Facts are friends
Recently, Matt Shumer – a tech founder and AI investor – published a widely shared piece comparing the current moment in artificial intelligence to February 2020: the deceptive window before a pandemic when most people hadn’t yet grasped what was coming. His argument was that AI had crossed a threshold in his own profession, software engineering, and was about to do the same in every other knowledge industry. He named law as one of the first sectors in the firing line.
He’s right about the direction. But he’s asking the wrong question.
The conversation about AI and the legal profession has so far been conducted almost entirely in terms of efficiency. AI drafts faster. AI reviews faster. AI researches faster. The implicit assumption is that legal services remain fundamentally the same product, just delivered at lower cost and higher speed. The debate, such as it is, concerns whether that efficiency gain will eliminate junior lawyers, compress fees or merely make existing practitioners more productive.
That framing is too small. It misses what is actually changing.
I’ve been a solicitor for nearly four decades. I advise on corporate transactions, debt restructuring and regulatory compliance – the kind of work that sits at the heart of commercial legal practice. Over the past year, I’ve integrated AI deeply into that work, not as an experiment but as a daily operating reality. And what I’ve come to understand is that the efficiency question – “can AI do legal work faster?” – is a distraction from the structural one: does AI change what clients are actually buying when they instruct a lawyer?
I believe it does. And the implications are far larger than a faster document review.
Clients don’t buy legal advice. They buy certainty.
This is the starting point the profession has been reluctant to examine.
No client has ever walked into a solicitor’s office wanting a contract. What they want is what they believe a contract gives them: an enforceable relationship, a framework for trust, a basis on which two parties can commit resources to a shared endeavour with reasonable confidence that each will do what they’ve promised. They want sustainable business relationships – arrangements that endure, adapt and generate value over time, not documents that sit in a drawer until something goes wrong. They don’t want a compliance audit; they want the assurance that their business won’t be shut down by a regulator. They don’t want a memorandum on the law; they want to know what they can safely do. Legal advice is the mechanism by which certainty has historically been delivered. It is not the end product. The end product is managed risk.
There is an emotional dimension to this that the profession understands instinctively but rarely articulates. A client instructing a solicitor is often acting from anxiety – a deal that must not go wrong, a regulation that must not be breached, a dispute that threatens the business. What they are buying is not merely a legal product but the reduction of fear: the reassurance that someone who understands the terrain is managing the danger on their behalf. Any salesperson will tell you that safety – often partly subconscious – is one of the most powerful drivers of a purchase decision. The question the profession must confront is whether that reassurance is permanently tied to the human relationship, or whether it flows from the quality and reliability of the risk management itself. If a system that monitors your compliance continuously and flags problems before they escalate makes you safer than a solicitor you speak to twice a year, the emotional logic follows the structural logic – even if it takes clients time to recognise it.
The legal profession, in its current institutional form, is a response to a pre-AI risk environment - one in which managing risk required episodic human intervention. If AI changes the structure of risk management itself, the institutional form that grew around the old structure cannot remain unchanged.
The legal profession exists in its current form because risk management has been static, episodic and constrained by the limits of human cognition and attention. Contracts are drafted at a point in time and reviewed when something changes or breaks. Compliance is assessed periodically. Advice is sought reactively. The entire model is built around moments of intervention separated by long stretches of unmonitored exposure.
Law firms sit in the gap between uncertainty and assurance. They charge for crossing that gap. And for decades, there has been no other way to cross it.
AI threatens that position not primarily by doing the crossing faster, but by closing the gap itself.
The architecture of risk is changing
Consider how the current model actually works – and how badly it serves the client’s real interests.
A contract is drafted at a fixed point in time, reflecting the parties’ understanding of their obligations and the risks they face at that moment. It is then filed away. The parties get on with business. Eventually – sometimes months or years later – one party becomes aware that the contract has been breached. Solicitors are instructed. Proceedings may follow. Substantial time and money are invested in seeking a remedy to put the injured party back in the position they would have occupied had the contract been performed. The only guaranteed winners, usually, are the law firms on each side. Even a successful claim is more of a Pyrrhic victory than a genuine restoration – the relationship is damaged, the management time has been consumed and the commercial opportunity has been lost.
