Architecture

A Semantic Layer, Thirty Years Later

The best way to understand where legal AI is going is to look at a precedent almost no one invokes, and which we believe is decisive.

There is, in the history of enterprise software, a category of moments that recur at roughly generational intervals. Each time, a technology of manifestly new power appears, the industry adopts it with enthusiasm, and for a few years it feels that nothing quite so spectacular has ever happened before. Then, without any official signal marking the moment, the centre of gravity of value shifts. The raw power, which had been the event, becomes the substrate; and something else, built on top of it, inherits the market, the margin, and the durability. That something else carries a name the industry always ends up discovering at least a decade after it first saw it emerge: the semantic layer.

The pattern is regular enough that it can be traced. And it is rare enough that it is forgotten between generations. Today, in the domain of legal work, that moment is opening again. For anyone who has watched the precedent play out, what follows is legible with almost unsettling precision. Provided one has watched the precedent.

1989

In the late 1980s, the big story in enterprise software is not the application. It is the relational database. Oracle, Ingres, Informix, then Sybase and Microsoft, have spent a decade perfecting a data organisation system that outclasses everything that came before. Major global companies pour their operational reality into it: their sales, their customers, their inventories, their payments, their employees. For the first time, the whole of what constitutes a company in the informational sense of the term is gathered in a single queryable place. The power of the object is striking.

And yet, almost no one inside these companies can use it. To ask a question of the database, one has to write SQL, a technical language that assumes knowledge of the exact structure of the tables, the precise names of the columns, the types of joins and the schema conventions. It is a language for specialists. The management controller who wants to know how the margin is behaving in the South-East region during the third quarter does not write SQL. They ask a developer, formulate their question as precisely as they can, wait three or four days, receive an answer that does not quite match what they wanted to know, and start again.

The data is there, the power is there, and the person who has to decide is separated from both by a language they do not speak.

This friction is not a small operational inconvenience. It is a structural blockage. A whole category of work, which consists of interpreting that data in order to decide, is suspended on the availability of a handful of technicians. The questions one would have wanted to ask are not asked, because the cost of formulating them exceeds the expected benefit of the answer. The system, viewed from outside, looks revolutionary. Viewed from inside, it looks strangely unusable.

It is in this period that, in France, an idea is born that will define an entire category of software for the next thirty years. The idea fits in one sentence. Rather than teaching business users to speak the language of the database, one builds a layer that speaks their language to them, and translates. This layer maps business names (revenue, margin, region, quarter) onto the underlying mechanics of tables, joins and keys. The business user no longer writes SQL. They pose a question in their own language, and the layer takes charge of turning it into a query. When the answer comes back, it too is expressed in that language.

The effect of this translation is radical, and it is not only a matter of convenience. When asking a question takes five seconds instead of three days, the questions one asks are not the same. The very nature of analytical work changes. A new category of product is born, which will be called business intelligence. The company that builds this idea is called Business Objects. The person who designs its core is called Jean-Michel Cambot. He does not yet know this on that particular day, but what he has just laid down is the archetype of an object enterprise software will rediscover several times: the semantic layer.

Raw power attracts attention. The layer above captures the market.

2010

Twenty years later, the precedent has replayed in several versions. Each time, the industry discovered with a certain surprise that value did not settle where it had been expected. When the web becomes a vast field of queryable APIs, it is not the providers of those APIs who capture the integration market, but the layers that orchestrate the calls and give meaning to the returns. When the cloud becomes a commodity, it is not the hyperscalers who take the margin on the governance of multi-cloud environments, but layers built on top, which speak the language of security and compliance teams rather than that of datacentres.

In each case, the mechanic is the same. A raw power appears, the industry celebrates it, and then, after a few years, one realises that this power is only usable at the scale of an organisation if someone builds the layer that translates it into the language of the work.

The layer is invisible because it succeeds. Visible, it would have failed.

That layer, because it is invisible and does not shine like the technology it exploits, never receives the press coverage of the layer below it. But it is the one that endures. Databases change; the semantic layer remains. API providers come and go; orchestration layers settle. Hyperscalers compete; governance layers prevail.

