| Analysis |
|---|
Spread financial statements- Why here
- CRAA drafts the spread; the reviewer owns it. Spreads feed the rating, so every AI-produced figure is recomputed against source — not sampled, recomputed.
- Verification
- Tier 1 — Full recalculation. Every AI-produced figure independently recomputed against source.
- If ignored
- A fabricated figure that foots correctly sits in the workpaper, and the rating rests on revenue the borrower never reported.
| A | C | · | R | · | · |
Build the EBITDA bridge- Why here
- The bridge is where aggressive add-backs hide. CRAA can assemble it and flag candidates; whether an add-back is credible is a judgment the reviewer must make and document.
- Verification
- Tier 1 — Full recalculation. Every AI-produced figure independently recomputed against source.
- If ignored
- A one-time add-back the machine carried forward without challenge overstates run-rate EBITDA, and leverage was understated in the file.
| A | C | · | R | · | C |
Flag quality-of-earnings anomalies- Why here
- Flags are leads, not findings. CRAA is good at surfacing anomalies at volume; the reviewer investigates each and owns both the pursued and the dismissed.
- Verification
- Tier 2 — Traced sample. A defined sample of AI figures traced to source, with a stated expansion trigger.
- If ignored
- “The tool didn't flag it” appears in a workpaper as if it were evidence of absence.
| A | I | · | R | · | · |
Test covenant compliance- Why here
- Covenant tests look mechanical and are not: definitions, cure periods, and amendment history all bite. CRAA computes; the reviewer verifies the definition, the period, and the math.
- Verification
- Tier 1 — Full recalculation. Every AI-produced figure independently recomputed against source.
- If ignored
- A covenant tested against the wrong period or the unamended definition reports compliance where there is a breach.
| A | C | · | R | · | C |
Recalculate DSCR / LTV / borrowing base- Why here
- Ratios feed ratings and report tables directly. Full recalculation against source is the price of using the machine's draft.
- Verification
- Tier 1 — Full recalculation. Every AI-produced figure independently recomputed against source.
- If ignored
- A DSCR the committee quoted turns out to be arithmetic on an extraction error, and the report table cannot be tied to any source document.
| A | C | · | R | · | · |
Assess collateral valuation adequacy- Why here
- CRAA can summarize appraisals and pull comparables; adequacy is a judgment about staleness, method, and market that the reviewer makes and the manager owns.
- Verification
- Tier 2 — Traced sample. A defined sample of AI figures traced to source, with a stated expansion trigger.
- If ignored
- A three-year-old appraisal summarized fluently by the machine reads as current support, and the collateral margin in the file is fiction.
| R | A | · | C | · | C |
Evaluate guarantor support- Why here
- CRAA can extract personal financial statement data; weighing contingent liabilities, double-pledged assets, and willingness to perform is reviewer judgment.
- Verification
- Tier 2 — Traced sample. A defined sample of AI figures traced to source, with a stated expansion trigger.
- If ignored
- Guarantor liquidity counted twice across related credits survives into the rating rationale.
| R | A | · | C | · | C |
| Conclusion |
|---|
Recommend a risk rating- Why here
- The reviewer forms an independent view before consulting CRAA's rating-relevant factors — sequence matters, because a machine suggestion seen first anchors the human who sees it (module 5). The machine's input is consultative, never the starting point.
- Verification
- Tier 3 — Editorial review. AI output is prose with no load-bearing figures; reviewer adopts it as their own words.
- If ignored
- Ratings drift toward the machine's suggestion file after file, and the rating rationale reads as boilerplate justifying an anchor.
| R | A | C | C | · | C |
Assign the final risk ratingHuman only- Why here
- The final rating is the review's conclusion. A named human assigns it, and a named human answers for it. CRAA holds no role — not because it couldn't produce a number, but because a number nobody owns is worthless in an exam.
- Verification
- Tier 0 — Independent. AI output not used; human performs from source.
- If ignored
- “The model assigned it” is offered to an examiner as the rating rationale. That answer has never once been accepted.
| · | R | A | × | · | C |
Determine accrual status / classificationHuman only- Why here
- Accrual and classification decisions carry direct accounting and regulatory reporting consequences. They are made from source by accountable humans.
- Verification
- Tier 0 — Independent. AI output not used; human performs from source.
- If ignored
- A nonaccrual call traced to a drafting tool becomes a financial reporting finding, not just a credit review finding.
| · | R | A | × | · | C |
Conclude on adequacy of the ACL / allowanceHuman only- Why here
- The allowance conclusion is an institutional opinion with board-level consequences. It is formed by humans from verified inputs, full stop.
- Verification
- Tier 0 — Independent. AI output not used; human performs from source.
- If ignored
- The review's allowance opinion cannot be defended line-by-line, and the finding lands on the review function itself.
| · | R | A | × | C | I |
| Documentation |
|---|
Draft workpaper narrative- Why here
- CRAA's strongest documentation use. The trap is fluency: a polished draft reads as verified. The reviewer edits until the narrative says what the reviewer concluded — then it is the reviewer's narrative.
- Verification
- Tier 3 — Editorial review. AI output is prose with no load-bearing figures; reviewer adopts it as their own words.
- If ignored
- A narrative nobody actually read end-to-end asserts something the workpaper evidence contradicts.
| A | I | · | R | · | · |
Cite evidence to source documents- Why here
- Citations are load-bearing: assertion → document → page. Language models fabricate citations under exactly the conditions that make them useful, so every citation is opened and checked.
- Verification
- Tier 1 — Full recalculation. Every AI-produced figure independently recomputed against source.
- If ignored
- The examiner pulls a cited page and it does not exist. Every other citation in the file is now suspect, and so is the function.
| A | I | · | R | · | · |
Draft the issue / finding- Why here
- CRAA can draft issue language from the workpaper. Severity, root cause, and what management must actually do are judgments the reviewer writes in their own words.
- Verification
- Tier 3 — Editorial review. AI output is prose with no load-bearing figures; reviewer adopts it as their own words.
- If ignored
- Issue language that overstates or understates severity survives because it was well-phrased, and the finding is negotiated against the machine's framing.
| A | C | · | R | · | C |
Approve the issue and its severityHuman only- Why here
- Issue approval sets what the institution is formally told to fix. That is an accountability moment, and accountability is human by construction here.
- Verification
- Tier 0 — Independent. AI output not used; human performs from source.
- If ignored
- A severity level nobody owns gets argued down in the management response, and the review function cannot say why it was set.
| R | A | I | × | · | C |