Skip to main content

Testing

Tests are how you confirm a flow behaves the way you intend — and keep it that way as you keep editing. The Copilot both authors tests and makes sense of their results, so you can close the loop between building a change and proving it works.

What a test case is​

A test case describes a scenario the AI should handle: the customer situation, the data the APIs should return (mocked from saved sample data so the test doesn't hit live providers), the run instructions, and an assertion — the specific outcome that must hold for the test to pass. A test fails only when the transcript violates that assertion, not for unrelated imperfections.

Draft test cases​

Create, update, copy, or draft test cases, scenarios, and run instructions to validate agent behavior. The fastest way to build coverage is to clone an existing case and vary one thing.

"Create a test case like the escalation case, but with status cancelled." "Write a test for the refund flow where the order is outside the return window — it should refuse politely." "Add a test where the customer uploads a damaged-item photo and the flow escalates to an agent." "Add a test where a warehouse child ticket returns Replacement approved, then verify the AI explains the next step." "Draft tests for each branch of the cancellation flow."

For a useful test, name: the customer scenario, what the mocked data should say, and the assertion — the exact outcome you expect (resolved, escalated, a specific message, a particular field set).

Review test results​

Analyze why a test passed or failed and get a concrete fix. The Copilot reads the transcript and the configuration together and identifies the root cause rather than just reporting the failure.

"Why did the refund test fail? How can I fix it?" "The escalation test passed but the message was wrong — what happened?" "Summarize which tests are failing and why."

A failure analysis distinguishes between causes so you fix the right thing:

  • Wrong configuration — an objective, transition, or condition is misconfigured. The Copilot points to the exact field to change.
  • Wrong test data — the customer payload, objective mocks, or child-ticket value didn't fit the scenario. The fix is the test, not the flow.
  • System issue — the config is correct but the behavior wasn't, which is worth flagging.

Close the loop​

Testing pairs naturally with building: make a change in Build & change flows, write or update a test here to lock the behavior in, and rerun it after each edit so a later change can't silently break it.