For auditors, and the teams they audit
A log is not evidence
Every agent framework can write a trace. Very few produce something an auditor will accept, and the difference has nothing to do with how detailed it is.
Start where the auditor starts
An auditor picking up an AI agent's record isn't asking whether it's thorough. They're asking whether they can rely on it without relying on you. That comes down to three questions, and they get to the bottom of what a system can prove very quickly.
Scroll horizontally to see all columns.
| The auditor asks | What a log gives them | What evidence gives them |
|---|---|---|
| Is this complete? | Whatever the system decided to write down. Gaps are invisible. | A record created at the policy check, as part of the same operation, so every checked action has one. |
| Could this have been changed? | Anyone who runs the system can edit it, and nobody can tell. | A signed inventory of the exported files. Change, drop, or add one and it fails verification. |
| Can I check it without you? | Only by asking the team or vendor that produced it. | A proof reference held under separate access, which the reviewer checks with their own tools. |
Why 'we have detailed logging' doesn't land
Detail doesn't touch any of those questions. A more verbose trace is still written by the party with something at stake, still editable by whoever runs the system, and still unverifiable from outside.
That isn't a knock on logging. Observability answers "is this working." Governance has to answer "was this allowed, and can you show me." Two different jobs, and the second doesn't come free with the first.
How the three answers are built
At the decision, not after it. The record is written when the policy is evaluated, before the action runs. Completeness stops depending on whether anyone remembered to log.
Signed when it leaves. An exported evidence package carries a signed inventory of its files. The reviewer checks that signature against a public key they got from the deployment operator, not one packed inside the export.
Checked against a ledger you don't write to. Each session's fingerprint goes into an append-only Azure SQL Database Ledger table. Verification runs with access separate from the access that records evidence, and checks the ledger against separately protected database digests. That is the default in the managed service. Private Enterprise, coming soon to Azure Marketplace, will run entirely in your own Azure subscription, that ledger included. When reviewers want a public reference, a SaaS environment can use the Solana blockchain instead, chosen when the environment is created. It only ever receives fingerprints, never decision content or personal data.
The auditor runs the verifier themselves. Nothing in that path requires calling us.
The question worth asking any vendor
If my company wanted to quietly change one of these records next year, what stops us, and how would anyone outside ever know? If the answer is "our internal controls," what you have is logging. That can be the right call for some systems. Make it on purpose rather than finding out during an inquiry.
Why the business owner should care as much as the auditor
If you own the agent, this is the part that decides how much work it's allowed to take on. An agent whose decisions you can defend with evidence is one risk and compliance will let do more. An agent whose history lives in an editable log gets held to the smallest scope anyone can watch by hand.
Agents make this sharper than earlier systems did, because of volume and speed. When something goes wrong across thousands of actions, reconstructing it afterward from a log nobody outside can check is not a position anyone wants to be in, least of all the person whose name is on the agent.
Check a record yourself
Download a sample evidence pack and run the checks an auditor would, or see what a reviewer gets in the browser.