Architecture
The Ceiling Is Not Technical
Every deployment follows the same curve: rapid rise, then stagnation. The ceiling is not technical.
Every serious legal AI deployment knows the same curve. The first months, usage progresses fast: users discover, the first time savings are visible, feedback is positive. Then, around the sixth or eighth month, the curve flattens. Usage does not collapse, but it no longer progresses. Users use the tool for what they master, and return to their methods for the rest.
This stagnation is a market phenomenon, not an individual failure. It repeats in almost every equipped organization, with a regularity that excludes particular explanations. When the same phenomenon recurs everywhere, whatever the tool, the organization, the sector, it has a structural cause, not a local one. It therefore deserves to be approached by elimination, methodically, because the cause one retains determines the solution one seeks, and a wrong cause sends one looking in the wrong place.
A plateau that repeats everywhere, whatever the tool, does not have a local cause. It has a structural one.
Eliminating technical explanations, one by one
Let us begin with the most common explanation: the stagnation would come from the quality of the models. If that were the case, it would resolve with new versions, more powerful. Yet new versions arrive, regularly, markedly better, and the curve stays flat. The ceiling does not move when the model improves, which suffices to rule out the hypothesis: if the cause were model performance, a better model would raise the ceiling. It does not raise it.
Second explanation: the stagnation would come from a training deficit. If that were the case, it would resolve with support programs. Yet these programs exist in serious organizations, often ambitious, and the curve stays flat. One can train the teams intensively and see the plateau persist, which rules out this second hypothesis. The problem is not that people do not know how to use the tool; they know perfectly, and it is precisely knowingly that they stop extending it.
Third explanation: the stagnation would come from the interface, the ergonomics, the product. If that were the case, it would resolve with better products. Yet products improve, become smoother, more pleasant, and the curve stays flat. One by one, the easy explanations fall, and they all fall for the same reason: they target technical causes, when the ceiling resists all technical improvements. When one has eliminated each technical cause and the ceiling remains, the conclusion imposes itself: the ceiling is not technical.
When all technical causes are eliminated and the ceiling remains, the ceiling is not technical.
What the stagnation really means
What the stagnation means is simpler, and deeper. The easy uses have been absorbed. In the first months, users carried AI onto everything it could hold without difficulty: one-off searches, first drafts, summaries, peripheral tasks. Those uses are now acquired, and there remains, beyond the plateau, only the difficult uses, those that would ask AI to hold what current tools do not hold.
To go further, AI would indeed have to integrate into contexts current tools do not hold: complex ongoing matters, sensitive matters, productions where traceability is critical, chains where coherence between versions counts. It is precisely there that users withdraw. The plateau is not the sign that they stopped progressing; it is the sign that they reached the limit beyond which the tool, as it is, cannot follow them.
And they withdraw not from distrust, but from operational lucidity. They know the tool will not hold that context, so they do not take it there. This point deserves emphasis, for it inverts the usual discourse on adoption. One often presents the withdrawing user as a brake, a laggard to convince. It is the opposite: they correctly assessed that the tool would not hold their complex matter, and they act as a prudent professional. The plateau is made of thousands of these lucid withdrawals, not of a lack of enthusiasm.
Users do not abandon AI on serious subjects out of fear. They abandon it because they are right.
Treating an architecture problem as such
This rereading changes the nature of the problem, and thus the solution. As long as one treats the stagnation as a buy-in problem, one persists on training, communication, change management, that is, on remedies aimed at users’ will. But the will is not at issue: users would like to extend usage, they cannot, because the tool does not hold the context. The moment one treats the stagnation as an architecture problem, one stops trying to convince people and builds what they lack.
What is lacking is not a better tool, but a layer above the tools that makes possible what they do not make possible separately: continuity, memory, methodology, traceability, governance. It is this layer that would let the user finally take AI onto their complex matter, because it would hold, in their place, the context the tool alone does not hold. The unblocking will therefore come not from below, from a more capable model, but from above, from a layer that makes the context holdable.
