Erosion or migration: how AI might rewrite the law
Before we start. Our arguments are drawn from public sources. We have checked the facts we rely on most, and we grade each source as we use it by how far we think it can be trusted. They were gathered by a writer outside the legal profession, working with an AI. Somewhat ironic given the subject matter. If you know better, tell us. We will correct the text and say so.
1. The client, the machine and the man who signed
A client used AI to invent the law, and the court held his solicitor responsible. In 2025 Hamad Al-Haroun was suing Qatar National Bank and an affiliate for £89.4 million over alleged breaches of a financing agreement. Material put before the High Court in his case cited forty-five legal authorities. Eighteen of them did not exist. Al-Haroun accepted that he had produced the citations himself, using publicly available AI tools among other sources. His solicitor had relied on his client’s research and put it before the court without checking it.
The court accepted the client’s apology and referred the solicitor to his regulator. In June 2025 a Divisional Court of the High Court heard the case together with another in which false citations had reached the court. It accepted that Al-Haroun had not meant to mislead anyone. It then held that “a lawyer is not entitled to rely on their lay client for the accuracy of citations of authority.” The solicitor had already referred himself to the Solicitors Regulation Authority (SRA); the court made its own referral as well. In the companion case, a barrister and the law centre instructing her were each ordered to pay £2,000 and were referred to their regulators.
The court addressed the whole profession, not just the two cases in front of it. It called on the bodies that regulate and lead the profession to take measures, and said: “Those measures must ensure that every individual currently providing legal services within this jurisdiction (whenever and wherever they were qualified to do so) understands and complies with their professional and ethical obligations and their duties to the court if using artificial intelligence.” The judgment is published in full by the judiciary.
Liability flowed past the machine and past the man who used it, to the one who held the professional title. The machine cannot be disciplined. The client had no professional duty to the court. The person who carried the error was the one person in the chain who had not produced it. That outcome looks strange until we ask what a client pays a lawyer for, and then it looks like precisely the point.
The impact of artificial intelligence (AI) on the law is unfolding as we speak. We expect it to mirror what has happened in other industries, but because of the stakes, and because of the nature of the law itself, we can only posit the detail of the outcomes. Rules are being written and rewritten. Insurance policies are being drafted and issued, and are likely soon to be contested. The pricing, of process and of answerability, is live and being fought over.
What this shows: one senior court, in one case, placed the whole weight of an AI-generated error on the professional who signed, not on the person who used the tool. What it doesn’t: how common such errors are, or what they cost. Al-Haroun ended in a referral, not a damages claim, and nobody’s insurer paid out.
2. The old playbook
AI is the latest round of an old fix: capital meets a cost it resents and buys a technology to move the problem somewhere else. David Harvey calls this the technological fix. When labour becomes expensive or unruly, capital substitutes machinery. The cost falls where the machine is installed, but the contradiction does not go away. It moves: into fewer wages to buy what is produced, new dependencies, new points of failure. The fix never resolves anything. It relocates it.
The large law firm already runs on a fix of this kind. Its economics rest on a pyramid. A small number of equity partners sit at the top. Beneath them sit associates and trainees, billed to clients at several times what they are paid, for the hours spent reviewing documents, working through due diligence and producing first drafts. The margin between the billing rate and the wage funds the partnership. When that base grew expensive, firms reached for cheaper hands: paralegals, then document processing sent offshore. Each round lowered the cost of the base without touching the top.
The tools now arriving are built to do the work at the base. Review, diligence, summarising and first drafts are what legal AI products are sold to do. A&O Shearman, which describes itself as the first firm to deploy generative AI across the whole firm, in late 2022, reports that its lawyers save two to three hours a week on drafting and summarising. That figure comes from the firm and from its software supplier, neither of them neutral, and it is modest. Its significance lies in the direction it points. If a week’s review becomes an afternoon, someone keeps the difference. Whether that is the firm or the client is the key question the rest of this piece follows.
