AI in mid-sized law firms: the managing partner's briefing

In January 2026 the Law Society published research it ran with MHA, surveying senior leaders at firms with eight to 50 partners across England and Wales. The headline finding was that 65% of mid-sized firms now use AI to drive productivity and growth (Strategic sector insights for the legal profession in 2026: mid-sized firms).

This briefing is for the partner who has to decide what that number means for their own firm. It covers what peer firms are using AI for and what your confidentiality duties require in practice, then sets out the order we’d move in. It ends with a pricing question a third of the surveyed firms haven’t asked yet. We give no legal advice and hold no legal accreditation; this is the technology and governance side, written for the person who has to sign it off.

Are we behind?

Probably not as far as the headline suggests, and the honest reading cuts both ways. The survey counts use, not depth, so a firm with one practice group on a paid research tool answers yes alongside a firm with AI running through every matter type. Being inside the 65% tells you little about whether the use is any good.

The direction of spend is harder to argue with. The same research found 94% of leaders believe the technology will boost productivity, and 65% of firms plan major IT upgrades in 2026, up from 48% a year earlier. The gap worth worrying about separates firms that govern and measure their AI use from firms where use just happens, and that gap doesn’t show up in an adoption percentage.

What firms your size are using AI for

The Law Society’s report names contract analysis and legal research as the common deployments, with predictive modelling appearing as firms push further. It also records cyber security rising as a concern in step with adoption, which is the survey’s way of saying firms know the data question is live.

Our own legal work sits in that first category. For BAiSICS we built a commercial lease review platform that took the first read of a lease from around two hours by hand to roughly ten minutes, at 96% extraction accuracy against historic leases senior partners had already marked up. GPT-4 and the leading legal-AI products missed the firms’ own accuracy bar on the same test. Every extracted field cites the lease line it came from, so the fee-earner checks the citation and makes the call, and the platform runs inside an AWS UK region. Its users save over £200,000 a year.

The lesson from that build is about where to point AI, and it applies to firms that will never review a lease. Document-heavy review pays back first because every answer can be checked against a document the firm already holds, so verification is part of the workflow rather than a separate research task. Work where the model would have to be trusted on its own authority is where adoption should come last.

What confidentiality requires in practice

The SRA regulates outcomes rather than tools, so the firm stays accountable for AI-assisted work the same way it is for a trainee’s draft (SRA risk outlook on AI in the legal market). In practice the duty comes down to controls rather than tool choice. Matter data stays inside an environment the firm controls, under contract terms that forbid training on your inputs. A qualified person verifies output against its source before anything reaches a client, and the firm keeps a record of where AI touched each matter.

We’ve set out the full position, with the SRA and Law Society sources and the case law on fabricated citations, in how UK law firms adopt AI safely under SRA duties. The short version is that the tool is rarely the risk; the controls around it are what make adoption safe.

The exposure the survey couldn’t measure

The Law Society asked leaders what their firms use. It couldn’t ask what individual fee-earners use on their own accounts, and that’s the gap most firm policies haven’t caught up with. Anyone under time pressure with a consumer AI tool on their phone can paste matter text into it, and a consumer account may retain the prompt or train on it. If the firm hasn’t issued a policy and a sanctioned tool, the default position is that each fee-earner decides privately, matter by matter.

The cost of unverified output is already in the law reports. In June 2025 the Divisional Court dealt with two cases in which fabricated authorities had reached the court, and referred the lawyers involved to their regulators (Ayinde v Haringey and Al-Haroun v Qatar National Bank). A ban doesn’t hold, because the time pressure driving the behaviour doesn’t go away. What holds is a sanctioned tool, a short acceptable-use policy, and an hour of training on what must never be given to a model.

The pricing question a third of firms haven’t asked

One finding in the report has nothing to do with technology and may matter more than any of it. 33% of firms haven’t considered pricing model changes despite AI efficiencies, and 49% still don’t use value-based billing. Set those against the most common partner charge-out rate of £301 to £400 an hour: when AI cuts a two-hour first read to ten minutes and the work is billed by the hour, the saving lands on the client’s fee note rather than the firm’s margin.

We’re not pricing consultants, and the right model differs by practice area. The point of the number is narrower. Adoption and pricing are one decision, and treating them as two is how a firm ends up funding an efficiency its billing model gives away.

The order of moves

None of this needs a transformation programme to start, and buying software is not the first step. The sequence that works runs governance first, then evidence, then build.

1. Close the policy gap. Write the acceptable-use policy, name the sanctioned tools, and train people on what must never leave the firm. It takes weeks rather than months, and it shuts the consumer-tool exposure before anything else happens.
2. Choose the first workflows on evidence. Count where fee-earner hours go and pick the document-heavy work where output can be verified against sources the firm holds. Resist starting with the most impressive demo.
3. Build against a measured baseline. Agree what the current process costs in hours before anything goes live, so six months later the question “did it work” has a numerical answer.

The first two steps are what our AI Audit and Transformation formalises: an independent assessment with no stake in recommending a build, run inside the firm within four weeks, ending in a costed and prioritised 12 month roadmap that names who owns each workflow. For how we handle legal work specifically, from privilege through to UK data residency, see AI for law firms and legal teams.