This reading has an immediate, discriminating practical consequence. Organizations that wait for models to improve to relaunch adoption will wait a long time, for that is not where the unblocking will come from: as we saw, the ceiling does not move when the model progresses. Those that build the orchestration layer will open the sequel without waiting for the next version of anything. Faced with the same plateau, some wait in vain, others act in the right place.
To wait for a better model to raise the ceiling is to wait where the unblocking will never come.
Why the withdrawal is a signal, not a brake
One must pause on the nature of this withdrawal, because the way one interprets it decides everything one will do next. A user who withdraws AI from their serious matters emits a precious signal: they indicate exactly the border where the tool ceases to hold. This signal, correctly read, is a map of the problem; misread, it becomes a pretext to blame the teams. The difference between the two readings separates the organizations that will progress from those that will bog down.
Read as a signal, the withdrawal says where to build. It designates, matter after matter, the contexts the tool does not hold, and thus what the missing layer will have to hold. An organization attentive to this signal has a free roadmap, drawn by the lucid choices of its own practitioners: where they do not take AI, there is exactly the work the layer will have to make possible. The withdrawal is not a problem to correct, it is information to exploit.
Read as a brake, on the contrary, the withdrawal sends the organization in the wrong direction. One takes it for resistance, one responds with persuasion, one relaunches training, one multiplies incentives, and nothing moves, because one treats a lucid symptom as if it were reluctance. Meanwhile, the real cause, the absence of a layer, stays intact, and the plateau with it. To confuse the signal with a brake is to condemn oneself to persisting on the people when the problem is in the architecture.
What the second wave would be
It is worth describing what the sequel would concretely be, for it does not resemble what we have known. The first wave of adoption was horizontal: many users, simple tasks, a broad but superficial use. The next will be vertical: the same users, but on their most complex, longest, most sensitive matters, those they still handle entirely by hand today. The unit gain there is far greater, because that is where the value of legal work concentrates, and it is precisely there that current tools do not go.
One then understands why the current stagnation is misleading. Seen from afar, it resembles a running out of breath, a disappointment, a sign that legal AI has reached its limits. Seen up close, it is the opposite: it marks the end of the easy phase and the threshold of the phase that counts. Organizations that interpret the plateau as a ceiling will reduce their ambitions at the wrong moment; those that interpret it as an architectural threshold will build what opens the sequel. The same flat curve leads to two opposite decisions, depending on whether one reads it as an end or as a threshold.
This is also why the usual adoption metrics mislead. Measuring the number of active users or requests per day describes the first wave, not the second. The right indicator is not how many people use AI, but on which matters they agree to take it. As long as AI stays confined to peripheral tasks, adoption can seem broad while remaining superficial. The depth of use, not its breadth, is the true sign that the missing layer has been built.
The right adoption indicator is not how many people use AI. It is on which matters they agree to take it.
A ceiling that does not rise by waiting
The current ceiling is therefore not a capacity ceiling, but an architecture ceiling. And an architecture ceiling does not rise by waiting for the next model; it rises by building the missing layer. It is a decisive difference, because it separates two attitudes before the same plateau: waiting for a progress that will not come from there, and building what really unblocks. The organizations that have understood it will not wait; they will build.
This is exactly the layer MAX was designed to carry: a Legal Semantic Layer that lets AI operate where, today, serious users do not take it. Not to add a capability to the tools, but to hold, above them, the context that makes their capabilities finally usable on the core of the work. It is this passage that opens the second wave of adoption, the one where AI ceases to be an accessory of easy tasks to become an instrument of the work that counts.
One must, finally, put the stagnation back in its true place. It is not proof that legal AI has failed; it is proof that it has exhausted what the first generation of tools could give. The plateau is not a wall, it is a step: the one that separates superficial use from deep use, and that one does not cross by pushing the tools, but by laying the layer above them. The organizations that see the plateau for what it is, a threshold and not an end, will be the ones that open the sequel.
The plateau is not a wall, it is a step. One crosses it not by pushing the tools, but by laying the layer above them.
