Test your agent before launch
Check answer quality, boundaries and the chat widget before making your agent available to customers.
Open copy menu
Test your AI agent with the questions and tasks your customers encounter. The aim is not to obtain one successful answer, but to find gaps before launch.
Before you start
Section titled “Before you start”Add the required data sources , update the knowledge base and wait for processing to finish. Save the agent’s instructions before testing responses. Use fictional test data rather than real personal information.
Run a repeatable test
Section titled “Run a repeatable test”Choose questions
Write down common customer questions and the expected answer or behavior. Include a question not answered by the sources and a request outside the agent’s task.
Try the preview
Open the agent editor. On a large screen, the preview sits beside the editor; on a small screen, open Preview. Send the questions and try a follow-up. Start independent tests in a new conversation so earlier messages do not affect the comparison.
Correct and repeat
Compare each answer with the expected result. Correct facts in the source or behavior in the instructions. Publish knowledge changes before retesting. Repeat both the failed question and a previously successful one.
What should you check?
Section titled “What should you check?”| Situation | Acceptance signal |
|---|---|
| Common question | The response matches current source information. |
| Missing information | The agent acknowledges uncertainty without inventing details. |
| Unclear question | A clarification helps the customer continue. |
| Out-of-scope request | The response stays within the agreed boundaries. |
| Another language | The answer is understandable and factually consistent. |
| Follow-up question | The agent takes the conversation context into account. |
Check the actual website too
Section titled “Check the actual website too”The preview does not prove that the website installation works. After installation , check that:
- The chat widget opens on a phone and a computer.
- Its name, welcome messages and language are appropriate.
- You can send a message and receive a response.
- Enabled human support or other actions also work from the visitor’s perspective.
If a test fails
Section titled “If a test fails”Change one thing at a time. Record the question, the actual answer and the expected result. First check the source content and knowledge update status, then the instructions. If a recent edit made things worse, restore the previous text and retest. AI responses can vary, so one successful repeat does not guarantee that every answer will be correct.