AI for legal teams: useful for the first pass, never for the final word

Obvious nonsense was never going to be your problem. Your problem is a clean, confidently worded summary of a clause that is not quite what the clause says. AI for legal teams is a harder conversation than AI for anyone else in the building, and not because lawyers are squeamish about new tools. Your function absorbs the consequences. Elsewhere an error produces an awkward correction. In yours it can reach a counterparty, a regulator or a client before anybody notices.

What it is worth using for

The list is narrower than the marketing suggests and still long enough to matter.

• Summarising a long agreement. Ninety pages of a supplier contract reduced to what each side must do, what triggers termination, and which obligations carry a deadline.
• Comparing a clause across a set. Hand it twelve customer agreements and ask how each treats limitation of liability. You get on Tuesday the comparison you were going to build by hand on Friday night.
• Translating policy into plain English. The note explaining what the retention policy actually means for someone in sales, written so that someone in sales will read it.

In every one of those cases the output is a first pass. You are reading faster, not reading less, and the second half of the sentence is the part that gets forgotten.

The plausible wrong answer is the one that gets through

The techniques that reduce this are real, they are Anthropic's own, and they are set out in the piece on why these systems get things wrong, which is the better place to learn them. What matters here is the caveat attached to them: even the published guidance stops short of saying the problem is solved, and asks you to validate anything critical yourself.

Length works against you too. Accuracy and recall degrade as the volume of text grows, which is how a context window behaves as it fills, and the ninety-page contract therefore deserves more scepticism than the four-page one, not less.

Before you write your function's rules, read the specific tasks these tools handle badly. Then read the situations where the answer is not to use one at all, because your function will meet more of them than most.

Confidentiality, which this article cannot settle

Whether you may put a particular document into a particular tool depends on four things at once, and they were not written to agree with each other. There is your organisation's contract with the provider, which governs what happens to the text after you paste it in. There are your own professional obligations, which do not soften because a tool is convenient. There are the client engagement terms you signed, some of which now carry an AI clause nobody has reread since it was negotiated. And there is the data protection position your organisation has taken, which may be a considered document or a paragraph somebody wrote in a bad week.

Those four rarely live in one place and are usually owned by four different people, which is the part I see teams underestimate. Everyone assumes somebody else has reconciled them. Nobody has, so the working answer becomes whatever the most confident person in the room believes, and that person is often not in your function. I cannot resolve any of it from here, and you should be wary of any article claiming it can. Nothing here is legal advice.

What I would do is put the question in writing to whoever owns the provider contract, name the other three documents in the same email so they cannot be answered separately, and do it before the tool is in daily use rather than after. An answer more conservative than you hoped is still an answer you can point at.

First pass, never the final word

Held to that boundary, this is one of the better uses of the technology in a professional services setting. Released from it, it is a professional risk with a friendly interface, and the friendliness is the part that does the damage.