How to Build an AI Knowledge Base Your Support Agent Can Trust

An AI knowledge base for customer support is the set of approved policies, product facts and how-tos an agent answers from. To make it trustworthy, keep one current source for each fact, write in clear standalone statements, leave live data like order status to tools, and test it with real customer questions.

The agent can only be as accurate as what it reads. Most wrong AI answers trace back to an old PDF, two policies that disagree or a rule nobody wrote down. This guide shows how to build a knowledge base that fixes those problems, whichever tool you use.

This separation solves a common failure. If you paste today's prices or stock counts into a document, the agent will quote them confidently months later. Facts that change often should come from a live system, and rules about behavior belong in the brief.

• One topic per article: An agent looks up the passage that fits. A 40-page policy PDF hides the answer.
• State the rule directly: "Unopened items can be returned within 30 days of delivery" beats "we want you to love your purchase".
• Include the conditions and exceptions: Sale items, international orders, opened products. Whatever your team would ask about.
• Make each passage stand alone: Replace "as above" and "this policy" with the actual name, since a passage may be read by itself.
• Use your customers' words: If they say "cancel my box" and your article says "terminate your plan", add both.
• Say what to do when the answer is no: "If the window has passed, offer to bring in the team."

Here is a before and after. Before: "Returns are generally accepted within a reasonable time." After: "Customers can return unopened items within 30 days of delivery for a refund to the original payment method. Opened items can't be returned unless they arrived damaged. For a damaged item, ask for a photo and the order number, then hand off to the team." (The 30-day window is an example; use your own.)

A hallucination is a confident answer the source doesn't support. A better knowledge base helps, but it takes a few other habits too:

• Tell it what to do without an answer: "If it isn't in our policies, say you're not sure and offer to bring in the team." An "I don't know" should be a normal result.
• Fix conflicts: Search for different answers to the same question and delete the old one. An agent can't know which is current.
• Keep risky answers to approved wording: Health, safety and legal topics should use text you have checked.
• Check its sources: Choose a tool that shows you what the agent read to reach its answer.

• 1. Collect real questions: Take 50 from past conversations, in the customers' own words.
• 2. Add edge cases: Sale items, overseas orders and questions you don't have an answer for.
• 3. Ask the agent and grade it: Right, partly right, wrong or properly unsure.
• 4. Fix the source: Change the knowledge rather than adding a one-off rule, unless the gap is about behavior.
• 5. Run it again: Keep the set and rerun it whenever you change a major policy.

Give every article an owner, and tie updates to the events that change the facts: a new shipping partner, a change in return policy, a new product line. Add a line to your launch checklist that says "update what the agent knows". Then use your support conversations to find gaps. When your team answers something the agent couldn't, that answer is a candidate for the knowledge base. Our post on how agents learn from your team explains how that can happen with less manual work.

Less than you think. Your return, shipping and subscription policies, plus facts for your best sellers, are enough for a first job like order questions. Don't delay a launch to document everything. Start narrow and let real questions tell you what to write next. For the next step, see how to write an AI agent playbook, which covers methods rather than facts.

You can use this shape for most policy articles. It gives an agent everything it needs in one place, and it helps a new teammate as well.

• Title: Use the words customers use: "Returning an item", not "Reverse logistics policy".
• Short answer: One or two sentences stating the rule.
• Conditions: What qualifies, with exact numbers.
• Exceptions: Sale items, opened items, overseas orders.
• What to ask for: Order number, photo, reason.
• What to do next: Which action to take, or when to hand off.
• Owner and date: Who keeps this current, and when they last checked it.

• Uploading everything: Old price lists, draft policies and internal debates become possible answers. Add only what you would be happy to see quoted to a customer.
• Burying facts in images: Size charts in pictures and policy text inside banners may be hard for an agent to read. Put the facts in text too.
• Writing for marketing: Warm, vague words are hard to act on. Have a separate, plain source for the facts.
• Leaving out the "no": Most articles describe what you do. Agents also need what you don't do, such as "we don't ship to PO boxes".
• Patching with one-off rules: A long list of "if the customer asks X, say Y" in the brief grows into a second, conflicting knowledge base. Fix the article instead.

Some information helps your team but shouldn't reach a customer: margins, supplier names, how you judge fraud, or how far you will go on an exception. Keep this apart and make sure a customer-facing agent can't read it. Don't rely on an instruction that says "don't mention this", since an agent can't leak what it can't see. Check how your tool separates audiences before you upload anything sensitive, and test by asking a customer-facing agent for the internal detail directly.

Concierge agents answer from what you approve: your policies, product rules, help articles and internal how-tos, not the open web or old notes. You manage it under Build in the Concierge admin, with Instant Answers, FAQs and Help Desk. Live details come from tools, such as orders and subscriptions after verifying the customer.

Short practices live in memory, and files go in your workspace Library. When a teammate gives an answer the agent didn't have, it can become a suggested memory for you to review. Glossary: knowledge base and grounded answers.

An AI knowledge base is the collection of approved information an AI agent reads to answer questions, such as policies, product facts and how-tos. It differs from a help center mainly in being written so that an agent can find and use a single passage correctly.

It is a reasonable start, but help centers often hold old or marketing-style pages. Test with real questions, then fix or remove anything that causes a wrong answer.

No. Prices and stock change often, so an agent should read them from your live systems. Anything pasted into a document will eventually be out of date, and the agent will still quote it confidently.

A first version covering returns, shipping, subscriptions and your best sellers can be small. It grows from real questions the agent couldn't answer, so plan on regular upkeep rather than one big project.