The impact of AI in the legal sector

Ask lawyers, and you get one answer. LexisNexis found 94% of UK legal professionals now use AI for legal work, 74% at least weekly and 34% every day (LexisNexis, 2 September 2026, vendor sourced, 543 respondents). Clio's UK and Ireland report found 89% use AI in some capacity and 70% started within the past year (Clio, 7 April 2026, vendor sourced). Thomson Reuters puts weekly use among professionals worldwide at 74% (vendor sourced).

Ask firms, and you get a different answer. LexisNexis polled 543 lawyers in July 2026 and found only 29% describe AI as "embedded" in their firm's strategy and operations, 43% say their firm is "experimenting but progress is slow", and 18% report no clear owner of AI governance (Legal Futures, 14 October 2026, vendor sourced). Clio's UK and Ireland report found only 27% of firms have embedded AI widely across the organisation (Clio, 7 April 2026, vendor sourced).

Both answers are true. The gap between them is the shadow AI problem, and the sharpest number on it comes from Access Legal's study published in May: nearly 60% of fee earners in small and mid-sized UK firms admit to using unapproved tools such as free ChatGPT, while 68% of firm leaders are confident they have full visibility and zero risk of unapproved AI being used for client work (Legal Futures, 30 July 2026, vendor sourced). Thomson Reuters found that a third of lawyers, accountants and compliance professionals use AI their organisation has not approved, rising to 41% among those who say their organisation is moving too slowly (Thomson Reuters, 22 June 2026, vendor sourced, 1,800 professionals in 62 countries).

Half of the professionals Access Legal surveyed would welcome AI built into their case management system. 25% of firms have it.

The uses are unglamorous. Legal research 69%, document summarisation 62%, drafting 53%, document review 53% (LexisNexis, 2 September 2026, vendor sourced). Richard Susskind's advice to small firms in September was to start with the unglamorous: recording and summarising meetings, marketing material, "undertaking basic admin, responding to emails". He also reports that some firms get their best results by giving staff a general purpose tool and telling them to "just use it for half an hour a day" (Clio, 17 September 2026).

What it saves is less clear than the vendors suggest. We have not found an independent, non-vendor measurement of hours saved by AI in small firms. We will add a figure here when an independent one exists.

What firms lose to the systems they already have is better documented. Access Legal's research puts it at 4.16 billable hours a week per fee earner lost to "ill-functioning technology", "well over half a day" (Legal Futures, 30 July 2026, vendor sourced). In the Bellwether survey of small and mid sized firms, 52% named administrative tasks as their biggest workflow issue, followed by case management at 41%, and only 9% rated their firm excellent on overheads (LexisNexis, 18 May 2026, vendor sourced). Small firms are losing at least as much time to broken tooling and admin as they are saving with AI.

The independent number underneath all of this is utilisation. The Hazlewoods benchmarking survey for the Law Society's Leadership and Management Section (121 firms, 2024/25 results, published April 2026) puts the median fee earner at 807 chargeable hours a year, "well below the generally accepted target of at least 1,000 to 1,100", at £147.33 an hour against a cost of £131.39, with lock up of 134 days (Legal Futures, 16 April 2026, independent). Half an hour a day of recovered chargeable time is worth about £16,000 a year per fee earner at that median rate, if it is billed and paid. That is many times what a typical case management licence costs per user. That is the business case, and it has nothing to do with which AI tool you buy.

Everything above comes from asking lawyers what they do. The AI companies now publish what people actually do, measured from the tools themselves. It is the companies' own research, so it sits in its own category here, and we are an Anthropic partner, which is one more reason to label it. But the newest Anthropic release has UK rows, and they say something the surveys cannot.

In May 2026, 1.1% of UK conversations on Claude.ai involved a task that belongs to a legal occupation: lawyers' work 0.77%, paralegals' 0.15%, arbitrators and mediators 0.08%, judicial law clerks 0.05%, and the rest smaller still. That is the share of conversations, not of people, and it does not say who was typing. The use case does. Of those UK legal conversations, 38% were work, 45% were personal and 17% were coursework (Anthropic Economic Index, June 2026 release, April and May 2026 data, platform data, Anthropic). Fewer than four in ten UK legal conversations with Claude are work. The rest are the public and students asking legal questions of a chatbot. Globally the work share is 60%, in the US 52%. Of the comparisons we ran (the US, Ireland and the global average), the UK has the highest public share of legal AI use. Richard Susskind has called this the latent legal market for thirty years. Here it is, counted.

The same release shows how UK legal conversations as a whole use it. UK legal conversations are 73% augmentation and 27% automation, against 65% and 35% globally. The commonest pattern is task iteration at 44%, then learning at 23%, and only 17% is a straight instruction to do the job. What comes out is documents and reports (44%), emails and messages (13%), analysis and summaries (13%) and explanations (15%). Most of it is drafting with the tool, not handing work to it. The API data, which is organisations building AI into their own systems rather than people chatting, is the mirror image: legal tasks there are 85% automation and 81% directive, and the commonest output is a summary. That is what embedded AI in a case management system looks like, and it is global data only.

