A company knows far more than it can say.
It knows why a particular customer is handled differently, and which promise must never be broken, and which step in the process is theatre and which is load-bearing. It knows who to ask. It knows what went wrong the last time someone tried this, and why the obvious fix is not the fix. Almost none of this is written down. It lives in the people who hold it, in the habits of teams, in the worn paths of how things are done here.
For most of the history of business, that was perfectly workable. Knowledge in human-shaped form is fine for humans. A new hire absorbs it slowly, by sitting near the right people, asking, watching, getting things wrong in survivable ways. It is inefficient, but it works, because the people who need the knowledge are the same kind of thing as the knowledge’s current host.
A model is not.
Context is the constraint, not capability
There is a widespread and costly misreading of where the current limit sits. Teams assume the constraint is the model’s capability: how clever it is, how current, how large. Occasionally that is true. Far more often, a model is asked to do real work inside a company and produces something confident, fluent and subtly wrong. The reason is not that it could not reason. The reason is that it was never given what it would have needed to reason well.
It did not know the unwritten rule. It could not see the last attempt. It had no access to the distinction that everyone in the room holds without thinking. It was working, in effect, like a brilliant new hire on their first morning, with no one to ask and no time to learn. It produced exactly what that person would produce.
A model is only ever as good as the context it can be given. The capability is largely bought in. The context is not. The context is the company, and almost all of it is in a form the model cannot read.
You cannot put a model in the loop of a company it cannot read.
The company brain
This is the work that has no glamour and cannot be skipped: making the company legible. Turning knowledge that is tacit, scattered, and human-shaped into knowledge that is explicit, structured, current, and addressable. A company brain. Not a wiki that decays, not a folder no one opens, but a maintained representation of what the company knows, kept close enough to the truth that a synthetic participant can act on it.
It is tempting to hear that as a technical project. It is not, mostly. The hard parts are organisational.
Surfacing tacit knowledge means asking people to externalise what makes them valuable, and to do it honestly, which cuts against a quiet instinct of self-protection. Documentation has no natural owner; it is everyone’s responsibility, which is to say no one’s, and it rots the moment it is finished. Much of what a company knows is contested: two senior people will describe the same process differently, and both will be partly right. And knowledge has a half-life. A representation that is accurate today is misleading in six months unless something keeps it true.
None of these are problems a tool solves. They are problems of ownership, incentive and discipline. Problems of how the company is run.
Knowledge as infrastructure
The shift that makes this tractable is to stop treating knowledge as documentation and start treating it as infrastructure.
Infrastructure has owners. It has a maintenance budget. It has a defined freshness: a sense of how current it must be to be trusted, and a process that keeps it there. It is built once, deliberately, and then kept, because the things that depend on it would fail otherwise. Companies already think this way about their systems and their finances. Almost none of them think this way about what they know.
A company that does treat knowledge as infrastructure gains something beyond the model. The knowledge stops walking out of the door when people leave. New people get up to speed against a representation instead of against the patience of their colleagues. Decisions stop quietly depending on whoever happened to remember. The company becomes legible to itself, which is valuable on its own, and was always worth doing.
But it becomes non-negotiable the moment the company wants synthetic intelligence to do anything that matters. Every later move (decisions made with the model in the loop, operations the model can act on) assumes a company the model can read. There is no version of being AI-native that skips this and survives contact with real work.
It is the first block, for a reason. Before a company can be rebuilt around intelligence, it has to become legible to it. A model cannot help you run a company it cannot see.