The AI Incident Response Runbook
Template
A short runbook for the first hours after an AI feature causes harm or behaves unexpectedly in production.
Contain
- Document exactly how to disable or restrict the AI feature.
- Name who can make that call at any time of day.
- Confirm the fallback experience for users when the feature is off.
Assess
- Identify which users and affected people were exposed.
- Classify the harm: individual, group, organisational or societal.
- Preserve prompts, outputs, logs and model versions for investigation.
Notify
- List when customers, affected individuals, the data protection regulator or sector regulators must be told.
- Name who approves external communication.
Investigate and learn
- Find the root cause: model, data, prompt, integration or use outside intended scope.
- Update the risk register, impact assessment and monitoring triggers.
- Record the decision to restore, change or retire the feature.
Start a conversation
Have a product or AI decision to make?
Useful first calls usually start with one unclear decision, a deadline and a team that needs a practical next move.
Tell us about it