Skip to main content

​
Workflows

Workflows let you design structured, multi-step conversation flows using a visual node editor. Unlike single-prompt agents, workflows give you fine-grained control over the conversation path — including branching, conditions, and step-by-step data collection.

​
When to Use Workflows

Use workflows when you need predictable, repeatable conversation paths. For flexible, open-ended interactions, a single-prompt agent with a knowledge base is usually sufficient.

​
Workflow Editor

The visual editor lets you build workflows by connecting nodes on a canvas. Each node represents a step in the conversation, and edges define the flow between steps.

​
Node Types

​
Start Node Properties

The Start node configures the initial behavior of every call in the workflow:

​
General

​
Text-to-Speech (TTS)

The Start node includes a full TTS configuration:
The TTS voice configured in the Start node applies to the entire workflow. All conversation steps will use this voice.

​
Background Sound

Add ambient audio to create a more natural call experience:

​
Conversation Node

Each Conversation node represents one step where the agent speaks and optionally collects information:
  • Step prompt: what the agent should say and how to respond to the caller
  • Data extraction: variables to extract from the caller’s response
  • Validation: conditions that must be met before moving to the next step

​
Condition Node

Condition nodes enable branching logic:
  • Evaluate variables collected in previous steps
  • Compare values (equals, contains, greater than, etc.)
  • Route to different paths based on the result
Example: After collecting the caller’s interest level:
  • If interest = high → route to appointment booking
  • If interest = low → route to email follow-up
  • Otherwise → route to general information

​
Transfer Node

Transfer the call to a human agent or external number:
  • Transfer target: phone number or SIP address
  • Transfer message: what the agent says before transferring (e.g. “I’m connecting you with a specialist now”)

​
End Node

Terminate the call gracefully:
  • Closing message: optional farewell message (e.g. “Thank you for calling. Have a great day!”)

​
Building a Workflow

1

Create a workflow agent

Navigate to Build > Agents, click New Workflow Agent, and name it.
2

Configure the Start node

Set the welcome message, language, background sound, and TTS settings.
3

Add conversation steps

Drag Conversation nodes onto the canvas. Define what the agent should say and what information to collect at each step.
4

Add branching logic

Use Condition nodes to route the call based on collected data. Connect the condition outputs to different Conversation, Transfer, or End nodes.
5

Connect nodes

Draw edges between nodes to define the flow. Every path must eventually reach an End or Transfer node.
6

Save and test

Click Save to persist the workflow. Test by calling the assigned phone number or using the built-in chat.

​
Global Prompt

Workflow agents have a Global Prompt (instead of a System Prompt). This prompt applies to all steps in the workflow and defines the agent’s overall behavior, tone, and constraints. Individual step prompts complement the global prompt with step-specific instructions.
Use the global prompt for personality, tone, and rules that apply everywhere. Use step prompts for step-specific instructions like “Ask for the caller’s name” or “Confirm the appointment details.”

​
Best Practices

Complex branching is powerful but harder to debug. Start with a linear flow (Start → Step 1 → Step 2 → End) and add branches only where the conversation genuinely needs to diverge.
Make test calls that exercise each branch. Untested paths are where bugs hide. Pay special attention to edge cases and fallback routes.
Name your nodes descriptively (e.g. “Collect Name”, “Qualify Interest”, “Book Appointment”) so the workflow is readable at a glance.
Add a catch-all path from condition nodes that handles unexpected responses. Don’t leave the caller in a dead-end.
Each conversation step should have one clear goal (ask one question, confirm one piece of information). Long, multi-part steps confuse both the LLM and the caller.