AI for lawyers: what actually works and what is hype

In 2026 the question is no longer whether your firm will use artificial intelligence, but for what. The commercial noise hasn’t died down, but the accumulated experience has grown: after a couple of years of trial and error, it’s now possible to say with reasonable precision which tasks AI handles well in a law firm — and which remain a brochure promise.

On the principles, we already wrote a few weeks ago: AI is a lever, not a substitute, and the lawyer always signs. This article is the other half of that conversation, the practical one. A concrete list of what works today, what doesn’t, and the legal framework worth having in front of you before you sign up for anything.

What actually works today

The uses that deliver value share a pattern: they work on material the firm already holds, they produce output that can be verified in minutes, and they leave the last word to the lawyer. On that condition, there are four tasks where the saving is real.

Summarising and extracting data from long documents. A land registry extract with three mortgages and an attachment entry, a fifty-page deed, a dense contract in a language that isn’t the client’s. AI locates and structures names, dates, charges, references and specific clauses, and produces a summary that guides your reading of the original. The lawyer checks against the source document, which is still right there; if an error slips in, it’s cheap to catch.

First drafts of routine communications. The monthly status email, the standard formal demand, the letter that goes out with an invoice. These are texts with a known structure where AI produces a decent starting point in seconds. It doesn’t replace your drafting: it brings it forward. The version that reaches the client goes out corrected and signed by you.

Searching your own matter history in natural language. Asking “what did we agree with this client about the payment schedule?” and getting an answer drawn from the matter’s documents and communications is faster than remembering which folder, email or note it lived in. The key is the perimeter: the search runs over what your firm controls, not over the internet.

Transcribing and structuring notes. Notes taken in a meeting, or dictated straight after a call, become an ordered entry: what was discussed, what was agreed, what tasks remain. It’s the difference between a sticky note that gets lost and a record anyone on the team can read six months later.

None of this is spectacular, and that is precisely the sign that it works. These are mechanical, bounded, verifiable tasks. The spectacular stuff, in legal AI, tends to be the other kind.

What is still hype

The “AI lawyer”. Every season a tool appears that presents itself as an associate who never sleeps. It isn’t one. A language model doesn’t weigh conflicting interests, doesn’t judge whether to litigate or settle, doesn’t know the judge or the client, and answers for nothing professionally. Legal judgment is not a feature awaiting the next release: it is exactly what the client is buying when they hire you.

Systems that promise outcomes or cite case law without a verifiable source. Courts in the United States have already sanctioned lawyers for filing briefs containing judgments the AI made up, and judges and bar bodies in Europe have issued warnings along the same lines. The practical rule is simple: a citation you cannot open and read in an official source does not exist until proven otherwise. A tool that doesn’t link its sources is asking you for faith, and faith is not something you can plead.

Vendors who won’t answer where your data goes. If, when you ask where documents are hosted, whether they’re used to train third-party models, and what data-processing agreement is on offer, the answer is vague or evasive, the evaluation is over. A law firm handles exceptionally sensitive data; vendor opacity isn’t a technical detail, it’s a professional risk that lands on you.

The legal framework is no longer optional

Using AI in a firm doesn’t happen in a regulatory vacuum, and in 2026 there are three pieces worth being clear about.

The first is the GDPR. Matter documents contain personal data, so the usual principles apply: data minimisation — don’t upload more to a tool than you need to — and a data-processing agreement with any provider that processes data on the firm’s behalf. Without that contract, there is no legitimate use to speak of.

The second is professional secrecy. What the client confides in you is protected by your duty of confidence, and that duty doesn’t evaporate because the intermediary is software. Before feeding matter information into a tool, the question is the same as with any third party: who can see it, and under what safeguards?

The third is the EU AI Act — Regulation (EU) 2024/1689 — in force since 2024, with obligations phasing in over time. For a law firm, the immediate effect is less about complying directly and more about demanding answers: a serious vendor should be able to explain how the Regulation affects them, which transparency obligations they take on and how they meet them. If they can’t answer, go back to the previous paragraph.

How to adopt it without tripping up

You don’t need a digital transformation plan. You need three modest decisions.

• Start with low-risk internal tasks. Summarising a document you were going to read anyway, tidying up meeting notes. Nothing that leaves the firm without passing through you.
• Human review always, no exceptions. Not as an abstract principle but as a workflow: AI proposes, the lawyer disposes. The day something goes out unreviewed “because it’s always right” is the day of the unpleasant surprise.
• Measure one workflow. Pick a single concrete task — summarising registry extracts, drafting status emails — and compare the time before and after over a few weeks. You don’t need industry statistics: you need your own stopwatch. If it doesn’t save time in your firm, it doesn’t save time, whatever the sales demo says.

AI in Mandato

In Mandato, artificial intelligence follows exactly this standard: it’s applied to concrete tasks, not to promises. It extracts data from and summarises documents inside the matter — the deed the client uploaded, the registry extract on the file — and everything it produces is presented as a proposal the lawyer reviews before accepting. As for where the data lives and under what safeguards, the answer is written down and published on our security page, which is how it should be with any vendor.

The practical conclusion of 2026 isn’t exciting, and that’s why it’s useful: the AI that works in a law firm is the kind that does boring tasks on material you control, under the eye of a lawyer who signs. Everything else, until it shows its sources and its contracts, is still hype.