Architecture
When Coordination Takes Production’s Place
With AI, part of the application logic moves down into the model. Value, in turn, moves up to the layer that coordinates. It is a change of floor, not of degree.
For thirty years, the application layer was where most of enterprise software’s value concentrated. It was what organizations sought, compared, bought: the application that produced the document, the calculation, the analysis. People reasoned in applications because that is where the work was done, and because producing that work was hard. The difficulty of production justified a whole market of specialized tools, each expert in its task. This obviousness is shifting, and the shift goes almost unnoticed because it does not look like a rupture. It looks like a mere improvement of the tools. It is something else entirely, and mistaking it for incremental progress leads to investing where value no longer is.
The gesture the model just swallowed
With generative models, a significant share of what made up the classic application logic moves down a notch, inside the model. Producing a text, analyzing it, transforming it, synthesizing it, reformulating it: these gestures, which yesterday justified the existence of dedicated applications, become base capabilities, available everywhere, almost mundane. It is not that applications disappear; it is that what they did specifically, production, stops being their distinctive value. When any model can draft a correct clause, knowing how to draft a clause is no longer what distinguishes a product, nor what justifies paying for it.
One must measure what this removes. Production was the primary justification of the application layer; it was the reason one bought one piece of software rather than another. By moving it down into the model, AI empties that layer of what made it rare. The work still gets done, better and faster; but the place where it gets done is no longer the place where value sits. It is a silent displacement, and it is that silence which makes it dangerous: nothing, in daily experience, warns that the center of gravity has moved. The tool still works, it even works better, and that is precisely why one does not see that it has ceased to be the place that matters.
When the model produces, the application’s value no longer lies in production. It lies in coordination.
This shift also changes the nature of competition. As long as producing was hard, a vendor could distinguish itself by the sole quality of its production, and defend that position for years. The moment production becomes a common capability, available to all at the same cost, that distinction evaporates: two tools now draft as well as each other, and the client has no reason to prefer the one that can only draft. The advantage moves to the one that can also hold the context, chain the steps, keep the matter’s memory. Competition no longer plays out on the quality of the gesture, but on the ability to inscribe it in a coherent whole.
Because producing a fragment was never enough. A firm does not live off a single well-drafted clause, but off the coherent chaining of hundreds of gestures: finding the right context, applying it to the right matter, checking consistency with positions already taken, sequencing the steps, respecting permissions, keeping track of who decided what. Each of these gestures was, until now, masked by the difficulty of production: as long as drafting took most of the time, one did not see that coordinating took just as much. Now that production becomes easy, the rest stands exposed, and it is the rest, henceforth, that is the real difficulty. The bottleneck is no longer producing the piece; it is holding the whole in which the piece makes sense.
Value has changed floors
This coordination has a precise architectural name: orchestration. And it is becoming the new application layer, where the value that production just vacated concentrates. To orchestrate is not to produce better; it is to hold together what production, alone, leaves scattered. It is the function that decides which model to call, with what context, in what order, under what rules, and that keeps memory of the result for the next step. Where the application produced an answer and stopped, orchestration holds a process and continues it, from one human intervention to the next.
This shift has very concrete consequences for the buyer, and they are costly for whoever does not see them. The organization that invests in AI still thinking in terms of applications, one tool for this, another for that, accumulates producers where it needs a coordinator. It ends up with a collection of tools, each capable in its corner, and no one to hold the whole. Each tool answers a question; none carries the matter. And the phenomenon worsens over time: the more tools it adds to fill the gaps, the more the coordination burden none of them assumes falls back on the teams, until the cost of that manual coordination exceeds the gain brought by the tools themselves.
We still judge each tool separately, at the moment when only what they do together matters.
An example makes this cost tangible. An organization deploys one tool for search, another for drafting, a third for review, a fourth for synthesis. Each, taken alone, does its work well and impresses at the demo. But between them, nothing flows: the context established in the first does not pass to the second, the position settled at review is not known to the drafting tool, and it is the lawyer who, at every transition, recopies, reformulates, re-explains. The stitching work between the tools, invisible in each demonstration taken separately, becomes in use most of the load. The organization bought four producers and inherited a fifth role, unbudgeted: that of coordinator, held by hand by its own teams.
This confusion is not a reasoning error, it is an effect of inertia. For thirty years, buying software well meant choosing the right applications, and that reflex was rewarded long enough to become second nature. It survives the change that invalidates it. But buying applications in a world tipping toward orchestration is investing in the layer the market is precisely leaving, and reproducing at great expense a logic that has ceased to be the right one.
One might think a sufficiently ambitious tool will eventually absorb this coordination function by widening. That is an illusion, and it is worth taking apart. A tool that expands stays a tool: it adds capabilities, it does not change nature. Coordinating requires standing above what one coordinates, having the whole in view, which none of the coordinated elements can have from its place, however vast it has become. A producer that grows stays a producer; it does not become, by accumulating functions, the layer that overlooks it. This layer is built as a dedicated infrastructure, with its own primitives, matters, memory, rules, and that is exactly where the MAX Legal Semantic Layer sits: not one more producer in the collection, but the layer that makes everything that produces hold together.
The next decade will be built where it has always been built: at the point where coordinating becomes more structuring than producing.
