An AI agent that answers only from approved knowledge is only as good as what you approve. The good news: the work is smaller than teams expect, and it compounds.
Start with the twenty questions
Your staff already know the questions that fill the phone — opening hours, prices, directions, the rules everyone asks about. Write those first, as plain question-and-answer pairs, in the language your customers use. Twenty precise entries typically cover the bulk of routine traffic; the rest arrives as demand data later.
Decide what the agent may do with each answer
Every entry carries an autonomy class: answer on its own, open a ticket, or hand straight to a human. Prices and hours answer autonomously. Complaints and anything medical, legal or financial should hand over — the entry then exists so the agent knows where to route, not so it can improvise.
Give time-limited facts an expiry
Today’s availability, this week’s offer, the seasonal schedule — enter them as dynamic entries with an expiry date. A lapsed entry makes the agent say it cannot confirm, which is the correct answer once the fact is stale. Never leave seasonal truth in a permanent entry.
Let the gap report write your backlog
Every question the agent cannot ground gets logged with its frequency. After the first weeks, that report — not a workshop — tells you exactly which entry to write next. Coverage grows where real customers push it.
What to skip
Marketing prose, aspirational claims, anything you would not want repeated verbatim to a customer at 3 a.m. The agent will say exactly what the entry says — write every entry as if it will be quoted, because it will be.