The underwriter stared at her screen at 3:15 on a Thursday afternoon, fifteen minutes before the credit committee convened. The AI-assisted underwriting tool had returned a recommendation: decline, with a composite risk score of 67 out of 100. The borrower was a 12-year owner of a regional landscaping company requesting a $480,000 SBA 7(a) term loan to purchase a competitor. Two years of federal returns showed healthy revenue growth, a 1.28 debt service coverage ratio (DSCR, the ratio of the business's net operating income to its total annual debt obligations, the primary underwriting metric for business loans), and a clean payment history on two outstanding obligations. The underwriter agreed with the score directionally. What she did not have was the model's reasoning. When the committee asked "why 67?" she had a number. She did not have an argument she could defend, challenge, or correct. That afternoon, the institution approved the loan on the underwriter's manual analysis. The AI tool's recommendation was noted in the file and then set aside. Three months later, the institution's model-risk officer flagged the workflow: the AI recommendation was being regularly overridden with no documented basis for the override, no feedback loop to the model, and no record that anyone had ever asked the tool to show its work. That is the chain-of-thought problem in lending, and this lesson is about solving it.

What Chain-of-Thought Means in a Credit Context

Chain-of-thought (CoT) prompting is a technique for instructing an AI model to produce its intermediate reasoning steps before delivering a final answer, rather than jumping from input to conclusion in a single opaque step. In general AI use, the technique improves accuracy on complex tasks by forcing the model to work through sub-problems explicitly. In a credit context, the technique does something more important than improve accuracy: it makes the model's reasoning visible, inspectable, and correctable by a human officer who holds accountability for the decision.

The distinction between a score and a reasoned analysis is the core of the ECOA (Equal Credit Opportunity Act, the federal statute prohibiting discrimination in any aspect of a credit transaction on the basis of race, color, religion, national origin, sex, marital status, age, or receipt of public assistance) problem. A score is an output. A reason is an argument. Under ECOA and its implementing regulation, Regulation B (Reg B, 12 CFR Part 1002, the Consumer Financial Protection Bureau rule that operationalizes ECOA's adverse-action, notification, and anti-discrimination requirements), any denial must be accompanied by specific, accurate reasons that the borrower can read and respond to. "The model scored you 67" is not a reason. "Your projected debt service coverage ratio of 0.97 after the proposed acquisition falls below the institution's minimum threshold of 1.20" is a reason. Chain-of-thought prompting is the mechanism that forces the model to produce the second type of output, not the first.

OCC Bulletin 2026-13, the April 2026 interagency model-risk guidance that superseded OCC 2011-12 and explicitly placed generative AI (GenAI) and AI-assisted decision tools under model-risk, fair-lending, third-party, and board-governance expectations, makes the inspectability requirement concrete. Under the bulletin, institutions must be able to demonstrate that AI-assisted decisions are explainable to the institution's own officers, to borrowers (through adverse-action reasons), and to examiners. A workflow where the AI produces a score and the human overrides it on intuition satisfies none of those three audiences. A workflow where the AI produces a reasoned analysis, step by step, that the human can read, test, and challenge satisfies all three.

A chain-of-thought prompt does not just improve the model's answer. It transforms the model's output from a score into an argument that a credit officer can challenge, correct, and sign off on.

Building a Chain-of-Thought Credit Prompt

A chain-of-thought credit prompt has three structural elements that distinguish it from a standard drafting prompt: an explicit step decomposition instruction, a policy-grounding requirement at each step, and a conclusion-last structure that prevents the model from anchoring the reasoning to a pre-determined outcome.

The step decomposition instruction. The prompt must explicitly tell the model to work through the analysis in defined steps, named in advance, before arriving at any conclusion. Without the step instruction, the model will produce a conclusion first and then construct supporting reasoning backward, which is called post-hoc rationalization and produces reasoning that looks complete but is not actually the basis for the conclusion.

A working step decomposition instruction for a commercial loan analysis:

Analyze this loan request in the following steps, completing each fully 
before moving to the next:

Step 1: Income and Cash Flow Analysis: Calculate the borrower's adjusted 
net operating income using only the figures provided in the tax returns 
I have entered below. Show the calculation. Cite every figure to its source.

