← Back to Dashboard
1. AI Incident Taxonomy2. Postmortems and Corrective Actions
AI Incident Taxonomy
📚 Incident Response for AI⏱ 10 min⭐ 120 XP
Classify Incidents by Failure Mode
AI incidents include quality regressions, safety policy breaches, latency spikes, cost explosions, and tool orchestration failures. Separate detection and response playbooks by incident type for faster triage - "the AI is acting weird" is not a pageable signal; "guardrail intervention rate tripled in 30 minutes" is.
The Five Failure Classes
| Class | Example | Detection signal | First response |
|---|---|---|---|
| Quality regression | Triage accuracy silently drops after model/prompt change | Online eval score dip, escalation rate rise | Correlate with release manifest; roll back candidate |
| Safety breach | Prohibited content served; PII echoed | Guardrail false-negative report, user report | Contain first: tighten guardrail/disable feature; preserve evidence |
| Reliability | p95 spikes, throttling waves, region impairment | Latency/error SLO alarms | Classic SRE: failover, shed load, engage backpressure |
| Cost explosion | Agent retry loop burns 40M tokens overnight | Token budget alerts, anomaly detection | Kill switch per tenant/route; find the loop in traces |
| Orchestration failure | Duplicate side effects, stuck plans, tool cascade | Tool error rate, idempotency-conflict logs | Pause affected flows; reconcile side effects from audit log |
Why the Taxonomy Pays Rent
- Different oncalls - safety breaches page the safety owner and may trigger legal/comms; latency pages the platform SRE. One generic "AI incident" page delays both.
- Different evidence - quality incidents need eval diffs and release manifests; cost incidents need token telemetry by route; safety incidents need preserved prompt/response pairs.
- Different severity ladders - a P3 latency blip and ANY safety false negative are not the same urgency, and your paging rules should encode that.
Detection completeness check: for each class, ask "what alarm fires, within how many minutes, paging whom?" A class with no answer is an incident you'll discover from a customer tweet.
🧪 Knowledge Check
Press 1-4 to select1 of 2
A useful AI incident taxonomy should:
Use one generic bucket
Separate quality, safety, reliability, and cost failure classes
Ignore policy issues
Exclude tooling failures