Observability for AI: What to Log, Trace, and Measure

Capture prompts, parameters, and key outputs with privacy controls. Correlate inference spans with upstream events and downstream user actions. Track latency distributions, token counts, and provider errors. With the right signals, flaky behavior becomes diagnosable, incidents shrink, and product teams gain confidence to iterate.