Step 2: Debt Obligations Inventory: List every existing debt obligation 
from the credit report and the financial statements. Calculate total annual 
debt service. Show the calculation. Cite every figure.

Step 3: DSCR Calculation: Divide the Step 1 NOI by the Step 2 total annual 
debt service (including the proposed new obligation at the terms I have stated). 
State the resulting DSCR. Identify whether it meets the institution's minimum 
threshold of 1.20.

Step 4: Collateral Assessment: Evaluate the collateral using only the 
values I have provided. Calculate LTV. Identify whether it meets the 
institution's maximum threshold of 75%.

Step 5: Risk Factor Summary: List every characteristic of this application 
that falls outside standard policy parameters. For each, state what the 
policy requires, what the application shows, and the magnitude of the gap.

Step 6: Recommended Next Steps: Identify the actions the reviewing 
underwriter should take before reaching a credit decision. Do not make 
a credit decision. Do not recommend approval or denial.

Notice the final instruction: "Do not make a credit decision." This is not a politeness; it is a governance control. Under OCC 2026-13 and under the longstanding principle that accountability for credit decisions belongs to the human officer who signs the file, the model's role is analysis and the underwriter's role is decision. The step structure forces the analysis to proceed in a logical order. The final instruction enforces the human-in-the-loop requirement.

The policy-grounding requirement. Each step in the chain must be anchored to a specific, stated policy standard rather than to the model's general knowledge of lending. "Does the DSCR meet the institution's minimum threshold of 1.20?" is a policy-grounded evaluation. "Is the DSCR adequate?" is not. The second formulation allows the model to apply whatever DSCR standard it learned from its training corpus, which may include standards from different institutions, different loan types, different vintages, or different regulatory environments. The first formulation forces the model to evaluate against the institution's actual standard, which the analyst has stated explicitly in the prompt.

Policy grounding at each step does two things beyond accuracy. First, it makes the reasoning auditable: an examiner reviewing the model's output can see whether the correct policy threshold was applied. Second, it creates a paper trail for OCC 2026-13 documentation purposes: the model's output shows, step by step, that it applied the institution's stated policy to the borrower's actual data. That is a model-risk record, not just a credit memo.

The conclusion-last structure. The most important structural choice in a chain-of-thought credit prompt is placing the summary or recommendation step last, not first. A common failure pattern occurs when the analyst adds a preamble like "I am sending you this file for credit analysis. The credit officer has already tentatively approved this." That preamble anchors the model's output to the approval framing; every step will be written to support it. In a lending context, this is not just a quality problem. If the model's reasoning was anchored to a conclusion the analyst suggested, and the borrower is subsequently denied on different grounds, the documented chain of reasoning is not actually the basis for the denial. That is an ECOA documentation problem.

The conclusion-last rule is simple: do not tell the model the outcome before it reasons. Do not describe the borrower favorably or unfavorably in the prompt header. Do not state what you expect the analysis to show. Provide the data, state the policy standards, define the steps, and let the reasoning produce the conclusion.

A Worked Example: SBA Acquisition Loan

Return to the landscaping company scenario from the opening. The underwriter who received a score of 67 and no reasoning had submitted a standard prompt: "Please analyze this SBA 7(a) loan request for the acquisition of Green Valley Landscaping by Ridgeline Services. The borrower has provided two years of federal returns and a financial statement. Recommend approval or denial." The model produced a score and a two-sentence summary. The underwriter had nothing to work with.

Here is the same scenario resubmitted with a chain-of-thought structure. The analyst's prompt, with actual figures entered from the source documents:

You are a credit analysis assistant for First Valley Bank. All outputs are 
subject to ECOA, Regulation B (12 CFR Part 1002), and First Valley Bank's 
SBA Preferred Lender credit policy, effective January 2026. Do not reference 
standards from any other lender or jurisdiction.

