Industry
What Does Not Change When the Model Changes
Two years ago, the debate was about one model. Today, about six or seven. For a firm that must decide now, only one question matters: what to stop betting on?
Two years ago, in legal organizations, the debate concentrated on one model. A year ago, on two or three. Today, one hears six or seven vendor names mentioned, and the debate over the best has run out of breath, for a simple reason: the best changes every three months. The raw performance of models now aligns faster than organizations’ purchasing cycles, and everything indicates this dynamic will accelerate rather than slow.
This article does not seek to explain why it is so, nor to draw a general law about the evolution of technologies. It starts from a very concrete situation, that of a firm leader who must, today, decide how to equip their organization, knowing that what they choose may be surpassed within a quarter. It is an uncomfortable position, and increasingly common. The question they ask is not theoretical: it is practical, urgent, and it calls for an answer no model ranking provides.
The best model changes every three months. No equipment decision can be made on a target that moves that fast.
One must take the measure of how unprecedented this instability is for a decision-maker. In most fields, one chooses a tool because it is the best, and it stays so long enough for the choice to make sense. Here, today’s best will be surpassed before the deployment is even finished, and the next will be in turn. The leader does not face a hard choice between several good candidates; they face a choice whose object structurally expires, whatever care they put into it. No diligence, no comparative analysis protects them, because the problem is not choosing well, it is that the choice itself has ceased to be the right level of decision.
What a firm can no longer afford to bet on
The first consequence, for this leader, is that they can no longer afford to bet on a vendor. Not out of excessive caution, but by simple arithmetic: if the ranking is redrawn every quarter, choosing a model today is choosing with a near-certainty of being wrong in six months. The bet on the right model is a bet one is almost sure to lose, not because one chooses badly, but because the object of the bet slips away underfoot at the very moment one makes it.
This impossibility of betting changes the nature of the decision. As long as one model dominated clearly, choosing was a strategic act, committing for years. Today, choosing a model has become an act of procurement, reversible, with no lasting scope: one takes one, one will change it, that is no longer where the organization’s future plays out. The lucid leader therefore stops asking “which model to choose” and starts asking “how to make the choice of model stop being a stake.” It is a complete shift of the question.
And this shift leads them to a new requirement, one they had not formulated as long as they reasoned in models: they must be able to change models without losing anything. To be able to replace the quarter’s vendor with the next quarter’s, or to combine several models by task, without that change costing the slightest regression in what the organization has accumulated. This requirement makes sense only if there is something not to lose, something that lives elsewhere than in the model and survives it. The whole question then becomes: where to house what one does not want to lose?
You no longer choose a model for years. You procure it for a quarter. The decision that lasts is elsewhere.
What remains when everything else turns
To answer, one must look at what, in AI-assisted legal work, does not change when the model changes. And the list is long, far longer than one imagines when the eye is fixed on model performance. Matters do not change vendor. Institutional memory does not get retrained. The firm’s methodologies depend on no model. Permissions, walls, access regimes are proper to the organization. Workflows, governance, traceability exist independently of the model that, at a given moment, generates the text.
What does not change, in sum, is everything that defines legal work as work, and not as mere text generation. The model brings the capacity to produce language; all the rest, what makes a text become a firm deliverable inscribed in a history, governed by rules, traced and defensible, exists outside the model and has no reason to vanish when one changes it. It is this stable whole that the leader seeks to protect, and it is there that they must house their lasting decision.
One then sees a sharp dividing line take shape between two natures of objects. On one side, the model: volatile, interchangeable, replaced every quarter, chosen as one chooses an energy supplier. On the other, the layer where everything that does not get retrained accumulates: stable, proper to the organization, built over time, impossible to buy back at once. The leader who has understood this line stops investing in the first object, which they consume, to invest in the second, which they own.
The model generates the text. Everything that makes a text a firm deliverable lives elsewhere, and has no reason to change with it.
