AI observability goes beyond standard application monitoring. Unique challenges: LLM outputs are semantic (not binary), costs scale with usage, hallucinations are hard to detect automatically. Key metrics: 1) Latency: time to first token (critical for UX), total generation time, retrieval latency for RAG. 2) Cost: per-request token usage, cost attribution per feature/user/model. 3) Quality: task-specific metrics (classification accuracy, retrieval precision, human evaluation scores). 4) Safety: rate of filtered outputs, user feedback on harmful outputs, prompt injection detection rate. 5) Reliability: error rate by error type (rate limits, context overflow, timeout), retry rate, fallback model usage. Tools: Langfuse (open source, self-hostable): rich trace visualization, prompt versioning, A/B test tracking, cost analysis, evaluation datasets. Helicone: request logging proxy layer (just change base URL), instant cost dashboards, prompt caching. Prometheus + Grafana: custom metrics instrumentation for production dashboards. Implementation: use OpenTelemetry for vendor-agnostic instrumentation, emit traces from every LLM call with model, prompt version, user segment, and feature tags. Alerting: P99 latency > 10s, error rate > 2%, cost per user > threshold, sudden accuracy drop on evaluation set.