Borrower: Ridgeline Services, LLC (acquiror)
Target acquisition: Green Valley Landscaping, LLC
Loan amount requested: $480,000
Proposed terms: SBA 7(a), 10-year term, 7.5% fixed rate
Annual debt service on proposed loan: $57,480

ACQUIROR FINANCIAL DATA (from federal returns, entered by analyst):
  2023 Form 1120S, Line 1a (gross receipts): $1,892,000
  2023 Form 1120S, Line 21 (ordinary income): $194,300
  2023 K-1 Line 1 officer compensation included above: $120,000
  2024 Form 1120S, Line 1a (gross receipts): $2,108,000
  2024 Form 1120S, Line 21 (ordinary income): $218,700
  2024 K-1 Line 1 officer compensation included above: $130,000

EXISTING DEBT SERVICE (from credit report and financial statements):
  Equipment loan A: $8,400/year (23 months remaining)
  Equipment loan B: $14,400/year
  Business line of credit: $0 current balance, $200,000 limit

TARGET ACQUISITION:
  Target 2024 gross revenue per seller-provided statement: $612,000
  Target 2024 EBITDA (earnings before interest, taxes, depreciation, 
    and amortization) per seller-provided statement: $87,400
  Purchase price: $480,000 (100% loan-financed)

FIRST VALLEY BANK SBA POLICY THRESHOLDS:
  Minimum combined DSCR (acquiror + target income, net of acquisition debt): 1.20
  Maximum LTV on business acquisition: no hard collateral; policy requires 
    SBA guaranty plus personal guaranty of majority owner(s)
  Minimum credit score for SBA preferred lender processing: 660
  Borrower's personal credit score (from credit report): 708

Analyze this request in the following steps:

Step 1: Income Normalization: Calculate the two-year average normalized 
owner cash flow for the acquiror using the figures above. Add back officer 
compensation to the ordinary income to arrive at owner benefit; then calculate 
the 2-year average. Show every step.

Step 2: Acquisition Income Credit: The institution's SBA policy permits 
crediting up to 75% of the target's EBITDA when supported by a CPA-prepared 
statement. The seller-provided statement above is not CPA-prepared. Apply the 
appropriate income credit per policy (no CPA review = 0% credit until CPA 
statement obtained). Note the exception.

Step 3: DSCR Calculation: Using Step 1 income only (Step 2 credit not yet 
available), calculate DSCR against all debt service: existing obligations 
plus the proposed loan. State the resulting DSCR and compare it to the 
1.20 policy minimum.

Step 4: Policy Compliance Summary: List every characteristic of this 
application that falls outside or conditionally within standard parameters. 
State the policy requirement, the application's current position, and 
what would be needed to bring the exception into compliance.

Step 5: Recommended Next Steps: State what the reviewing underwriter 
and loan officer should obtain before this file proceeds to committee. 
Do not make a credit decision.

Here is the type of output this chain-of-thought prompt produces, showing the model's step-by-step reasoning:

STEP 1: INCOME NORMALIZATION

2023 Analysis:
  Ordinary income (Form 1120S, Line 21): $194,300
  Add back: officer compensation (K-1, Line 1): $120,000
  Total owner benefit 2023: $314,300

2024 Analysis:
  Ordinary income (Form 1120S, Line 21): $218,700
  Add back: officer compensation (K-1, Line 1): $130,000
  Total owner benefit 2024: $348,700

Two-year average owner benefit: ($314,300 + $348,700) / 2 = $331,500

STEP 2: ACQUISITION INCOME CREDIT

Per First Valley Bank SBA policy: acquisition EBITDA credit requires 
CPA-prepared financial statement. Documents provided are seller-prepared.
Policy credit: $0 (pending CPA statement).
[EXCEPTION NOTED: Target income cannot be credited until CPA-prepared 
statement is obtained. This is a required document gap, not a 
discretionary item.]

STEP 3: DSCR CALCULATION

Qualifying income (Step 1, acquiror only): $331,500
Total annual debt service:
  Equipment loan A: $8,400
  Equipment loan B: $14,400
  Proposed SBA loan: $57,480
  Total: $80,280

DSCR = $331,500 / $80,280 = 4.13 (acquiror income only, without target credit)

[Note: If CPA-prepared target EBITDA is obtained and 75% credit is applied:
  Target income credit: $87,400 x 0.75 = $65,550
  Combined qualifying income: $331,500 + $65,550 = $397,050
  Combined DSCR: $397,050 / $80,280 = 4.95]

Both scenarios well exceed the 1.20 minimum. DSCR is not the binding 
constraint in this file.

