The training crisis

We've now watched a lot of people use AI — lawyers drafting contracts, analysts building models, researchers running literature reviews, consultants putting together strategy decks. Some get extraordinary results. Others get mediocre output and conclude the technology is overhyped. The gap between those two people is real, and it isn't one thing. It's four things.

To get a good AI outcome, you need four ingredients working together.

Start with model capability. This sounds obvious, but it matters more than people think. The Harvey founders describe the shift from GPT-3.5 to GPT-4 as the moment legal AI became viable. The jump in reasoning quality was that significant - Anthea remembers it very clearly. More recent advances, like Claude Code + Opus 4.6, and GPT 5.4 or GPT-5.3 codex, have been important for agentic work, where the AI doesn't just answer questions but executes multi-step tasks. Not all models are equal, and matching the right model to the right task is itself a skill. If you don't pay for frontier models, people are often using less good freely available models. Or they are using ones like ChatGPT that actually have multiple models under the hood and that might be routing your inquiry to a faster but less bright model. Models matter and understanding which model to use when also mattes.

Then there's context. Models are general. Advice needs to be specific. A contract review is useless without the deal context: the counterparty history, the client's risk appetite, the jurisdiction, the commercial relationship. And all of that needs to be in a form the AI can actually use, which means machine-readable documents, structured information, accessible files. Agents thrive on a filing system filled with clearly indexed markdown files and code, not a mess of word documents and pdfs spread across a myriad of systems. This isn't just about data, either. Providing the right level of context is harder than it sounds: too little and you get generic output, too much and you drown the signal in noise. This is what legal AI companies are trying to solve with matter-centric workspaces and firm systems integration, an infrastructure play for context delivery. But the context problem goes well beyond any single platform, and it also requires AI skills to get it right.

That brings us to the third ingredient, which is AI skills — prompt engineering, context engineering, iteration. Knowing how to structure a request, when to break a task into steps, how to recognize when the AI is confident versus confabulating, how to provide feedback that actually improves the next output etc. When you look at someone who is a superuser of AI systems, they will have built a complex layered system of context files, custom agent instructions, custom commands for repeat actions, custom workflows for repeat sequences, and iterative feedback loops that transforms how AI performs for them. The gap between their results and a first-time user's isn't the tool. It's the technique. Anthea wrote about this in Tools vs. Technique as the gap that explains why the same tools produce wildly different results for different people.

And the fourth, the one this piece is really about: domain expertise. Taste, judgment, the kind of knowledge that comes from years of doing the work. Knowing whether this force majeure clause is right for this deal, in this jurisdiction, with this counterparty. Knowing what a good answer looks like before you see it. In our experiential piece, we called this metis — practical wisdom that can only be built through practice. There we argued that domain experts need to develop AI metis. Here we ask the reverse: how is domain metis produced, and what happens when that production process breaks down? You get it from drafting hundreds of clauses, sitting in dozens of negotiations, having a senior partner tear apart your work and explain why. You get it from being a junior lawyer in a firm or doing a clerkship with a judge.

When it comes to creating compounding value, these four ingredients multiply, they don't add. A brilliant model with no context produces generic answers. Perfect context with no domain expertise means you can't evaluate what comes back. But when all four elements exist together, they create a powerful multiplicative force.

Fast Metis and Slow Metis

Given this mental model, here's where the legal profession has a problem it hasn't solved.

Ingredients three and four are both forms of metis. You learn AI skills by using AI, not by reading about it. You learn legal judgment by doing legal work, not by studying it. Both are practical knowledge that can only be developed through practice. But, at least from our observation, they compound at very different speeds.

AI skills — ingredient three — are the fast metis. They can be developed with weeks and months of deliberate practice, not years and decades. And some people find it more natural than others. In Strange Attractors, Anthea explored how certain cognitive styles and AI seem to have converged on the same processing patterns — nonlinear, associative, comfortable holding multiple threads, drawn to productive sprawl. In her observation of many superusers, they often have similar cognitive profiles: certain types of cognitive diversity like ADHD, highly curious information seekers, an exploration rather than exploitation mindset, comfort with uncertainty and probabilistic systems, openness to not just learning but unlearning and relearning constantly. The AI skills gap isn't just about effort or exposure. It's partly about cognitive style. Some minds are already wired in ways that make AI collaboration intuitive. But whatever the individual variation, this is a gap that closes in months.

Domain expertise — ingredient four — is the slow metis. It compounds over years. A first-year associate who spends weeks reviewing thousands of documents in a due diligence exercise isn't just completing a task. She's building pattern recognition — learning what normal clauses look like, how transactions work, how parties negotiate, where risks hide. She's training the judgment that will make her, a decade later, the lawyer whose opinion a client trusts on a bet-the-company decision. That judgment can't be shortcut. It's the product of doing the work.

This creates an asymmetry that we think defines the next decade of the legal profession. We've mapped it as a quadrant — domain expertise on one axis, AI skills on the other — and the picture is stark. You can explore the interactive version here.

AI-native juniors have ingredient three — they grew up with these tools, can prompt-engineer fluently, iterate naturally. But have they drafted a force majeure clause from scratch? Sat in a negotiation and watched a counterparty's face when a term lands wrong? Had a senior partner tear apart their work and explain why? If they haven't, are they able to develop the kind of domain expertise and metis that will give them the senior judgment they need to direct and review AI outputs? Will they develop the expertise and taste that takes something from AI output that looks plausible but may be generic and flawed to being actually excellent in form and substance?