Two earlier Anthropic findings fit the same picture. Its March 2026 labour market report put legal occupations at 89% of tasks AI could do in theory against 20.4% seen in actual use (Anthropic, 5 March 2026; the two figures as reported by Euronews, 14 March 2026). Its April survey of 81,000 users found legal and scientific professions reporting the mildest productivity gains, with some lawyers' complaint being precision: given exact rules for reading a document, the tool "diverges every time" (Anthropic, 22 April 2026). OpenAI's first "Work at the Frontier" report in July found that, once general tasks like writing and summarising are set aside, 56% of US legal workers' remaining work messages concern tasks from other occupations, not law (OpenAI, 27 July 2026, US users). OpenAI's own data release, Signals, has no legal breakdown for the UK, but its US figures agree with Anthropic's: the strictly legal work activities (researching law, advising on legal matters, preparing legal documents, consulting legal materials) are about 1.3% of US work messages in June 2026, and "advise others on legal or regulatory matters" has doubled since July 2024 (OpenAI Signals data download, v2.0, data through June 2026, platform data, consumer accounts only; since superseded by v3.0). All of it says the same thing: the tools are good at the first draft and the summary, poor at the rule you gave them, and used far more around the legal work than on it.

This is the biggest change of the last eighteen months. AI governance in a law firm is now a supervision and compliance matter, not an IT preference.

The courts went first. In Ayinde v Haringey and Al-Haroun v Qatar National Bank, decided together on 6 June 2025, the Divisional Court dealt with five nonexistent authorities put forward by a pupil barrister and 18 fictitious cases out of 45 cited by a solicitor who had relied on his client's AI research. Wasted costs had already been ordered at an earlier hearing in Ayinde; the Divisional Court referred both lawyers to their regulators ([2025] EWHC 1383 (Admin)). In November 2025, the Upper Tribunal in UK v Secretary of State for the Home Department [2026] UKUT 81 (IAC) said that to put client letters and Home Office decision letters into an open tool such as ChatGPT is to place them in the public domain, and so to breach client confidentiality and waive legal privilege (Upper Tribunal, 17 November 2025). In May 2026, in Cork v Smith [2026] EWHC 1199 (Ch), a junior solicitor at Pinsent Masons used AI to draft two misleading letters to the High Court, quoting text attributed to the Insolvency Rules which, the judge found, the rule "says nothing of the sort". The judge said the junior had "almost entirely outsourced the thinking process to the program" and that "there appears to have been a failure to supervise". The firm self-reported to the SRA (Cork v Smith, 22 May 2026; Legal Futures, 26 May 2026).

Then the SRA. It added AI sections to its guidance on effective supervision in June 2026. On 17 August 2026 it published a warning notice on misuse of AI, having received 42 reports of possible misuse between July 2025 and July 2026. The notice ties AI failures to the Principles and to named paragraphs of both Codes of Conduct, says that "reliance on an output of AI would not be a suitable defence", that putting client information into open tools "will likely breach client confidentiality", and that supervisors "remain accountable for the work carried out through them" (SRA warning notice, 17 August 2026; Legal Futures, 18 August 2026). On 8 October it opened a consultation to write technology and AI into the Statement of Solicitor Competence, in three places: keeping up with developments in technology and AI, research and drafting, and supervising work done by others using AI. It closes on 3 December, with the final statement due in April 2027 and in force from September 2027 (SRA consultation). Competent practice will shortly include competent use of AI, by definition.

The scale is now measurable. Damien Charlotin's database of court decisions involving fabricated AI output stood at 2,202 on 19 October 2026: 1,501 in the United States, 234 in Canada, 112 in Australia, 76 in the United Kingdom, 6 in Ireland. Lawyers were responsible for 872 and litigants in person for 1,267 (damiencharlotin.com, independent, updated 19 October 2026).

Two things follow for a small firm. First, the question an insurer, a client or the SRA will ask after a problem is no longer "did you use AI?" but "what did you have in place?". Second, the SRA's warning notice says what it expects to find: supervision that covers AI assisted work, proper systems and controls, and checks that every authority cited is real. Our one page policy post and the court decisions post cover both.

UK and Irish firms are already mixing in fixed fees. Clio's UK and Ireland data has fixed or flat fees at 53% of matters and hourly billing at 32% (Clio, 7 April 2026, vendor sourced). HSBC's October survey of over 100 senior decision makers at firms with turnover of £18m to £1bn found clients "calling most frequently for fixed fees, according to 68% of law firm leaders" (Legal Futures, 15 October 2026). The share of UK lawyers expecting law firms to change how they bill because of AI rose from 40% in January 2025 to 55% in September 2026 (LexisNexis, vendor sourced).

