The interesting question about AI in a law firm is not what it can do. It is what it should be permitted to do, and how anyone would know it stayed inside that boundary.
We have opinions about this because we had to write them down before we could ship anything.
Most firm work is not judgement
It is worth being precise about what actually consumes a day in a personal injury practice, because the debate usually skips this.
Requesting records. Noticing that a request went unanswered. Assembling a package from documents already in the file. Telling a client what happened this week. Extracting a date of loss from a form and putting it somewhere else. Checking whether the thing that was supposed to happen has happened.
None of that is legal judgement. All of it requires care, most of it is tedious, and doing it badly loses cases — but the skill involved is diligence rather than analysis.
That is the work worth automating, and it is a large fraction of the total.
The line we drew
Three categories, and the boundaries are enforced in the product rather than described in a policy.
Draft. A machine may produce something a person will review before it leaves the firm. A letter, a records request, a summary of a medical report. The output is a starting point and it is obvious that it is one.
Propose. A machine may say "this looks like it needs attention" and put it in front of somebody. Flagging a case that has gone quiet, noticing a document that contradicts another, suggesting the next step. It cannot act on its own proposal.
Decide. A machine may not choose whether to take a case, what a claim is worth, whether to accept an offer, or anything that binds the firm or the client. Not because a model would necessarily be bad at some of these, but because the person carrying the professional responsibility must be the person who decided.
The categories matter less than the fact that they exist and are checkable. A system where the boundary is a matter of interpretation is a system where it moves.
Why a suggestion must carry its evidence
A proposal a person cannot check is not a proposal, it is an instruction with extra steps.
If the system says a case needs attention, it should say what it looked at. If it drafts a demand from the file, the figures should point at the documents they came from. Not as a footnote for auditors — as the primary way anyone interacts with it, because the reviewing lawyer's job is to disagree when the reasoning is wrong, and they cannot disagree with a conclusion.
This has an unglamorous consequence: it makes the software slower to build and less impressive to demonstrate. A confident answer demonstrates better than an answer with its workings attached. The workings are the point.
The failure mode we are actually worried about
Not a dramatic hallucination. Those get caught, because they are obviously wrong and somebody notices.
The dangerous version is plausible and slightly wrong, in a document nobody re-reads. A summary that inverts a clause. A date transcribed from the wrong page. A figure that is right for a different claimant. It passes review because it reads correctly, and the error is discovered at the point where it costs something.
The defence is not a better model. It is keeping the human in the position where noticing is easy: showing the source next to the claim, keeping generated text visibly generated until somebody accepts it, and never letting output become input to something else without a person in between.
What this means in practice
If you are evaluating any tool in this space, including ours, the questions I would ask are:
- What can it do without a person approving it? Ask for the list, not the philosophy.
- When it suggests something, can I see what it read?
- What happens to its output if nobody reviews it — does it stop, or does it proceed?
- Is any of this enforced, or is it a setting somebody can change on a busy day?
A vendor that has thought about it will have answers immediately, because these are the questions they had to settle before building. A vendor that has not will talk about the model.