Test conversations and evaluations
Turn representative customer requests into repeatable checks of answers, actions and voice behaviour.
Start with Save draft, then choose Saved draft → Test draft in the assistant header. Select Live version → Test live to test the published voice configuration. In Analysis → Evaluations, create repeatable scenarios and choose Saved draft, the current publication or a specific published version. Save changes before testing the draft. Each run uses a fixed snapshot. Evaluations use configured providers and can consume AI credits and provider usage.
Build a small regression suite
| Scenario | Check |
|---|---|
| Ordinary request | The assistant completes the intended task. |
| Missing evidence | It clarifies or uses the configured fallback without inventing support. |
| Required field omitted | The flow stays in collection. |
| Branch answer supplied | The first matching branch is taken; an unmatched answer uses the default route. |
| Collected answer corrected | The review stage returns to collection and confirms the updated value. |
| Outbound permission refused | An outbound starter goes to the caller exit without collecting further details. |
| Action approved | The expected tool receives the intended arguments. |
| Action declined or failed | No success is claimed and the flow does not advance incorrectly. |
| Language switch | The answer and voice use the requested language. |
| Consent declined | Conversation continues according to policy without prohibited artifacts or memory. |
| Caller interrupts or leaves | Speech stops and the call releases its resources. |
Use concrete caller messages and expected outcomes rather than a broad criterion such as “be helpful.” Include a regression scenario for every confirmed production failure.
After editing in the Flow designer, resolve its validation items and save before running a draft test. A gallery preview shows the stages without executing them. Test any assigned external action separately from the template's collection path, and inspect collected values and the stage history in Logs. Correction and refusal routes must be configured in your flow; the added gallery starters include them.
Choose text or voice mode
Text checks answers, permitted retrieval and tool contracts through the selected model. Voice synthesizes scripted caller speech in an isolated audio room and exercises transcription, the model and voice synthesis. Voice messages are limited to 1,000 characters each.
Select browser or phone capabilities. Phone simulations exercise applicable tool schemas; they do not place carrier calls.
Supply tool fixtures
External actions in evaluations use JSON fixtures, including transfers and DTMF. Define expected tool names, arguments and ordered flow stages. Missing fixtures or tool failure should not count as successful completion.
Fixtures make regression runs repeatable, but do not prove a real integration works. Dashboard tests with approved tools and a controlled customer-channel test can execute actual external actions; use test accounts and intentional approvals. Test live uses dashboard test permissions, so verify the intended public channel separately.
Review failures before rerunning
Inspect stage results, citations, argument mismatches and latency. Correct and save the draft, then rerun the relevant suite against that draft before publishing. Confirm the published version also passes. An interrupted evaluation is a failure, not an automatic paid retry.
After simulated checks, test the real website, phone or WhatsApp path before launch.