CRAA gives a community bank or credit union an always-on, independent loan-review function — consistent risk ratings and examiner-ready documentation on every credit — for less than the cost of a single outsourced engagement. A qualified human always owns the final call.
AI can produce the work. Only a human can own the conclusion.
The 2020 Interagency Guidance on Credit Risk Review Systems requires a qualified, independent review function. A small institution cannot realistically staff a bench of seasoned reviewers.
And it arrives once or twice a year as a static report — a snapshot that is stale the week after it lands.
Grades vary by which officer originated the loan. When that person retires, the judgment built over a career walks out the door.
Inaccurate or untimely risk ratings are the most common way a community institution gets criticized in an exam — and they flow straight into the CECL allowance.
CRAA is not one black box. It is three layers, each doing what it is genuinely best at — which is what makes the output defensible to a model-risk team and an examiner.
Ratios — DSCR, leverage, coverage, LTV — are computed from the raw financials, not trusted from free text. Hard policy gates are enforced in code, not by persuasion: a DSCR below 1.0 cannot be Pass; an unwaived covenant breach cannot be Pass. Same inputs, same result, every time.
Within those guardrails, the model performs the qualitative work that is inconsistent and un-staffable today: the five-Cs synthesis, red-flag detection, accrual-status reasoning, and a draft narrative in examiner style — each assertion tied back to a specific data point and a cited regulatory passage.
The reviewer accepts, or overrides with a documented rationale. Nothing is a rating until a qualified person makes it one. Every decision — AI recommendation, human decision, override reason — is stamped and written to an immutable audit trail.
When the examiner asks “why is this still a Pass?”, the evidence, the cited classification, the five-Cs rationale and the override note are already there — consistent across the whole book.
Financials are spread for you and the first-draft narrative is written. You spend your time on judgment and exceptions — and a junior produces senior-quality drafts.
An immutable trail, a rating-migration view of the portfolio, and a validation record behind the tool itself — the artifacts that make an exam a formality.
The core classification — Pass, Special Mention, Substandard, Doubtful, Loss — is the shared interagency standard used by the OCC, FDIC and Federal Reserve, drawn from the OCC Comptroller’s Handbook. The regulatory layer is a per-institution profile, so grounding matches your charter.
UCS classifications, five-Cs framework, accrual-status determination and examiner-defense notes, grounded in embedded interagency guidance.
Aligned to NCUA supervisory expectations where they diverge from the interagency banking framework — validated before we claim coverage, never assumed.
Configurable profiles for funder- and state-specific review requirements. One rule throughout: the tool never fabricates a citation.
This is the first question a security team asks, so here is the direct answer — no marketing around it.
Reviews run on Anthropic’s commercial API under zero-data-retention terms. Borrower financials are not retained and never become training data.
Route inference through your own Anthropic account, governed and billed under your contract — your data never touches ours.
Customer-managed keys (BYOK/CMEK) for data at rest. You hold the key; you can revoke it.
TLS on every request, with a signed Data Processing Agreement available for every customer.
Each institution’s data is strictly partitioned — essential for review firms serving many client banks.
Formal SOC 2 attestation is on the roadmap, with the control evidence trail built from day one.
No install, no infrastructure lift. Sign in and review your first credit the same day.
Every logged review rolls up into a book-level view — the artifacts examiners and your board actually ask for.
| Pass | SM | Sub | Doubt | |
|---|---|---|---|---|
| Pass | 88 | 9 | 3 | 0 |
| SM | 12 | 61 | 24 | 3 |
| Sub | 2 | 15 | 70 | 13 |
| Doubt | 0 | 0 | 19 | 81 |
Sample data · plus concentration, staleness & override-rate views
Loan-review and consulting firms aren’t just competition — they’re the sharpest users. CRAA turns a fixed-capacity practice into a scalable one.
The tool does the first-pass grind, so reviewers spend billable hours on judgment — expanding margin without expanding headcount.
Offer always-on review between annual engagements — something no manual practice can deliver today.
Strict multi-tenant isolation per client institution, under your own identity and workflow.
Because a human owns every conclusion, CRAA is a documentation and consistency assistant — not an autonomous decision model. And where scrutiny applies, we bring evidence.
A curated set of historical credits across C&I, CRE and stressed segments, run on every prompt or model change to catch regressions before release.
Rating-distribution shifts, segment override spikes and hallucination trends are tracked, with defined investigation thresholds.
Every review stamps the exact prompt, model and regulatory-knowledge version that produced it, with a changelog and rollback path.
Bring a deal you already know the answer to. In one session you’ll see the rating, the evidence, the narrative — and exactly where your judgment still decides.
Advisory only. CRAA’s recommendations are for informational purposes and require reviewer judgment. The tool does not constitute a formal credit determination and is not a substitute for the independent judgment of a qualified credit professional. Every conclusion is owned by a person, not the software.