What Business Objects showed in 1990, and what the semantic layer continued to show for three decades, is that with each technological generation, durable value moves slowly but resolutely upward in the stack. Power attracts attention, therefore investment, therefore the race for performance. The layer above, more discreet, attracts usage, therefore margin, therefore durability. The original object, acquired by SAP in 2007, still equips a significant share of the world's largest companies today. A fashion lasts five years. A well-laid layer lasts a generation.

2026

The precedent, today, is replaying. This time, the raw power is generative AI. The language models produced by the major laboratories are, objectively, one of the most impressive technological advances of the decade. They write, they reason, they translate, they synthesise with a fluency that would have seemed to belong to science fiction five years ago. The industry celebrates them, rightly. Investment pours in. The performance race is in full swing. Benchmarks are published every three months, and each new model is presented as a generational leap.

And yet, where these models meet real work, the pattern of 1989 can be observed again. Inside a law firm, the power of the models is manifest, but the value of that power remains locked behind a boundary of language. The models speak the language of sessions, prompts, tokens. Legal work speaks the language of matters, positions, precedents, responsibilities, validation chains. Between the two, there is a structural gap that no further progress of the model will suffice to close, because that gap is not a lack of capability. It is a lack of translation.

When a lawyer wants to seriously use a language model for their work, they find themselves in the position of the 1989 management controller facing the relational database. The power is there. The data is there. The methodology they would like to apply exists. But the way the model understands a question is not the way a lawyer frames a problem. The way the model structures an answer is not the way a matter should accumulate. The way the model handles memory between two interactions is not the way an affair evolves across weeks. What is missing, between the models and the work, is what databases eventually received: a layer that speaks the language of use.

Thirty-seven years after the relational database, the lawyer of 2026 lives exactly what the management controller of 1989 lived.

The similarity between the two situations goes further than a simple analogy. In both cases, the raw power is a universal system, designed to serve any use case, and therefore optimised for none in particular. In both cases, real work has a specific grammar, its own vocabulary, requirements that are not anticipated by the universal system. In both cases, the solution is not to make the universal system more powerful still, but to build on top of it a layer that encodes the grammar of the work. And in both cases, the industry begins by failing to see that it is the layer, and not the system, that will define the category.

That layer, in the legal domain, is what is called the Legal Semantic Layer. It is what MAX is building. The conviction driving this project is not an intuition about the future, it is a serious reading of the past. When the same pattern plays out four times in forty years, in four different technological contexts, one is no longer in the realm of betting. One is in the recognition of a historical regularity. The semantic layer for the database lasted thirty years. The semantic layer for APIs is still alive. The semantic layer for the cloud is only beginning to show its value. The semantic layer for legal language models will settle in the same durability range, because it obeys the same mechanic.

Four times in forty years, the same pattern. At this level of regularity, it is no longer a bet on the future. It is a reading of the past.

One question remains: why this regularity? Why does value always move, in the end, toward the layer above raw power? The answer rests on an asymmetry that has not changed in forty years. Raw power, by construction, is universal. It is designed to serve all use cases at once, and therefore it cannot be optimal for any in particular. The layer above, conversely, speaks the language of a specific use. It can therefore encode what a use demands, that a universal system could not, by definition, anticipate. And the more demanding the use, the wider the gap between what raw power offers and what the use requires. Law, by the precision of its vocabulary, the rigour of its methodology and the weight of its responsibilities, is a particularly demanding use. It is therefore a use where the layer will have particularly much to say.

This asymmetry has another consequence, subtler but just as important. Raw power evolves in abrupt leaps, as laboratories release new models. The layer above evolves by accumulation. It learns the methodology of the firm that uses it, it takes on the rules specific to each practice, it accumulates the memory of matters handled. After a few years, this accumulation becomes something that neither another vendor nor a newer model can catch up with quickly. This is what makes the layer endure, where raw power of the moment, despite its brilliance, is only a stage.

The Business Objects moment of legal AI has begun. It will not be announced by a press release. It will not make the cover of technology magazines. It will be recognisable by quieter signs: a gradual shift in conversation, which will stop being about the best model and start being about the best layer; a change in the nature of requests for proposals inside firms, which will stop asking for assistants and start asking for architectures; a silent migration of value, from the visible to the load-bearing. When that migration is complete, the world will say it was obvious. By then, it will be. Today, it is only obvious to those who know where to look.