Here is the uncomfortable truth that the profession rarely acknowledges: many breaches are not the result of deliberate bad faith but operational failure, misalignment or drift. A party fails to anticipate a risk, trips over an unforeseen hurdle or simply loses track of an obligation in the noise of running a business. The breach is accidental, the consequences are disproportionate and the legal remedy is an expensive, adversarial attempt to unscramble an omelette that need never have been made.
The entire structure is reactive, episodic and built around failure rather than prevention.
Now consider what becomes possible if AI can deliver continuous compliance monitoring rather than periodic review, real-time regulatory tracking rather than quarterly updates, automated policy adjustments rather than ad hoc advice, and audit-ready documentation generated as a byproduct of business operations rather than as a separate professional service. Consider predictive modelling – not merely of litigation exposure after a dispute has crystallised, but of the operational failures that lead to disputes in the first place. If a system can monitor contractual performance in real time and flag where a party is drifting towards non-compliance, the failure can be addressed before it becomes a breach. Action taken quickly limits the scope of loss. The risk of litigation is not merely predicted – it is reduced, because the underlying cause is caught and remedied before it escalates into a claim.
In that world, risk management migrates from episodic advisory work into embedded infrastructure. “Legal services” as currently constituted contract dramatically. Not because law disappears – it doesn’t – but because law becomes integrated into software systems and insurance products. The need for external counsel in routine documentation, compliance and interpretive advisory work falls away. Much of what law firms bill for today exists to compensate for the friction of human-managed risk – the static contracts, the periodic reviews, the reactive instructions when something has already gone wrong. Remove that friction, and the revenue category shrinks.
I want to be precise about this claim. I am not arguing that adversarial law disappears. Courts will still exist. Disputes will still arise. High-stakes negotiation, fiduciary responsibility and the exercise of judgment in genuinely novel situations will remain human domains for the foreseeable future. Power dynamics, strategic ambiguity and the irreducible complexity of contested facts are not susceptible to automation in any timeframe I can credibly predict.
But the revenue-heavy middle layer – documentation, monitoring, routine compliance and interpretive advisory work, which constitutes a very large proportion of current law firm income – is structurally capable of migration into automated, insurable infrastructure over time.
The most defensible version of this claim is not that lawyers disappear. It is that the economic centre of gravity of legal risk management is shifting from professional services to platform infrastructure. Law does not vanish. It relocates.
Capability is real – and accelerating
I should ground this in evidence, because the structural argument only holds if the capability is genuine.
METR, an AI evaluation organisation, has been systematically measuring what they call “task-completion time horizons” – the length of real-world tasks, measured by how long they take a human expert, that AI agents can complete autonomously with 50% reliability. Their peer-reviewed research, first published in March 2025, describes the trend as “exponential,” with the time horizon roughly doubling every seven months over a six-year period. Their live tracker, most recently updated on 3 February 2026, shows no sign of that trajectory flattening.
Consider what falls within a five-hour professional window: a first draft of a share purchase agreement; a due diligence review of a medium-sized data room; a regulatory impact memorandum; a detailed analysis of case law on a contested point. These are not marginal tasks. They are the daily substance of commercial legal practice.
When I feed a current-generation AI model a shareholders’ agreement and ask it to identify every clause creating asymmetric risk for a minority investor, it doesn’t simply flag the obvious problems. It maps the interactions between provisions – the drag-along, the board composition mechanics, the consent thresholds – and explains how their combined effect strips a minority holder of meaningful protection. This is work that would take a competent associate half a day. The AI does it in minutes and explains its reasoning clearly enough for efficient senior review.
I should be clear: these capabilities are very recent. The models that produce this quality of analysis have arrived in the past few months, and they still require human verification. The output must be checked. But the nature of that checking is evolving in a way that matters.
Think of it as a crash dummy model. In the early phase – where we are now – the human solicitor is the safety mechanism. You test the AI’s output, verify its reasoning, catch its errors and build confidence in its reliability across different types of work. The AI is the vehicle; the human is the crash test that proves it’s safe. But as the models improve and the evidence of reliability accumulates, the dynamic inverts. The AI becomes the primary performer, and the human becomes the quality assurance layer – intervening on exception rather than reviewing by default. That transition doesn’t happen overnight, and it shouldn’t. But it is the direction of travel.