Once the fix has run, capital still pays humans for two things: connection, and carrying risk. Connection is the premium for dealing with a person rather than a process. Risk is the premium for having someone to hold responsible when things go wrong. In most trades the two can be told apart. In law they cannot. The client pays for the relationship because the partner carries the risk, and the relationship is how the client comes to trust the person who carries it.
What this shows: the work AI tools are designed for is the work that funds the pyramid, and the earlier fixes in law followed the same pattern. What it doesn’t: that firms have yet cut the base. As we will see, so far they have not, and the time saved has largely been absorbed.
3. What the signature buys
The payment is not for what they do, or who they are, but the risk they shoulder. Why do you pay a professional to do something? Because they say they have the skills, the qualifications, the experience, the judgement? Or because you want someone on the hook if it goes wrong? Skill plays its part. It is not what you are paying for. You can check. You can test. You can trust. But you can never be absolutely sure. When the interpretation of contestable texts has consequences for your money or your liberty, it is answerability that matters. The credentials might fail. Answerability is what stands behind them.
Economists explained the reason sixty years ago: the client cannot judge what they are buying. Kenneth Arrow, writing about medicine in 1963, argued that professional arrangements exist where the buyer cannot assess the quality of the service. A patient cannot tell good diagnosis from bad; a client cannot tell sound advice from a confident error until it is too late. So the client relies on substitutes. The credential stands in for a skill the client cannot check, and the profession’s rules, its regulator and its compulsory insurance stand behind the credential in case it fails.
This is where the best-known prediction about the professions may have missed its target. In The Future of the Professions (2015), Richard and Daniel Susskind described a grand bargain: society grants the professions exclusive rights to certain work, in return for trusted expertise. They argued that technology would dissolve the bargain by making expertise widely available. If the bargain was for expertise, they are right, and AI ends it. If the bargain was mainly for answerability, AI leaves it standing and strips out the labour beneath it. The rest of the evidence in this piece points to the second reading.
Anthony Giddens gives the signature its place in modern life: it is an access point. We trust systems we cannot inspect, such as banking, medicine and law, and we meet those systems through particular people. The partner is where the legal system acquires a face and a name. The question AI raises is whether trust survives if the face is removed, or whether the system needs a named person at the point of contact for the trust to exist at all.
The Court of Appeal has now written this separation into law. In September 2025 the High Court held, in a case called Mazur, that staff who are not authorised under the Legal Services Act 2007 could not conduct litigation, even under supervision. For six months the everyday model of the litigation firm, with trainees, paralegals and caseworkers doing much of the work, was arguably unlawful. On 31 March 2026 the Court of Appeal, led by the Master of the Rolls, unanimously reversed that ruling. Unauthorised staff may do the work of litigation under supervision, because the authorised person who keeps responsibility is the one who, in law, carries the litigation on. The work can be done by anyone. The responsibility cannot.
The regulator applies the same rule to machines. When the SRA authorised Garfield.Law in May 2025 as the first law firm to deliver regulated services through AI, it set conditions: the system may not propose case law, it may take no step without the client’s approval, and named solicitors remain accountable for everything it produces. In August 2026, after receiving 42 reports of possible AI misuse in a year, the SRA issued a warning notice that put the point in two sentences: “The use of AI does not diminish or transfer your professional responsibilities. You remain accountable for all work and advice delivered to clients, whether or not AI has been used in its preparation.” Commentators read the notice as a restatement of the supervision guidance the SRA issued after Mazur. Paralegal or machine, the answer is the same: someone authorised must answer for it.
The senior judiciary has also said why the machine itself cannot be the one to answer. In July 2026 the UK Jurisdiction Taskforce, chaired by the Master of the Rolls, published a legal statement on liability for AI harms. AI has no legal personality, it said, so liability cannot attach to an AI system. A careless user of AI is likely to be liable for foreseeable harm. And in appropriate circumstances, the duty of care may require a professional to use AI. The answerable human is now exposed from both sides: negligent for trusting the machine without checking, and potentially negligent for not using it at all.
