Building an AI Customer Service Knowledge Base: Start with the 20 Most-Asked Questions

How well an AI customer service answers depends largely on the information you give it. The good news is that you don't need a complete handbook from the start.

First, identify the most frequently asked questions

Look through your customer service records, emails, and social media DMs from the past month or two, and rank the questions by how often they appear. For most online stores, the top questions fall into these categories:

• Orders and shipping: How soon will it ship? How long does delivery to a certain country take? How is shipping calculated? How can I track my shipment?
• After-sales support: How do I return an item? How long does a refund take to arrive? What should I do if there's a problem with the product?
• Products: How do I choose a size? Is it in stock? What material is it made of?
• Payment: Which payment methods are supported? Can I get an invoice?

Write answers to the first 20 questions, and you'll cover a large share of inquiries.

Four principles for writing answers

State the conditions and numbers clearly. “Free shipping on orders over $59; otherwise shipping is $6, with remote areas charged separately” is far more useful than “some orders qualify for free shipping.” The AI answers only based on what you write; vague information leads to vague answers.

One answer per question. Write “shipping time” and “shipping cost” separately so the AI can find the right one more easily.

Phrase questions the way customers do. Customers ask “How soon will I receive it?”, not “shipping time policy.” The more closely a question matches a customer's own words, the easier it is to match.

Spell out exceptions. Custom items can't be returned, and shipping is delayed during holidays. These exceptions are the most likely to cause disputes, so be sure to include them.

Put longer materials in documents

For content such as return and exchange policies, size charts, and warranty terms, just upload a file or paste the original text, and the system will split it automatically. For help pages already on your website, enter the URL and it will fetch the body text.

Check with test chats

Once your materials are ready, put yourself in the customer's shoes and ask a few questions in a customer's tone, including some tricky ones. If an answer is wrong, it usually means the information is missing or not specific enough.

Keep adding to it after launch

After launch, questions the AI can't answer will be recorded. Spend ten minutes each week reviewing them and add the answers. After a few weeks, you'll notice that noticeably fewer questions need human handling.

For more operational details, see the documentation: Knowledge base, Test chats and adding answers.