STEP 4: POLICY COMPLIANCE SUMMARY

Item 1: Target acquisition income - EXCEPTION
  Policy requires: CPA-prepared seller financial statement for income credit
  Application status: Seller-prepared only
  Required to cure: CPA-prepared statement for 2024 fiscal year
  Impact if not cured: No income credit for target; acquiror-only DSCR 
    applies (4.13, above threshold, but income qualification is incomplete)

Item 2: Credit score - MEETS POLICY
  Policy minimum: 660
  Borrower score: 708 (Source: personal credit report, dated [date on report])

Item 3: Collateral - MEETS POLICY
  Policy: SBA guaranty plus personal guaranty required; no real property 
    collateral required for this loan type
  Application status: SBA guaranty will attach; personal guaranty of sole 
    majority owner (100% ownership confirmed on K-1) required at closing

STEP 5: RECOMMENDED NEXT STEPS

1. Obtain CPA-prepared financial statements for Green Valley Landscaping, 
   LLC, fiscal year 2024. The file should not be submitted to committee 
   without this document if the institution intends to credit target income.
2. Confirm personal guaranty commitment with borrower and obtain signed 
   guaranty agreement before closing.
3. Review the 23-month remaining term on Equipment Loan A: this obligation 
   reduces in less than 2 years, improving DSCR further. The underwriter 
   should note this in the credit memo as a favorable factor.
4. All figures above are as entered by the analyst. The reviewing underwriter 
   should independently verify the K-1 officer compensation figures against 
   the original returns before submitting to committee.

No credit decision is made or implied by this analysis.

This output gives the credit committee something the score of 67 never did: a specific reasoning trail they can test. The committee can ask "does the policy actually require a CPA statement for this loan size?" and look up the answer. They can ask "what if the CPA statement shows lower EBITDA?" and recalculate. They can note that the DSCR at 4.13 is well above threshold even without the acquisition income, which changes the risk profile materially. The underwriter who presents this at 3:15 Thursday afternoon has an argument to make, not a number to defend.

Inspecting and Challenging the Chain

Producing the chain-of-thought output is step one. Using it correctly is step two. A chain-of-thought output from an AI model is not a verified analysis; it is a structured draft that a trained underwriter must read critically, step by step, before it becomes the basis for any credit discussion.

The critical reading process has three checkpoints.

Checkpoint one: verify the arithmetic. Chain-of-thought prompting does not make AI models more accurate at arithmetic. It makes errors more visible. In the example above, the model correctly adds $314,300 and $348,700 and divides by two. But a model operating on slightly different figures could produce a plausible-looking calculation that is wrong by a margin that matters. The reviewer's job at Step 3 is not to confirm the format of the DSCR calculation. It is to verify that the income figure in the numerator and the debt service figure in the denominator are both correct before accepting the ratio. The chain makes it easy to trace exactly which figures were used. Use that traceability.

Checkpoint two: test the policy citations. The model applies whatever policy the analyst stated in the prompt. If the analyst stated the wrong threshold, the model will apply it correctly. If the credit policy has been updated since the system prompt was written, the model will apply the outdated version. Policy citations in the chain output must be checked against the institution's current credit policy document, not just accepted as accurate because they appeared in the model's output. In the example, the reviewer should open the SBA Preferred Lender policy and confirm that the CPA-prepared statement requirement applies to this loan size and type before treating Step 2 as a binding exception.

Checkpoint three: identify missing steps. A chain-of-thought output that does not address a risk factor the underwriter knows about is not a safe output. It is a gap. In the acquisition example, an experienced underwriter might note that the chain never addressed the borrower's management capacity to run two landscaping operations simultaneously: the target company may have had a key-person dependency on its former owner, and that risk does not appear in the financial statements. The chain correctly analyzed the financial data and correctly identified the missing document. It did not know to ask about operational risk that requires underwriter judgment. The underwriter's job is to add what the chain omitted, not just to verify what the chain included.

These three checkpoints are not a criticism of chain-of-thought prompting. They are the reason chain-of-thought prompting is more useful than a black-box score: the chain gives the underwriter something to verify, challenge, and augment. A score gives them nothing but a number. Under OCC 2026-13, being able to demonstrate that the reasoning was reviewed, tested, and supplemented by human judgment is the institution's defense in a model-risk examination.

