Reviewing With Machines
AI in credit review, taught for practitioners: what the machine is, where it fits, how it fails, and what has to be true for AI-assisted work to survive an examiner. About four hours, self-paced, free and unwalled. You leave with artifacts you can take to your own governance committee.
AI can produce work. Only a human can own a conclusion. In credit review, the AI is never the accountable party — not for a risk rating, not for an issue, not for a sign-off.
- M1
What the machine is actually doing
Plain-language mechanics of an LLM: next-token prediction, context windows, non-determinism, and why it sounds confident when it is wrong. A fluent junior analyst with no memory and no accountability.
- M2
Where AI fits in the credit review lifecycle
The lifecycle end to end — scoping to examination support — with each step graded high fit, conditional fit, or off-limits. Off-limits is a real category with real entries.
- M3
The RACI
The centerpiece. Who is Responsible, Accountable, Consulted, and Informed for every task in the lifecycle — and the one assignment that is locked by construction.
- M4
Failure modes
A field guide with worked examples: fabricated figures that foot correctly, citations to pages that do not exist, OCR corruption, silent extrapolation, anchoring, drift, and irreproducibility.
- M5
Automation bias, and the reviewer’s job under it
Why a polished draft suppresses challenge, why the second reviewer defers to the first machine, and why “I checked it” degrades to “it looked right.” Introduces verification tiers.
- M6
Controls and evidence
What has to be true for an AI-assisted workpaper to survive an examiner: number-level citations, four-eyes on AI drafts, prompt and model-version logging, and QA of AI output as its own population.
- M7
Governance and regulatory context
AI-assisted review inside existing supervisory expectations — model risk management, loan review guidance, third-party risk, fair lending, and the emerging AI frameworks — mapped to your jurisdiction.
- M8
Standing up AI in your own shop
Pilot design, scope limits, a challenge log, the first 90 days, metrics that actually detect degradation, and when to switch it off.
The centerpiece is live now: the RACI — an interactive matrix over the credit review lifecycle that you can read, adapt to your institution, and export as a policy exhibit. One cell in it is locked. That cell is the whole course in one sentence.