Deciding for the layer, not the model
This distinction transforms the equipment decision. It no longer consists in choosing the right vendor, a bet lost in advance, but in choosing an architecture that makes the choice of vendor indifferent. The right question to ask a legal AI tool is no longer “which model do you use,” but “what happens to everything I have accumulated the day I change models.” If the answer is “nothing moves,” the tool protects the organization; if the answer is “everything must be redone,” it exposes it to the very risk it sought to avoid.
This criterion overturns the way of evaluating. A leader who judges a tool on the performance of its model judges on what will change anyway, hence on nothing durable. A leader who judges a tool on its independence from the model judges on the only thing that will decide the value of their investment in three years. The first optimizes the volatile component; the second secures the stable layer. And over a few years, these are two positions with no common measure, even though they appear equivalent at the moment of purchase.
One must measure how liberating this approach is for a firm. As long as it believed it had to bet on the right model, it lived in the anxiety of being wrong, condemned to restart its choice at each new announcement. The moment it shifts its decision toward the layer that makes models interchangeable, this anxiety disappears: models may well succeed one another, change, surpass each other, that no longer concerns it strategically, because its value no longer depends on them. It watches the model race with a user’s interest, no longer a bettor’s anxiety.
Do not ask a tool which model it uses. Ask it what you lose the day you change.
The trap of the vendor who integrates everything
A leader might object that there is another way: to wait for a model vendor to eventually integrate everything, the layer included, and entrust the whole to them. The idea is tempting, because it promises the simplicity of a single interlocutor. It is also the best way to reconstitute, in another form, the bet one sought to avoid. To entrust the layer to the model vendor is to reintroduce a dependence on that vendor, at the precise moment one had understood that one had to break free of it.
For a layer owned by a model vendor is no longer neutral. It is tied to that vendor, it espouses their interests, it locks the organization into their ecosystem. The day a better model appears elsewhere, the organization can no longer freely benefit from it, because its layer is captive. It has traded the volatility of the model for lock-in with a vendor, which is not progress but a displacement of the problem. The lucid leader wants the opposite: a layer that belongs to them and stays free to plug, beneath it, any model, including that of their current vendor’s competitor.
This is why the independence of the layer is not a technical detail, but the heart of its value. A layer makes sense, for an organization, only if the organization owns and controls it, independently of any model vendor. Failing that, it is not a protection against the volatility of models, it is another form of it, more insidious because harder to undo once one has settled into it. The only layer that truly protects is the one that belongs to none of those it coordinates.
Model rotation as an asset, not a threat
This shift has a counterintuitive consequence: for an organization endowed with the right layer, the rapid rotation of models ceases to be a threat and becomes an asset. Each new model, each improvement, each cost reduction benefits the organization immediately, without it having to redo anything, because its layer knows how to welcome the newcomer as it welcomed the old. Where the poorly equipped firm suffers each announcement as a challenge, the well-equipped firm banks each announcement as a benefit.
It is a complete reversal of the relation to model progress. Without a stable layer, the acceleration of the model race is exhausting: it renders obsolete, every quarter, what one thought acquired. With a stable layer, that same acceleration becomes a flow of free benefits: the faster models progress, the more the organization profits, without effort and without risk. The same market dynamic is experienced as a threat or as a windfall, depending on whether one placed one’s value in the component that changes or in the layer that remains.
Without a stable layer, each new model is a challenge. With it, each new model is a free benefit.
Building the layer that remains
It is precisely this position that MAX occupies: a layer above the models, that makes them interchangeable and makes the organization, not the vendor, the owner of the durable value. MAX does not take sides in the model race; it makes that race indifferent to the firm, by housing in a stable layer everything the firm does not want to lose, and by treating the model for what it has become: a replaceable component, plugged in beneath the layer, never above it.
The leader who must decide today therefore does not need to guess which model will prevail, an exercise in divination no one succeeds at. They need to make a choice of another nature, safer and more lasting: that of the layer in which they will accumulate what does not get retrained. Models will keep succeeding one another at a sustained pace; value, for its part, will remain where the firm placed it, provided it placed it in what remains, and not in what passes.
You do not bet on the component that changes. You build the layer that remains, and watch the models go by.
