Build an AI chatbot knowledge base that gives accurate answers

Your AI chatbot knowledge base needs approved answers before the bot meets a customer. In a 2024 Gartner survey reported by CX Today, 61% of service leaders had knowledge articles waiting for edits. Start with these seven content groups, then test the gaps.

1. Why should you add shipping and service details first?

Add the delivery, collection and service-area details that customers need before they decide to buy. These answers usually come from several places, so gather them into one approved page instead of making the bot piece together an old banner, a checkout note and a staff document.

For a store, cover where you ship, how delivery options differ, when an order leaves the warehouse and what happens after a missed delivery. For a service business, replace shipping details with coverage areas, appointment availability, travel fees and what the customer must prepare before a visit.

Write the exceptions beside the main rule. If oversized products use another carrier or weekend appointments cost more, say so in the same source. A technically correct general answer still disappoints the customer when it hides the exception that applies to their order.

Accurate and regularly maintained content matters because both automated support and customer self-service depend on it, according to Gartner. Assign each operational page to the person who already controls that policy, such as the store manager, dispatcher or service lead.

2. Why should you write return and cancellation rules plainly?

Write returns, refunds, cancellations and warranties as direct rules with their exceptions. Don’t expect the bot to infer whether a damaged item, discounted purchase or late cancellation qualifies from several paragraphs of legal wording.

Separate each policy into the questions a customer actually asks. State who qualifies, what evidence you need, which costs are refundable, how the request begins and which cases need staff approval. Add exchange rules separately if they differ from refund rules.

The cost of a wrong answer is more than an awkward chat. In a Canadian case described by Manatt, Phelps and Phillips, a chatbot told a passenger that he could request a bereavement refund within 90 days, although the airline didn’t allow retroactive applications. The tribunal ordered the airline to pay about $482 plus other costs.

That case gives you a useful rule for chatbot FAQ content: don’t let the bot make exceptions, approve compensation or interpret unclear policy language. It can explain the published rule and collect the facts. A person should make the decision when money, eligibility or legal rights depend on the answer.

3. How should you add product and service facts?

Add the facts customers use to choose, compare and use what you sell. Product names alone aren’t enough. Include dimensions, materials, compatibility, care instructions, variants, current availability and anything that disqualifies the product for a common use.

For a service business, document what each service includes, what it excludes, how long it usually takes, where it is available and what affects the quoted price. Keep estimated outcomes separate from promises. If staff must inspect the job before confirming a price or date, make that limit part of the answer.

The knowledge base shown above accepts website URLs, PDF documents and question-and-answer pairs as knowledge sources for product training. Use a page or document for stable information with useful context. Use a short question-and-answer entry when you need an exact, approved response to a narrow question.

A short approved question-and-answer entry is easier to control than a full product page when the answer has exceptions. Hand-written entries are best for questions where small wording changes affect what the customer expects. The longer source can still explain the background.

Don’t copy every catalog field into prose. Choose the details customers mention in messages, then link related terms to the same answer. For example, customers may use “navy,” “dark blue” and a product’s official color name for the same variant.

4. Which repeat questions should you write by hand?

Write direct answers for frequent questions that have one dependable response. Start with questions your team repeatedly answers from memory, especially when the wording on your website is scattered or written for a different purpose.

Useful candidates include payment methods, invoice requests, booking changes, account access, gift cards, product care and what happens after a customer submits a form. Phrase the question as a customer would. Put the answer first, then add the condition or next action.

Keep each entry focused on one intent, meaning the result the customer wants from the question. “Can I change my address?” and “Why was my address change rejected?” look similar, but they need different answers. Combining them encourages the bot to give a general reply when the customer needs a specific next step.

Self-service still struggles when the source material doesn’t match the question. A Gartner customer survey reported by CX Today found that 14% of issues were fully resolved through self-service, rising to only 36% for issues described as very simple.

Review your recent inbox, saved replies and search terms to find the real phrasing customers use. If you need help turning those answers into clean source entries, follow the process for training a chatbot with a knowledge base.

5. Which questions should you send to a person?

Send emotional, unusual, account-sensitive and high-risk questions to a person. Your knowledge base should explain where the bot stops as clearly as it explains what the bot can answer.

Write handoff rules for damaged orders, disputed charges, threats, safety concerns, personal data requests and customers who directly ask for a person. Add any case where staff must inspect evidence or use judgment. The bot can collect an order number or short description, but it shouldn’t decide the outcome.

Speed alone doesn’t solve the customer’s problem. A 2026 Glance study reported by CXM found that 75% of consumers had received a quick automated response that still frustrated them, while 68% ranked complete resolution as the most important part of support.

Keep human help visible. Nearly 90% of consumers in the same CXM report said removing human support entirely reduced their loyalty to a business.

The handoff shown above can start when the system lacks a confident answer or the customer asks for a person, with different behavior during and outside operating hours. Whatever system you use, test the entire chatbot human handoff. Check the customer message, the staff notification and the conversation history that reaches the agent.

6. What should you leave out of the knowledge base?

Leave out outdated, duplicated, private and unapproved material. More content doesn’t automatically produce better answers. One current policy is safer than several versions that disagree.

Don’t upload internal notes that contain passwords, private customer details, unpublished prices or comments never meant for customers. Remove draft policies and expired campaign pages. If staff need the information but customers shouldn’t see it, keep it outside the chatbot’s source collection.

Watch for conflicts between public pages and uploaded documents. A website may show the latest delivery rule while an old staff PDF still shows the previous one. The bot doesn’t know which source your team considers authoritative unless your setup provides a clear priority.

This cleanup isn’t a minor launch task. The 2024 survey covered by CX Today found that 61% of service leaders had a backlog of knowledge articles needing edits, and more than one-third lacked a formal process for updating old content.

Create a source register with the page name, owner and event that triggers an update. A price change should trigger a pricing review. A carrier change should trigger a shipping review. Event-based checks catch errors sooner than relying on someone to remember.

7. How should you test the bot before launch?

Test the bot with real customer wording, policy exceptions and questions it should refuse to answer. A correct reply to the neat version of a question doesn’t prove that the bot will handle abbreviations, missing details or a frustrated message.

Build the test set from recent conversations. Include straightforward questions, follow-up questions, misspellings, vague requests and two topics in one message. Check whether each reply answers the question, states the right condition, points to the right next action and avoids making a promise your team can’t keep.

Test failure paths as carefully as successful answers. Ask about a product that doesn’t exist, an expired promotion and an exception the policy doesn’t cover. Then ask for a person. The safest result is a clear limit and a working handoff, not a confident guess.

Give one person responsibility for the final decision to turn the bot on. Knowledge management is becoming a defined support skill. In a survey of 321 service leaders published in February 2026, Gartner found that 58% planned to develop agents into knowledge management specialists.

After launch, record every unanswered question and wrong answer. Fix the source before adding clever instructions around it. Start with returns, cancellations and other policies that can cost the customer money, then add the high-volume questions that consume your team’s time.