60-Second Guide · 20

AI Model Updated. Do You Need Revalidation?

Assess whether a model or service update changes approved use, failure modes, acceptance criteria, human controls, or previously demonstrated performance.

Author
Sandip Thorat
Published
12 September 2026
Last reviewed
12 September 2026
Duration
About 1 minute
Content type
Video learning
Topics
AI change assessment · Revalidation · Critical misses

Transcript

The supplier version label is not a risk rating

An A I provider calls a model update minor quality improvements. That label does not determine revalidation. The relevant question is whether the update can affect your approved intended use, supported population, critical failure modes, source behavior, human review, or previously demonstrated performance.

Compare the controlled configuration

Review model and service version, prompt, reference corpus, retrieval settings, parser, guardrails, tool permissions, workflow, roles, and interfaces. Execute the locked reference set and representative human workflow. Compare correct findings, false positives, false negatives, critical misses, source correctness, overrides, and prohibited actions.

Retest, restrict, roll back, or approve

If evidence shows no relevant effect, document the basis. If performance changes, run affected regression, strengthen controls, restrict use, or roll back where possible. Define monitoring and revalidation criteria. Revalidation is not required because the version number changed. It is required when the change challenges the approved use or its supporting evidence.