Product
Why Every New Model Makes MAX Stronger
When you build on top of AI, you always hear the same objections. Here they are, and here is why they invert when you build not an assistant, but a layer.
Every investor who meets an AI startup asks, at some point, the same question. What happens when the next model absorbs what you do? This question is legitimate. It has, in fact, killed several companies over the past two years, which thought they had built a defensible product and found themselves competing head-on with a model they could not match. For MAX, this question comes up regularly, and it deserves a serious answer, not a reassuring formula.
The honest answer is that the question inverts depending on what one builds. For a product that lives in the same layer as the models, the risk of absorption is real and permanent. For a layer that lives above the models, the same progress that would threaten an assistant becomes an asset. The difference is not a commercial argument, it is a consequence of position. And the best way to demonstrate it is to take the most serious objections one hears, one by one, and look at what they become when applied to a layer rather than an assistant.
OBJECTION. The next model will absorb what you do, the way GPT-4 absorbed a whole generation of tools.
RESPONSE. Absorption happens when a product does something that the model, as it improves, eventually does better natively. A better drafting assistant is absorbed by a model that drafts better. But MAX does not draft, nor summarise, nor reason. It orchestrates the best available models, holds the memory they structurally lack, and governs the execution they cannot govern alone. No improvement of the model will give it these properties, because they are not in the model's layer. When a better model arrives, MAX simply has a better engine to orchestrate. The layer does not have to be rebuilt. The memory it holds is still valuable. The governance it enforces is still required.
MAX does not run in the model race. It sits in the stand above the track and decides which runner to use.
OBJECTION. OpenAI or Anthropic will add memory and orchestration to their model, and you will be short-circuited.
RESPONSE. This is the most serious objection, and it deserves a precise answer. Model labs can add, and do add, memory and orchestration functions. But their memory is universal, designed to serve all use cases. The memory a law firm requires is specific: structured around the notion of the matter, governed by the professional rules of the bar, traceable to meet European regulatory requirements, and organised according to the firm's particular methodology. No generalist lab will build this specificity, because it optimises for the average of its uses, and the average does not need this precision. The specific layer is exactly what the generalist lab cannot build at a reasonable cost, and that is precisely what MAX is.
OBJECTION. Your value depends on the models. If access conditions change, you are dead.
RESPONSE. This objection actually inverts our position. Because MAX owns no model and is bound to none, it is us, and with us the user firm, that are not locked into a single provider. The value the firm builds up, its encoded methodology, the memory of its matters, its governance, lives in the layer, not in the model. If a provider changes its conditions, one substitutes the model without losing anything. This substitutability is not an option added to the product, it flows from the architectural position of the layer. By contrast, a product that has trained a specialised model, or that depends on the particularities of a model family, is locked in. And it is they who die, not us, the day access conditions change.
A layer that depends on no single model leaves the firm depending on no single company. Independence is an architecture, not a promise.
OBJECTION. If your value is in the layer, anyone can copy it in six months.
RESPONSE. The layer, seen from afar, seems reproducible. Seen up close, it is not. The reason lies in what accumulates inside it. A legal semantic layer is not a static program. 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, the trade-offs, the positions held, the firm's quality standards. After a few years, this accumulation becomes an asset that neither another vendor nor a newer model can rebuild quickly, because it is the product of years of use under specific conditions. The layer is easy to sketch in six months. It is very hard to rebuild after five years of accumulation at the client.
OBJECTION. You still depend on the evolution of the models. If the fundamentals change, your architecture is obsolete.
RESPONSE. This objection assumes that the layer is designed for specific models. It is not. The layer is designed to translate legal work, and that translation remains valid regardless of how the models work underneath. When models evolved from recurrent architectures to transformers, and then to the mixed architectures we see today, the semantic layers above databases were not affected, because they did not depend on the internal particularities of the databases. They depended on the grammar of business data, which had not changed. The legal semantic layer depends on the grammar of legal work. That grammar does not change every six months. It does not even change every ten years. The notions that structure how a contract is written, how an argument is built, how a position is held, how a matter is handled, have remained stable across generations of practitioners. It is this stability of the legal substrate that ensures the durability of the layer, independently of the underlying technical cycles.
What happens when a better model arrives
For a product that competes with the models, the arrival of a new model is a threat to be survived. For MAX, it is simply a better engine to orchestrate. The same dynamic played out with the semantic layer over databases a generation ago. Faster databases, new query engines, entirely new storage architectures came and went underneath the semantic layer, and the layer endured through all of it, because it was never in competition with the thing underneath. It was the translation between that thing and the people who needed it. The engine improved; the layer remained the point of contact.
MAX occupies exactly this position in the world of legal AI. The pace of AI progress, then, is not the threat to MAX that it would be to a model-bound product. It is the tailwind. Every improvement below the layer is an improvement the layer can use, and every cycle that obsoletes a model-bound competitor leaves the layer exactly where it was: above the models, holding the work together.
The engine improves. The layer remains the point of contact. That is what it means to be built to last in a field that changes monthly.
The inversion between threat and asset is not a communication trick. It is the direct consequence of the decision not to compete with the models, and to place oneself deliberately above them. This decision has an immediate cost, which has to be paid to buy the right to occupy the position. But once the position is occupied, it durably protects against the dynamic that destroys products that made the other choice. It is precisely where MAX's defensibility is built, and it strengthens with every new model, because every new model confirms that durable value does not lie in the race for capabilities, but in the position above it.
For the firm choosing today which product to rely on, the strategic question reformulates in the light of this inversion. Choosing a product that depends on a model is betting on the trajectory of that model. Choosing a layer above the models is betting on the stability of one's own profession, which is a much safer bet. Models will continue to change, at a pace no one predicts with precision. Legal work, by contrast, is structured by notions that have lasted for centuries. It is this stability that makes the layer defensive, and it is this defensiveness that makes the firm's investment durable across the cycles to come.
This reasoning, seen from the firm's side, is also the one any serious investor looking at the legal AI market will run. The question they ask themselves is not which of the current companies is the best today. It is which one will cross the next three model cycles without having to rebuild itself. Products that depend on a particular model, or that position themselves as assistants in direct competition with the native capabilities of the models, fail this test, because their commercial lifespan is mechanically tied to the lifespan of their category. A layer above the models passes this test, because its value does not depend on the model it leans on at a given moment, but on what it has accumulated for the firm that uses it. This property is what makes MAX defensible in the eyes of a demanding investor, and it is what makes the choice of MAX rational in the eyes of a demanding buyer.
The models are replaceable. Governed legal execution is not.
