Architecture
Three Levels, and Only One That Counts
You do not guess the 2028 legal AI stack, you deduce it. Three levels, and only one will concentrate value.
No one can predict with certainty the exact form the legal AI market will have in three years. But by looking carefully at what is falling into place today in mature organizations, one can describe with reasonable confidence the structure of the legal AI stack in 2028. Not because it would already be built, but because the constraints that determine it are already all present, and no visible force opposes their consolidation. To project here is not to guess; it is to extend lines already drawn.
In 2028, the legal AI stack of a serious organization will comprise three clearly distinct levels, and their hierarchy of value will be the inverse of the one imagined today. What impresses most now will count least then, and what goes unnoticed today will become the decisive asset. It is this inversion, more than the description of the levels, that is the real subject, because it commands the decisions an organization has an interest in making today.
You do not guess the 2028 stack. You extend lines all the previous waves have already drawn.
Three levels, from the bottom up
At the lowest level, several foundation models. Not one: several, chosen and combined by use, substitutable from one quarter to the next without traumatizing the organization. These models will be, essentially, perceived as commodities, negotiated at market price as cloud hosting is today. No firm, no legal department will build its strategy around the performance of a particular model, because that performance will change too fast and too little to justify a structuring investment.
At the intermediate level, a semantic layer proper to the organization. It is this level that does not really exist today, and that constitutes the great worksite of the next eighteen months. This layer will hold the memory of matters, the methodology of the firm or department, the permissions, the governance, the orchestration of workflows. It will speak the language of the legal craft, survive the changes of models beneath it, and become, in the intellectual and operational balance sheets of organizations, their true AI asset.
At the highest level, the everyday tools, messaging, editor, document management, matter management, which will not outwardly resemble AI products, but will discreetly integrate, in the background, the capabilities of the semantic layer. The user will not perceive that they use AI; they will perceive that they work better, faster, more in coherence with their organization. AI will have ceased to be an object one manipulates to become a property of the work environment.
It is worth pausing on why one can describe this stack with such assurance, when it does not yet exist. The method is not clairvoyance, it is the observation of regularities. At each wave of enterprise software, the same three-level structure imposed itself: a base technology that commoditizes, an intermediate layer that speaks the language of the craft and concentrates the value, surface tools that make it accessible. The mainframe, the database, the cloud all followed this pattern. There is no serious reason to think AI will escape it, and many reasons to think it will follow it faster.
The three-level structure is not a prediction. It is the regularity all the previous waves have already followed.
It will be objected that models resemble nothing we have known, and that their power could suffice to absorb everything, including the layer above. It is the argument heard at each wave, about the technology of the moment, and it proves false each time for a constant reason: a general technology cannot, by construction, carry the specific context of a particular organization. The more powerful and general a model, the more it needs, above it, a layer that specializes it. The power of the model does not remove the need for the layer, it makes it more glaring.
Why the power of models changes nothing
One must take seriously the objection that models could absorb everything, for it is sincere and widespread. Those who hold it observe, rightly, that each new generation of models does what seemed, the previous generation, reserved for a specialized layer. They conclude that the intermediate layer is only a temporary palliative, destined to disappear when models are powerful enough. The reasoning seems solid, and it is nonetheless false, because it confuses general capability with particular context.
A model, however powerful, knows law in general; it does not know that organization’s matters, its past decisions, its methods, its access rules, its own way of handling files. This context is not a question of power, it is a question of belonging: it belongs to the organization, it lives in its history, and no model trained on the whole world can contain it, because it is not in the whole world, it is in that organization. The intermediate layer is not there to compensate for a weakness of models; it is there to carry what, by nature, cannot be in a model.
This is why the progress of models, far from making the layer superfluous, makes it more necessary. The more capable a model, the more it can do with the context it is given, and thus the more the quality of that context becomes decisive. A mediocre model wastes a good context; an excellent model fully brings it out, provided a layer supplies it. The growing power of models increases the yield of the layer, it does not remove it. The two levels are not competitors, they are complementary, and the progress of one raises the value of the other.
The only level that does not commoditize
Of these three levels, only one is neither a commoditized model nor a surface tool: the intermediate level. This is why it will concentrate, as at each previous cycle of enterprise software, the bulk of the durable value. The model is replaced, the tool is changed, but the semantic layer accumulates the organization’s own context, and it is this accumulation that one does not reproduce by changing vendors. The hierarchy inverts: what impresses today will be worth least, and what goes unnoticed today will be worth most.
One can draw from this structure a practical consequence for today’s decisions. If the durable asset constitutes itself at the intermediate level, then investing massively in the bottom level, by attaching oneself to a model, or in the top level, by multiplying surface tools, amounts to investing in what will commoditize. The right placement is neither the model nor the tool, but the layer between the two, the one built slowly and not bought back. It is counterintuitive today, because the model is what impresses and the tool what is seen; but it is precisely what is seen least that will count most.
It is exactly this three-level stack that MAX builds as a Legal Semantic Layer. Not by betting on an uncertain future, but by building today what will be, in three years, the intermediate level, the only one around which the durable asset will constitute itself. To build this level now is to arrive at 2028 with a layer already constituted, when others will arrive with a stack of tools to reconfigure. The difference will not be caught up in a few months, because a semantic layer is the product of time, not of a purchase.
One must clearly see what it means, concretely, to arrive at 2028 without this layer. It means having excellent models and excellent tools, but with nothing linking them, no common memory, no shared governance, no overall coherence. Each tool will be capable in its corner, and the organization will spend its time piecing together by hand what its tools ignore about one another. The individual performance of the components will not buy back the absence of the layer that holds them together, just as a collection of very good musicians does not make an orchestra without a conductor or a score.
In 2028, the incoherence of a stack without a layer will no longer be bought back by each tool’s performance. It is what will separate those who saw it coming from the others.
