How to prepare and set up a knowledge base for AI customer support

An AI agent answers from your knowledge base, so your content caps the quality of its answers. To prepare it: inventory what you have, give it a clear structure, rewrite key articles so each answers one question, fill gaps from real tickets, then connect the sources and test with real questions before launch.

Here are the seven steps at a glance:

Why the knowledge base decides how well AI answers

An AI agent reads a customer's message, works out what they want, and looks for the answer in your content. When the content is current and clear, the answer is right. When two articles disagree, or the answer only lives in someone's head, the AI either picks the wrong one or has nothing to say. Gaps and contradictions in the source are a common cause of confident wrong answers.

The same content serves your agents too. CoSupport AI Assistant drafts replies for your agents using your knowledge base and past tickets, so every fix you make improves both the answers customers get from AI and the drafts your agents start from.

Step 1. List everything you already have

Most support teams have more documentation than they think, spread across more places than they'd like. Start with an inventory.

Content types to look for:

• FAQs and troubleshooting articles
• Policies: returns, refunds, billing, shipping, warranties
• Product manuals and setup guides
• Macros, saved replies, and the best past ticket answers
• Internal guides that agents use but customers never see
• PDFs, spreadsheets, and slide decks with product details

And the places they live: your helpdesk's help center, Notion or Confluence, shared drives, internal wikis, Slack threads, and email. Put it all in one list with an owner and a last-updated date for each item. That list is the map for the next steps. Flag content the AI won't be able to read as it stands: scanned PDFs, password-protected folders, and answers that only exist in Slack threads.

Step 2. Give it a structure the AI can follow

Group the content into a small number of categories, with two or three levels at most:

• Getting started
• Account and access
• Billing and payments
• Orders, shipping, and returns
• Product features, for example "Product features" > "Dashboards" > "Custom views"
• Troubleshooting
• Integrations

A clear structure does two jobs. It shows you where two articles cover the same thing, and it helps the AI work out which article fits a question.

Then keep one version of each answer. If the returns window is 30 days in one article and 60 in another, the AI can quote either one. Merge duplicates, then redirect or retire the outdated version so old links still work. Keep internal-only rules, like refund approval limits, out of any knowledge base the customer-facing AI uses.

Step 3. Write articles the AI can answer from

How an article is written matters as much as whether it exists. Seven rules cover most of it:

• One question per article. An article called "Account settings" that covers passwords, email changes, and deleting the account gives the AI three half-answers. Split it into three.
• State the question in the customer's words. Customers tend to write "where's my refund", while the article is called "Refund policy". Use both in the title or the first line.
• Make each answer complete on its own. The AI may use one section without the rest of the article, so don't write "as mentioned above". Expand abbreviations in every section that uses them, and use one name for each feature, plan, and setting everywhere.
• Answer directly, then add the detail. "Yes, you can change the delivery address before the order ships. Here's how." works better than a bare "Yes".
• Spell out conditions. Which plan, which region, which product, what's excluded, and what happens after a deadline.
• Use numbered steps for procedures and short paragraphs for everything else, with headings that say what the section answers.
• Describe what's in images and videos in text. The AI can't rely on a screenshot to explain a setting, so put the button names and steps in the article itself.

Step 4. Fill the gaps from real tickets

Your tickets show what customers actually ask. Pull a month of them, group them by topic, and rank the topics by volume. For each of the top topics, check whether an article answers it completely. Where it doesn't, write one, starting with the most frequent topic.

Two sources help here. Your best agents' saved replies are often the most accurate answers you have, so turn them into articles. Remove names, emails, and order numbers when you turn tickets into articles. And once the AI is live, the questions it escalates are your gap list: each one is either a case that should stay with people or a missing article.

Step 5. Decide who keeps it current

A knowledge base goes stale the week a policy or product changes. Name an owner for each category and agree on a review rhythm, for example monthly for the 20 most-used articles and quarterly for the rest. Add one rule on top: when the product, a price, or a policy changes, the matching article changes the same day.

The owner assigns each article to a person and tracks when it was last checked.

Step 6. Plan for more than one language

If you support customers in several languages, start with the content you have. CoSupport AI detects the customer's language and answers in it, in 60+ supported languages (120+ compatible). Many teams keep one source language for most articles, and the test in Step 7 shows whether that works for the languages you serve.

Translate the articles where exact wording matters: terms, policies, legal notices, and anything a customer might quote back to you. Keep product names, plan names, and error codes the same in every language, so the AI and the customer are talking about the same thing.

Step 7. Connect the sources and test before launch

Before you connect anything, write down your goals and today's numbers: which categories the AI should handle, your current first response time, and the share of tickets agents close. Decide which topics always go to people, such as payment disputes, legal requests, and account security, and tell your team what the AI will handle and what will reach them.

Then connect the knowledge base through your helpdesk integration or by uploading content directly, and test before customers see anything:

• Ask the 50 most common questions from last month's tickets, worded the way customers wrote them.
• Add the awkward ones: two questions in one message, a policy exception, an angry customer.
• Check each answer against the source article. A wrong answer usually points to a conflicting or incomplete article. Fix the article and the answer follows.
• Test in each language you serve.

One low-risk way to run this is to let CoSupport AI Assistant draft answers for your agents first, so the team sees the AI's answers before customers do.

CoSupport AI is ISO 27001 certified and GDPR and CCPA compliant, with details on the security page. You still decide what goes into the knowledge base, so keep personal data out of the articles.

How to tell if it's working

Track five things after launch:

• AI resolution rate by category. Low categories point to missing or unclear articles.
• Escalations by reason. Questions the AI couldn't answer are your next articles.
• Reopen rate on AI-resolved tickets. It shows whether "resolved" was real.
• Articles touched in the last 90 days. Old articles on busy topics are the next risk.
• Customer feedback on AI conversations. Compare it with the same score for your agents.

Count tickets the AI actually resolved. A conversation where the customer simply stopped writing isn't one. The difference is explained in what AI for customer support is.

How to set up a knowledge base in CoSupport AI

In CoSupport AI, knowledge bases live in their own section of the dashboard, under AI Knowledge Bases. Inside each one you'll find its sources, Add Folder and Add Source buttons, and a search.

Add your sources

Click "Add Source" and choose the type:

• FAQ. Write question-and-answer pairs directly, for example "How do I reset my password?".
• File. Upload documents with text content, such as manuals, policies, or internal wikis.
• Website. Add one or more URLs. Turn on "Sync nested pages" to include every page under an address, and exclude the pages you don't want the AI to use. The list shows an auto-sync status for each source.

Connect your helpdesk

Two helpdesks connect as sources directly:

• Zendesk: your Help Center articles and tickets. Real-time sync can be switched on or off.
• Freshdesk: your tickets.

From Zendesk, the AI reads your tickets, macros, and Help Center articles. If your tickets are connected, the same rule applies as in Step 4: articles stay free of personal data, and the security page explains how ticket data is handled.

Organize and assign

Group sources into folders, and select several at once to move or delete them. When the knowledge base is ready, assign it to the projects that should use it, and those projects start answering from it.

Start with the knowledge base you have

Cover your top ticket topics, keep one version of each answer, name one person who owns updates, then run the test from Step 7. CoSupport AI Agent responds using your knowledge base and business systems and learns from your tickets and help center articles, so every article you fix here improves what it says. It works with Zendesk, Freshdesk, Intercom, and Zoho Desk through its integrations.

For the bigger picture, read the AI customer support strategy ebook. The same content work pays off outside support too, as generative AI in sales shows.