What this shows: the courts, the regulator and the law of negligence all place responsibility on a named, authorised human, and all treat that responsibility as separable from the work. What it doesn’t: that clients pay for answerability rather than labour, or that this settlement will hold. Those are questions for the money and the insurance, which come next.
4. The supervision problem
These rulings may appear to have settled the legal reality of responsibility for AI, but they swerve the economic reality. Once again we meet a contradiction. The courts and the regulator now say the relevant work can be done by anyone, human or machine, so long as an authorised human supervises it. But supervision means different things depending on what is being supervised.
Human juniors fail in human ways, and they learn when to ask for help. They get tired, cut corners and panic before a deadline. But they learn when they are out of their depth, and when to seek supervision. The machine never falls asleep on the job and never stops looking, at least if the looking is framed correctly. But it does not know its limits, and it does not reliably ask for help, even when told to.
If supervision has to be total, the saving shrinks; if it isn’t, answerability becomes a formality. If supervising AI properly means checking every citation and every claim, much of the time saved goes into the checking. If the checking is not done, the saving is kept, but answerability becomes a name on work nobody verified. The human is needed to supervise, but supervision may eat the very savings that justified the machine.
The contradiction has not been resolved. It has moved. Answerability, whatever the rulings say, may be a convenient fiction, even if that answerability is backed by insurance.
What this shows: the legal settlement depends on supervision, and the machine’s way of failing makes supervision harder and more expensive than it is for human staff. Al-Haroun is what an unchecked case looks like, and the SRA’s 42 reports in a year suggest such cases are not rare. What it doesn’t: how much checking firms actually do, or how much of the saving it consumes. Nobody publishes that.
5. The insurance nobody has called on
Behind every solicitor’s signature sits an insurance policy the law requires them to hold. Every firm regulated by the SRA must carry professional indemnity (PI) insurance on terms the regulator sets, the Minimum Terms and Conditions, bought from an insurer that has signed up to them. The minimum cover for any one claim, excluding defence costs, is £3 million for firms set up as companies or limited liability partnerships (LLPs), and £2 million for all others. Brokers describe the terms as very broad, and the exclusions insurers may apply are written into them. This is answerability made concrete: when the signature fails, the policy pays.
The risk AI adds to that policy is covered, but there is no sign that it has been priced. When a solicitor’s AI-assisted work goes wrong, the error is in law the solicitor’s, so the solicitor’s policy responds. Insurers have a term for a risk like this: silent cover, meaning a risk a policy ends up carrying because it neither includes nor excludes it, and so was never written or priced for it. The head of professional indemnity at RSA, one of the UK’s larger commercial insurers, told the trade paper Insurance Post that PI insurers are carrying an unpriced exposure to generative AI. That is one senior underwriter’s view, not a market survey, but nobody we found says otherwise. Underwriters surveyed at the start of 2026 named professional indemnity as the line most exposed to a major AI loss.
The last time this happened, with cyber, the regulator had to rewrite the policy. “Cyber” here carries the insurers’ meaning rather than the everyday one: losses from hacking, data theft, ransomware or computer failure that turned up in claims under ordinary policies, such as property or professional indemnity, which had never mentioned them. That was silent cyber. In 2019 the Prudential Regulation Authority, which supervises insurers, and then Lloyd’s of London pushed insurers to deal with it, by excluding it or covering it explicitly. For solicitors, the SRA told insurers not to add exclusions that might conflict with the minimum terms until the terms themselves were changed. After a long and technical consultation, it added an explicit cyber clause in 2021. Our reading, which no source states directly, is that AI would follow the same path: an insurer cannot simply exclude AI from a solicitor’s compulsory cover without the SRA first changing the terms. The analogy has a weak point, and we should say so: cyber was a new kind of loss, while AI mostly makes old kinds of loss, bad advice and missed errors, faster and larger. That difference may make AI harder to carve out of a policy than cyber was.
