How should an organisation prove who initiated, approved and executed an AI agent action?

In plain language

In a bank transfer, the person who gives the instruction, the person who approves it and the system that moves the money may all be different. One line saying 'transaction completed' does not show who requested it or what was approved.

Why this matters

If the roles are not separated, it may be impossible to detect retrospective approval, reused permission, false human attribution or approval that was granted for one item and used for another.

Do not confuse

• Initiation: the event that creates the action or proposal.
• Approval: permission for particular content, a particular target and a defined scope.
• Execution: technical performance of the action.
• Target effect: the result that actually occurred in the external system.

What should you do?

1. Assign a unique action_id to every action.
2. Record the initiator and evidence of initiation.
3. Preserve the approver's identity and authority, together with the target, content, duration and single-use status of the approval.
4. Link the executing agent, runtime, tool and technical-principal identities.
5. Require fresh approval if the content, target or amount changes after approval.
6. Obtain evidence of the result and the before-and-after state from the target system.
7. Keep retries, partial execution, rollback and failure in the same evidence chain.

How do you audit it?

• Can the identities of the initiator, approver and executor be verified?
• Does the approval apply to the exact version that was executed?
• Was the person or policy giving approval authorised for this action?
• Was the same approval reused or applied to another target?
• Does the source log match the result in the target system?
• Are the timeline and chain of actors complete?
• Could the record have been altered later?