Audit trail and explainability for AI agents, what we ship by default

Audit and explainability are often treated as 'nice to have.' For regulated customers, they're table stakes. Here's what we ship by default.

What's logged for every agent action:

- Timestamp, user, agent identity

- Full prompt (with PII masked at the storage layer)

- Full response and tool calls

- Model version and routing decision

- Retrieval sources used in RAG

- Cost in tokens and dollars

- Policy version active at time of execution

What's exposed for explainability:

- Conversation replay UI for compliance review

- 'Why this answer' trace showing retrieval and reasoning steps

- SIEM export for the customer's existing security tooling

The lesson, audit isn't just logs. It's the ability to answer 'why did this happen' six months later when no one remembers.