Skip to main content
Start with one decision: for example, whether a visitor needs a new appointment or help with an existing one. A flow defines paths and actions, while the assistant still interprets natural language and can need clarification.

​
Build the first path

  1. Open the intended client account’s Flow Builder and create a named flow.
  2. Use a Conversation node to ask one question. Add destination nodes for the expected answers and connect each condition to its destination.
  3. Add a path for unclear answers so the assistant can ask again. Use Global Prompt for instructions that apply throughout the conversation.
  4. Connect the ending paths to End Convo. Wait for Saved.
  5. Open Test, then select Start conversation.
  6. Test each expected answer, an ambiguous answer, and a failure in any connected tool. Check the highlighted node and the resulting record for external actions.
The test uses the draft. The Voice/Chat and direction selectors do not currently change test execution or place a phone call. Test tools can still perform real external actions, and Reset does not undo them.

​
Publish the version the assistant will use

Click Publish when the draft is ready. In the assistant’s model selector, choose the published flow from the same account. Publish the assistant, then run a test on its real phone or website channel. If the flow is missing from the selector, check its account and publication state. If the assistant uses old behavior, confirm the latest flow was published and assigned. If the editor cannot run a model, check LLM Config instead of assuming a personal API key is always required. Continue with Flow Builder setup, node guides, and testing.

​
Settings used in this example

This example intentionally uses a small subset. For the full set, see all node types, Conversation settings, and all flow variables.