AI Audit Trails: Recording Who Approved What, and When

An AI audit trail is a record of what AI produced or changed, who approved it, which version they approved, what the approval covered and when. In a growing business it lives next to the work. AI may keep it tidy, but never fills in an approval, a date or a link itself.

This is general information, not legal advice.

What is an AI audit trail?

An AI audit trail is a dated record of AI work and the human decisions around it: what was drafted, which version was approved and by whom, what was published or changed, and how a change can be reversed. It lets anyone reconstruct later what happened and who agreed to it, without relying on memory.

Most businesses have fragments of one. There is a chat thread where someone said "fine", an email with an attachment, and a tool's activity log. None of them, alone, says which version went out or who took responsibility for it.

A proper trail joins those fragments. It does not need special software. A status field, an approver field and a date on each piece of work, plus a change log for anything AI does in company systems, covers most of it.

Why does a growing business need an AI audit trail?

Because questions arrive long after the work. A customer disputes a claim, a figure turns out wrong, or a campaign went to the wrong list. Without a record, nobody can say who approved it or which version they saw. With one, the answer takes a minute, and the fix goes to the right step in the workflow.

Australian guidance asks for it. According to the National AI Centre's Foundations guidance, businesses should "Keep clear records of the actions you take under each practice". Its implementation guidance asks for records of "governance decisions, testing, incidents and monitoring".

Security guidance names the gap too. According to OWASP, lack of audit and telemetry is one of the top ten risks for AI systems connected through the Model Context Protocol. An agent that can act in company systems should leave a record of every action.

There is a learning benefit as well. A trail shows which drafts were approved first time and which were rewritten. Patterns in the rewrites point to the instruction or fact that needs fixing.

What should an AI audit trail record?

Record six things: what the work was, which version was approved, who approved it, what the approval covered, when, and where it went. For any change AI makes in a company system, add what changed, what it looked like before, and one line that undoes it. Each entry should make sense to someone who was not there.

Scope is the field most often missed. Approving the words of an ad is not approving its budget, and approving a draft is not approving its publication. Writing down what the approval covered stops one decision being stretched into another. The guide to approval gates covers what counts as approval in the first place.

Where should an AI audit trail live?

Keep it next to the work, in files and systems the business owns. A status and approval block at the top of each draft, a change log beside each system AI works in, and a register of AI tools and connections cover most needs. Avoid scattering it across inboxes and chat threads, where it cannot be found.

Tool logs help but answer a different question. Business AI plans increasingly include audit logs. For example, Anthropic lists audit logs among the features of Claude's Enterprise plan. Those logs show what users and the system did. However, they do not show that a person read version 4 and agreed to publish it.

Comparing plans? The guide to Claude Team and Enterprise covers which controls each one includes.

What should AI be allowed to write in an audit trail?

AI may record what it did and what it observed, and may mirror decisions a person has already recorded. It must never fill in an approver, an approval date or a published link on its own. It should also log a system change as applied only after reading it back, because a test run is not a change.

The temptation is real. An agent finishing a task wants to tidy up: mark the draft approved because the user said it looked fine, or fill in today's date. Each of those creates a false record, and a false record is worse than none, because people trust it.

Three rules keep the trail honest:

1. Mirror, never originate. AI copies an approval from where the person recorded it, word for word.
2. Read back before recording. A change is logged as applied only once the system confirms it.
3. Leave gaps empty. A missing published link stays missing until a person supplies it.

The same thinking applies to access. The guide to MCP security explains why connections should start read only, with write access added one action at a time.

How do you review an AI audit trail?

Review it monthly in a short session. Look for work that went out without a recorded approval, approvals with no named person, system changes with no revert line, and drafts that were heavily rewritten. Each gap points to a workflow step to fix. Then check the AI register quarterly for connections nobody uses.

The review should be short and practical. It is not an inspection of people; it is a check on the system. If approvals keep going unrecorded, the approval step is probably in the wrong place or too awkward to use.

Keep the trail for as long as the work stays live. A published article or a running campaign needs its record for as long as anyone might ask about it.

Skein runs its own business on a Company OS, and every report, plan and piece of content it delivers goes through it, with a named person approving anything published, sent or spent; the shared knowledge layer page shows how it is organised. That suits a business where several people and agents produce work under one name. A founder producing everything personally may need only a dated approval note on each piece.

Frequently asked questions

What is an AI audit trail?

An AI audit trail is a dated record of AI work and the human decisions around it. It shows what was drafted, which version was approved and by whom, what the approval covered, where the work went and how any system change can be reversed. It lets anyone reconstruct what happened later.

What should an AI audit log include?

Include the item and its version, the named approver, what the approval covered, the approval date and where the work went. For changes AI makes in company systems, add what changed, what it looked like before and a single step that undoes it. Each entry should stand on its own.

Are vendor audit logs enough?

No. Vendor audit logs show what users and the system did inside a tool, which is useful for security. They do not record human decisions, such as who read a draft and agreed to publish it. A business needs both: tool logs for actions, and its own records for approvals.

Can AI keep its own audit trail?

Partly. AI can log what it did and what it observed, and can copy approvals a person has recorded. It should never fill in an approver, an approval date or a published link itself, and should record a change as applied only after confirming it in the system.

How long should AI approval records be kept?

Keep them for as long as the work stays live, and longer where contracts, industry rules or record-keeping obligations require it. A published article or running campaign needs its approval record for as long as someone might question it. Check any specific retention rules with your adviser.

Related reading