A knowledge base your AI agent can actually use
Your help centre was written for people who are already reading it. An agent needs the same facts arranged very differently.
The usual first move is to point the agent at the help centre. It is already written, it is already approved, and it is right there. Then the agent starts giving answers that are technically present in the documentation and useless on a phone call.
The help centre was written for someone who has already decided to read. An agent needs the same facts arranged very differently.
Four things that break retrieval
- Answers that depend on where you are
- "Click the button in the top right" means nothing when the article is retrieved out of context. Anything that assumes the reader is mid-journey retrieves badly.
- One article covering six situations
- A long "Billing" page containing refunds, invoices, payment methods, plan changes and tax will be retrieved for all six and answer none of them well. Split it.
- Facts buried in prose
- "We generally aim to respond within one working day, although during busier periods this may extend." What is the SLA? A person can extract it. Retrieval returns the paragraph.
- Stale pages nobody deleted
- The 2024 pricing page still exists on your site. The agent has no way to know it is old, and "it's on the website" is a defence you will not enjoy making.
Write for retrieval instead
Three habits, and they are not much work:
- One question per document. Title it as the question a customer would ask — "How much does a consultation cost?" not "Consultation Pricing Information".
- Answer in the first two sentences. Then elaborate. If the agent only uses the opening, it should still be right.
- State facts as facts. "Response time: one working day." Then the caveats, in their own sentence, so a partial retrieval cannot turn a caveat into the answer.
Say what is not there
The most valuable thing in a knowledge base is often an explicit absence. Add short documents that say what the agent should not attempt:
Custom and enterprise pricing is not published. If asked, say pricing beyond the published plans is quoted individually, and offer to arrange a call.
Without that, the model reaches for the nearest published number and improvises. With it, the refusal is itself an answer, and a good one.
What to actually load
In rough order of value:
- The twenty questions you are asked most. Ask whoever answers the phone; they will list them in about four minutes.
- Published prices and what is included. Exactly as published, nothing inferred.
- Hours, locations, coverage. Boring, constant, and asked constantly.
- Policies with numbers in them. Refunds, cancellation, notice periods.
- The refusals. As above.
Notice what is absent: your brochure, your about page, your blog. They are prose about a company and they crowd out the facts. A tight base of forty good documents outperforms four hundred pages of website every time.
One base, many agents
The same facts serve your chat bot, your voice agent and your after-hours agent. If each has its own copy, they will disagree within a month — and you will find out from a customer.
Keep one base and point several bots at it. In Convarza a knowledge base shows a Used-by list of the bots relying on it, so nobody deletes one out from under a live agent. Where knowledge sits on a bot.
Keep it honest
- One owner. Not a committee. A base nobody owns goes stale by default.
- Review when the fact changes, not on a quarterly calendar. Price changes should trigger it; a date on a spreadsheet should not be the mechanism.
- Read the questions it could not answer. That log is your content roadmap, and it is more honest than any keyword tool.
Most "the AI is not very good" complaints are knowledge problems wearing a model costume. Fix the base and the same agent gets noticeably better overnight.