This connects directly to the insurance question. If AI-monitored contracts reduce the incidence and severity of breaches – because failures are caught early, losses are mitigated before they compound and the scope of exposure is narrowed – then the risk profile of AI-assisted legal work actually improves over time. Insurers, who price risk on outcomes rather than on methods, will eventually recognise that a well-implemented AI compliance system generates fewer claims than a traditional human-only model. At that point, the economics of professional indemnity begin to favour AI adoption rather than resist it. Insurers will move cautiously – but they ultimately price outcomes, not tradition.
“But I tried ChatGPT in 2023 and it hallucinated.”
I hear this regularly. You were right – those early models did confect authorities and cite cases that didn’t exist. But that was, in AI terms, a different geological era. The current models reason, self-correct and flag uncertainty rather than manufacturing false confidence. They are imperfect – but imperfect in the way a newly qualified solicitor is imperfect, not in the way a random number generator is imperfect.
The critical distinction, however, is between capability and adoption – and I want to resist the technological inevitabilism that characterises too much commentary on this subject. The fact that AI can do something does not mean it will be adopted, or that its adoption will follow a smooth curve. Which brings us to the friction.
Regulatory friction is real – but it shapes the path, not the destination
Unlike software engineers – the profession Shumer knows best – lawyers operate inside a regulated liability framework. AI output is not the end of professional responsibility. It shifts the locus of it.
The Solicitors Regulation Authority has been clear: professional duties apply to all work carried out by a solicitor, regardless of whether AI was used. The SRA Principles and Code of Conduct – the duties of competence, effective supervision and upholding the administration of justice – apply in full. The High Court’s 2025 judgment in Ayinde v London Borough of Haringey underlined the consequences for solicitors who fail to verify AI outputs. The Law Society has called for urgent SRA guidance on how AI intersects with the Mazur ruling and the Legal Services Act 2007, particularly in relation to the conduct of litigation. Professional indemnity insurers have not yet fully priced the implications of AI-assisted practice.
These are genuine constraints. A partner at a provincial firm is right to ask: what happens to my PI cover if AI-assisted due diligence misses something material? What are the SRA’s supervisory expectations for AI-generated work product? What disclosure obligations do I owe my client about how their matter was staffed?
But regulatory friction slows adoption; it does not prevent it. The question is not whether AI will be integrated into legal practice but how – and how quickly. History suggests that when a technology offers a genuine cost and quality advantage, regulatory frameworks adapt to accommodate it, sometimes slowly, sometimes clumsily, but they adapt. The firms that treat regulatory uncertainty as a reason to wait will find themselves overtaken by firms that engage with it proactively. Developing an AI use policy, documenting supervision processes and talking to your insurer now is not premature. It is prudent.
None of this happens automatically. Institutions resist change. Culture lags capability. Clients tolerate inefficiency longer than economists predict. But once a cheaper, higher-quality and more reliable mechanism for delivering certainty becomes visible, the pressure on the old model becomes structural rather than cyclical.
The leverage model is structurally exposed
The economic implications of this shift are severe – not because they arrive overnight, but because they attack the foundations of how law firms make money.
Large and mid-tier firms operate on a leverage model. Partners generate and supervise work; associates and trainees execute it. The profit margin sits in the gap between what the client pays for an associate’s output and what the firm pays the associate. Multiply that across a pyramid of fee-earners and you have the economics that fund partner drawings, trainee recruitment and offices in expensive postcodes.
AI does not simply threaten to replace associates. It collapses the economic logic that sustains the pyramid.
If a partner can produce the same output with AI that previously required three associates, the firm doesn’t need the associates – but it also cannot charge the client for three associates’ worth of hours. The billable hour model, which has survived every previous wave of legal technology, is not resilient to a tool that compresses the time required for substantive legal analysis from days to minutes.
