Skip to main content

​
Overview

Actions transform your AI agents from conversation-only assistants into powerful business automation tools. Beyond answering questions, your agents can schedule appointments, transfer calls, update systems, and integrate with your existing business workflows - all while maintaining natural conversation flow. Think of actions as your agent’s toolkit. Just as a human employee uses different tools to complete tasks, your AI agent performs actions to execute real work during customer interactions.
Actions execute automatically based on conversation context and the instructions you provide in your agent’s prompt. You define when and how actions should be used through prompt engineering.

​
What Are Actions?

Actions are pre-configured tools that your agent can invoke during conversations to accomplish specific tasks. When a customer requests something that requires action - like booking an appointment or speaking to a specialist - your agent can execute the appropriate action seamlessly.

​
How Actions Work

During conversations, your agent autonomously decides whether to respond directly or execute an action: The agent uses action names and descriptions to understand what each action does. These are crucial for helping the LLM select the right tool. Best practices:
  • Give actions clear, descriptive names (e.g., “book_appointment” not “action1”)
  • Write detailed descriptions explaining what the action does
  • Add explicit instructions in your agent instructions (system prompt) about when to use each action
While action names and descriptions tell the agent what an action does, your agent instructions (system prompt) should specify when to use it. For example: “When a customer asks to speak to a human, use the transfer_to_support action.”

​
Action Categories

Actions are organized into two categories in the dashboard:

​
Call Control

Transfer Call

Transfer to another AI agent, phone number, or SIP address

End Conversation

Have the agent end the conversation with a custom message

​
Integration

Custom Action

Create a custom API or webhook action

Calendar Booking

Schedule appointments using Cal.com integration

​
When Actions Execute

Actions execute during the conversation when triggered by your agent based on its instructions. Unlike traditional IVR systems that follow rigid scripts, AI agents use contextual understanding to determine when actions are appropriate.

​
Trigger Mechanisms

Instruction-Based Triggers: 
When the customer asks to speak to a human, use the 'Transfer to Support' action.

After successfully booking an appointment, use the 'End Call Politely' action.

If the customer says they need time to check something, use the 'Pause Inactivity Timer' action.
Conditional Triggers: 
If the customer reports a billing issue:
1. Use the 'Lookup Account' action to retrieve their information
2. If balance is overdue, transfer to billing department
3. If balance is current, troubleshoot the issue

{% if contact.contact_status == "vip" %}
Always offer to transfer VIP customers to dedicated support immediately.
{% endif %}
Multi-Step Workflows: 
Appointment Booking Flow:
1. Gather required information (name, email, preferred date)
2. Use 'Check Availability' action to query Cal.com
3. Present options to customer
4. Use 'Book Appointment' action to confirm
5. Use 'Send SMS Confirmation' action if customer provided phone
6. Use 'End Call Politely' action to conclude
Actions execute in real-time during the call. Ensure your endpoints are reliable and respond quickly to avoid awkward pauses in conversation.

​
Configuring Actions

All action configuration happens in the agent editor under the Abilities → Actions tab.
1

Navigate to Actions

Open your agent in the editor and click Abilities → Actions tab
2

Choose Action Type

Browse the Call Control and Integration categories and click Add on the action you want to configure
3

Configure Action

Fill in the action-specific configuration form:
  • Name: Give your action a clear, descriptive name
  • Description: Explain what this action does
  • Action-specific settings: Configure parameters based on the action type
4

Save

Save the action to add it to your agent’s toolkit

​
Action Status Indicators

After adding actions, you’ll see them listed in your agent’s Actions tab with status indicators:

Configured

Action is properly configured and ready to use

Incomplete

Missing required configuration or credentials

Error

Configuration error or authentication failure

​
Referencing Actions in Instructions

To use actions, reference them by exact name in your agent’s instructions:

​
Direct Reference

When a customer asks to speak to someone about billing,
use the 'Transfer to Billing Department' action.

