Bank regulators ramp up AI scrutiny

US bank regulators are folding artificial intelligence into every routine examination, Reuters reported on June 12, citing people familiar with the matter. In those exams, the Office of the Comptroller of the Currency and the Federal Reserve have started pressing banks to spell out where AI already runs in sensitive work: loan decisions, know-your-customer identity checks and sanctions screening.

The questions get specific. Examiners want to see how a bank leans on outside vendors, how it protects customer data, and whether it can hit a “kill switch” to stop a model fast. The governance probe runs wider still, covering guardrails, the human kept in the loop, exposure through subcontractors, and what the plan is when a system fails. One source said AI now surfaces in every bank exam. Nobody is writing rules yet; the agencies say they are still mapping how the technology is really used.

Why this matters

This is fact-finding, not a rulebook, but it sets the direction. The areas examiners flag are the same ones most firms have not nailed down. American Banker found that 72% of bankers rated their own firm least ready on exactly those points: the power to halt a model on demand (34%) or to flag an AI failure to regulators (38%).

You do not need a bank charter to borrow the checklist. Knowing where AI touches important work, keeping a human in the loop, and being able to drop a vendor fast are the same defenses behind our guide to vendor lock-in exit ramps and the AI customer-data privacy baseline. The cost of skipping them is what sank KPMG’s pulled AI report: output nobody checked.