What 'vertical AI' actually means in a profession that bills by judgement

“Vertical AI” is having its moment, and legal is one of the verticals everyone points at. The pitch is familiar: horizontal models are general, professions are specific, so the winners will be the companies that take AI and make it legal. As far as it goes, I agree. The problem is what people think the verticalising actually consists of, because most of them are pointing at the wrong layer.

The common assumption is that a vertical AI product is a foundation model that has been trained or tuned on legal data, wrapped in a legal-looking interface, sold to legal buyers. A model with a wig on. That is the cheapest, most replicable, and least defensible version of the idea - and in a profession that bills by judgement, it isn’t where the vertical lives at all.

The model is the most horizontal thing in the stack

Start with an uncomfortable fact for anyone whose whole strategy is a fine-tuned model: the model is the most horizontal, most commoditising layer in the entire stack. It gets cheaper and better every few months, for everyone, simultaneously. Whatever edge you got from tuning it on contracts last year is largely erased by the next general release, which is better at contracts than your tuned version and better at everything else.

If your “vertical AI” company’s defensibility is “we trained it on legal,” you have built your moat out of the one material that the foundation labs are actively, expensively eroding on your behalf. That’s not a vertical. That’s a feature with a countdown timer.

The real vertical - the part that is genuinely specific to law and genuinely hard for a horizontal player to replicate - sits above and below the model, in the places the foundation labs have no reason to go.

Where the vertical actually lives

Three layers, none of them the model:

The workflow. A legal task is not a prompt. It’s a sequence with handoffs, approvals, exceptions, and a specific point where a human must take responsibility. A real-estate completion, a disclosure exercise, a matter intake - each has a shape that took the profession decades to settle, and that shape is full of hard-won detail an outsider gets wrong. Encoding the actual workflow, including the parts that look irrational until you understand why they exist, is vertical work no general model does for you.

The judgement boundary. This is the one that matters most and that horizontal players consistently miss. In law, the value isn’t the output - it’s knowing precisely where the machine should stop and a responsible human should take over. Draw that line in the wrong place and you’ve built something that is either useless (too cautious) or dangerous (too confident). Knowing where the line goes, for each task, is not a model capability. It’s domain judgement encoded into product design, and it is deeply, irreducibly vertical.

The trust and accountability layer. Law is a regulated profession where someone’s name and licence sit behind the advice. A vertical legal product has to make its work checkable - show its sources, expose its reasoning, fail loudly rather than quietly, and slot into a chain of professional responsibility. None of that comes from the model. All of it comes from understanding what it means to be on the hook for an answer.

Why this matters for who wins

If the vertical lived in the model, the foundation labs would win legal by default - they own the models and they’re getting better at everything. The reason they probably won’t is that the defensible layers are made of exactly the things a horizontal company has no incentive to build: the specific workflow, the precise judgement boundary, the professional accountability. Those require sitting inside the work, watching where it actually breaks, and caring about the fifteen per cent of edge cases that a general product treats as rounding error and a law firm treats as a negligence claim.

This is also why I’m sceptical of legal AI companies whose teams contain no one who has done the work. You cannot encode a judgement boundary you’ve never had to hold. You’ll put it where it demos well, not where it belongs, and the gap between those two places is precisely where a regulated profession gets hurt.

The practical test

If you’re a firm evaluating a “vertical AI” tool, or a builder deciding what to build, here’s the test I use. Ask: if a frontier model got twice as good tomorrow, what about this product would still be valuable?

If the answer is “nothing - it would just be obsolete,” you’re looking at a wig on a model, and you should rent it, not marry it. If the answer is “the workflow, the judgement boundary, and the accountability would all still be valuable, and would just get better,” you’re looking at something genuinely vertical - something that treats the model as an ingredient that keeps improving for free, rather than as the product itself.

The vertical in legal AI was never the model. It’s everything the model can’t do for you: knowing the work, knowing where judgement lives, and being willing to stand behind the answer. That’s not a thing you train. It’s a thing you earn.