Start with a completed agent action and ask whether another person can reconstruct it. A successful tool response proves execution, but it does not by itself show the permission, policy or review behind it.
The Gap Between Logging and Audit Evidence
Compare the execution record with the permission decision. Link them through a stable event reference so a reviewer can distinguish an allowed action from an action that merely succeeded.
What Auditors Actually Look For
Agree what the review is meant to establish and who can inspect the evidence. Ask for the relevant action, actor, time, policy version and outcome; this is a practical schema, not a universal statutory field list.
EU AI Act Article 12: What the Law Actually Requires
Article 12 requires high-risk systems to support automatic event recording over their lifetime, with traceability appropriate to their intended purpose. Relevant events support risk identification, post-market monitoring and operational monitoring. The specific usage-period, reference-database, matched-input and verifier-identity fields in Article 12(3) apply to the remote biometric identification systems in Annex III point 1(a).
Three Structural Gaps to Review
Inspect explanation, integrity and policy history as separate concerns. A record can be readable yet incomplete, or protected against later changes yet inaccurate when first written.
Gap 1: Missing Decision Rationale
Record the decision inputs and rule needed to explain the action, using protected references where possible. Do not treat a generated explanation or the model's private reasoning as verified evidence of how the decision was made.
Gap 2: No Tamper Evidence
Choose integrity controls from the threat model and applicable obligations. Restricted access, retention controls, append-only storage and cryptographic verification can help. Article 12 does not prescribe Ed25519, hash chaining or a separate storage service. Signatures help detect changes only when verification and key management work; they do not prove that the original record was truthful or that an audit will pass.
Gap 3: No Policy Reference
Keep the policy version that applied at the time of the action. A link to the current policy alone cannot explain a decision taken before that policy changed.
What an Audit-Ready AI Agent Audit Trail Looks Like
Define a stable event identifier and connect the proposed action, permission result, execution outcome and any human review. Record failed and cancelled actions as well as successful ones; restrict access to sensitive supporting material.
How to Diagnose Your Current Logs
Select completed, denied and interrupted actions. Ask a reviewer to reconstruct each from the available records, then list missing links, inconsistent timestamps and permissions they cannot verify. Fix the highest-impact gaps before increasing log volume.
FAQ
Can a signature make an audit pass? No. It can support integrity checks, but the review also depends on accurate records, applicable controls and evidence that those controls worked.
Review updated 1 October 2026: the AI Omnibus entered into force on 27 July 2026. The high-risk requirements in Chapter III Sections 1-3 apply from 2 December 2027 for Article 6(2)/Annex III systems and 2 August 2028 for Article 6(1)/Annex I systems, subject to the Act's transitional provisions. Other obligations have separate application dates.
EU AI Act: consolidated text, 27 July 2026 | European Commission: AI Omnibus entry into force | GDPR: Regulation (EU) 2016/679