What an auditor, or Sekit's evidence engine, asks for.
Change and release management records
The process to review and approve changes to production systems before applying them, and how new versions are released in a controlled way.
From the Sekit evidence catalog
In practice
Deploying an AI system should follow the same discipline as any production change: reviewed, approved, with a rollback plan if it goes wrong. The change and release management records show this for the AI system specifically, not only the surrounding application. A common gap is a vendor AI feature that gets toggled on through a settings panel with no change ticket at all, because it "isn't code." Auditors ask whether a recent AI-related change, a new model version or a new prompt template, went through the same approval as any other production change.
Common gaps
Enabling a vendor AI feature through a settings panel bypasses the change management process because it isn't seen as code.
There is no rollback plan for an AI system, so a bad model update stays live while a fix is worked out.
Release readiness criteria exist for software generally but were never adapted for AI-specific risks like a degraded model.
Questions your auditor will ask
Does turning on a vendor AI feature go through change management?
Yes, enabling or reconfiguring an AI feature is treated as a production change requiring the same review and approval as any other.
What is the rollback plan if an AI system misbehaves after deployment?
A documented rollback step, reverting the model version or disabling the feature, recorded in the release management records.
Who authorizes an AI system release?
The named approver defined in the change management process, the same authority used for any production release.
How do you know a release was genuinely tested and approved before going live?
The change and release management records show the test result, the approver, and the deployment date together.
Where regulation demands it
NIS2 art. 6.4 (Change management, repairs and maintenance) covers this, which applies to AI system deployment the same as any production change.
Related controls
Via the shared Sekit CSF topic, not the framework's own index.