Senior partners have ingredient four — twenty-five years of judgment, the ability to spot a bad clause in seconds, a sense for how deals go wrong that no dataset can replicate. In our experiential piece, we argued that this domain expertise is what makes human direction of AI valuable, and that the AI skills gap is crossable for those who invest the time. But the more troubling question runs in the other direction. A motivated senior can develop meaningful AI fluency in months. You cannot compress a decade of apprenticeship into a few months of AI review. And it is the slow metis — the one that takes years to build — that the profession is now at risk of losing.

The Apprenticeship at Risk

And this is where the training crisis enters. The profession is undermining the very process that produces the judgment it depends on. It is killing the goose that lays the golden egg. The apprenticeship — the years of doing the work that build the taste and judgment AI cannot replicate — is being devalued and may be eliminated. And if it is, who will be able to check the answers AI gives? Who will have the expertise to know when a plausible-sounding output is subtly wrong?

AI is automating the work through which domain expertise was historically developed. Document review, legal research, first-pass drafting, due diligence — these aren't menial tasks endured on the way to real lawyering. They are the apprenticeship. They are how a law graduate who knows doctrine but not practice becomes a lawyer who can exercise professional judgment. When we automate those tasks, we don't just gain efficiency. We lose the production process for the scarcest ingredient in the formula.

This connects to the pyramid economics we described in Harvey's Strategic Evolution. AI is automating the base of the pyramid — but the base is also where the apex was produced. Automate it and you don't just threaten the economics. You eliminate the production process for the expertise at the peak.

Each firm's decision to automate this work is individually rational. No managing partner would defend assigning associates to review documents by hand when AI does it faster. But collectively, those rational decisions destroy the pipeline. And here's what makes it worse than a normal coordination failure: firms don't need to train lawyers. They need trained lawyers. They'd happily hire them from someone else. Indeed, this is exactly what in-house legal departments have been doing for years. But when everyone free-rides on training, no one trains. The resource disappears.

There's a paradox that sharpens this further. As AI improves, its errors get subtler. A hallucinated citation is easy to catch. A plausible but subtly wrong contract clause is not. The judgment needed to catch these errors increases precisely as the training that builds that judgment decreases. We're building a system that depends on human verification capacity we may be setting ourselves up to systematically fail to produce.

We introduced the training crisis in The Scrambled Competitive Map as something that cuts across every force in the legal market — incumbents, challengers, and clients alike. This is why. It isn't about any one vertex of the triangle. It's about the scarcest ingredient in the formula, the one that takes the longest to develop, being quietly removed from the production process.

The Counter-Argument

There is an obvious counter-argument, and it deserves to be taken seriously: if the traditional apprenticeship is being automated, can domain expertise be built through other means?

The possibilities are real. AI-powered simulations could put junior lawyers through hundreds of negotiation scenarios and contract variations in a fraction of the time it would take to encounter them in practice. Structured mentorship programs could pair senior judgment with junior execution more deliberately than the current sink-or-swim model. And AI itself could serve as a training tool — generating exercises, evaluating drafts against expert benchmarks, surfacing the reasoning behind good and bad clauses. Some of these approaches are already being explored. Alternatives to traditional apprenticeship exist.

But there is a difference between simulated experience and real experience — and it may matter more than the alternatives literature assumes. Reviewing a contract generated for training purposes is not the same as reviewing one where a client's money, reputation, or freedom is on the line. The stakes change how you learn. The pressure of real consequences sharpens attention in ways that practice exercises may not. The senior partner who tears apart your work on a live deal is teaching you something different from an AI that scores your draft against a rubric. You learn differently when it counts — when the error has consequences, when the client is in the room, when the deadline is real. That difference between training and practice is precisely the kind of thing that's easy to dismiss in theory and yet is obviously important in practice. Even occupations like pilots that have extensively used simulations for years would never think of allowing someone to fly a plane solo after only training in a simulator. There is a reason why pilots must be co-pilots first.

The question isn't simulation versus apprenticeship. Nor are we saying that the current modes of apprenticeship training are necessary, or even particularly effective at producing deep domain expertise. As David often points out, no one has ever demonstrated that bleary-eyed associates reading documents at 3AM in a windowless warehouse in Phoenix Arizona in August is an efficient or effective way of teaching how to spot attorney-client privilege material, let alone doing this multiple times over the course of years. Nor, when apprenticeship is left largely in the hands of individual partners whose incentives are to get the work done as quickly as possible is there any guarantee that the real training opportunities required to develop meaningful domain expertise will be equitably distributed. Instead, the question is what combination — accelerated simulations, deliberate mentorship, real-world exposure, AI-assisted feedback — can produce sufficient domain expertise in a sufficient number of junior lawyers if the volume of traditional junior work shrinks. The profession hasn't seriously grappled with this question yet, and it needs to. The old model is eroding. Whether a new one can be built — and built fast enough — is one of the most important open questions in legal education.

The Binding Constraint

In a world where model capability, context infrastructure, and AI skills are all improving rapidly, the constraint on the legal profession isn't technology. It's whether we can still produce people with the judgment to use it well.

Without domain expertise to multiply against, AI skills produce fluent mediocrity. And if we stop building domain expertise — if we automate the apprenticeship that produces it — the whole formula collapses. Domain expertise is one of the multipliers, and right now it's the one we're losing.

Law schools and the profession need to produce multidisciplinary problem solvers capable of functioning in volatile, uncertain, and complex environments — not just technically proficient researchers. That requires domain expertise as a foundation. The adoption of AI is likely to be close to universal. The question is whether we're also investing in the human capacity that makes that adoption valuable.