What an AI audit log should record in a law firm
By the ChamberOne team 4 minute read
Sooner or later someone will ask a question about AI in your firm. A client may want to know whether their documents were put through a tool. A regulator may ask how you supervise its use. A partner may want to know why a draft contained an odd paragraph. If your answer is that you think someone used it last month, you have a gap. An audit log closes it. The trick is to record enough to answer real questions, without building a second archive of confidential material.
Key points
- Record user, matter, time, tool, model and what a person did with the output.
- Keep full inputs and outputs under the matter's own access rules and let the log point to them.
- The log should record automatically and ordinary users should not be able to switch it off.
- Use it for monthly supervision questions, not only after something goes wrong.
Why the log matters more for AI than for most tools
Most firm software does the same thing every time. AI does not. The same request can give a different result tomorrow, and the provider may change the underlying model without telling you. That makes it hard to reconstruct what happened after the event unless you wrote it down at the time.
A log also turns a policy into something checkable. A policy that says only approved tools may be used on client matters is a sentence. A log that shows which tools were actually used, by whom and on which matters is evidence that the sentence is being followed, or a warning that it is not.
The core fields
Think of each AI request as an event. A useful record of that event answers the following questions without anyone needing to guess.
- Who: the named user, not a shared login. If the request was made by a system on someone's behalf, record both.
- Which matter or client, so that the use can be traced to the work it supported and, if needed, discussed with that client.
- When: a precise time, in one consistent time zone.
- Which tool and which model, including the version where the provider states one. This is the field people forget and later miss most.
- What kind of task: summarising, drafting, extraction, question answering.
- The size of the request, such as the amount of text sent, which helps with cost checks and unusual activity.
- What happened next: whether a person accepted, edited or discarded the output.
Inputs and outputs: record them, or only a reference?
This is the hardest design choice. Keeping the full text of every prompt and answer gives you the best reconstruction. It also creates a store of privileged and personal material that must itself be secured, retained and, in some cases, deleted on request.
A middle course suits many firms. Keep full inputs and outputs where they already live, attached to the matter, and have the log hold a reference to them. Then the same access rules that protect the matter protect the evidence, and a restricted matter's AI history is as restricted as its documents. If your tool stores prompts in a separate place, find out who can read that place.
- Gives the best reconstruction
- Creates a store of privileged and personal material
- That store must be secured, retained and sometimes deleted on request
- Inputs and outputs stay attached to the matter
- The matter's access rules protect the evidence too
- A restricted matter's AI history is as restricted as its documents
Make the log hard to quietly switch off
A log that users can disable, or that records only when they remember to tick a box, will be missing exactly the entries you most want. Ask whether recording happens automatically on every request, whether an ordinary user can turn it off and whether administrators can edit or delete entries.
Also ask who reads the log. A read-only role for a compliance lead or auditor, separate from the people who do the work, is a sensible arrangement. If the only people who can see the log are the people it records, it is less useful as a control.
How the log supports supervision
Supervision is where the log pays for itself. A responsible partner can ask a few simple questions each month and get answers in minutes.
- Which matters used AI this month, and were any of them restricted or particularly sensitive?
- Which people use it heavily, and which not at all? Both groups may need a conversation.
- Were any models used that are not on the approved list?
- For a sample of entries, was the output reviewed and changed, or accepted as it came?
The last question matters. If nearly everything is accepted without edits, either the tool is very good or the review is thin. A log cannot tell you which, but it tells you where to look.
- 1Which matters used AI, and were any restricted?
- 2Who uses it heavily and who not at all?
- 3Were any unapproved models used?
- 4Sample entries: was output reviewed or accepted as it came?
- 5Talk to the people the answers point to
- Then back to step 1.
Answering a client or a regulator
Suppose a client's in-house counsel asks whether anything from their matter was sent to an AI provider. With a log that records matter and model, you can answer in an afternoon and attach the facts. Without one, you are asking each fee earner to search their memory.
Regulators and bars differ in what they expect, and expectations are still forming. Check the guidance from your regulator or bar. In general terms, being able to show what you used, who used it, what controls applied and how you supervised it is a better position than having to reconstruct it under pressure.
Retention and privacy of the log itself
Decide how long you keep AI records, and write down why. Many firms align this with how long they keep other matter records and audit data. A log that is deleted after a month cannot support a complaint raised a year later. A log kept forever, with full text, becomes a liability of its own.
Remember that a log names staff and sometimes client individuals. Treat it as personal data under whatever data protection rules apply to you.
A note on shared accounts and integrations
Two situations quietly weaken a log. The first is a shared login, used by a whole team because it was cheaper or quicker to set up. Every entry then shows the same name and tells you nothing. The second is an integration that calls an AI service in the background, for example when a document is uploaded. If the integration does not record the request, a whole category of use will be invisible. When you review a tool, ask for a demonstration of both cases and look at what appears in the log afterwards.
What a log cannot do
A log records that an event happened. It does not show that the output was right, or that someone read it carefully. It will not catch a lawyer who pastes client material into a free consumer tool outside your systems. For that you need clear rules, training and a culture in which people say when they have used something. The log is one control among several, not a substitute for judgment.
AI records in ChamberOne
- Every AI call is recorded in the audit log with the model and token counts.
- Every read and write is recorded automatically with who, when, IP address and before and after values, and users cannot switch it off.
- The log can be filtered by person and action, and exported.
- Audit events are kept for 7 years by default, on every plan.
- A read-only Auditor role exists.
What it does not do
- Tamper-evident hashing is on the roadmap. It is not in the product today.
- The log shows that an AI call happened. It does not show that the output was correct, and AI output needs a lawyer's review.
This article is general information about the law, legal practice and technology, current when it was published. It is not legal advice. Laws and guidance change: check the sources linked above and take advice on your own circumstances from a lawyer qualified where you are.
Read next
Corporate Transparency Act: who must still file beneficial ownership reports
FinCEN's August 2026 final rule made permanent the end of BOI reporting for US-formed companies; only certain foreign entities still file.
5 minute readClient confidentiality and AI tools: a checklist before you paste anything in
The questions to answer about retention, training use, access, hosting and sub-processors before client information goes into any AI tool.
5 minute readShadow AI in law firms: staff are already using it, so what next?
Why bans fail, how to find out what is being used, and how to offer an approved alternative with clear rules.
5 minute read