Chain-of-Thought for Consumer Credit Files

The same chain-of-thought structure applies to consumer credit analysis, with adjustments for the different metrics and regulatory constraints of consumer lending. For mortgage underwriting, the critical chain steps are income qualification (verifying the income calculation method against the applicable guidelines, such as Fannie Mae Selling Guide income calculation rules), debt-to-income (DTI) calculation (listing every debt obligation from the credit report and calculating the proposed DTI), loan-to-value (LTV) calculation (dividing the loan amount by the lesser of the appraised value or purchase price), and reserve requirement verification (confirming liquid asset sufficiency against the applicable reserve standard).

Consumer chain-of-thought prompts require one additional element that commercial prompts can sometimes omit: an explicit ECOA and Reg B instruction at the conclusion step. For consumer files, adverse-action reasons must be drawn from the chain's specific findings. The Reg B instruction tells the model that any characteristic it flags as a policy exception must be stated in terms specific enough to appear in an adverse-action notice under Regulation B Section 1002.9. A working instruction for the consumer context:

In your Policy Compliance Summary step, state each exception in language 
specific enough to serve as an adverse-action reason under Regulation B 
Section 1002.9. Each reason must identify the specific factor, its value 
in this application, and the standard it fails to meet. 
Example of compliant language: "Debt-to-income ratio of 47.1% exceeds
the maximum permitted 43.0% under [Bank Name] conventional mortgage
guidelines (Source: total monthly obligations $3,192 / gross monthly
income $6,784 = 47.1%)."
Example of non-compliant language: "DTI too high."

This instruction accomplishes several things at once. It forces the model's compliance findings to be specific enough for adverse-action use. It creates a natural workflow: the chain's Step 4 output becomes the draft adverse-action reasons, which the underwriter then verifies against the file before the notice is prepared. And it creates a documented connection between the model's reasoning and the adverse-action notice, which is exactly the type of audit trail OCC 2026-13 requires for AI-assisted credit processes.

For consumer files involving protected-class applicants or any file where a denial is possible, the chain-of-thought approach is not just a quality technique. It is the mechanism that makes the adverse-action process auditable. A regulator reviewing an ECOA complaint can read the chain output and verify that the stated denial reasons actually appear in the model's analysis of the borrower's file, that the analysis used the correct figures, and that the figures traced to the source documents the analyst supplied. Without the chain, the regulator has a denial notice and a score. With the chain, the regulator has a reasoning trail.

Common Failures and How to Prevent Them

Chain-of-thought prompting in lending workflows fails in predictable ways. Understanding the failure modes allows analysts and underwriters to design prompts and review processes that catch them before they affect a credit file.

Failure mode one: the model skips a step. When the chain includes six defined steps and the model produces five, a quick reading may not catch the missing one. The model will not announce that it skipped Step 4. It will either skip it silently, collapse two steps into one, or absorb the skipped analysis into an adjacent step without labeling it. Prevention: the review checklist for every chain output should verify, explicitly, that every defined step appears in the output before the reviewer reads the substance of any step.

Failure mode two: the conclusion bleeds back into the chain. The model, trained to produce coherent and helpful outputs, will sometimes write Steps 1 through 4 in a way that sets up a specific conclusion it has already decided on, rather than letting the reasoning produce the conclusion. This is the post-hoc rationalization failure mode, and it produces reasoning that looks internally consistent but is actually constructed backward from the endpoint. Prevention: after reading the chain, test whether the data in Steps 1 through 4 actually support the conclusion in Step 5. A simple test: if you changed one figure (for example, the DSCR drops from 1.31 to 1.09), does the chain's conclusion change? If the model's conclusion would not change regardless of the input figures, the chain is not producing real reasoning.

Failure mode three: the model confabulates policy. When the analyst does not explicitly state a policy threshold, the model applies a threshold from its training data. In the SBA example above, if the analyst had not stated the 1.20 DSCR minimum, the model might have applied 1.15, 1.25, or some other threshold it learned from generic SBA lending guidelines in its training corpus. Any threshold the model states that the analyst did not explicitly provide in the prompt should be flagged and verified against the institution's current credit policy before the output is used. Prevention: state every relevant policy threshold explicitly in the prompt. Never rely on the model to supply policy standards from its training data.

