Blog
What should we record on the file when AI is used on a matter?
Record what the AI was asked to do, what a fee earner checked, and what the fee earner concluded. The file needs to show human judgement, not a transcript of every prompt.
Alesis · · 5 min read
Record enough to show who did what, and on what basis. In practice that means a short note of the task the AI was used for, the sources or documents it worked from, what the fee earner verified, and the conclusion the fee earner reached. You are not required to keep a transcript of every question you typed, and doing so usually adds noise rather than protection.
Why this matters more than it looks
A file record is not paperwork for its own sake. It is what you rely on when something is questioned months or years later, and the people questioning it are rarely interested in the software.
Four audiences may read the file:
- A supervisor reviewing work in progress, who needs to see that checks happened before the work went out.
- A client or the Legal Ombudsman looking at a complaint about the service, including whether the client was told how the work was done.
- Your insurer and its solicitors, if a claim is made about advice or a missed date.
- A costs assessor or an opponent, where the time claimed has to match the work described.
In every case the useful record is the same: a human being considered this, on this material, and decided this. The involvement of AI is a detail within that, not a separate story.
What to record
Keep it short and specific. A few lines on an attendance note or file note will normally do.
- The task. "Used the AI assistant to pull the key dates out of the correspondence bundle and produce a first chronology." Not "used AI".
- The material it worked from. The bundle, the client's documents, named legislation or official guidance. If the answer depended on documents, say which.
- What you checked. The point that matters most. "Checked each date against the underlying letters at pages 14, 22 and 40. Checked the limitation calculation against the statute myself."
- What you concluded, and what you did next. Your view, in your words, and the step you took.
- Anything the AI could not answer. If a question came back as unsupported or the papers were silent, that is a real finding and belongs on the file, because it usually points at a document you still need.
Where a draft was produced with AI assistance and then revised, the revised draft is the record. Keep the version that was reviewed and signed off, in the usual way you keep drafts.
What not to record
Do not paste long chat transcripts into the file as a substitute for a note. They are hard to read, they bury the reasoning, and they invite a reader to pick over discarded lines of enquiry that never formed part of your advice.
Do not record anything that suggests the AI made the decision. "The system says the claim is in time" is a weak line in a claim file. "I calculated the limitation date and checked it against the statute; the AI working is at the back of the note" is a strong one.
Do not create a parallel AI log outside the matter file. Two systems means two versions of events, and the one you cannot find is always the one you need. If your firm wants a central view of AI use for supervision or training, take it from the matter files, not instead of them.
Where the record lives
The matter file. The same place as attendance notes, research notes and drafts. If your case management system has a note type for research or a chronology, use it.
Two practical points:
- Retention. AI-assisted notes are part of the file and follow your ordinary retention policy. Do not invent a shorter or longer period for them.
- Privilege and disclosure. Treat a research note produced with AI assistance the way you treat any other internal note. If you would not have put a stray thought in a research memo, do not put it in an AI note either.
When the AI gets something wrong
This is where records earn their keep. If a fee earner catches an error before it goes anywhere, the ordinary file note is enough: the check found the problem, the fee earner corrected it, the work went out right. That is the system working.
If an error reaches a client, a court or another side, treat it as you would any other error. Correct it, tell whoever needs to know, and record the correction on the file with the date and who made the decision. Consider whether it needs reporting internally under your firm's own procedures, and whether your insurer's notification requirements are engaged. The fact that AI was involved does not change the obligation, and it is not a defence.
It is worth logging near misses somewhere your supervisors will see them, because patterns in what the tool gets wrong are how a firm learns what to check hardest.
Where Alesis fits
Alesis answers questions about a matter from the matter's own papers and names the page each answer came from, so the note a fee earner writes can point at the same pages. Where it counts key dates and figures, the working is shown step by step with the rule that allows each step, and it says the firm's diary system still governs. When it cannot find support for a point it says what is missing instead of guessing, which is often the thing most worth recording. It prepares drafts for a qualified person to review and sign off, and it does not provide legal advice.
