A water utility caught a wrong shutoff-notice draft during review and needs to reconstruct how the agent produced it. Which responsible-AI practice should the team use?

Audit of agent traces and provenance metadata is the responsible-AI practice that lets the water utility reconstruct how a wrong shutoff-notice draft was produced. Trace and provenance records typically capture prompts, system instructions, retrieved policy or account context, tool calls, intermediate reasoning steps, and the final draft. Replaying that path after a near-miss enables root-cause analysis—wrong customer segment, stale shutoff policy chunk, or misapplied template—and supports corrective controls without guessing. Auditing output token logits and sampling parameters from the incident shows low-level decoding detail and randomness settings but does not reconstruct the draft path; it does not reveal the prompts, retrievals, and tool calls behind this draft. Auditing retrieved passages and relevance grades from the incident covers only the retrieval slice and does not reconstruct the full path that produced this failed draft. Auditing guardrail decisions and policy version IDs from the incident shows whether safety checks fired but does not trace the prompts, retrieved context, and tool calls that produced this particular draft. Exam caveat: provenance for responsible AI is about reconstructability and accountability, not optional telemetry you enable only after a breach. Operational check: confirm that shutoff-notice generations retain immutable trace IDs linking prompt, retrieval hits, tool results, and reviewer disposition for a defined retention window aligned with utility records policy.