Guide

Model risk and generative AI: what reviewers will ask for

A practical checklist for getting generative AI systems through model risk review.

Model risk frameworks were designed for scorecards and statistical models. Generative AI challenges many of their assumptions, but the underlying questions are the same: what does the system do, how do we know it works, and what happens when it does not?

Purpose and scope

Reviewers will want a clear statement of the use case, the decisions the system supports, and the decisions it must never make on its own.

Evidence of performance

  • Evaluation sets that reflect real inputs, including difficult and adversarial cases
  • Metrics tied to business risk, not just generic benchmarks
  • Results broken down by segment to reveal uneven performance
  • Known limitations, documented in plain language

Controls

Expect questions about guardrails, human review, logging and the ability to switch the system off. Show how outputs are monitored in production and who responds when something goes wrong.

Change management

Models and prompts change. Reviewers will ask how changes are tested and approved. Continuous evaluation in your pipelines turns this from a burden into a routine.

Teams that prepare this evidence from the start find model risk review becomes a checkpoint rather than a wall.

Let's discuss where AI can take you.

Tell us what you are working on. A senior member of the team will respond to arrange a conversation.

Contact us