How to set production AI alerts

Watch latency, quality, and cost after you ship.

1. Log p95 latency, error rate, token spend, and quality proxy (citation / groundedness)

2. Set baselines from last 7 days of stable traffic

3. Page on 2× latency or 5xx spikes; rollback on −30% quality vs baseline

4. Link alerts to trace IDs for one-click investigation

5. Review false positives weekly — tune thresholds to reduce alert fatigue

**Try the lesson:** `production-monitoring-lab` in Lane D