Outside the solicitors’ protected market, insurers are already excluding AI. In 2025 the US insurer W. R. Berkley introduced what it called an “absolute” AI exclusion for its directors and officers, errors and omissions and fiduciary liability products, removing cover for any claim connected to the use, deployment or development of AI. Insurance lawyers report that at least three insurers have since applied very broad AI exclusions across several lines, and the insurance law firm Kennedys reports exclusions of silent AI appearing in professional indemnity cover beyond solicitors’ policies. Not everyone thinks this can last: one insurance lawyer argues that because AI mostly amplifies existing risks rather than creating new ones, broad exclusions will be hard to apply.
Where insurers do cover AI explicitly, they insure the machine as a product, not as a professional. Armilla, a Lloyd’s coverholder, sells standalone AI liability cover, including for losses caused by errors and hallucinations. But its trigger is a performance benchmark: a single mistake does not automatically pay out, and cover responds only if the tool performed significantly below what was expected of it. Munich Re’s aiSure, offered since 2018, insures the performance guarantees that AI vendors give their customers. These products can be thin. A broker at Lockton gave the example of a general policy covering $5 million carrying a sublimit of just $25,000 for AI. And they are selective: the Lloyd’s insurer behind Armilla’s policies says it will not cover systems it judges too prone to failure. (Armilla and Munich Re are describing their own products here.)
So the market has split trust in two, and priced each half differently. A solicitor’s indemnity is a promise about judgement and duty, with no threshold below which an error is forgiven. A performance guarantee is a promise about error rates, which tolerates mistakes up to a benchmark. One insures the person who answers. The other insures the product. For now, the insurance market has not moved answerability from the human to the machine. It has split liability between them: product risk to the vendor’s warranty, professional risk to the human’s policy.
And the human’s policy has never been tested on an AI error. We found no reported UK case in which a client sued a solicitor over an AI error and the indemnity insurer paid. That may simply reflect how negligence claims work: most settle privately, and insurers do not publish what they pay. A professional negligence specialist wrote in June 2026 that the route to such claims is now clear. The scale of what could come is visible in older figures: the SRA reported in 2016 that solicitors’ insurers had paid around £2 billion on some 142,000 claims in the ten years to 2014, of which about one in five succeeded. The signature, as we said at the start, is rarely checked against an actual error. Across the whole profession, the AI risk sits inside a promise that, as far as the public record shows, has never been called on.
What this shows: answerability for AI errors currently sits, unpriced, inside the human’s compulsory insurance; the specialist market insures AI as a product, not as a professional; and outside the solicitors’ protected market, insurers are retreating. What it doesn’t: how the first real AI claim will be handled, whether the SRA will change the terms, or what it will do to premiums. The claim that the minimum terms block an AI exclusion is our reading of the cyber precedent, not a stated rule.
6. Following the money
If clients paid for the work, AI should be pushing fees down. So far, it isn’t. This is the most direct test of the thesis. Average hourly rates for associates at the largest US firms reached $798 this year, according to the pricing firm Persuit, up 33% since 2023, with partner rates up 29%. Big US firms raised rates by more than 10% last year, according to Wells Fargo. These figures are American, but the direction in London is the same: profit per equity partner at Linklaters, Clifford Chance and A&O Shearman was around £1 million higher in 2025/26 than a decade earlier, and double-digit profit rises are once again normal across the City.
The time AI saves has so far been absorbed, not passed on. A study commissioned by Harvey, the leading legal AI supplier, found that lawyers using its tools were not working fewer hours; they were moving the time saved into strategy and client work. A vendor-commissioned study is not a neutral witness, but its finding fits the headcount: since January 2023 the average midsized US firm has grown its headcount by more than 8%. A practising US lawyer writing in Bloomberg Law suspected that firms were investing in AI quietly and keeping the gains, pointing to double-digit profit growth against only modest growth in billable demand. Thomson Reuters’ index of US firms shows the other side: in early 2026, expenses rose sharply, partly from AI investment, and productivity, the hours billed per lawyer, fell. That fall could be the first trace of AI shrinking billable time, or simply a slowdown in deal work. The data cannot yet tell us which.
