AI Auditability
An AI audit trail should explain who initiated an interaction, which configuration and data influenced it, which action was proposed or executed, and who approved the action.
Evidence Model
| Event category | Evidence to retain |
|---|---|
| User interaction | User, organization, agent, conversation, timestamp, request and response metadata |
| Agent configuration | Agent version or change record, model connection, tools, skills, memory, data collections |
| Retrieval | Collection and source identifiers, relevant retrieval metadata, authorization context |
| Tool execution | Function, sanitized arguments, result, status, duration, connection, target system |
| Approval | Request, approver, decision, timestamp, reason when available |
| Administrative change | Actor, entity, changed properties, timestamp, correlation ID |
| Provider operations | Deployment, region or data zone, quota, service logs, request correlation where available |
Avoid writing credentials, access tokens, complete secret values, or unnecessary sensitive payloads into logs.
Use the Right Evidence Source
- Tool Executions explains what an agent attempted through a tool and whether it succeeded.
- Audit Log explains platform configuration and access changes.
- Conversation and feedback provide user context and quality evidence.
- Workflow history provides changes to repeatable automation.
- External provider logs confirm what occurred in the target system.
- Azure or model-provider monitoring provides deployment, quota, latency, and provider-side evidence.
Use correlation identifiers and timestamps to join these sources during an investigation. Define retention based on legal, security, support, and privacy requirements rather than keeping every payload indefinitely.
See Audit Log, Tool Executions, and Incident Response.