"Where is AI used in your credit process?"
This is an inventory question. The answer requires knowing every point where a model touches the file — document classification, data extraction, spreading, drafting, research — and being able to show that each runs through a defined, controlled path rather than an ad-hoc tool an analyst pasted into.
"How do you know the output is grounded?"
The strong answer is structural: generated claims carry citations to sources — documents, accepted data, regulations — and the system declines to generate when no source exists. The weak answer is "analysts check it," with nothing that distinguishes checked from unchecked output.
"Who approved this, and what exactly did they approve?"
Examiners are alert to rubber-stamping. The artifact that answers this is a record separating the model's proposal from the human's action — accepted as-is, edited, rejected — per output, per file. If the system cannot make that distinction, neither can the lender.
"The model has been upgraded since this decision. Explain the decision."
This is where ungoverned adoption fails silently. The answer requires generation snapshots: the model version, inputs, source context, and output frozen at decision time, alongside the versioned policy and rules then in force. With those, last year's decision is explainable as last year's decision. Without them, the lender is re-arguing history with a different model.
"Show me."
The closing question is always a live one: pick a memo claim and walk it back to its source; pick a decision and replay it. Preparation for that moment is architectural — either the trail exists in the system of record, or a team spends the week before the exam rebuilding it by hand.
The pattern across all five: examiners are not asking whether the lender uses AI carefully. They are asking whether the system enforces the care — because assurances retire with the people who gave them, and artifacts do not.
Common questions
Do examiners prohibit AI in underwriting?
The scrutiny is on control and accountability, not the technology itself. The consistent expectations: human ownership of decisions, explainability, and records that survive model changes.
What is a generation snapshot?
A frozen record of a model run — model version, inputs, source context, prompt package, output, and the human action taken on it — preserved with the credit file.
How should a lender prepare for the first AI-focused exam question?
Run the drill internally: pick a memo claim, trace it to source; pick an AI-assisted decision, replay it. If either takes more than minutes, the gap is architectural, not procedural.