Product
The Choice of Primitive Decides Everything
A product’s primitive, the unit around which everything organizes, determines what it will be able to do and what it will never be able to do. Legal AI chose the prompt. MAX chose the workflow.
In product development, one choice says more about what a product will become than all the feature lists combined: the choice of primitive, that is, of the elementary unit around which everything else organizes. This primitive determines what the product will be able to do naturally, and what it will never be able to do except through contortions. And the primitive chosen by nearly all current legal AI is the prompt.
What a primitive is, and why it decides everything
The word primitive deserves a pause, because it is in understanding what it covers that one grasps the stake. A product’s primitive is its unit of thought, the elementary brick in terms of which all the rest is conceived. Every product has one, whether chosen consciously or not, and that choice is the most structuring of all because it precedes and conditions every other. You can change almost everything in a product, its functions, its interface, its performance; you do not change its primitive without rebuilding it, because it is what everything else rests on.
This is why the choice of primitive is not caught up by adding features. Once the primitive is set, it draws a horizon: what is in its natural extension is built easily, what is foreign to it is built only through detours that end up hitting a limit. Two products that start from different primitives will never truly converge, however many functions are added to them, because they do not organize the world the same way. The choice of primitive, often made early and without measuring its reach, therefore decides what the product will be able to become years later.
The deceptive primitive
The prompt is an elegant primitive: a question, an answer, a short cycle that lends itself to experimentation, demonstration, fast iteration. It is this elegance that explains its massive adoption: it is easy to understand, easy to show, easy to iterate. But it has a structural characteristic that makes it, for legal work, a deceptive primitive: it supposes the user knows, at each interaction, how to formulate what they need. The prompt performs when you can ask your question well. It becomes inoperative when the question is itself part of the problem.
Yet in real legal work, formulation is rarely the starting point. What presents itself first is not a well-posed question, it is a workflow: a set of steps to chain in order to produce a deliverable, with dependencies among them, time constraints, validation rules, human contributions at different moments. The question, in a workflow, is only a local moment. The real subject is the chaining, and it is precisely this chaining that prompt-centered tools do not take charge of.
This limit of the prompt is not a quality defect a better model would correct. It owes to what the prompt takes charge of, and to what it leaves out. It takes charge of translating a clear intention into an answer; it leaves out everything that precedes the clear intention, that is, the essence of legal work: deciding what to do, in what order, under what constraints, mobilizing who and what at each step. A more powerful model will answer the posed question better, but it will not pose the question for you, and it will not chain the steps the question did not name.
A better model answers the question better. It does not decide, for you, what the next step is.
One must see that this is not a nuance, but a frontier. As long as the need lets itself be formulated as a question, the prompt suffices, and is even excellent. As soon as the need is a process, whose formulation is discovered while advancing rather than preceding the action, the prompt ceases to be the adequate tool, not because it answers badly, but because the thing to do is not to answer. This is where most current tools meet their ceiling, and that ceiling is not in the model, it is in the primitive.
The prompt assumes you know how to ask. Legal work often begins where you do not yet.
Why the market chose the prompt
If the prompt is a primitive so poorly fitted to legal work, one may ask why it imposed itself so massively. The answer is not that it was chosen for its merits on real work, but that it was chosen for its merits on something else: ease of construction and demonstrative force. Building a tool around the prompt is fast; showing it is spectacular; iterating it is easy. For a vendor who must ship fast and convince in a meeting, the prompt is the primitive of least resistance.
This default choice has a consequence few measure at the moment they make it: it locks the product into a category before anyone has even thought about what it should do. The prompt is adopted because it is the standard, because it is what others do, because it is what demonstrations expect, and one inherits, without having wanted to, all the limits this primitive imposes. The market did not so much choose the prompt as slide into it, and that is why it hits, collectively, the same ceiling today.
Making another choice therefore requires going against the current, and accepting its cost. Choosing the workflow as primitive means giving up the prompt’s ease of construction, accepting less immediate demonstrations, investing in a coordination engineering the prompt does not require. It is a slower and costlier choice at the outset, justified only if one aims at real work rather than the demonstration effect. But it is precisely because it is costlier that it is defensible: what you build on the right primitive, others will not obtain by improving the wrong one.
Changing the primitive changes the product
It is this observation that determined one of MAX’s most structuring design choices: the primitive around which everything organizes is not the prompt, it is the workflow. A matter triggers a workflow. A workflow contains steps, which can call models, sources, human collaborators, validation rules, different permissions. The semantic layer knows how to coordinate this chaining, hold the context from one step to the next, apply the rules between steps, produce the corresponding traces. The prompt still exists, it remains useful at certain moments, but as a local component, never as a structuring primitive.
This choice profoundly changes what the tool can do, and one must see that it is not an improvement but a change of category. An AI centered on the prompt will always remain, structurally, a tool for assisting one-off tasks: you can improve it indefinitely, it will not become, by accumulation, able to carry a process. An AI centered on the workflow can, without a change of architecture, take charge of complex, long, multi-step, multi-actor processes, with operational coherence preserved from one end to the other. They are two families of objects, not two levels of quality.
You do not improve a prompt tool into a workflow tool. You change the primitive, hence the category.
One can illustrate the reach of this choice with an example. Preparing an opinion on a transaction is not a question, it is a sequence: frame the request, identify the relevant sources, check their currency, confront the positions, draft a first version, submit it to review, integrate the feedback, have it validated by a senior, archive with the trace. A prompt tool can help on each of these steps taken in isolation, provided you formulate it. A workflow tool knows the sequence, knows where you stand, prepares the next step, applies the rules between steps, and does not force the user to replay the conductor at every transition.
The workflow organizes human judgment instead of removing it
It must be added that this choice does not oppose the workflow to human judgment, it organizes it. In a workflow-centered layer, the points where a human decision is required are not interruptions suffered, they are foreseen, situated, traced steps. The workflow does not take the hand away from the lawyer; it gives the hand back at the right moment, having prepared all that could be prepared before. It is the opposite of blind automation: an orchestration that reserves for the human what is theirs and takes charge of the rest.
This distinction matters, because it is often objected that structuring work into a workflow would rigidify it, removing from the lawyer the suppleness of their judgment. The opposite happens. Unstructured work leaves no more room for judgment; it drowns it under the burden of holding everything oneself, recalling each step, forgetting nothing. By carrying the coordination, the workflow frees the practitioner’s attention for what truly requires their judgment, instead of scattering it on the mechanics of the chaining. Structure does not oppose the freedom to judge; it creates its conditions.
The workflow does not take the hand away from the lawyer. It gives it back at the right moment, the rest being already prepared.
Building MAX around workflows rather than prompts was not a marketing choice, but the reflection of a reading of the work: legal work happens in chains, not isolated interactions. A product that does not respect this reality cannot durably serve the work, whatever the performance of its components taken separately. You can have the best model in the world and remain useless on a real matter, for lack of knowing how to chain the steps that lead to the deliverable.
It is on this primitive, and not on a visible feature, that the difference between a surface legal AI and a structural legal AI is built. The first shines on the isolated question; the second holds the entire process. And it is the process, not the question, that defines legal work.
A surface AI answers the question. A structural AI drives the process to the deliverable.
