AI incident response playbook
AI incidents need product, engineering, risk, legal, and support coordination. This playbook keeps the first response concrete.
Triage
Classify severity, affected users, active harm, model or data scope, and whether containment is required.
Contain
Pause a feature, disable a model route, add human review, roll back a version, or narrow access.
Learn
Record root cause, evidence, user impact, corrective controls, and monitoring updates.
- Keep escalation contacts current.
- Separate detection time from resolution time.
- Feed incident lessons back into validation and governance checklists.