The clients have noticed, and the pricing fight has begun. The clearest example so far is one bank. In early September 2026 the Financial Times reported that Citigroup had begun asking law firms to bid for work and to explain how much they save through AI, and that it expects a new working model within a year. Morgan Stanley’s general counsel told the Financial Times that top lawyers have long been paid on the back of associates billing long hours, which is the pyramid described by one of its biggest customers. Yet Citi’s own figures show its legal spending still rose 9.6% in the first half of 2026. That rise may say more about volume than price: the same figures show demand for legal work rising sharply over the period, so the spending number cannot tell us whether rates have moved. And most in-house lawyers, according to a survey by their professional association, do not know whether the firms they hire use AI on their work.
The price is beginning to split into two parts: work and answerability. The model emerging in the trade press divides fees: fixed or capped prices for clearly bounded work, and premium pricing for high-stakes judgement. Nick Rowles-Davies, founder of Lexolent, a legal finance fund based in London and Dubai, put the logic bluntly to CNBC: “You can’t charge 16 hours for something that takes 16 seconds.” The bounded work is the base of the pyramid, and its price can fall. The premium survives where the stakes, and so the answerability, are highest.
Two explanations fit what we see, and transparency will decide between them. If clients are paying for answerability, the price of bounded work will fall once clients can see it, while fees for high-stakes work hold. If firms are simply keeping the saving because clients cannot see it, transparency will push all fees down. Financial services has run this experiment before. In 2018 European rules forced fund managers to pay for investment research separately from trading. The price became visible, and research budgets fell. By 2024 the UK regulator was consulting on letting firms bundle the payments again. Whether unbundling also cut research coverage is disputed, but it showed that once a bundled price is made visible, it gets repriced. Citi is now trying to do the same to legal fees.
Marx explains why firms are keeping the saving, and why it won’t last. Strip away the software and we are back in a Victorian factory. In Capital Volume 1, in the chapter on relative surplus value, Marx argues that the first capitalist to adopt a new machine still sells its output at the old social value. That value is set by the labour time everyone else still needs. The gap between the old social price and the adopter’s new, lower cost is extra surplus value. The capitalist pockets it at first, but it disappears once the method spreads and the social norm resets. The billable hour is exactly that social norm, priced in human time. A firm whose AI cuts the real time while it still bills at the human rate is pocketing Marx’s extra surplus value. Competition, or a client like Citi forcing transparency, is what resets the norm. It is a theory with real explanatory power, as David Harvey’s work has shown for fifty years, and it explains the work half of the price. Answerability is something else.
Marx’s categories also describe the line the insurers have drawn. Marx distinguished dead labour, the machine, which passes its own value on to what it produces but creates no new value, from living labour, which does. The insurers price AI as dead labour: a product, insured against a performance benchmark. They insure the solicitor as living labour: a professional, answerable for judgement and duty. But the machine is not replacing a lathe. It is replacing thought and service, which sit on the living side of Marx’s line. The insurance market is classifying a substitute for thought as a tool, and that classification decides who pays when it fails.
Follow the two halves of the price and they turn into different things: the work becomes surplus value, and answerability becomes rent. The saving on the work is extra surplus value of the kind Marx described: real, but temporary, competed away as the method spreads and clients see through the old price. Answerability does not behave like that. The partner who signs may have done none of the work, so her return is not a return on labour at all. It is a return on a position that only an authorised person may occupy, because the law reserves the conduct of litigation and other legal work to authorised people, and backs them with compulsory insurance. That is rent in the classical sense: an income from a scarce, legally protected position rather than from effort. Harvey calls rent of this kind monopoly rent. Competition cannot erode it, because the law decides who may compete. The commodification of thought may finally separate the two halves cleanly enough to show which was which. The work’s premium is temporary; the answerability premium lasts as long as the law keeps the machine from answering. The machine is new. The fight over who keeps its saving is not.