​
With Conditions

If the customer's issue cannot be resolved:
1. Apologize for the inconvenience
2. Explain you're connecting them to a specialist
3. Use the 'Transfer to Support' action

​
With Parameters

After gathering the customer's email and preferred date,
use the 'Book Consultation' action to schedule the meeting.
Action names are case-sensitive and must match exactly as configured. If you rename an action, update all references in your instructions.

​
Configuration Best Practices

Begin with basic actions before adding complex integrations. Add one action at a time, test thoroughly, then add the next.Example progression:
  1. Add End Conversation action
  2. Add Transfer to Support action
  3. Add Booking action
  4. Add Custom API actions
Action names are critical - the LLM uses them to understand what each tool does. Use descriptive, action-oriented names that clearly convey the function’s purpose.Why this matters: The model relies on function names and descriptions to detect when a function needs to be called and choose the right tool for the task.Good names:
  • “Get Customer Account” - Clear action verb + specific target
  • “Transfer to Billing Department” - Specific destination included
  • “Book 30-Minute Consultation” - Includes relevant details
Poor names:
  • “Action 1” - No context about what it does
  • “Transfer” - Too generic, unclear where
  • “API Call” - Doesn’t describe the action
Action descriptions help the LLM understand what the action does. The description should explain the action’s purpose, what it returns, and what parameters it uses.Best practices from OpenAI function calling:
  • Clearly describe what the action does and what it returns
  • Explain what parameters or data it uses
  • Use precise language that guides the model’s understanding
  • Keep it concise yet comprehensive
Example:
Name: Get Customer Account
Description: Retrieves customer account data from Salesforce CRM using
their phone number. Returns account status, balance, and recent orders.
Note: Describe what the action does in the description. Specify when to use it in your agent instructions (system prompt).
Test each action in the agent test interface before going live:
  • Verify action executes correctly
  • Test success scenarios
  • Test failure scenarios
  • Verify error handling
  • Check conversation flow
Configure fallback behaviors for when actions fail. Instruct your agent what to do when actions don’t work.
If the 'Book Appointment' action fails:
1. Apologize sincerely
2. Offer to have someone call back to schedule
3. Collect their preferred callback number
4. Use the 'End Call Politely' action
Ensure agents gather required data before executing actions. Don’t attempt to book appointments without email addresses or transfer calls without explaining why.
Before using the 'Book Consultation' action:
1. Confirm the customer wants to schedule
2. Ask for their email address if not in contact record
3. Discuss their preferred dates and times
4. Explain what the consultation will cover
5. Only then execute the booking action
Use appropriate authentication for all custom actions. Never expose API keys or credentials in URLs or unencrypted fields.
  • Use Bearer tokens for API authentication
  • Use Basic auth over HTTPS only
  • Store sensitive credentials securely
  • Rotate credentials regularly

​
Testing Actions

Before deploying agents with actions, thoroughly test in the dashboard test interface:
1

Open Test Interface

Click Test Call in the agent editor’s top-right corner
2

Start Web Call

Click Start Web Call to begin a test conversation
3

Trigger Each Action

Run through scenarios that trigger each configured action
4

Verify Execution

Check that actions execute correctly and handle responses appropriately
5

Test Failure Cases

Simulate failures to verify error handling works as expected
6

Review Transcript

Examine the conversation transcript to ensure flow is natural and actions integrate smoothly

​
What to Test

For Transfer Actions:
  • Transfer executes to correct destination
  • Hold music plays if configured
  • Transfer message is appropriate
  • Cold vs warm transfer modes work correctly
For Booking Actions:
  • Availability is retrieved correctly
  • Booking confirms successfully
  • Email/SMS notifications send properly
  • Timezone handling is accurate
For Custom Actions:
  • API calls succeed
  • Authentication works
  • Response data is available to agent
  • Error responses are handled gracefully
For All Actions:
  • Agent references action by correct name
  • Agent gathers required information first
  • Conversation flow remains natural
  • Failures don’t break the conversation
Test calls use real integrations. If you’re testing a booking action, it will create real appointments in your Cal.com account. Clean up test data afterward.

