How do you use AI with internal knowledge nobody wrote down?

Every business runs on things that are not written anywhere: which customer is difficult, why the process has that extra step, which supplier to call when the usual one cannot deliver, what the exception is that nobody documented. These are not accessible to any system, and the failure mode when a system is asked to work without them is not silence — it is a confident, reasonable answer that ignores the constraint.

The instinct is to have the system infer what it can from the data. This works for patterns that are genuinely in the records and fails for the ones that are not, and the two are indistinguishable in the output. A system inferring that a customer prefers morning appointments from their booking history is doing something sound; one inferring why a process step exists is producing a plausible story, and the story will be believed because it is coherent.

The productive move is to lower the cost of capture instead. The reason this knowledge is undocumented is not that nobody values it; it is that writing it down is tedious and never urgent. A recorded conversation, transcribed and drafted into a structured document, converts an afternoon of someone talking into a written process. The person then corrects a draft rather than facing a blank page, which is a fundamentally easier task and one people will actually do.

The material worth capturing first is the exceptions. General process descriptions are usually available somewhere or reconstructable from the artefacts. What is not available is the list of cases where the general process does not apply and what happens instead, and that is exactly the knowledge that determines whether an automation works or produces defensible wrong outcomes.

Capture should be triggered by events rather than scheduled. When someone is about to leave, when a process is about to be automated, when the same question has been asked three times, when something went wrong because a constraint was not known. Each of these is a moment when the value is obvious and the knowledge is fresh, and scheduled documentation exercises produce comprehensive material nobody reads.

Once written, the material has to live where it will be supplied. A captured process in a document nobody attaches has not changed anything. It should be in the folder the systems read, in the project instructions, or in the shared drive that the assistant is connected to, and the capture exercise should end with putting it there rather than with the document existing.