AI in UK Legal Services

Why this briefing

The SRA received 42 reports of potential AI misuse between July 2025 and July 2026 and has investigations open. No other sector in this series has a number like that, because no other sector has a forum that catches the error in public and on the record.

Read what the courts actually referred. The Divisional Court referred a barrister to the Bar Standards Board and included in the matters for consideration her supervision during pupillage, how work was allocated to her, and how her competence was assessed. It referred the instructing solicitor to the SRA over the steps he took to satisfy himself she had sufficient experience and competence. It referred a second solicitor for relying on a lay client for legal research. The Upper Tribunal referred a solicitor to the SRA for supervisory failures. Four separate findings, one common factor: nobody checked the work. The technology is how the unchecked work got there. It is not why it went unchecked.

That matters commercially, because of what is happening on the other side of the firm at the same time. Complaints to the Legal Ombudsman rose 37 per cent in 2025/26, to 14,259. Complaints resolved fell 1 per cent. Redress paid stayed flat. Communication and delay together account for 46 per cent of what consumers complain about. Both pressures meet at the same thin layer of the firm: the junior lawyer whose review work is simultaneously what the tool is bought to replace and what the quality control depends on.

The human moment

Someone selling a house who cannot get an answer. Residential conveyancing is 36 per cent of all accepted complaints and rose 41 per cent in a year, and poor service was evidenced in 78 per cent of investigated conveyancing complaints. A quarter of consumers received no final response from their firm at all. The complaint is almost never about the law. It is about silence. This is the person the efficiency case is supposed to be for, and the one nobody in the vendor material is talking about.

The finding that surprised the research

One firm has tested models against English law questions and published the marks. Between October 2023 and February 2025 the best score rose from 4.4 to 6.4 out of ten, and hallucinated authorities fell from 47 of 150 answers to 9 per cent. The tools got substantially better across two model generations. The firm's recommendation did not move: they should not be used for English law legal advice "without expert human supervision."

The same conclusion arrives independently from the rulebook. The Civil Justice Council consulted on whether AI needs its own rules for the documents lawyers file, and reported in June 2026 a high degree of alignment among respondents that no additional formal requirements are currently necessary. An independent judicial advisory body asked the profession whether the gap was in the rules, and the profession said it was not.

What's inside

Sixteen pages, written for a managing partner or COO picking up the firm's AI question:

• A note on the numbers — what is included, what was excluded, and why
• What the regulator has counted, and what its warning notice actually cites
• What the courts referred, and the supervision analogy the Divisional Court used
• The conduct picture, from the best complaints dataset in this series so far
• Where AI actually is — and why the measurement problem is worse here than anywhere
• The privilege and confidentiality point a tribunal has now raised, framed as contested rather than settled
• The threat running in two directions, including AI used against firms' own client due diligence
• Review capacity, and why the workforce argument is made from the Ombudsman data instead
• An action plan aimed at the supervision layer rather than at tool selection

The discipline behind it

Vendor noise is more severe in this sector than in any other covered so far, and the exclusions are correspondingly expensive. Every headline UK legal AI adoption statistic located in the research was vendor material — including one carried on the Law Society's own site as paid partner content — and none of it appears. A trade body's widely repeated figure for the proportion of lawyers adopting AI is excluded because the underlying survey is not named; it would have been the most quotable number in the briefing.

Three date traps were caught in a single sitting, and all three would have flattered the argument. The SRA's only dedicated AI research publication is dated 20 November 2023 and reads as current in search results. A widely quoted law firm benchmark score is from the 2023 version, two model generations old — obtaining the 2025 version changed the finding, and improved it. A major firm's AI trial report is dated June 2024, with the trials themselves run between November 2023 and March 2024.

Where the research could not obtain something, the briefing says so rather than treating it as absent. A trade body's benchmarking survey returned an access error and is recorded as not obtainable, not as evidence of nothing.

Who it's for

A managing partner, a COO or a head of risk in a firm of solicitors in private practice in England and Wales, being asked whether to buy AI, or being asked what the firm's position is. It assumes legal practice is well understood and AI is not.