Originally written in French. Translated by AI — the meaning has been preserved, not the prose.
Angle
A company information system holds the objects — the ticket, the customer, the date, the status — and loses what surrounded them: why this ticket mattered, which decision it answered, what it will have taught. That gap stayed bearable as long as the reasoning lived in the heads of the teams; it becomes the limiting factor as soon as an assistant is plugged into the material, since a model cannot restore a qualification nobody wrote down. The shift now opening therefore does not consist in replacing the applications, but in writing beside them what they were never able to carry, and the hard point is not the platform: it is the governance of that material and the daily upkeep work that no organization today counts as work.
Synthesis
The usual diagnosis is a filing diagnosis: what is missing would be a better tree structure, a better nomenclature, a better search engine. That diagnosis misses, because finding better what was never written yields nothing. What is missing is not findable — it does not exist.
The cause is old and involves no negligence. Tables, forms and pages were chosen because things had to be stored, indexed and stabilized; they are answers to machine constraints, and nobody thinks in tables. Every entry is therefore a translation from a connected thought into a structure that is not, and the translation loses something every time — silently, since what remains is accurate.
The arrival of AI inside the tools does not correct this. An assistant in the CRM, a chatbot in support, a summary in the ticketing system speed up entry and reading without changing what is preserved: the information stays in the form and within the borders where it was. That is a gain, and it is that gain which serves as an alibi, because it can be demonstrated in a meeting when the underlying work cannot.
The real shift consists in organizing the material around contexts — product, support, customer, source code, standard — that cut across the applications instead of replacing them. Each gathers around a purpose its sources, its rules, its decisions and its limits, drawing on several tools. These contexts do not merge: support has no need of the product's memory, management no need of the detail of execution. What replaces centralization is connectability — a product context that draws on support's signals, a source code context that serves as the reference for regenerating documentation.
Two conditions decide what follows, and neither is technical.
The first is governance, and it works against intuition. AI is expected to dispense with structuring; it is AI that makes structure decisive. An assistant dropped onto a shared folder does not distinguish what is approved, dated, abandoned or hypothetical, because nothing says so; disorder with a model added to it simply becomes faster at hallucinating. Naming, dating, settling on a source of truth, marking a limit — this work with no visible return becomes the quality condition of everything that will be produced afterwards.
The second is the daily work that holds this material. Contributing, correcting an interpretation, validating a source, flagging a contradiction, deciding what is kept: that is where the gap plays out between two teams equipped with the same tool, not in the wording of prompts. This work has no deliverable, and therefore no place in a job description, and that is precisely why the brake is cultural. Accepting that the delivered document is not what you keep, that raw information has to be reworked, that memory cannot stay in a few heads — these renunciations touch established positions, and no technical demonstration answers them.
The shift is ultimately measured in a single place: who supplies the context of the question. Today it is the person, who knows where to look, in which tool, under which vocabulary, and who brings along without thinking about it what they are trying to do, what has been decided and what is reliable. An invisible load, paid by every newcomer, and lost with every departure. What changes hands is therefore not a tool, it is that load.
Tensions / contradictions
One tension runs through the angle without being resolved: the same material has to be governed — hence named, dated, qualified, bounded — and stay flexible enough to receive what no field was expecting. Yet governance proceeds by stable forms, and it is the requirement of stable form that produced the flattening denounced at the outset. Nothing says where useful structure stops and the form begins again.
A second tension, on scope: connectability between contexts presupposes a shared vocabulary at least on what crosses between them, and that shared vocabulary is a form of centralization — the very one that separating the business prisms was meant to avoid.
Questions
- What volume of context can an organization maintain before upkeep costs more than the memory returns?
- How do you spot that a context has started to age, when nothing in its content signals obsolescence?
- What, inside a context, must remain readable by a human once nobody reads it through any more?