Incident Response
Plan for AI failures before they happen: severity levels, a kill switch, communication, and a blameless review that leads to real fixes.
Objective
By the end of this lesson, you can write an incident response runbook for an AI feature, including severity levels, a kill switch, who communicates with whom, and how to run a blameless review.
AI incidents are different
Traditional incidents are often loud: the site is down. AI incidents are often quiet: the assistant gives wrong eligibility answers in Spanish for three days, and nobody notices until a family is turned away. That is why monitoring (last lesson) and a clear way for staff and clients to report problems matter so much.
Examples of AI incidents:
- Wrong answers that could lead someone to miss a deadline or a benefit.
- Personal data shown to the wrong person.
- Harmful or offensive output.
- An agent taking an action nobody approved.
- A successful prompt injection.
Severity levels
Decide levels in advance so nobody argues during the incident:
| Level | Meaning | Example | Response |
|---|---|---|---|
| SEV 1 | Harm to people or data exposure | Client records shown to another user | Kill switch now; leadership and legal notified |
| SEV 2 | Wrong answers on important topics | Wrong ID rules in one language | Disable affected topic or language; fix within a day |
| SEV 3 | Degraded but safe | Slow responses, fallbacks working | Fix in normal work |
The kill switch
A kill switch turns off the AI feature quickly and sends people to a safe path. Requirements:
- One action, by a named role, with no code deployment needed (a feature flag).
- The fallback is ready: FAQ page, fixed replies, phone line.
- It is tested on a schedule. An untested kill switch is a hope.
- Partial switches are useful too: turn off one tool, one topic, or one language.
The runbook
A runbook is the step-by-step guide for the person on call:
- Detect and declare. Anyone can raise an incident. Name an incident lead.
- Contain. Use the kill switch or partial switch. Stop the harm first, investigate second.
- Assess. Who was affected? What data? Over what time? Logs answer this (last lesson).
- Communicate. Tell staff what to say. Notify affected people honestly and in their language. Follow legal notification rules for data exposure.
- Fix and verify. Add the failure to the test set. Confirm the fix passes before turning the feature back on.
- Review. Hold a blameless review within a week.
Blameless reviews
A blameless review asks how the system allowed the failure, not who to punish. People report problems honestly when reporting is safe. The output is a short write-up: timeline, impact, root causes, and action items with owners and dates.
A national example
Incident readiness
0 of 8 doneTabletop drill
Go deeper: Systems Administration, Simplified: A Practical Blueprint for Uptime, Security, and Scale
Discussion
Ask. Answer. Build together.
Loading…