AI for Law Firms in the UK: How to Evaluate It Properly

Start with the problem, not the shortlist

The usual evaluation runs backwards. A partner sees a competitor's press release, someone books demos for the three best-known products, and the firm picks whichever confuses people the least. Twelve months later the licences renew, and nobody can say what changed.

Run it the other way. Before you see a single product, write down the three places in your firm that time goes to die: work that is necessary, repetitive and hated. First drafts of standard documents. Bundling and summarising. Chasing information that already exists somewhere in the firm. Then ask each vendor to show you their product doing that work, on documents like yours. If they can't map the product to one of your named problems within the first 10 minutes, the rest of the demo is theatre.

Know what you're actually buying

Legal AI in 2026 comes in three layers, and knowing which one you're looking at explains most of the pricing.

• Raw models. Claude, ChatGPT, Gemini and Copilot, bought directly from the companies that make them. Cheapest per seat, most general, and now capable of a great deal of legal work if someone in the firm learns to drive them properly.
• Legal platforms. Harvey, Legora and their rivals. These sit on top of the same models and add legal workflows, security controls, document-system integrations and an interface lawyers recognise. You're paying a premium for packaging and trust, and whether that premium is justified is the whole evaluation.
• Point solutions. Single-job tools for disclosure review, contract analysis, dictation and the rest. Judge them like any other software purchase: on the job, not on the AI.

The platforms deserve one extra note, because the market is moving underneath them. Most are what the industry calls wrappers: products built on models they don't own, as I set out in Is Legora just a wrapper? Wrapper status doesn't disqualify a product, but it should sharpen your questions about supplier risk, because the model makers can absorb a wrapper's features, and in Harvey's case largely have.

Minimum seats, and the small-firm question

One question comes up constantly from smaller firms: is there decent law firm AI without minimum seat commitments? The enterprise platforms mostly price per seat with minimums attached, and the economics that make sense at two hundred lawyers make none at twelve.

The honest answer is that the raw models are the small firm's friend. A business-tier Claude or ChatGPT subscription, a written usage policy and a fortnight of deliberate practice will take a five-partner firm further than most platform pilots, at a fraction of the cost. Start there, learn what the technology is good at on your own matters, and let the platforms earn their premium later, if they ever do.

Six questions that expose a weak product

Whatever you're shown, these travel well:

• Which underlying models does this run on, and what happens to the product when those models change?
• What does it do that the raw model, driven well, does not?
• Where is our client data processed and stored, and is any of it used to train anything?
• What does it cost against buying the same capability directly from the model maker?
• Can we start with five users, or is there a minimum commitment?
• If we leave in two years, what do we take with us?

A good vendor answers all six without flinching. The ones that reach for the deck instead are telling you something.

The UK homework: SRA, confidentiality, privilege

Nothing about AI suspends your existing duties. The SRA hasn't produced an AI rulebook because it doesn't need one: competence, confidentiality, supervision and acting in the client's best interests already cover the ground. Responsibility for AI-assisted output sits with the solicitor who signs it, and the courts have shown no patience with made-up citations.

Three practical checks before anything touches client data. First, the data processing terms: where the servers sit, whether UK GDPR transfer rules are met, and an explicit commitment that your data isn't used for training. Second, privilege: material pasted into a consumer-grade tool on someone's personal account constitutes a confidentiality incident waiting for a file number, so give people a sanctioned route before they invent their own. Third, your engagement terms: decide whether and how you tell clients that AI is used on their matters, because some institutional clients now ask first.

What works once it's in

The pattern I set out in AI in law firms: what actually works still holds. The AI that delivers improves a workflow the firm already has, runs on data the firm already generates, and puts its output in front of someone who can act on it. Document work has matured into genuine value. The most underused ground remains client-facing data: calls, complaints, feedback, the record of what clients actually experience. Almost nobody analyses it, which is exactly why it's an opportunity.

The licence is not the edge

Evaluate carefully, buy what fits, and then remember what the purchase can't do. Every firm that licenses the platform you choose gets the same capability; a shared tool keeps you level, and level is all it keeps you. I've made the longer argument in AI won't differentiate your law firm, but the short version fits in a sentence: the edge comes from what your firm understands about its clients and builds for itself, on top of whatever everyone else is renting.

So run the evaluation, and run it hard. Just don't confuse passing it with strategy. The tool you pick this year decides your costs. What you build around it decides your position.