Audit trail for AI agents

How it works

1. 01

   Bind every action to an identity

   Each Helix agent has exactly one sponsor (a human with a verified account) and exactly one approval policy. The audit log is keyed on the identity, so revoking it or rotating its scopes is reversible and traceable.

2. 02

   Capture the approval decision, not just the send

   Approval queue entries record the agent's reasoning, the matched rule, the approver, and the time-to-decision. Approve, edit, and skip events are first-class — not derived from inbox state.

3. 03

   Export for SOC 2 and internal review

   Pull audit events from the dashboard or the REST API. Records carry provider, message ID, identity, approver, and rule match — enough for SOC 2 evidence collection and internal incident reviews.

In plain prose

The most-cited blocker to letting AI act on a user's behalf is not capability — it is accountability. "Why did the agent take this action? Was the action authorised?" is the question every compliance officer and security reviewer asks. Native MCP server logging was designed for debugging, not for evidence collection; getting an audit trail out of a hand-rolled MCP server is non-trivial.

Helix bakes the audit trail in. Every send, every calendar write, every approval decision is captured with the agent identity, the sponsor, the matched approval rule, and the recipient or attendee list. The log is per-identity, so revocation cleanly bounds blast radius.

For SOC 2 evidence collection, exports include the rule match (or "fallback" when no rule fired), the approver, and a deterministic identity ID. For internal incident review, the audit timeline is enough to reconstruct what an agent did during any window without replaying the full message body.

FAQ