Skip to main content
AI assistance can improve a Conversation node’s prompt and suggest outgoing conditions. Start with a specific task and review the proposed wording before publishing.

Improve a prompt

  1. Open a Conversation node in Prompt mode.
  2. Write the task, the information to collect, and the action that marks completion.
  3. Click Improve prompt with AI.
  4. Read the replacement text. Remove invented business hours, prices, policies, promises, or actions your integrations cannot perform.
  5. Test the node with a normal answer, an ambiguous answer, and an unrelated request.
For a booking enquiry, an appropriate instruction is “Collect the preferred date, then check availability.” A prompt alone should not invent available slots or claim an appointment has been booked.

Suggest paths

Use the transition suggestion control to generate candidate conditions from the current prompt. Add only conditions that describe distinct outcomes, then connect each added path to its destination. A suggested condition without a connected destination does not create a complete journey.

Revert a prompt change

After a successful improvement, Revert to previous prompt restores the prompt text held immediately before that improvement in the current editor session. It does not undo every flow edit, remove created external records, or replace a published version automatically. For larger changes, keep a record of the original behavior and review version history before relying on a rollback. Publish only the version you have tested.

When generation fails

Confirm model access and retry after reviewing the error state. Keep your existing prompt available. A returned prompt is an editing suggestion, not evidence that your business rule or integration works. Continue with conversation nodes and testing.

AI editing controls

A flow assistant’s “verified” or “goal met” message describes its editing checks. It does not prove that an actual caller reached the correct destination or that an external booking succeeded. Run your own test scenarios, then test the published channel.