What should an audit trail contain for an AI tool?

Enough to place events in order and attribute them: who initiated something, when, against which system, and what the system did in response. Whether the full content of inputs and outputs belongs in it is a judgement rather than a default, because storing that material is itself a data decision worth taking deliberately.

Is a log the same thing?

Not quite. Logs are usually operational, kept for debugging, and often editable or short-lived. An audit trail is kept deliberately as evidence, which implies durability, restricted access and a defined retention period. Many products have the first and describe it as the second, so it is worth asking which you are being offered.

Who normally owns it?

Whoever administers the account, which is why unadministered accounts are the practical problem: the record either does not exist or belongs to the individual. Getting people onto administered tiers of tools they already use is usually the single largest improvement available to an organisation's evidence position.

How does this relate to unapproved tool use?

Directly, and it is the reason the two terms sit together. The exposure from staff using unapproved tools is only partly about what was sent; the larger part is that no record exists, so the organisation cannot answer questions about it afterwards even in good faith and with everybody cooperating.

Glossary

Audit trail

An audit trail is a durable, tamper-resistant record of who did what and when inside a system, kept so that a question asked afterwards can be answered with evidence rather than recollection.

In plain terms

It is the log that lets somebody reconstruct events months later. Who asked, what they asked, what came back, what the system did next, and when. Ordinary business software has kept these for decades and nobody thinks about them until a question arrives that cannot be answered any other way. At that point the difference between having one and not having one is the difference between an answer and an apology.

Why it matters

Because with AI tools the questions arrive more often and are harder to answer from memory. Somebody wants to know whether a customer's document was ever sent to a particular product, which staff have been using a tool and for what, or why a system produced a specific output on a specific day. None of those can be reconstructed after the fact without a record kept at the time, and the moment you discover you need one is always after the window in which you could have started keeping it.

How it works

It records actors, actions and time, at minimum. Who initiated something, what they did, what the system did in response, and when each of those happened. The value comes from being able to place events in order, which is why timestamps and identity matter more than volume of detail.

Durability is the property that makes it evidence rather than telemetry. A log people can quietly edit answers questions only as reliably as the person answering wants it to, so the useful ones are append-only, retained deliberately and held somewhere separate from the system they describe.

AI tools complicate it in a specific way. What went in and what came back are both content rather than simple events, and both may be sensitive, so recording enough to reconstruct a decision can mean storing exactly the material you were trying to be careful with. That tension is real and is usually settled by recording references and metadata rather than whole payloads.

Coverage is decided by the account, not by the tool. Individual accounts of a product typically produce a record only that individual can see, if any, which is why the administered organisational tier is usually the thing that makes an audit trail possible at all.

It is only useful if somebody can actually query it. A record that exists but cannot be searched, filtered or exported answers no questions in practice, and discovering that during an incident is the worst possible time. It is worth asking to see the interface, not only whether logging is on.

What one entry needs to be useful

A single entry is worth looking at closely because it shows what the record is trading off. It has to be specific enough to reconstruct events, which means identity, time, action and the thing acted upon. It also has to avoid becoming a second copy of the material you were being careful with, which is why the source appears as a reference rather than as the document's contents. The account line is the one people leave out and then regret: the same action taken through a personal login and through an administered one produce entries with entirely different evidential weight, and often the personal one produces no entry at all. Note also what is absent. There is no verdict here, no assessment of whether the action was appropriate. An audit trail records what happened; judging it is a separate activity, and mixing the two tends to produce a record people are reluctant to write honestly.

Seen in the wild

Ask which product tier records who asked what, and whether an administrator rather than the individual can see it.
ChatGPT
Check what an enterprise search product records about who searched for what, given it can reach material across many systems.
Glean
Look at what an automation platform keeps about each run, which is the record of what a system did unattended and on whose behalf.
Make

Common misconceptions

People assume

The tool logs things, so we have an audit trail.

In fact

Most products log something for their own operational purposes, which is a different thing from a record you can query and rely on. The questions worth asking are who can see it, how long it is kept, whether it can be altered, and whether you can get it out. Logging existing answers none of those.

People assume

We can turn it on if we ever need it.

In fact

A record only answers questions about the period it covers, so switching it on the day a question arrives produces nothing about the months anybody is asking about. It is one of the few controls that is worth having before there is a reason, precisely because it cannot be applied retrospectively.

Questions

What should an audit trail contain for an AI tool?: Enough to place events in order and attribute them: who initiated something, when, against which system, and what the system did in response. Whether the full content of inputs and outputs belongs in it is a judgement rather than a default, because storing that material is itself a data decision worth taking deliberately.
Is a log the same thing?: Not quite. Logs are usually operational, kept for debugging, and often editable or short-lived. An audit trail is kept deliberately as evidence, which implies durability, restricted access and a defined retention period. Many products have the first and describe it as the second, so it is worth asking which you are being offered.
Who normally owns it?: Whoever administers the account, which is why unadministered accounts are the practical problem: the record either does not exist or belongs to the individual. Getting people onto administered tiers of tools they already use is usually the single largest improvement available to an organisation's evidence position.
How does this relate to unapproved tool use?: Directly, and it is the reason the two terms sit together. The exposure from staff using unapproved tools is only partly about what was sent; the larger part is that no record exists, so the organisation cannot answer questions about it afterwards even in good faith and with everybody cooperating.

Key takeaways

An audit trail records who did what and when, durably enough to be evidence.
Logging existing is not the same as having a record you can query and rely on.
It cannot be applied retrospectively, so it is worth having before there is a reason.
AI tools complicate it because inputs and outputs are content, not simple events.
Coverage follows the account: unadministered accounts usually produce no usable record.

Tools that use this

ChatGPT
Where the recording differs between individual and administered accounts.
Glean
A product reaching across systems, where who searched for what is worth recording.
Make
Run history as the record of what ran unattended, and on whose behalf.

Last checked July 2026

All glossary terms