And wherever the saving lands, it stops where competition stops. If Citi wins its pricing fight, its legal costs fall. But legal fees are overhead for a bank. Cutting them improves the bank’s cost-to-income ratio, and the bank prices its loans and fees on competition and market power, not on its legal bill. The likeliest destination for the saving is the bank’s shareholders. At each link in the chain the saving passes on only if competition forces it to. It pools where market power sits: with partners today, perhaps with banks tomorrow, and in neither case with the wage earners whose spending the economy runs on.
What this shows: so far, firms are keeping the saving, and the price is starting to separate into work, which can fall, and high-stakes judgement, which holds. Marx gives that split a mechanism: the saving on work is temporary extra surplus value; answerability earns a rent protected by law. What it doesn’t: prove the mechanism. The data so far is consistent with it, not a demonstration of it. Rising fees fit “clients pay for answerability” and “clients can’t see the saving” equally well. Much of the rate and productivity data is American, and the pass-through argument is our reasoning, not measured fact.
7. The justice gap
AI is widening access to everyday law, but not always to someone who will answer for it. Everything so far has been about the top of the market: City firms, bank clients, partners’ profits. The bottom of the market looks different. Here the question is not who keeps a saving, but whether people who could never afford a lawyer get legal help at all, and what kind of help it is.
At one end is cheap help that still comes with someone to answer for it. Garfield.Law, the AI firm the SRA authorised in May 2025, works mostly on small debt claims, the kind people abandon because a solicitor would cost more than the debt. The firm says it has handled more than 600 claims and recovered around £500,000 for clients, almost all settled before a hearing; those are its own figures. On 14 May 2026 one claim went to a contested trial at Wandsworth County Court. A freelance HR consultant, owed £7,000 by a hospitality business, paid Garfield about £400. The firm’s system prepared the witness statements, disclosure and trial bundle. The other side had a firm of solicitors who instructed a barrister. After a three-hour trial she won her claim and defeated the counterclaim against her. Even here, a human stood up in court. The barrister Garfield instructed, Dominic Li, told The Independent that AI can now do the heavy lifting of preparation competently and at a fraction of the usual cost, but “the advocacy that wins a contested trial is still done by a human on their feet in court.” He told the Guardian the same thing more briefly: the advocacy at trial “remained essential and a fundamentally human exercise.” Cheap, AI-prepared and answerable: a regulated firm, named solicitors, an insured policy and a barrister.
At the other end is the same help with nobody to answer for it. Since legal aid was withdrawn from most private family and civil cases in April 2013, far more people go to court without a lawyer. The Law Society says private family cases in which neither side has a lawyer have almost trebled since then, and in 2025 neither party was represented in almost half (47%) of private family cases. Research for the Ministry of Justice in 2014, a small qualitative study and now dated, found a mix of reasons: many people pursuing small sums were content to act for themselves, while others wanted a lawyer, could not afford one, and went ahead alone rather than drop the claim. These are the people now turning to AI.
The courts treat the same mistake very differently depending on whether a professional stands behind it. In December 2025, in D (A Child) (Recusal), the Court of Appeal heard a child arrangements case in which a mother, representing herself, had put documents before the court citing authorities that did not exist; she accepted she had used AI to help prepare them. Lord Justice Baker was sympathetic. It “is entirely understandable” that litigants in person turn to AI, he said, and he accepted she had not meant to mislead the court. But he did not lower the standard: every party, represented or not, owes a duty to the court to ensure that the cases it cites are genuine. Set this beside Al-Haroun, where we began. There, a client’s AI invented the law and a professional carried the consequence. Here there was no professional, so nobody else could carry it. The error sat inside a mother’s own case about her child.
Even where you may act for yourself, the party with the most money at stake buys answerability. In England and Wales there is no legal requirement to use a solicitor or licensed conveyancer to buy or sell a home. In practice only cash buyers can do it themselves. Where there is a mortgage, the lender will almost always insist that a solicitor or licensed conveyancer acts for it, and usually the same professional acts for the buyer. AI could make much of the paperwork trivial, and the lender’s demand would remain. Answerability is bought to protect the lender’s security over the property. The borrower gets it because the lender needs it, not because the borrower does.
