What happens when it breaks at three in the morning. Who is called, what they check first, how the problem is communicated, and what is written down afterwards. AI systems add their own failure modes, including quality degradation that no alarm detects because everything is technically working.
Technical view
Requires severity definitions, on-call rotation, runbooks per known failure class, and a declared incident channel. AI-specific classes include provider outage and deprecation, quality regression after a model update, guardrail false positive spikes, cost anomalies from runaway loops, and prompt injection incidents. Blameless post-incident review with tracked actions is what converts an outage into a durable improvement, and quality incidents need a detection story since they are silent by default.
Related
Part of AI Terminology Index, section: Running It in Production.