But AI itself has barely touched a price list. 86% of solo firms and 78% of small firms "have made no pricing changes" for AI, and fewer than 33% have increased revenue with it (Clio, 4 May 2026, vendor sourced, US survey). So far, the saving is going one of two places: into unrecorded time on hourly matters, or into margin on fixed fee ones. Firms paid per matter are ahead without having done anything.

The competition is starting to price it in for them. Arceus, a venture backed US firm that pairs lawyers with AI agents, raised $17m in October on "a flat price that is quoted before a matter begins" for commercial contract work (SiliconANGLE, 1 October 2026). It will not take a complex matter from a good firm. It does set a price the client now has in mind.

Our post on pricing AI assisted work has the practical answer: price on value, not on your new lower cost, say in the engagement letter how you use AI, and move repeatable work to fixed fees per stage.

The practical choice for a firm under 50 fee earners is a cloud platform with AI bundled, a legacy UK system with AI at the edges, or Microsoft 365, plus whatever people use at home. In our experience, most small firms sit in the third camp, which is what the shadow AI figures above describe. For mid-market firms, the problem is different: 40% cite integration difficulties with new tools, against 23% in smaller practices (Clio, 7 April 2026, vendor sourced).

What small firms say they want is independent help choosing. The Law Society's Small Firm Manifesto of 14 October asks government for SME technology adoption programmes providing "independent technology assessments, support and targeted financial assistance for essential legal technology tools", and says the barriers to adopting legal technology are "as much about knowledge and confidence as cost" (Legal Futures, 14 October 2026, independent). That reflects the conversations we're hearing. The Law Society said it first.

Recruitment is the problem small firms name first. 85% of small firms "saw recruitment as their biggest challenge" in the Law Society's research (Legal Futures, 14 October 2026, independent). The two are linked. 22% of UK lawyers would consider leaving an organisation that does not embrace AI, up from 16% a year earlier, and 46% say their career would be negatively affected at one, up from 28% (LexisNexis, 2 September 2026, vendor sourced).

The junior question is the one with no neat answer. Judgement comes from doing the work, and the work being automated first is the work juniors learn on. Susskind's own view, in the 30th anniversary edition of The Future of Law, is that "the judgement objection falls away" once the question is whether machines can handle uncertainty, and that lawyers face two strategies: compete with AI, or build it (Legal Futures, 28 September 2026). Our post on how juniors learn when AI does the junior work sets out what good firms are doing now.

Clients are complaining more, and about the same two things. The Legal Ombudsman received 14,259 complaints in 2025/26, up 37% on the year before, and accepted 8,412. Residential conveyancing was 36% of those accepted, wills and probate 14%, personal injury 14%. Poor communication was 24% of complaint types and delay 22%, nearly half between them (Legal Ombudsman, 2025/26 annual data, independent). Both are workflow problems before they are legal ones, and both are what a client portal and an automatic update exist to fix.

The firm numbers tell the rest. There were 8,899 solicitor firms in England and Wales in September 2026, down from 10,885 in July 2010. Sole practitioners fell from 4,030 to 1,306 over the same period (SRA regulated population statistics, independent). The profession is consolidating into fewer firms, and the ones that make the move are the ones that run as businesses.

The compliance officer split is approved, and paused. The Legal Services Board approved SRA rules on 17 August stopping anyone who can make significant management decisions on their own from also being COLP or COFA, in firms above £600,000 turnover or £2m client money. The SRA estimated around 1,660 firms, 18% of the total, and 431 sole owner manager firms would need to act, from January 2027 (Legal Futures, 17 August 2026). At the end of September the SRA paused implementation. "We are pausing to take stock," said executive director Aileen Armstrong, citing concerns "particularly by smaller firms". It will consult on options "over the coming weeks". No new date has been given, and the rules have not been withdrawn (Today's Conveyancer, 5 October 2026). The SME and Boutique Law Firm Alliance, formed to fight the change, still wants a vote of no confidence in the SRA board and understands the Law Society council may revisit the question in December (Legal Futures, 16 October 2026).

April 2027: new accountants' report rules, approved by the LSB in the same decision and not mentioned in the pause.

September 2027: AI in the competence statement, if the SRA's consultation lands as drafted.

The longer view. Susskind's position in the 30th anniversary edition of The Future of Law, published in September, is that in the short term AI in law means "automating, improving and optimising traditional legal practice inside the existing business model", and that the generative AI "we rightly use today with caution will be reliable for most practical purposes" by the early 2030s (Legal Futures, 28 September 2026). His record over thirty years is right on direction and early on dates. For small firms specifically he now says they "can move ahead more quickly" while large firms suffer "supertanker syndrome: hard to change direction, hard to slow down, hard to stop" (Clio, 17 September 2026). That is a reversal of a decade of pessimism about the high street, and it is the most useful thing he has said for a firm of 20.

Six things, in order. None of them is "buy an AI tool".

Everything above is independent of which vendor you use, and most of it costs time rather than money. If you want a baseline before you start, that is what our Assessment is for. We do not sell or resell software, which is the only way to give the independent answer small firms told the SRA they wanted.