AI does not just widen or narrow access. It sorts people by their starting position: sophistication, money and tolerance for loss. Credentials and answerability are the reasons the buyer of a legal service turns to a professional. Doing it yourself, with or without the help of AI, means you carry the risk of acting on poor advice. If you think you can judge its quality, have the time and resources to check it, or can afford to bear the risk if it all goes pear-shaped, this may work for you. If you can’t, you face a loss: of money maybe, or worse, of your home or your children. This already happens. AI may make it worse.
Rawls gives the test any reform of access to justice should pass. His difference principle asks whether an arrangement improves the position of the least advantaged. Cheap answerable help, of the Garfield kind, plausibly passes it. Cheap unanswerable help may give the sophisticated a discount and the unsophisticated a risk.
What this shows: AI-prepared help can be cheap and still answerable, as Garfield’s trial shows; the courts hold litigants in person to the same standard as lawyers but have no professional to place the consequence on; and in conveyancing, answerability is bought for whoever has the most money at stake. What it doesn’t: whether people without lawyers who use AI actually do worse, or by how much. That is the mechanism’s prediction, not a measured result. Garfield’s trial is one case, and the evidence on why people go unrepresented is a small, dated study.
8. Who signs in 2040
The mechanism that would thin the pipeline is in place; the evidence that it has started is not. Partners learn judgement by doing the junior work: the review, the diligence, the first drafts, the long nights finding the clause nobody else saw. If the tools take that work, firms need fewer juniors, and in time there are fewer people fit to become the partners who sign. Every fix in this piece depends on a named human at the top. The fix itself uses up the route that produces that human.
So far, the numbers have barely moved. Legal Cheek’s survey of more than a hundred leading firms found trainee numbers down 2% in 2026. Commentators put most of that fall down not to AI but to the growth of solicitor apprenticeships, which let school leavers qualify while working, over five to six years, rather than through a degree followed by a training contract. The measure is itself shifting: since the Solicitors Qualifying Examination was introduced in 2021, the training contract is being replaced by a broader requirement for qualifying work experience, so a fall in “trainees” may partly reflect a change in how people qualify. In the US, as we saw, the average midsized firm has added staff since 2023, not cut them. The pyramid has not yet shrunk. It has grown more profitable.
The apprenticeship may be the first trace of the old playbook, or it may not. Firms present apprenticeships as a way to open the profession to people who could not afford the traditional route, and that may be the whole story. But an apprentice is also a cheaper route to a qualified lawyer: paid less while training, with the training itself paid for through the apprenticeship levy, and doing the base work of the pyramid from the start. From January 2026, government funding for these apprenticeships has been restricted largely to candidates aged 21 and under, which may change the arithmetic again. Read through the history of fixes in this piece, it looks like the latest in a line that runs through paralegals and offshoring. We cannot show which reading is right. We note that both fit.
Supervision is where the limit will be set. The Court of Appeal’s ruling in Mazur allows unauthorised people to do litigation work only under the supervision of an authorised person who keeps responsibility. The Law Gazette noted that supervision is hard to do properly in high-volume work, and that the ruling may push regulators to revisit what adequate supervision means. Whatever answer they reach will decide how many authorised humans a firm needs per unit of work, human or machine. If one partner may supervise a great deal of AI output, the pyramid can narrow sharply. If supervision has to be close, it cannot.
What this shows: the logic of erosion is clear, the economics that would drive it are in place, and the supervision rules that will govern its pace are being revisited. What it doesn’t: that erosion has begun. Trainee numbers have barely fallen, and the main explanation offered for that fall is not AI. The apprenticeship reading is our interpretation.
9. Where it could go
Trust may stay with the answerable professional, move to insurers and vendors, or fall back on the individual. These are not three futures to choose between. They are three places the contradiction can be squeezed into, and on the evidence so far all three are happening at once, in different parts of the market.
