Architecture

Every Profession Will Produce Its Layer

The pattern taking hold in law will reproduce in every profession of expertise. By necessity, not by chance.

An interesting hypothesis is starting to circulate in serious conversations on the future of enterprise AI, and it is simply put: the pattern falling into place in legal AI will reproduce, at different rhythms but by the same logic, in every sector of expertise. Medicine, finance, engineering, accounting, consulting. Each time, the same sequence: generative assistants in the first wave, rapid fragmentation, a demand for governance that emerges, and the appearance of a domain-specific semantic layer, that makes generic models usable in the complexity of the craft.

This hypothesis is not an arbitrary projection. It follows from a simple architectural reasoning: generic models speak the general language, but no craft of expertise functions in the general language. Each craft has its own objects, its own constraints, its own governances, its own methodologies. And each needs a layer that translates generic capabilities into its own operational capabilities. This layer cannot be horizontal, because it speaks a craft language; it must be vertical, by sector.

No craft of expertise functions in the general language. That is why the layer will be vertical, never horizontal.

A layer per domain

If this reading is right, one will observe over the next five years the emergence of several distinct categories. A Legal Semantic Layer for law. A clinical layer for medicine. A financial layer for finance. And so on. Each will be an infrastructure category in its own right, with its specialized players, its emerging standards, its own craft depth, its specific defensible advantages. The horizontal market of generic assistants will keep existing, but it will be relegated to the status of base layer, on which the vertical semantic layers will rest without reducing to it.

The strength of this hypothesis lies in its resting not on a belief in AI, but on a property of the crafts of expertise themselves. A craft of expertise is defined precisely by the fact that it does not reduce to the general language: it took a practitioner years to master its objects and rules, and it is this very irreducibility that calls for a dedicated layer. The more expert a craft, the larger the gap between the model’s general language and the craft’s language, and the more the layer that fills this gap has value. Law, precisely, is one of those crafts where the gap is maximal.

This segmentation has already begun, and one can observe it outside law. In healthcare, players are building what generic medical copilots are not: a layer that speaks the language of patient records, clinical references, pharmacovigilance rules. In finance, layers emerge around the language of instruments, positions, regulated risks. In law, it is precisely the category MAX builds as a Legal Semantic Layer. The movement is the same everywhere, only staggered in time.

The generic wins first because it is there. The vertical wins next because it is right.

What “speaking the language of the craft” means

The expression “speaking the language of the craft” deserves precision, for it is the heart of the argument and not a convenient metaphor. A generic model knows the words of law, but it does not know the objects of law as they function in real practice: what a matter is for this firm, how a position is built, what distinguishes a useful precedent from a dangerous one, which verifications are non-negotiable. To speak the language of the craft is not to handle its vocabulary, it is to manipulate its objects with the right constraints, which no general model does without a layer to teach it.

This distinction between vocabulary and objects explains why the vertical layer cannot be approximated by a simple tuning of the generic model. One can give a general model detailed instructions, examples, a legal tone; one obtains a model that talks like a lawyer, not a system that reasons within the constraints of the craft. The difference appears the moment the work becomes serious: the well-instructed model produces plausible text, the craft layer produces work conforming to the tacit rules of practice. The first resembles the craft, the second holds it.

This is why the depth of a vertical layer is not made fast. It supposes having encoded, over real usage, a multitude of constraints no one ever writes explicitly because they are self-evident to a practitioner: the expected chains, the implicit verifications, the cases where one departs from the general rule. This encoding is a long, cumulative work, impossible to rush, and it is precisely what makes it a defensible asset: what takes years to build is not caught up in a few months, even with the best model in the world beneath it.

A regularity already observed one level below

This logic of verticalization follows a regularity already observed at the previous layer. Enterprise software began horizontal, with generalist suites meant to serve all crafts, then verticalized sector by sector, because craft depth always ended up prevailing over generality. What was true of management software, customer-relationship tools, analytics platforms will be true of AI semantic layers. The generic serves first because it is available; the vertical wins next because it is right. AI does not depart from this movement, it replays it one level up.

One must understand why this regularity is so robust, for it illuminates what is coming. At the start of a wave, generality is an advantage: it allows serving a broad market fast, without waiting to understand each craft in detail. But as the market matures, users no longer settle for a tool that understands them roughly; they want a tool that speaks exactly their language. Generality, which was an advantage at entry, becomes a handicap at depth, and the vertical, slower to build, ends up winning because it is the only one to reach the precision the craft demands.

One can therefore read the present of legal AI as a repetition, at the upper floor, of a well-known movement. We are at the moment when generality has served its bootstrapping role, and where craft depth begins to claim its due. This moment is precisely the one when vertical layers emerge, before the market has named them, carried by players who bet on depth against the ease of generality. To recognize this moment is to understand why sector layers are not one possibility among others, but the next step already underway.

Generality is an advantage at the entry of a wave. It becomes a handicap the moment the craft claims its depth.

What this imposes on organizations

This dynamic has an important strategic consequence for legal organizations. The semantic layer proper to their craft will not come from the large horizontal model vendors: it will come from players specialized in the craft, who will have built in depth the necessary operational understanding. It is a category apart, and it would be vain to wait for horizontal vendors to do this work, because it is not in their logic to do it.

One must understand why horizontal vendors will not fill this void, for it is the heart of the matter. Building a vertical layer supposes investing years in the fine understanding of a craft: its objects, its methodologies, its tacit rules, its edge cases. This investment makes sense only for whoever devotes themselves to that craft; for a horizontal vendor, serving ten sectors at once, it would be both too costly and too specific. The generality that makes the model’s strength is exactly what prevents it from descending into the depth of a particular craft. The vertical void is therefore not a lag to fill, it is a place structurally left to the specialists.

The right architectural choice of legal organizations is therefore not to wait, but to invest in the semantic layer of their craft at the moment it is being built. To wait would amount to betting that horizontal vendors will end up descending into the depth of law, which their economic model forbids them; it is therefore to wait for a thing that will not come. The emergence of vertical semantic layers is one of the great movements of enterprise software of the next five years, and the place MAX builds in the legal segment is a direct expression of it.

It remains to measure what this reading changes for an organization deciding today. It shifts the question from the choice of a vendor to the choice of a category. An organization still reasoning in terms of “which assistant to adopt” stays in the horizontal layer, the one that will become a commoditized base; an organization reasoning in terms of “which craft layer to build” places itself in the category where value will concentrate. It is not a nuance of vocabulary, it is the choice between building in the layer that rises or staying in the one that commoditizes.

Your craft’s layer will not come from a horizontal vendor. It will come from whoever knows the craft in depth.