The disruption will likely arrive not as sudden mass redundancy but as a gradual, compounding economic squeeze. Billable hours per matter will fall. Output per partner will rise. Clients – who are increasingly experimenting with AI themselves and are ever less willing to pay for work they suspect a machine could do – will demand fee reductions. Mid-tier firms, with thinner margins and less pricing power than the Magic Circle, will feel this first. Profit per equity partner will come under pressure. The response will be predictable: fewer trainees, fewer associates, more technology, more concentration of value at the top.
I should stress-test this honestly. It is possible that AI increases total legal demand rather than reducing it. Lower cost per transaction may expand the addressable market – more businesses able to afford proper legal advice, more deals that become economically viable, greater access to justice for individuals currently priced out. Law firms have adapted to previous waves of technology and emerged largely intact. Legal work remains relationship-driven, and clients value the team they trust, not merely the output they receive.
The strongest version of this counter-argument is not that law firms will simply adapt, but that the trust layer – the relationship between a General Counsel in the client and the partner who has guided the business through multiple restructurings – carries value that no platform can replicate, and that even well-capitalised competitors with deeper enterprise distribution have historically failed to displace trusted advisors in high-stakes professional services. While that is true, it is also based in the “old model”.
All of that may prove true. But even in the most optimistic scenario, the distributional question persists: who captures the surplus? If AI doubles a partner’s output while halving the need for associates, the firm may be more profitable – but for fewer people. The total pie may grow, but the slices will be cut very differently.
The training pipeline is breaking
This brings me to the consequence that troubles me most.
The traditional model by which junior lawyers develop professional expertise depends on them doing precisely the work that AI is automating. Associates learn to draft by drafting. They develop judgment by reviewing documents, spotting issues and having their work corrected by seniors. The grunt work is not merely a cost centre – it is the profession’s training infrastructure.
Training contract numbers at City firms peaked in 2024 at 2,267 and have since declined by approximately 8%, according to Legal Cheek’s research tracking 67 firms over eight years – with economic uncertainty, reduced deal flow and AI cited among the contributing factors. That headline figure still masks the deeper structural question: if the substantive junior work disappears, how does the next generation develop the instinct, the pattern recognition and the commercial judgment that come from years of doing it?
The profession does not yet have a credible answer. “We’ll cross that bridge when we come to it” is not one. Every firm that employs trainees should be thinking now about how to restructure training around AI-augmented practice rather than AI-replaced tasks. If we fail to do this, we will find ourselves in a decade’s time with a generation of senior lawyers and no pipeline of talent behind them.
The Engels’ Pause and the question of distribution
The economic historian Robert C. Allen coined the term “Engels’ Pause” to describe the period from 1790 to 1840 during the British Industrial Revolution. GDP per worker expanded rapidly. The profit rate doubled. Capital’s share of national income grew at the expense of labour. Technical progress was the engine – but its benefits flowed overwhelmingly to capital owners rather than workers, and did so for decades before the gains were shared more broadly.
The structural resemblance to the current moment is striking. AI is dramatically increasing professional output while compressing demand for the human labour that previously produced it. The gains accrue to those who own the technology and the client relationships – the capital holders, in Allen’s framework. Associates and mid-level lawyers occupy the position of the skilled weavers rendered redundant by the power loom. Their expertise is not being augmented. It is being replaced.
There are reasons to hope that modern institutions – tax policy, equity participation, regulation, new ownership models – may respond more swiftly than Georgian-era mechanisms did. We have better data, more responsive markets and, one hopes, more political will. But the legal profession, with its partnership structures and longstanding resistance to external investment, is not well placed to redistribute the gains of AI quickly – even if the wider economy does.
The critical question is therefore not primarily technical. It is distributive. If embedded legal certainty is delivered through AI platforms and insurance-backed systems rather than through law firms, the surplus shifts from partnership structures to technology providers, insurers and capital owners. That concentration of value is not inevitable – but it is the default outcome unless the profession consciously designs an alternative.
In the draft of my forthcoming book, AI, Abundance and the Future of Work – written last year, before many of the capabilities I describe above had fully materialised – I explored this dynamic at length: the way technological disruption creates abundance in aggregate while generating acute dislocation for specific groups. What struck me in revisiting that draft is how much of what felt like medium-term prediction has already become present-tense reality. The legal profession is now a case study in exactly the phenomenon I set out to examine.
