Architecture
A Matter Is a Narrative, Not a Stack
Ask a lawyer to describe a matter they have carried for six months. They will not give you a list of documents. They will tell you a story.
Ask a legal professional to describe what, concretely, a matter they have carried for six months is. Their answer will never resemble a list of documents. It will resemble a narrative. The matter began with such an exchange; such a first analysis led to such a strategy; such a decision was taken three months ago for such a reason; since then, such an event changed the orientation; at such a moment, such a position was defended with the client; and it is on this basis that the next step is being prepared. The matter is not a binder, it is a trajectory.
The word narrative is not a decorative image; it describes exactly the form of the matter. Like a narrative, the matter has an order in which things happened, and that order is not indifferent: what was decided before constrains what can be decided after. Like a narrative, it has a causality: each decision answers what preceded it and commits what will follow. And like a narrative, it has a thread the one who carries it keeps in mind, and which gives the individual pieces their meaning. Out of that thread, the pieces are only paper; it is the thread that makes them a matter.
This trajectory is precisely what makes legal work coherent beyond individual moments. Without it, each interaction becomes autonomous again, each decision is made out of context, each deliverable is only an answer to a given instant, detached from what precedes. No serious professional works this way, because no serious matter can be handled this way. And it is exactly the absence of this trajectory, in current AI tools, that makes them so hard to use on real matters, the ones that last and that count.
Legal work has never been a sequence of questions and answers. It is, and will remain, a work of trajectories.
What a tool that sees only a stack sees
To measure the gap, one must look at what a matter becomes when it is stripped of its narrative. A tool that treats the matter as a stack of documents does not see a story; it sees a set of files, which it can index, retrieve, compare, but whose order, causality and thread it does not perceive. For it, the March note and the September note are two equivalent documents; it does not know that the second extends the first, nor that the decision contained in one orients all that follows. It has the matter, but it has lost the form.
This loss is not trivial, because it is in the form that most of the meaning resides. To know all the pieces of a matter without knowing their sequence is like having all the chapters of a novel shuffled: you possess every page, but you have lost the story. A lawyer who resumes a matter does not ask “which documents exist,” they ask “where do we stand,” and that question has an answer only in the narrative, not in the stack. A tool that sees only the stack can answer the first question and stays mute on the second, which is nonetheless the only one that matters.
One can give a concrete measure of what this blindness to narrative costs. When a lawyer reopens, with a tool that has no trajectory, a matter left three weeks earlier, they must reconstruct the whole thread by hand: reread their own notes, recover the order of decisions, recall why such an option was discarded, re-explain to the tool what they already knew last time. This reconstruction work brings no new legal value; it merely repays a debt created by the tool’s inability to hold the narrative. Multiplied by the number of matters and resumptions, this permanent repayment ends up canceling the gain promised by AI.
A tool without a trajectory charges, at every resumption, a context debt it created itself.
Why the narrative is not reconstructed on demand
One might think it enough, at each resumption, to ask the tool to reconstruct the narrative from the pieces: since all the documents are there, the story should be deducible from them. This is an illusion, and it deserves to be taken apart, because it underlies most current tools. A matter’s narrative is not deduced from the pieces, because a large part of what constitutes it is written nowhere: why such an option was discarded, what the client said on the phone that changed everything, the trade-off that left no documentary trace but orients all the rest.
The narrative is therefore richer than the sum of the documents, and this richness is lost each time the tool forgets. To reconstruct the story from the pieces alone is to obtain an impoverished version, amputated of all that was not written, that is, often of the essential. This is why the lawyer who resumes a matter cannot rely on a tool that re-deduces everything: they know that tool will have lost precisely what does not appear in the files, and that it falls to them, the human, to reinject it, again and again.
Keeping the narrative alive, rather than trying to reconstruct it on demand, is therefore the only approach that takes the measure of the problem. It supposes capturing the thread at the moment it forms, not rebuilding it afterward from incomplete traces. A layer that follows the trajectory records decisions and their why as they are made, keeps the thread instead of re-deducing it, and returns at resumption not a tidy stack, but a continuous story. It is a difference of method that decides everything else.
Following the thread rather than filing the pieces
This observation is one of the explicit starting points of MAX’s design. MAX is not a tool that answers questions, it is a layer that follows trajectories. A matter, in MAX, is not a file in the documentary sense; it is a living object, endowed with a history made of exchanges, decisions, versions, trade-offs, and this history is kept active in each new interaction. The user who returns to a matter after several weeks does not find a blank page: they find a narrative that has not been interrupted, that knows where the work stopped, what was discussed, what remained open, what was settled and why.
This choice changes the very nature of the product. MAX does not present itself, at each interaction, as an assistant ready to receive a new isolated request. It presents itself as a layer that continues a thread. It is this thread, the matter’s trajectory, that structures its functioning, and it is this that makes AI truly usable on complex, sensitive, long matters, the ones where current tools quickly prove limited because they treat each solicitation as an event with no past and no future.
One must measure what this difference implies for design. Following a trajectory supposes modeling the matter as an object that has its own memory, states, a chronology, actors, decisions linked to one another. It supposes keeping this narrative alive from one interaction to the next, making it available without the user having to reconstruct it, updating it as the matter advances. It is engineering of an entirely different order from that of a tool that merely files and retrieves pieces, and it is precisely this engineering that separates a demonstration AI from a real-work AI.
Following a trajectory is not one more feature. It is taking the work seriously as it is.
There is, in this choice, a consequence for the trust granted to the tool. A professional hesitates to entrust a sensitive matter to an AI that, by construction, loses the thread of what played out in it: they know they will have to remain the sole guardian of coherence, and that the tool will only assist them in fits and starts. A layer that follows the trajectory changes this relation: it becomes itself the keeper of the matter’s coherence, and the lawyer can lean on it rather than watch it. It is this inversion, from the watched tool to the tool one leans on, that opens AI use on the matters that count.
This way of conceiving legal AI follows from a simple stance: take the work seriously as it is, and not as it would be convenient for the technology. An AI that cannot follow a trajectory is not a serious legal AI, whatever the occasional quality of its answers. It is on this principle that the MAX Legal Semantic Layer was built, and it is this that separates an AI one tries from an AI one truly works on.
An AI that does not follow a matter’s trajectory stays brilliant in flashes, and unusable over time.