If trust stays with the professional, the risk is erosion. The partner keeps the rent on answerability for as long as the law keeps the machine from answering. But the stock of people fit to answer is drawn down without being replaced. The accumulation sits with the current generation of equity partners, and then runs out.
If trust migrates, it goes to whoever prices it. Product risk is already moving to vendors’ warranties and the specialist insurers behind them. If the SRA one day rewrites the minimum terms to deal with AI, as it did for cyber, the price of answerability will be set by insurers, and the firm becomes, in part, a wrapper around someone else’s policy. The accumulation then sits with insurers and the owners of the models.
If trust falls back on the individual, the risk lands on whoever can least bear it. Cheap help without answerability hands the judging back to the buyer. The sophisticated get a discount. The rest carry the loss.
Whichever way it goes, the saving comes out of wages, and wages are what buy the output. Each step in the chain is rational for the firm, the bank or the vendor taking it. The saving pools with partners, shareholders, insurers and model owners, not with the people whose spending the economy depends on. Every fraction of capital is competing for a market that its own success is shrinking.
Underneath all of this is a question about what we were trusting all along. The word “solicitor” reads as “safe”, in the way Roland Barthes described a myth: a sign whose history has been made to look natural. AI does the work without the arrangements, and in doing so shows that the arrangements were what we trusted: the regulator, the compulsory policy, the protected position. Societies have always needed ways to make strangers trustworthy enough to deal with. The law has been one of the most successful. AI tests whether that trust can be extended to something that cannot answer for itself.
10. What to watch
The direction of travel is clear; these are the signs that will show how far and how fast.
- The pricing fight. Whether clients like Citi push down the price of bounded work while fees for high-stakes work hold. If they do, clients are paying for answerability. If all fees fall, they were paying for hours.
- Billable hours. Thomson Reuters’ Law Firm Financial Index for the third quarter of 2026, expected shortly, and those after it. A sustained fall in hours billed per lawyer would be the first hard sign of AI shrinking the base. (US data.)
- Trainee and apprentice intake, and whether the mix keeps shifting towards apprentices.
- The insurance terms. Any move by the SRA to add an explicit AI clause to the minimum terms, and any change in what solicitors’ insurers ask about AI at renewal.
- The first claim. The first reported case in which a client sues a solicitor over an AI error and the insurer pays.
- Supervision. Whether the SRA tightens or loosens what counts as adequate supervision, and what comes of its investigations into the 42 reports of AI misuse.
- The bottom of the market. Whether firms like Garfield grow, and whether the courts see more AI-generated errors from people representing themselves.
Methodology
This piece was written as at early November 2026, about a market, legal services, that is moving as we write. It was researched and drafted by a human writer, outside the legal profession and with no inside knowledge of it, working with an AI. The AI searched for and summarised public sources and drafted sections of text. The human set the argument, wrote or rewrote the paragraphs that carry it, and tried to check the facts the argument relies on most. Each source is described in the text by what it is, a judgment, a regulator, a trade paper, a firm or a vendor, so the reader can judge its weight. Where we infer rather than report, we say so. Each section ends with what we think the evidence shows and what it does not.
We kept a log of every correction made in the course of the work: what was wrong, who caught it and how. Some were the AI’s: an overstated description of how a court treated two lawyers, an overclaim about the effect of a financial regulation, a case cited before it had been checked. Some were the human’s: losing the thread, rushing to a conclusion, rhetorical flourishes.
Nothing of commercial value hangs on this blog, so neither of us had any reason to fudge the record. If you find an error we missed, tell us, and we will correct the text and say so.
But imagine if something of real importance did depend on the work we have done here. AI is plainly an unprecedented enabler. It is also, like its human counterpart, fallible, though it fails in different ways and does not always know when it has. This is the contradiction at the heart of how this blog is now written. The same contradiction is worming its way into the legal profession. It won’t be the only one.
Leave a Reply