AI incident classification and playbooks

What it does

It prepares the organisation for the incidents AI actually produces. The incident taxonomy gains categories for harmful output, discriminatory outcomes, data disclosure through a model, prompt injection, agent misbehaviour, supplier model change and others, with criteria for severity based on harm to people as well as to systems. Each category has a playbook that says how to contain, who to involve and what to preserve. The head of incident management owns the taxonomy; AI system owners write the playbooks. It is maintained continuously and used in every AI incident.

How to implement it

• Add AI categories and severity criteria to the incident taxonomy.
• Write a playbook for each category, with system-specific steps for containment.
• Name the roles needed, including someone who understands the model, and put them on the rota.
• Include the reportability check and the customer impact assessment as standard steps.
• Train responders and run exercises on AI scenarios.
• Update categories and playbooks after each incident and exercise.

Evidence it operates

Design: the taxonomy and playbooks; rota arrangements. Operation: incidents classified under AI categories; playbooks referred to in incident records; training and exercise records; revisions made after incidents.

Where it fits

Prepared in advance and applied within AI incident response and rollback. It is the primary control, with that procedure, for AI incident response gap, and with regulatory incident notification for serious incident reporting failure, because an incident that is never classified is never assessed for reporting.

Limits

Playbooks cover the incidents someone anticipated, and AI failures are often novel. Classification depends on the person who first sees the problem recognising it as an incident. Severity criteria based on harm to people need information that is not available in the first hours. Playbooks that are not exercised are not known.