What the profession should actually do
Let me be practical, solicitor to solicitor.
Redefine what you are selling. If what you sell is document production and hourly expertise, AI will eat your margins within years. If what you sell is judgment, accountability, trusted relationships and the willingness to own an outcome, you have a durable business – but it requires a fundamentally different pricing model and a fundamentally different value proposition. Start having that conversation with your partners and your clients now, not when the revenue starts falling.
Use these tools seriously. Not the free version. Not a five-minute test you can use to confirm your prior scepticism. Subscribe to Claude Pro or ChatGPT Plus, make sure you’re using the most capable model available and feed it your actual work. Give it a real contract and ask it to draft a counterproposal. Give it messy heads of terms and ask what’s missing. When the first attempt isn’t perfect, iterate – more context, clearer instructions, another attempt. The people getting extraordinary results are not the technically gifted. They are the ones who’ve learned to communicate effectively with the tool. And remember: if it partly works today, today is the least able AI will ever be.
Engage with the regulatory questions proactively. Develop an AI use policy. Document your supervision processes. Understand your obligations under the Code of Conduct. Talk to your PI insurer. Don’t wait for the SRA to produce comprehensive guidance – firms that waited for the regulator to bless e-signatures were years behind those that adopted them thoughtfully.
Redesign your training model. If junior work is being automated, training must be restructured around supervised AI-augmented practice: reviewing and improving AI output, learning to spot what the machine misses, developing the judgment that comes from understanding why a clause matters rather than merely how to draft it. This is a solvable problem, but only if it’s treated as urgent.
Think about who owns the infrastructure. This is the question most lawyers are not yet asking, and it is the most important one. If legal certainty migrates from professional services to embedded platform infrastructure, the profession must decide whether it governs that infrastructure or merely feeds it. The choice between being a governance layer overseeing intelligent systems and being a legacy cost centre displaced by them is still open – but the window in which the profession can shape the answer is narrowing.
If I am wrong, the profession will absorb AI as it has absorbed previous technologies and continue much as before. If I am right, we are living through the early phase of a structural redesign that will redefine what it means to be a lawyer.
Where this lands
I want to end not on fear but on redesign.
The legal profession has historically monetised complexity and uncertainty. AI reduces both. That is not a temporary efficiency gain that can be absorbed by the existing model. It is a structural shift in the economics of risk management.
The profession can defend the current pyramid – fewer trainees, more technology, the same billing model, a slow squeeze on margins until the model collapses under its own weight. Or it can reimagine itself as something different: a governance layer, a quality assurance function, a fiduciary overseer of intelligent systems that deliver legal certainty at scale.
That reimagination requires honesty about what is actually being sold, courage to restructure institutions that have been profitable for generations, and a willingness to ask the distributive question before the market answers it by default.
But I want to end on something broader than the profession itself, because what is happening in law is not unique to law. It is a signal of what is coming for every knowledge industry.
We may be entering a period of labour displacement across multiple professions simultaneously – something the modern economy has not previously had to absorb at once, and something that risks long-term redundancy on a scale that threatens established societal norms. Not the redundancy of a single sector – coal, steel, textiles – where displaced workers could retrain into adjacent industries. This is the potential displacement of cognitive work across professions simultaneously
If that happens, the economic consequences are severe. But the human consequences may be deeper still. For many people, professional identity is not merely a source of income. It is a source of meaning, status, community and purpose. When the work disappears, the salary is not the only thing that goes with it. People will need to find new ways to contribute, new sources of dignity and new answers to the question that every redundancy forces into the open: what am I for, if not this?
That is not a question the market will answer well on its own. It requires institutional imagination, political courage and – I would argue – a moral seriousness about human dignity that our current discourse lacks.
AI is generating abundance in legal output, and in every other form of cognitive output. The question – for the profession and for society – is whether we redesign our structures to share that abundance, or allow it to concentrate at the top while millions of skilled, educated people search for a new purpose that the economy has not yet created for them. That is not merely an economic question. It is a moral one.
And it is the question I wrote my book to ask.
