When an organization connects a knowledge base to an AI system, the expectation is that the system will answer correctly based on what's written. In practice, a large share of wrong answers doesn't come from the system but from how the procedure is worded.
The first rule: condition before action. A procedure written as a continuous story is hard to break down. A procedure written as "if X — then Y" translates into a precise answer even when the question is phrased differently.
The second: no vague references. A phrase like "as in the previous case" or "per standard policy" means nothing when the system retrieves a single paragraph. Every section has to stand on its own.
The third: a date and an owner for every document. A document with no effective date can't tell you whether it overrides the one before it, and the system will choose arbitrarily between two conflicting versions.
The fourth: spell out the exceptions. Long-timers know the exceptions by heart, which is why they're never written down. Those are exactly the cases where a wrong answer is expensive.
This comes to light at the cleaning stage, before the system learns. In most organizations it's also the stage where it turns out there are two active price lists and nobody knew.
The bottom line
Before connecting AI to your organizational knowledge, it's worth going through your procedures and asking of every clause: is it clear to someone who reads only that clause?