Model Risk Management

Model risk management asks what could go wrong when a business relies on model output, including output from a technically sound model used for the wrong purpose. The banking agencies’ revised 2026 guidance organizes this around model risk, development and use, validation and monitoring, and governance. It explicitly describes guidance rather than enforceable standards (Federal Reserve, Supervisory Guidance on Model Risk Management). Do not present that document as a universal legal requirement for AI products.

For example, an internal assistant that drafts a summary and a system whose output sets a lending decision have different consequences when they are wrong. For your own application, record the intended use, who relies on the output, the known limitations and who can suspend it. Define which changes require another validation pass, such as a new model or a broader permitted use.

Discussing the governance questions enterprises bring to agent deployments, Vikram Chatterji says, “I have auditability, and I have all of these other responsibilities” (episode 29, 21:52).

For builders, turn a policy into evidence you can inspect: a reviewed test set, approval records and a monitoring plan with an owner. Audit trail covers reconstructing actions; how to govern AI agents covers operating responsibilities. The enterprise AI hub connects these decisions to deployment and business outcomes.