An audit trail is a chronological record of events and actions that makes a GenAI system’s behavior reviewable. It should help reconstruct what the model saw, what safeguards did, and who or what was affected. section 4.3.1 section 2.6.2
For sensitive GenAI systems, capture enough to trace a request and review consequential actions:
- Record timestamps, user and session identity, model version, token counts, tools invoked, and whether actions were allowed or blocked; link events with a shared request identifier. section 4.3.1
- Include retrieval sources, prompt versions, guardrail decisions, human reviews, data accessed, external calls, and policy overrides where relevant. section 4.3.1 section 5.6.4
- Redact or tokenize personal and sensitive information before storage, and alert on behavior such as unusual data access or attempts to use disallowed sources. section 4.3.1 section 5.6.4
Reviewing an agent action
A smart-irrigation agent records its action ID and outcome after running a zone, alongside the checks for valve status, water budget, and authorization. That evidence supports review of who acted, when, and under which policy. section 2.6.2




