Industry
Why the Market Is Solving the Wrong Layer
The legal AI market has spent two years optimising what is no longer the problem. To understand why, you have to unpack the stack, layer by layer.
Three years of legal AI. A scene that has solidified around a certain way of thinking about the problem, and that way of thinking is becoming a trap. To understand where value will lodge in the rest of the cycle, one has to look coldly, layer by layer, at what this market has actually built, what it has forgotten to build, and where the real constraint now lies.
Layer 1 — Generation, which was the focus of the whole race
The first thing the industry optimised was the quality of raw generation. The first legal AI products competed on the ability to draft a warranty clause in proper English, to summarise a court ruling without misinterpretation, to answer a legal question with a level of precision that would have seemed impossible two years earlier. This race made sense, because the quality of generation conditions everything else.
This race is largely over. The models available today generate at a level that exceeds what almost all common legal tasks require. Raw quality has become an entry condition, not a differentiator. Competition is shifting elsewhere, and the market, by habit, continues to fight on the ground it knows.
When everyone has access to excellent generation, excellent generation stops being an advantage. It has become a floor.
Layer 2 — Assistance, which became the product reflex
The second thing the industry built is the assistant. The format has become so dominant that it is used as a synonym for legal AI. A chat, a window in which a lawyer asks questions and receives answers. The economic model is understandable: an assistant shows easily, sells easily, renews easily.
But the assistant has a structural flaw that has long been chosen to be ignored. It is, by nature, transactional. A question, an answer, and the transaction closes. An assistant has no institutional memory, no matter continuity, no ability to carry an affair across weeks, because the format does not provide for it. As long as legal work let itself be sliced into short transactions, the assistant held its role. When the work exceeds that unit, the assistant shows its seams. And almost all serious legal work exceeds that unit.
Layer 3 — Integration, which was left to the user
The third thing that should have been built, but was not, is integration between the assistants. Vendors have shipped products each excellent in their silo. One tool for drafting. Another for review. A third for research. None knows what the other has done.
The result, in firms that have adopted AI seriously, is an accumulation of tools whose addition does not form a system. The lawyer becomes the integration layer themselves. They copy a paragraph from one tool to paste it into another. They re-explain context to an assistant that has forgotten it. They check that one tool has not taken a position contradicting what another established. This work appears in no budget, but it consumes precisely the part of the work that technology was supposed to free.
Each tool answers a question. None of them holds the matter together. That gap is where the work breaks down.
Layer 4 — Coherence, which has become the real constraint
Above integration, there is an even rarer layer, that of operational coherence. It is not just about making tools work together.
Making tools work together is a question of integration. Making sure a matter remains a matter is a question of architecture.
It is about ensuring that a matter remains a matter through time, through people, and through tools. It is about ensuring that what was decided on Monday remains true on Friday, that what a partner has settled remains settled when an associate picks up the matter, that the methodology of the firm is applied equally by every AI use across the practice.
This layer exists practically nowhere in current products. The reason is simple: to build it, one has to accept not being an assistant. One has to accept not having a visible interface, because coherence lodges between interfaces. One has to accept being an infrastructure, not a product.
Why the market continues to optimise the wrong layer
Three reasons, which reinforce each other. The first is commercial. Selling an assistant is easy: it shows, it has a unit price. Selling an infrastructure is hard: it does not show in a demo, its return on investment is measured in years. For a growing vendor, the assistant is the natural format; for a buying firm, the assistant is the reassuring format. The market agrees on a format that is simpler for both sides, even if it does not solve the real problem.
The second is cognitive. Buyers judge a product by what they can see in a demo. Now, what can be tested in a demo is generation and assistance. Coherence, which is a property of use over time, does not test in a demo. It is lived in practice, after several months of use. As long as purchases remain guided by what can be demonstrated in an hour, they remain guided by the layers the demo rewards, which are not the layers where durable value lodges.
The third is historical. The enterprise software industry has always taken time to recognise infrastructure layers. When the relational database arrived, it took a decade for the market to understand that the semantic layer above would be more durable than the databases themselves. When the cloud arrived, it took several years for governance layers to be recognised. It is normal that the legal AI market, three years into its existence, still reasons in assistants. What is not normal is to fail to see that this phase is coming to an end, and that those who linger in it for too long will build products that do not cross the next cycle.
Firms do not need more AI tools. They need coherence across the tools they already have.
Where value will lodge now
Two things will happen in parallel. Vertical assistants will not disappear, but they will tip into commodity status. Their unit price will fall, their differentiation will fade, their commercial lifespan will shorten. Vendors who have built their identity around a vertical assistant will find themselves in a situation the industry knows well: holding a product that works, but on which the margin compresses, in a market that is already looking elsewhere. And economic value will shift toward the layer that makes the stack coherent, that is to say, toward what MAX calls the Legal Semantic Layer.
This layer is not an additional assistant. It is an infrastructure that makes all assistants usable together. It carries the memory of the matter, the methodology of the firm, the governance of access, the trace of operations. It speaks to all tools without taking their place. It does not replace the drafting assistant, it gives it the context it did not have. It does not replace the research tool, it links its outputs to the matter in progress. It is what turns a stack of tools into a system.
For a serious buyer, the strategic question changes. It is no longer about choosing the marginally better assistant, because the margin is shrinking and the choice made in three years will be regretted. It is about choosing the architecture that will allow any assistant to be used under conditions of operational coherence. This architecture is what lodges between the models and the work. It is what defines the new category, and it is what will capture, in the next three years, the durable value that the assistant race has, so far, wasted on the surface. A firm that picks its assistant today is betting on one product. A firm that picks its architecture today is betting on a whole generation of products, including those that have not yet been written.
The market continues to run on the wrong layer, and that will last as long as the reflex to optimise what one knows how to optimise lasts. The buyers who will look back in five years and feel they made the right decision will not be those who chose the best assistant of 2026. They will be those who chose the layer above it, the one that makes any assistant durable, useful, and accountable inside their firm. That is where the work lives, and that is where the value will lodge.
Choosing an assistant is a bet on a product. Choosing a layer is a bet on a category.
