The uncomfortable prerequisite
An AI employee can only be as good as what it can read. Companies expecting to buy capability discover instead that their written knowledge is partial, contradictory, and years out of date — and that this was already costing them with human staff.
Start from questions, not from documents
Don’t audit the wiki. List the questions the role must answer, then find which have a written answer that is current and correct. The gap list is your actual scope, and it is usually far smaller than reorganising everything — the same logic as scoping a corpus by question.
You are not documenting your company. You are answering thirty specific questions in writing.
Contradictions are worse than gaps
A missing answer produces an escalation. Two documents that disagree produce a confident wrong answer, because retrieval picks by similarity rather than correctness. Resolving conflicts is the highest-value pass and should come before writing anything new.
Capture the unwritten rules
The decisions people make without thinking — when to make an exception, which accounts get handled carefully — are the material that separates sounding right from acting right, and they exist only in conversation until someone interviews for them.
Assign a maintainer before you start
A knowledge base with no owner returns to being wrong within months, which then silently degrades the AI employee reading it. This is precisely the job that goes unassigned when nobody owns the system.
The by-product is worth the work
Even if the AI project is cancelled, you end up with current answers to the thirty questions your business is most often asked. New human hires get productive faster on the same material, so the work is defensible on its own.