US oversight remains sectoral and use-case based. Banking agencies revised model-risk guidance in 2026, while fair-lending, adverse-action, privacy and consumer-protection laws continue to apply to algorithmic decisions.
The 2026 interagency model-risk guidance is risk-based but explicitly excludes generative and agentic AI. Agencies have signalled further work on those technologies.
Banks according to prudential supervisor and asset profile; consumer-finance obligations apply by activity and product.
Summarised from the primary instruments below. Open the source before relying on an obligation for legal or compliance decisions.
Use materiality to calibrate model development, validation, ongoing monitoring and governance for in-scope models.
Provide specific adverse-action reasons and comply with fair-lending and consumer-protection law regardless of model complexity.
Understand and validate vendor models, including design, data, performance, limitations and ongoing fitness for purpose.
The original authority remains the source of truth. Dates below are publication and effective dates, not OneBench ingest dates.
Federal Reserve, OCC and FDIC
Replaces prior guidance with a materiality-based framework; generative and agentic AI are outside its stated scope.
Consumer Financial Protection Bureau
States that creditors cannot use model opacity as a reason for failing to provide specific adverse-action reasons.