What AI cannot do in legal work today
By the ChamberOne team 5 minute read
Most writing about legal AI lists what it can do. That is half the picture, and for a partner setting expectations the less comfortable half is more useful. If the team believes a tool can do something it cannot, the cost arrives later, usually at a bad moment. These are the parts of legal work that, as things stand, belong to people.
Key points
- Responsibility for the work stays with the lawyer, whatever tool produced the draft.
- Credibility, rapport and negotiating judgement remain human skills.
- Unrecorded local practice is invisible to a tool, so do not rely on it for that.
- Anything you will rely on must be checked against an official or trusted source by a person.
- Use tools where output can be checked quickly and a mistake is contained.
It cannot take responsibility
A tool can produce a letter, a summary or an argument. It cannot answer for it. When a client complains, a court asks a question, or a regulator calls, the person on the end of the line is a lawyer. Professional responsibility attaches to people, and no software contract changes that.
This is more than a technicality. Responsibility is what makes a lawyer read the draft properly. If the team starts to think of the tool as the author, review gets thinner exactly where it should be thickest.
It cannot assess a witness or a client
Credibility is a human judgement made from a mixture of consistency, demeanour, motive and plausibility against everything else you know about the case. A tool can compare two statements and point out where they differ. It cannot sit across a table and notice that the answer was too smooth, or that the client's account has changed each time the same question is asked in a different way.
The same applies to instructions. Clients often say one thing and mean another, or leave out the fact that matters most because they are ashamed of it. Drawing that out is a skill built on trust. A document analysis will not do it for you.
It cannot read a room or negotiate a relationship
Negotiation is rarely just about the positions on paper. It depends on what the other lawyer wants to avoid, how the judge responds to a certain tone, and whether this is a one-off dispute or one of many between the same parties. Much of that is never written down.
A tool can help you prepare: list the issues, draft options, tabulate what each side has conceded. It cannot judge when to pause, when to push, or when a concession now buys goodwill that will matter in a different case next year.
It does not know unrecorded local practice
Every court, registry and tribunal has habits that are not in any rulebook: which clerk to call, how long a listing really takes, which judge dislikes lengthy bundles, how a particular office interprets a form. This knowledge is passed on from one practitioner to another, and models trained on published text will not have it.
This matters most for lawyers in smaller jurisdictions or specialised courts, where there is little published material in the first place. A tool will still answer confidently. That fluency can hide the fact that it has no real basis.
It cannot guarantee the law it states is real or current
General tools can invent authorities and misstate real ones. Tools tied to a source database reduce that risk, but they still summarise, and a summary can drop the qualification that changes the meaning. Law also changes, and a tool's information may be out of date, or may draw on the wrong country's rules.
The working rule is simple: anything you will rely on must be checked against the official text or a trusted source, by a person, before it leaves the firm.
It cannot see what was never given to it
A tool works with the documents and instructions it receives. It does not know about the phone call where the client changed their mind, the email that never reached the file, or the side agreement that was made in a corridor. Where the file is incomplete, the summary will be incomplete, and it may read as if it were complete.
This is one reason a lawyer who knows the matter should read the output. They can tell when something obvious is absent, which is a kind of review a second tool cannot do.
What this means for how you use it
None of this makes the tools unhelpful. It tells you where to put them. They fit best on work where the output can be checked quickly and the cost of a mistake is contained.
- Good fits: first drafts of routine documents, summaries you will verify, lists of issues, extraction of dates and names, restructuring your own notes.
- Poor fits: deciding strategy, judging a witness, giving advice to a client directly, choosing whether to settle, or anything where nobody will check the result.
- Ask each time: who will check this, against what, and how long will that take? If the answer is 'nobody', do not use the tool.
- First drafts of routine documents
- Summaries you will verify
- Lists of issues
- Extraction of dates and names
- Deciding strategy
- Judging a witness
- Advising a client directly
- Choosing whether to settle
Setting expectations inside the firm
Suppose a four-lawyer firm is deciding whether to let a tool prepare the first draft of a settlement proposal. The tool can set out the heads of claim and compare the figures. It cannot know that the other side's solicitor will take offence at a certain phrase, or that the client would rather settle for less than face a hearing in the month their child is born. Those are the considerations that decide the outcome, and they come from the people in the room.
Say these limits out loud. Junior staff in particular may assume that a fluent output has been verified by someone. Trainees who have never seen a bad AI draft may not know what one looks like. Show them an example where a plausible paragraph is wrong, and make clear that finding the error is part of the job, not a sign of failure.
Partners should also be wary of the opposite mistake: dismissing the tools because of their limits. A tool that cannot judge credibility can still save an hour on a chronology. The aim is an accurate map of what it does well, so that people's time goes where only people can help.
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
How a small law firm can start with AI without a big project
Pick two low-risk uses, set review rules and measure before expanding, with a list of what to postpone.
5 minute readA 90-day AI pilot plan for a law firm
A week-by-week plan to choose a use case, set measures, run a small trial, review errors and decide.
5 minute readChoosing AI models and firm-level controls: what a practice should decide
Model choice, an off switch and a usage record are three decisions a firm should make deliberately.
4 minute read