Workflow during the call

This feature is not enabled by default. Please contact support to activate it on your account.

A live call workflow gives your agent an action it can run in the middle of a conversation. The agent decides when to use it, provides the parameters you defined, and uses the result in its next answer. Typical uses:

Look up customer information from an identifier given by the caller
Check the status of an order or a request
Create or update a record in your systems

A pre-call workflow runs once, before the call starts. A live call workflow runs during the call, when the agent decides to use it.

How it works

During the conversation, the agent decides from the action name and description that it needs the action.
The agent fills in the properties you defined, for example a customer_id, using what the caller said.
The workflow runs: the trigger receives the properties, and the intermediate actions call your systems.
The last step, Respond to the agent, returns the result, and the agent uses it to continue the conversation.

A live call workflow generally consists of:

Live call trigger: defines when and how the agent can use the action.
Intermediate actions: use connectors to perform actions.
Respond to the agent: returns a result to the agent. It must be the last step of the flow.

The caller is waiting while the workflow runs, and the maximum execution time is 15 seconds. Keep the workflow short, limit the number of calls to external systems, and return only what the agent needs.

Timeout

A live call action has a maximum execution time of 15 seconds. This limit doesn’t apply to transfers. If the workflow doesn’t return a response within 15 seconds, it stops and the call continues, without the result of the action.

Configure the live call trigger

The live call trigger is the component that exposes the action to your conversational agent during the call.

Define the action

Set up the basic information of the action:

Action name: a unique identifier that the agent uses to invoke this action, for example get_customer_info. Use lowercase words separated by underscores, starting with a verb.
Action description: a clear explanation of the purpose of the action, for example Retrieves customer information based on a given identifier.

Configure dynamic properties

Dynamic properties define the parameters that the agent must provide to run the action:

Click Add Item for each required parameter.
For each property, specify:

Property name, for example customer_id.
Description, for example The unique identifier of the customer to look up.
Type: String, Boolean, Integer, and so on.
Required: whether the parameter is mandatory.

Prepare sample data

Sample data lets you test the action:

Load a sample data set and copy it into the Sample data field.
Modify the sample data by adding the action parameters.
Click Retest.

The action description should clearly explain when and why the agent should use this action. A precise description helps the agent determine the right moment to use it.

Type each property correctly to avoid runtime errors. For example, if you expect a numeric identifier, use the Integer type.

Write descriptions the agent can act on

The agent relies on the action description and the property descriptions to decide when to run the action and what to pass. Say when to use the action, and what each value looks like.

Element	Too vague	Specific
Action description	`Customer info`	`Retrieves a customer's profile and open orders from the CRM. Use it when the caller gives their customer number and asks about their account or an order.`
Property description	`id`	`The customer number stated by the caller, digits only`

Use data from previous steps

Each step can use the data produced by the steps that come before it. To insert a value:

Click in any input field. The data selector opens.
Expand the items to check what they contain, then click a value to insert it at the cursor position. You can mix fixed text and dynamic values in the same field.

Test each step before adding the next one. A step that has not been tested has no data to offer in the selector. If a value is missing from the selector, write its path manually:

Data path syntax

{{step_name['output']['path']['to']['property']}}

To get the step_name['output'] part right, right-click the step on the canvas and select Copy reference.

You can access pre-call context data in your action during the call. This allows you to reuse previously retrieved information as well as system variables.

Configure the response

In the Respond to the agent step, define the response structure in JSON format, and use the data selector to insert values from previous steps.

To return a result that the agent can use during the call, the last step of the flow must be Respond to the agent.

Use clear key names and return only the information the agent needs to answer. Also plan for the case where nothing is found, so the agent can tell the caller instead of guessing.

Result found
No result

Result found

{
  "found": true,
  "customerName": "Camille Martin",
  "lastOrder": "Order 10482, delivered on 3 November",
  "openTicket": "Delivery address change requested on 4 November"
}

No result

{
  "found": false,
  "message": "No customer found with this number"
}

Complete example: look up a customer

This workflow lets the agent retrieve a customer’s information when the caller gives their customer number.

Live call trigger

Set the action name to get_customer_info and describe when to use it. Add a required property customer_id of type Integer, described as the customer number stated by the caller.

Look up the customer

Add a step for your CRM and search for the customer using the value of customer_id. Test the step to see the fields the CRM returns.

Respond to the agent

Build the response from the CRM fields. If the lookup finds no customer, return a response with found set to false.

Test and debug

Test the whole workflow before attaching it to a live agent.

Use sample data that includes the action parameters, then click Retest to check the result of the trigger.
Test each step to confirm that the data you need is available.
Check that the response of the last step is valid JSON.
After real calls, open the workflow run history to review each run. Each step shows whether it succeeded or failed, with its input and output.

Common issues

The agent never uses the action

The action description may not make clear when to use it. Describe the situation in which the agent should run the action, and check that the action is set up on the agent.

The agent passes wrong or missing values

Check the property descriptions and types. Describe the format of each value, and mark as required the properties the workflow cannot run without.

The agent doesn't use the result

Check that the last step of the flow is Respond to the agent, and that the response contains the information the agent needs, under clear key names.

The call continues without the result of the action

The workflow most likely exceeded the 15-second limit without returning a response. Open the run history to find the slow step, and reduce or speed up the calls to external systems.

The test fails with a JSON error

The sample data or the response contains invalid JSON. Look for a missing comma, an unclosed quote, or a dynamic value inserted without quotes in a text field.

Workflows

Workflow during the call

How it works

Timeout

Configure the live call trigger

Write descriptions the agent can act on

Use data from previous steps

Configure the response

Complete example: look up a customer

Test and debug

Common issues

Next steps

Configure actions during the call

Pre-call workflow