The bottleneck is almost never AI

This is the stage where AI projects die in traditional companies. Almost everyone expects the problem to be the model, but it is usually somewhere else: the company has never written down what it knows.

The knowledge exists. It is in the head of someone who has been doing the work for ten years, in a PDF no one opens, in a tacit agreement between two teams. None of that is machine-readable, and much of it is not even readable by a new employee.

Start with the ontology, which is the formal name for a simple thing: writing down what words mean inside this company. In a child sponsorship organization, "child" may mean a sponsored child, a child served without a sponsorship relationship, or a child from a family served by the organization. Three different entities that the same report calls by the same name.

Until that is written down, every agent will make mistakes. And every human will make mistakes in a different way, which is worse, because you cannot even correct them systematically.

Then come the policies. They already exist: discount limits, what cannot leave the company, LGPD, brand voice. What is missing is form. A rule written in prose inside a manual does not become a guardrail; a structured rule (field, reason, exception, responsible owner) does.

Sign that you are in this stage and did not know it: every project meeting gets stuck in a discussion about what a term means.