Before you start
You need a tested, published flow, an assistant in the same subaccount, working voice settings with a standard voice (GPT Live voices cannot be combined with a flow), and a number configured for the intended inbound or outbound call path.Connect the flow
- Publish the flow after checking its branches in Test.
- Open the assistant’s model selector and choose that published flow.
- Review the assistant’s voice in Voice Studio.
- Publish the assistant. This applies the assistant configuration to its voice agent.
- Use your normal inbound or outbound setup to run a controlled call.
Verify on a real call
Ask something that takes a recognizable branch. Listen for the greeting, response timing, handling of interruptions, and the next question. Test a missing answer and an integration failure as well as the successful path. If the flow transfers calls or writes to another system, check the transfer destination or external record. A correct spoken confirmation by itself does not prove the action completed.Reduce avoidable pauses
Keep node tasks focused, avoid loops through several model-powered steps, and use a short neutral filler for a slow tool. Keep dependent actions synchronous. Model choice affects behavior and response time; test it with the actual voice settings.Troubleshooting
- Call reaches a different assistant: check the number’s routing and assistant assignment.
- Assistant follows an old flow: publish the latest flow, confirm its selection, then publish the assistant again.
- Editor test passes but call audio fails: inspect the voice and number setup; the editor runs text turns and does not test voice transport.
- Flow pauses after a tool or branch: check the tool result and connected default paths.
Which settings belong where
Publish a changed flow to update its conversation logic. Publish an assistant when applying its changed voice configuration. Test both directions you intend to use; an editor selector does not configure inbound or outbound telephony.