Failure mode four: the chain omits the exception. The most consequential failure is a chain that works through the defined steps but misses a policy exception that matters for the credit decision. This typically occurs when the exception involves a relationship between two data points that the analyst did not highlight in the prompt. In the SBA example, a chain prompted with "analyze income and coverage" might miss the CPA-statement requirement if the analyst did not include that policy provision explicitly. Prevention: before submitting the prompt, the analyst should review the institution's credit policy for the loan type and list every relevant threshold and exception standard in the policy-grounding section of the prompt. The chain's job is to apply that list; the analyst's job is to make the list complete.

Chain-of-Thought as a Model-Risk Artifact

OCC Bulletin 2026-13 changed the governance requirements for AI-assisted credit workflows in a specific way that makes chain-of-thought prompting not just a quality technique but a compliance requirement. The bulletin requires institutions to maintain records that demonstrate the AI tool was used in a controlled, documented manner, that human judgment was applied at the decision points, and that the AI's outputs were reviewed before they influenced a credit decision. A chain-of-thought output satisfies all three requirements in a way that a black-box score does not.

Consider what is in the file when chain-of-thought is used: the prompt (which shows the policy standards the analyst applied), the chain output (which shows the reasoning the model produced), and the underwriter's annotations and credit decision (which show the human judgment applied to the model's output). An examiner reviewing this file has a complete record: what data the analyst provided, what policy the analyst applied, what reasoning the model produced, and what the human decided. That record is the institution's defense in a model-risk examination under 2026-13.

Compare this to the file that existed before chain-of-thought was introduced: a score of 67, an override decision, and no documented reasoning trail. The examiner reviewing that file has a number and a human decision that does not explain itself. That file fails the OCC 2026-13 documentation standard. It also creates a fair-lending exposure: if the same model regularly produces scores that are overridden by different underwriters in different ways for different borrower profiles, and if the override patterns correlate with protected-class status, the bank cannot demonstrate that the process was consistent and non-discriminatory. The chain-of-thought record, applied consistently, creates the consistency the fair-lending defense requires.

For institutions that use AI tools built by third-party vendors (the loan origination platforms, the credit scoring services, the automated underwriting systems that most banks now use), OCC 2026-13's third-party risk requirements add another dimension. The bulletin holds the institution responsible for understanding and being able to explain the AI tool's behavior, even when the tool is a vendor product. A vendor that provides a score but no reasoning is providing a tool the institution cannot fully govern under 2026-13. Institutions should require, in their vendor contracts and their ongoing model-risk management, that vendor AI tools support chain-of-thought or step-by-step explanation outputs, and that those outputs be logged as part of the credit file. This is not a technology request. It is a model-risk governance requirement.

Key Takeaways

• Chain-of-thought prompting instructs an AI model to work through defined reasoning steps before reaching a conclusion, making the reasoning visible, inspectable, and challengeable by the human officer who holds accountability for the credit decision.
• The technique transforms the model's output from a score into an argument. Under ECOA, Regulation B, and OCC Bulletin 2026-13, a credit decision that can only be expressed as a score is not sufficient: specific, accurate reasons are required, and chain-of-thought prompting is the mechanism that forces those reasons to be produced and documented.
• A well-constructed chain-of-thought credit prompt includes a step decomposition instruction (naming the steps in advance), a policy-grounding requirement at each step (applying the institution's specific thresholds, not generic standards), and a conclusion-last structure (preventing the model from reasoning backward from a pre-determined outcome).
• The three checkpoints for reviewing a chain output are: verify the arithmetic against source documents, test the policy citations against the current credit policy, and identify risk factors the chain did not address that the underwriter's judgment must supply.
• Common failure modes include the model skipping a step silently, reasoning backward from a pre-determined conclusion, confabulating a policy threshold the analyst did not explicitly state, and omitting an exception that requires a data-point relationship the analyst did not highlight in the prompt.
• A chain-of-thought output preserved in the credit file (prompt, reasoning chain, and underwriter annotations) satisfies OCC 2026-13's documentation requirements for AI-assisted credit workflows, and creates the consistency record needed to defend the process in a fair-lending examination.
• For third-party vendor AI tools, institutions should require chain-of-thought or step-by-step explanation outputs as a condition of the vendor relationship: a vendor that provides only a score is providing a tool the institution cannot fully govern under the 2026 interagency model-risk standards.