pattern in CRA examination data that invites further scrutiny. Consistent application of the verification workflow across all borrower segments is both a credit quality and a CRA hygiene matter.

Key Takeaways

• AI-assisted spreading has two distinct components that carry different risk profiles: extraction tools (which read source documents and populate templates, with high accuracy on well-formatted returns) and generative tools (which draft commentary from figures the analyst provides, without direct access to source documents). Treat each appropriately.
• The DSCR is the highest-risk figure in commercial real estate spreads, and AI tools systematically bias it upward through vacancy and management fee errors, and through incorrect debt service denominators. Verify the DSCR calculation independently every time: correct NOI numerator, correct all-in debt service denominator, correct policy-defined measurement approach.
• The anchor-and-trace verification method requires the analyst to confirm each figure in the spread against a specific source document, tax form, and line number before accepting the spread as a deliverable. This is not a general review; it is a line-by-line confirmation.
• Global cash flow analysis for small-business loans, combining business and personal return data for a pass-through entity, is the most complex spreading task and the one where AI errors are most common. Verify the combination methodology explicitly: confirm there is no double-counting, the correct personal expense deduction is applied, and the institution's qualifying income definition is satisfied.
• AI ratio commentary tends toward optimism and coherence, reflecting the patterns of approval-track credit analyses in its training data. The qualitative verification of commentary (whether the framing of risk and trend is analytically accurate, not just whether the numbers are correct) is as important as the quantitative figure check.
• The analyst who submits the spread and commentary owns the conclusion. The AI's calculation is a starting point; the analyst's verification is the product. When the verified figures do not support the commentary's conclusion, the conclusion is corrected, regardless of how well the AI draft reads.
• OCC Bulletin 2026-13 requires model-risk documentation for AI-assisted spreading tools: intended use, performance characteristics, known limitations, and evidence of the verification workflow on individual credits. A clean, documented workflow is both the compliance requirement and the institution's defense if a spreading error is later challenged.
• Inaccurate spread data that contributes to a denial creates ECOA, Reg B, and potentially UDAAP exposure, because adverse-action reasons drawn from an erroneous spread are themselves erroneous. The verification workflow is the fair-lending control that prevents a spreading error from becoming a civil rights violation.