Quentin Solt is a solicitor, legal consultant and author of the forthcoming AI, Abundance and the Future of Work. I write about the intersection of technology, professional life and faith at https://factsarefriends.substack.com. Follow me on LinkedIn at www.linkedin.com/in/quentin-solt.
Sources
References are listed in the order in which they are mentioned above.
Matt Shumer’s original piece
Matt Shumer, “Something Big Is Happening,” published on X, February 2026. https://x.com/mattshumer_
METR task-completion time horizons research
METR, “Measuring AI Ability to Complete Long Tasks,” blog post, 19 March 2025. https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/
Kwa, T. et al., “Measuring AI Ability to Complete Long Tasks,” arXiv preprint, March 2025 (revised March 2025). https://arxiv.org/abs/2503.14499
METR, “Time Horizon 1.1” (updated methodology and expanded task suite), 29 January 2026. https://metr.org/blog/2026-1-29-time-horizon-1-1/
METR, “Task-Completion Time Horizons of Frontier AI Models” (live tracker, most recently updated 6 February 2026). https://metr.org/time-horizons/
METR homepage, with research overview and autonomous AI capability trackers. https://metr.org/
SRA regulatory framework and AI guidance
Solicitors Regulation Authority, SRA Principles and Code of Conduct for Solicitors, RELs and RFLs (the primary regulatory framework for professional duties including competence, supervision and use of technology). https://www.sra.org.uk/solicitors/standards-regulations/
Solicitors Regulation Authority, “Risk Outlook report: The use of artificial intelligence in the legal market.” https://www.sra.org.uk/sra/research-publications/artificial-intelligence-legal-market/
The Law Society, “Generative AI - the essentials” (practice guidance). https://www.lawsociety.org.uk/topics/ai-and-lawtech/generative-ai-the-essentials
Ayinde v London Borough of Haringey
R (on the application of Ayinde) v London Borough of Haringey; Al-Haroun v Qatar National Bank QPSC [2025] EWHC 1383 (Admin), judgment of the Divisional Court (Dame Victoria Sharp P and Johnson J), 6 June 2025. https://www.judiciary.uk/judgments/alharoun-v-qatar-national-bank-and-ayinde-v-london-borough-of-hackney-order/
The earlier wasted costs judgment: R (on the application of Ayinde) v London Borough of Haringey [2025] EWHC 1040 (Admin), Ritchie J, 30 April 2025.
Mazur ruling and AI in litigation
Mazur v Charles Russell Speechlys [2025] EWHC 2341 (Admin), Sheldon J, 16 September 2025.
The Law Society, “Mazur and the conduct of litigation” (practice note), October 2025. https://www.lawsociety.org.uk/topics/civil-litigation/mazur-v-charles-russell-speechlys-litigators
Legal Futures, “Law Society calls for urgent SRA advice on impact of Mazur on AI,” 26 November 2025. https://www.legalfutures.co.uk/latest-news/law-society-calls-for-urgent-sra-advice-on-impact-of-mazur-on-ai
Solicitors Regulation Authority, “Mazur and conducting litigation” (guidance page, updated 20 October 2025). https://www.sra.org.uk/home/hot-topics/conducting-litigation/
Training contract numbers
Legal Cheek, “Exclusive research reveals training contract numbers remain stable - despite talk of a ‘graduate jobpocalypse,’” 6 November 2025. https://www.legalcheek.com/2025/11/exclusive-research-reveals-training-contract-numbers-remain-stable-despite-talk-of-a-graduate-jobpocalypse/
Engels’ Pause
Allen, R.C., “Engels’ pause: Technical change, capital accumulation, and inequality in the British Industrial Revolution,” Explorations in Economic History, Vol. 46, No. 4, pp. 418-435, 2009. https://www.sciencedirect.com/science/article/abs/pii/S0014498309000199
Full paper (PDF): https://www.nuff.ox.ac.uk/Users/Allen/engelspause.pdf
Broader context
Frey, C.B., The Technology Trap: Capital, Labor, and Power in the Age of Automation, Princeton University Press, 2019. (Draws extensively on Allen’s Engels’ Pause framework in the context of AI and labour markets.)