AI strategy for law firms: a set of decisions, not a list of tools

An AI strategy for a law firm is a short set of decisions: which uses of AI you will explore, test, defer or reject; in what order; who owns each one; and what would make you change the plan. If what you have is a list of tools, licences and pilots, that's activity. It's a useful start. It isn't a strategy yet.

Few firms are short of AI activity now. Someone in the property team is drafting with ChatGPT. There's a pilot running in one practice group. IT has rolled out a batch of licences. A partner has seen a demo and wants one of those too.

None of that is wrong. The problem is that it often gets mistaken for a plan.

What is an AI strategy for a law firm, and what isn't it?

Let's start with what it isn't.

• It isn't a tool list. "We use Copilot and we're trialling a legal platform" describes your software estate, not your direction.
• It isn't an AI policy. A policy sets rules for use. It doesn't tell you where AI should be used first, or at all.
• It isn't a vendor's roadmap. A supplier's 12-month plan will, quite reasonably, put their product at the centre. Yours should start with your work.
• It isn't a training programme. Training has a place. But if nobody has decided what people are being trained to do, it can turn into innovation theatre with slides.

What it is: a record of decisions that a firm's leadership has actually made, and can defend. It should be short enough that a partner will read it, and specific enough that someone can act on it on Monday.

The gap between use and strategy shows up in the research. LexisNexis's UK survey (The AI culture clash, 1 September 2025) found that 61% of lawyers were using generative AI in their work, but only 17% said AI was fully embedded in their firm's strategy and operations. Two-thirds described their organisation's AI culture as slow or non-existent. Separately, City AM reported in July 2025 that only 24% of UK legal professionals said their firm had an AI strategy, citing Thomson Reuters research.

Both surveys are now more than a year old and both come from organisations that sell legal technology, so treat them as a signal rather than a precise measurement. The direction is hard to argue with, though: plenty of use, much less decision.

Why does more AI activity make the decision harder?

You might expect a firm with lots of AI activity to be further ahead. Sometimes it is. Often it just has more to untangle.

Think of it as a ladder:

1. First drafts with ChatGPT. Useful enough to create interest. Unstructured enough to create questions about what's being pasted where.
2. A platform pilot in one practice group. A real experiment, but not yet a shared decision about what comes next.
3. Copilot licences nobody's measuring. The tools are moving faster than the basis for deciding whether they matter.
4. Several teams building their own thing. More activity can mean more competing assumptions, not more direction.

Each rung adds sunk cost, internal sponsors and expectations. By the fourth, stopping anything feels political, and starting anything new feels like it should wait for the last thing to finish. That's when firms most need a strategy, and when it's hardest to write one, because every decision now has someone attached to the other answer.

If you're still on the first rung, our practical guide to AI implementation covers getting started. This article is for the firms further up.

What decisions does a law firm AI strategy have to make?

We'd suggest six. You can phrase them differently, but if any of them is missing, the strategy probably has a hole in it.

1. Which work are we trying to improve? Start with the problem, not the tool. Name the work: first drafts of reports on title, pitch preparation, matter opening, first-pass review of disclosure. "Productivity" isn't a piece of work.

2. For each candidate, explore, test, defer or reject? This is the heart of it. Explore means find out more. Test means run something bounded with an end date. Defer means not now, and here's what would bring it back. Reject means no, and here's why. The reject list is the one firms most often skip, and it's the one that saves the most time later.

3. In what order? Twelve months is short. Two or three tests done properly will teach you more than eight started at once.

4. Who owns each decision? One name per line, not a committee.

5. Where are the boundaries? What client data can and can't go where, which work always needs human review, and which matters are off-limits for now. Your professional obligations on confidentiality and supervision don't pause for a pilot. This article isn't legal advice, and boundaries should be agreed with whoever holds compliance responsibility at your firm.

6. What would change the plan? Name the evidence that would make you scale something, stop something, or bring a deferred idea forward. Without this, the strategy is a snapshot that's out of date by spring.

Notice that none of these decisions names a product. Products come in at decision two, as candidates, and they're judged on the work they improve.

Our working view: when AI comes up at a partners' meeting, the questions are rarely hostile. Partners aren't anti-technology. They're anti-vague. A strategy built on the six decisions gives them something concrete to agree or disagree with.

How should a firm prioritise AI opportunities?

