Skip to main content
Business knowledge gives your agent a source for the facts and policies it shares with customers. It is the right place for information such as opening hours, delivery areas, returns conditions, product guidance, and how-to articles. Good knowledge preparation is an editorial task as much as an upload task. The agent needs content that is current, internally consistent, and specific enough to answer the customer’s question.

​
Prepare a reliable source of truth

Choose the policy or article your team treats as authoritative. Remove superseded versions and make exceptions explicit. If a delivery rule changes by location, describe the locations and their rules instead of saying only that delivery is “usually available.” Assign an owner to each important topic. That person should review the agent’s answers when the underlying policy changes and decide whether a missing answer needs a new article or a clearer existing one.

​
Add content to Knowledge

Use the Content tab to create and organize articles. Use Sources to upload approved files and inspect their indexing state. The documented upload formats include PDF, DOC, DOCX, TXT, CSV, and Markdown. Keep related information together and use descriptive article titles. A policy with clear sections is easier for both your team and the agent to interpret than a long document mixing unrelated topics. Notion and Confluence imports are read-only snapshots. When the original source changes, use the supported sync workflow and verify the imported content before assuming the agent has the update.

​
Enable the saved article for your agent

Save the article, then turn on AI Agent in its details or the content list. An enabled draft is available in Sandbox only, so you can test it before customers use it. Choose Publish to make that saved version available in Production as well. If you later edit and save a published article, Sandbox uses the latest saved draft while Production continues using the last published version. Publish again when those changes are ready. Turning AI Agent off removes the article from both environments.

​
Test the questions customers actually ask

Try a direct question, a paraphrase, and a question involving an exception. For a fictional store, these might be “Where do you deliver?”, “Can you send an order to my area?”, and “Does the same rule apply on public holidays?” Compare each answer with the approved source. If the answer goes beyond what the source supports, tighten the content or instructions and test again. Include a question for which there is deliberately no answer; the agent should follow your clarification or escalation policy.

​
Keep live records out of static policy content

An uploaded document is not a substitute for a live order or account system. Do not expect a spreadsheet uploaded last month to provide a current payment status. Use knowledge to explain a policy and a connected lookup to retrieve the current record. For example, the policy can explain what “Awaiting dispatch” means, while the order API confirms whether that is the customer’s current status. Read connected data and actions when your workflow needs information that changes for each customer.