Building a knowledge base your AI agents can trust

Garbage in, confident garbage out

A language model does not know your return policy, your delivery times, or your contract terms. It knows what your knowledge base tells it. When that base is outdated or contradictory, the agent does not get cautious, it gets confidently wrong, because it has no way to tell a stale answer from a current one.

That is why the knowledge base, not the model, is where most of the quality of your automation is decided. It deserves the same care you would give your pricing page.

Start with one source of truth

Most companies store answers in five places at once: a help center, internal wikis, PDFs, macros in the ticket system, and the heads of two long-serving colleagues. Each copy drifts on its own schedule. Ask three people about your refund window and you may get three answers.

Before automating anything, run an inventory. Collect every place an answer lives, decide which version is correct, and move it into one base. Then delete or redirect the copies. One question, one entry, one owner: everything else is a contradiction waiting to be served to a customer.

Write for machines and people

Agents read the way skimming humans do: they find well titled, focused articles and struggle with sprawling documents. A few writing rules make your knowledge dramatically more reliable:

• One question per article. Split a ten-page policy into entries that each answer a single thing.
• Put the answer in the first sentence, details after.
• Use concrete values: write 30 days, not a reasonable period.
• Name the exceptions explicitly. Unstated exceptions become wrong answers.
• Date entries that change over time, like prices, deadlines, and promotions.

Let real conversations show you the gaps

You will not guess every question customers ask, and you do not have to. Every conversation the agent cannot answer, and every handoff to your team, is a pointer at a missing or unclear entry.

Make reviewing those cases a weekly routine. Write the missing article, sharpen the ambiguous one, and the same question is handled automatically from next week on. Over a few months this loop, not the initial writing, is what builds a base that covers your actual support reality.

Review, approve, repeat

Treat the base as a living product. New entries go through a short draft and approval step so nothing unchecked reaches customers. Stale entries get retired, and every entry has an owner who is responsible for keeping it current.

A quarterly full review sounds like overhead, and it is far cheaper than the alternative: an agent politely serving your customers last year's prices.