​
Common Use Cases

​
Customer Support Workflow

Agent Configuration:
- Transfer to Support (for complex issues)
- Lookup Customer Account (custom API)
- Create Support Ticket (custom API)
- End Call Politely

Instructions:
When a customer calls:
1. Greet them warmly
2. Use 'Lookup Customer Account' to retrieve their information
3. Ask about their issue
4. If you can resolve it, do so using knowledge base
5. If it's complex, use 'Create Support Ticket' and provide ticket number
6. If customer requests human, use 'Transfer to Support'
7. When resolved, use 'End Call Politely'

​
Appointment Booking Workflow

Agent Configuration:
- Book Consultation (Cal.com booking)
- Send SMS Confirmation (optional)
- Transfer to Scheduling (fallback)
- End Call Politely

Instructions:
When a customer wants to book:
1. Ask what type of appointment they need
2. Collect email address if not in contact record
3. Discuss their preferred dates
4. Use 'Book Consultation' to show availability and confirm
5. If booking succeeds, confirm details verbally
6. If booking fails, use 'Transfer to Scheduling'
7. Use 'End Call Politely' to finish

​
Sales Qualification Workflow

Agent Configuration:
- Lookup Company Data (custom API)
- Update CRM Lead (custom API)
- Transfer to Sales (for qualified leads)
- End Call Politely

Instructions:
For outbound sales calls:
1. Introduce yourself and purpose
2. Use 'Lookup Company Data' to retrieve firmographics
3. Ask qualifying questions (budget, timeline, authority)
4. Use 'Update CRM Lead' with qualification status
5. If qualified, use 'Transfer to Sales' with context
6. If not qualified, thank them and use 'End Call Politely'

​
Troubleshooting Common Issues

Problem: Agent doesn’t use the action even though it should.Solutions:
  • Verify action name in instructions matches exactly (case-sensitive)
  • Check action status is “Configured” not “Incomplete” or “Error”
  • Make instructions more explicit about when to use action
  • Test in isolation by explicitly asking agent to use the action
  • Review conversation transcript to see agent’s reasoning
Problem: Agent verbally confirms it’s executing an action (e.g., “I’m transferring you now”) but the action doesn’t execute until the next conversation turn.Why this happens: The agent generates a response and executes the action in the same turn, but only one can happen per turn.Solution: Prompt the agent to ask for user confirmation before executing actions:
Before using the 'Transfer to Support' action:
1. Explain why you're transferring them
2. Ask "Would you like me to transfer you now?"
3. Wait for confirmation
4. Once confirmed, execute the 'Transfer to Support' action
This ensures the agent executes the action immediately after receiving confirmation, not in the same turn as announcing it.
Problem: Custom actions fail with 401/403 errors.Solutions:
  • Verify credentials are correct and not expired
  • Check authentication type matches API requirements
  • Ensure Bearer tokens include “Bearer” prefix if needed
  • Test endpoint independently with Postman or curl
  • Review API documentation for authentication format
Problem: Cal.com booking action returns errors.Solutions:
  • Verify Cal.com integration is connected
  • Check event type exists and is active
  • Ensure meeting platform is configured in Cal.com
  • Verify scheduling windows allow requested dates
  • Check timezone configuration
  • Test booking manually in Cal.com to verify availability
Problem: Transfer action executes but doesn’t connect.Solutions:
  • Verify destination is in correct E.164 format (e.g., +15551234567)
  • Ensure phone number is reachable (for phone transfers)
  • Verify SIP address is correct (for SIP transfers)
  • Test destination independently

​
Next Steps

Call Control Actions

Configure Transfer, End Conversation, and Inactivity Management actions

Transfer Details

Master transfer destinations, modes, and advanced settings

Booking Integration

Set up Cal.com appointment scheduling with email and SMS notifications

Custom API Actions

Connect your agents to external systems and APIs

Instructions Guide

Learn how to write effective agent instructions

Prompt Engineering

Master advanced prompting techniques for reliable agent behavior