Once you have a list of candidate uses, you need a way to sort them. We use five lenses:

• Value. If this worked, what would genuinely improve: speed, consistency, client experience, risk, or capacity? Be specific about who would notice.
• Feasibility. Can current tools do this reliably enough on your documents, not just in the demo?
• Risk. What happens when it's wrong? A weak first draft of an internal note is a different risk from a wrong answer in front of a client.
• Effort. What will it take to set up, test, train people and maintain? Most of that effort never appears on a quote.
• Readiness. Do you have the source material, the people with time to test, and a sponsor who will make a call at the end?

Readiness is the one that trips firms up. A use case can score well on value and feasibility and still fail because the precedents are out of date, or the only person who understands the workflow is in court three days a week.

You can score each lens from one to five if that helps. Just don't let the numbers do the thinking. The point of scoring is to get the disagreement into the room early: the partner who rates something a five on value and the ops lead who rates it a one on readiness are both telling you something useful.

What does a 12-month AI sequence look like?

Here's an illustration. This is a hypothetical firm, not a client. It's a composite designed to show the shape of a sequence, not a template to copy.

Hypothetical firm: a regional firm with four practice groups. Some fee earners use public chatbots for first drafts. The commercial property team is trialling a legal AI platform. IT bought Copilot licences for one office last year. Nobody is quite sure who's in charge of any of it.

Months 1–3: decide. Gather what's already happening, including the informal use. Agree interim boundaries on public tools. Run the five lenses across a long list of candidate uses with a small group from leadership, practice, operations and IT. Come out with the six decisions written down: two things to test, a handful to explore, some to defer and at least one to reject. In this example, a client-facing chatbot that answers legal questions goes on the reject list for now, because the risk is high and the firm has no way of reviewing its answers at scale.

Months 4–6: test. Run the two tests. One is the property team's platform trial, now with exit criteria and an end date rather than an open-ended "see how it goes". The other is a contained build for the private client team: what we call an AI brain, meaning a bounded set of instructions, approved material and workflows for one team, tested against real examples of their work. Meanwhile, start measuring the Copilot licences against the work they were meant to help.

Months 7–9: decide again. Each test ends with one of three calls: scale it, fix it, or stop it. Stopping is a legitimate outcome, not a failure. Revisit the deferred list: has anything changed? Decide what happens to the licences based on use, not on the renewal date.

Months 10–12: embed and review. Put ownership and maintenance in place for whatever is scaling. Update the boundaries based on what the tests taught you. Then write next year's decisions, starting from evidence rather than enthusiasm.

Your sequence will look different. A firm with no activity yet would spend longer exploring. A firm with five teams building their own tools would spend more of the first quarter deciding what to stop. What shouldn't change is the rhythm: decide, test, decide again.

This kind of written sequence is what our Legal AI Roadmap produces, alongside a leadership readout.

What are the signs you've got activity rather than strategy?

A quick self-check. If three or more of these sound familiar, you have activity.

• You can't name anything you've decided not to do. No reject list, no deferred list.
• Pilots have sponsors but no end dates. They just carry on, quietly.
• Licence count is the success measure. "We've rolled out to 200 people" says nothing about whether the work improved.
• Nobody can say who decides whether a pilot continues. Or three people think they do.
• The strategy document names more tools than pieces of work.
• Every new model release reopens the whole plan. If your direction depends on which model launched last week, it's built on the tool rather than the work.
• It lives in one person's head. If your innovation lead left tomorrow, would anyone know what had been decided?

None of these is a criticism of the people involved. They're what happens when activity grows faster than the decisions around it. That's normal. It just needs catching.

What to do next

Try the one-page decision test. One person can run it in an afternoon.

1. List every AI activity you know about across the firm: tools, licences, pilots, informal use, things people have asked for. Aim for completeness, not detail.
2. Next to each one, write three things: the decision (explore, test, defer or reject), the owner (one name), and what would change that decision.
3. Count the blanks.

The blanks are your strategy gap. Take the page to your next leadership or partners' meeting and ask for decisions on the blanks only. Don't ask for approval of the whole page. Don't present tools. Ask for decisions.

If the meeting fills in most of the blanks, you're closer to a strategy than you thought. If it can't, that tells you something too, and it's far cheaper to learn it now than after another round of licences.