Incidents and what to do

Why AI incidents are different

A server outage is obvious. An AI incident often is not: a model starts hallucinating in a high-stakes workflow, drifts after a data change, or is tricked by a prompt injection. Standard IT playbooks (re-image the server, rotate the key) may not fix it. You need a playbook written for behavior.

A simple playbook: seven stages

1. Detect. Define what counts as an AI incident so people know to report it.
2. Triage. Assign a severity.
3. Contain. Pause, restrict or switch the system to a fallback.
4. Assess. Who is affected, how many, which regions, what data.
5. Notify. Inside the organization first. Then consider possible external notifications.
6. Fix. Correct the cause and re-test before restoring service.
7. Learn. Hold a blameless review and update policy, tests and monitoring.

Severity: a plain scale

Level	Meaning
S1	Serious harm to people or rights, or a likely legal notification
S2	Significant harm or many people affected
S3	Limited harm, contained
S4	Near miss

Some programs use three tiers (critical, significant, advisory) instead. What matters is that the scale reflects impact on people, compliance and the business, not how hard the fix is. A simple wording change that closes a critical gap goes ahead of a complex advisory redesign.

Escalation

Decide in advance who does what: operational leads contain, the AI governance lead assesses regulatory exposure, and critical events go straight to the steering group and legal.

Notification: a prompt, not an answer

Some incidents may require telling a regulator, customers or affected people. A personal data breach, for example, can carry short deadlines under data protection law in several regions. Serious incidents involving high-risk systems may be reportable under the EU AI Act. The rules, triggers and timing differ, and this site cannot tell you whether a duty applies to you. Confirm with counsel and the primary source.

Example

A lender’s credit model started declining applications from a group at a higher rate after a change in applicant data. Standard IT checks showed everything healthy. Because a playbook existed, the model owner paused automated declines, switched to manual review, the lead assessed exposure with legal, and the review changed the monitoring to track decline rates by group.

What to do now

• Download the incident checklist and agree severity levels.
• Name who will make the call to pause a system.
• Write down the contacts you would need, and have counsel review them before you need them.