Voice AI Workflows

When a customer speaks with a Voice AI agent, the conversation is captured as a transcript—a single text string containing the entire exchange. Workflows parse this transcript, extract structured data, and push it to your claims system, AMS, or any API.

How Voice Transcripts Work

During a call, Voice AI captures everything said by both the caller and the AI agent:

Assistant: Thank you for calling Acme Insurance. How can I help you today?
User: Yeah, I need to file a claim. My basement flooded.
Assistant: I'm sorry to hear that. I can help you file a claim right now. Can I get your policy number?
User: It's HO-12345678.
Assistant: Thank you. And what's your name?
User: John Smith.
Assistant: Got it, John. When did the flooding occur?
User: This morning, around 6 AM.

This transcript is passed to a workflow as one big string. The workflow’s job is to:

Parse the transcript into structured data
Transform that data into the format your system expects
Send it to your claims platform, AMS, or API
Return a response so the AI can confirm with the caller

Example: FNOL to Snapsheet

Let’s walk through a real workflow that takes a First Notice of Loss conversation and creates a claim in Snapsheet (a claims management system).

Start → AI Prompt Node → Mapping Node → Task Node → End

Receive the transcript

The workflow receives the full conversation transcript from Voice AI

Parse with AI

An AI Prompt Node extracts structured data from the unstructured text

Transform the data

A Mapping Node formats the data for Snapsheet’s API

Create the claim

A Task Node sends the data to Snapsheet and receives the claim number

Return confirmation

The response bubbles back to Voice AI: “Your claim number is CLM-98765”

AI Prompt Node

The AI Prompt Node lets you interact with LLMs directly from your workflow. Use it to parse transcripts, extract entities, classify intent, or generate responses.

Parsing a Transcript

Configure the AI Prompt Node to extract structured data:

Prompt Configuration
Sample Output

Parse the following transcript and extract all claim information 
as a JSON object. Include:
- claimType (e.g., water, fire, theft, auto)
- claimDetails (description of what happened)
- callerFirstName
- callerLastName
- policyNumber
- dateOfLoss
- timeOfLoss
- propertyAddress (if mentioned)
- contactPhone (if mentioned)

Transcript:
{{context.transcript}}

Respond with only the JSON object, no additional text.

The AI returns structured data extracted from the conversation:

{
  "claimType": "water",
  "claimDetails": "Basement flooded",
  "callerFirstName": "John",
  "callerLastName": "Smith",
  "policyNumber": "HO-12345678",
  "dateOfLoss": "2025-02-13",
  "timeOfLoss": "06:00",
  "propertyAddress": null,
  "contactPhone": null
}

The AI Prompt Node uses the same technology that powers Voice AI—you can configure LLM interactions for both real-time conversations and backend processing.

Mapping Node

The Mapping Node transforms data using JSONata expressions. Use it to restructure the AI’s output into the exact format your target system requires.

Example: Format for Snapsheet API

Input (from AI Prompt)
JSONata Expression
Output (for Snapsheet)

{
  "claimType": "water",
  "claimDetails": "Basement flooded",
  "callerFirstName": "John",
  "callerLastName": "Smith",
  "policyNumber": "HO-12345678",
  "dateOfLoss": "2025-02-13",
  "timeOfLoss": "06:00"
}

{
  "claim": {
    "type": claimType,
    "description": claimDetails,
    "loss_date": dateOfLoss & "T" & timeOfLoss & ":00Z",
    "policy_id": policyNumber
  },
  "reporter": {
    "full_name": callerFirstName & " " & callerLastName,
    "relationship": "policyholder"
  }
}

{
  "claim": {
    "type": "water",
    "description": "Basement flooded",
    "loss_date": "2025-02-13T06:00:00Z",
    "policy_id": "HO-12345678"
  },
  "reporter": {
    "full_name": "John Smith",
    "relationship": "policyholder"
  }
}

Task Node

The Task Node calls external systems via HTTP. This is how you push data to Snapsheet, Guidewire, your AMS, or any API.

Configuring a Connection

Before using a Task Node, create a Connection to your target system:

Navigate to Connections in the left sidebar
Click Create a connection
Configure authentication and endpoints for QA and Production

Example: Snapsheet Connection

Setting	QA Environment	Production Environment
Name	Snapsheet Claims API	Snapsheet Claims API
Type	REST - Bearer Token	REST - Bearer Token
Base URL	`https://api.staging.snapsheet.com/v1`	`https://api.snapsheet.com/v1`
Token	`sk_test_xxxxx`	`sk_live_xxxxx`

Task Node Configuration

Setting	Value	Description
Connection	Snapsheet Claims API	Select your pre-configured connection
Method	POST	HTTP method for creating a claim
URL	`/claims`	Appended to base URL
Request Body	`{{context.claim}}`	Data from the Mapping Node

Handling the Response

Configure where to store the API response:

Outcome	Property	Usage
On Success	`snapsheetResponse`	Stores claim ID, confirmation number
On Error	`snapsheetError`	Stores error message, validation details

The response is added to your workflow context, making it available to subsequent nodes or to return to Voice AI.

Returning Data to Voice AI

After the Task Node succeeds, the claim number is in your context:

{
  "snapsheetResponse": {
    "claim_id": "CLM-98765",
    "status": "created",
    "message": "Claim successfully submitted"
  }
}

The workflow returns this to Voice AI, which can then say:

“Great news, John. Your claim has been submitted successfully. Your claim number is CLM-98765. Is there anything else I can help you with?”

Error Handling

Every node has success and error paths. Hover over a node to see the + (success) and ⚠ (error) connectors.

Branching on Errors

Task Node (Snapsheet)
    ├── Success → End Node (return claim number)
    └── Error → Error Handler → End Node (return error message)

Common Error Scenarios

Error Type	Cause	Response to Caller
Validation Error	Missing or invalid data	”I’m missing some information. Could you repeat your policy number?”
API Error (500)	Snapsheet system issue	”I’m having trouble submitting your claim right now. Let me transfer you to an agent.”
Auth Error (401)	Invalid credentials	Escalate to human agent
Timeout	API not responding	Retry or escalate

Configure retry strategies on Task Nodes for transient errors (500, 502, 503). Use exponential backoff for production resilience.

Additional Node Types

Beyond the core nodes, Liberate offers specialized nodes for complex logic:

Code Node

Write and execute custom JavaScript or Python code when JSONata isn’t enough:

// Example: Calculate days since loss
const lossDate = new Date(context.dateOfLoss);
const today = new Date();
const daysSinceLoss = Math.floor((today - lossDate) / (1000 * 60 * 60 * 24));

return {
  ...context,
  daysSinceLoss: daysSinceLoss,
  isUrgent: daysSinceLoss <= 1
};

Decision Table Node

Quickly map inputs to outputs using a visual table—no code required:

Claim Type	Amount	Priority	Assignment
water	greater than 50000	high	senior_adjuster
water	50000 or less	medium	auto_adjuster
fire	any	high	senior_adjuster
theft	any	medium	auto_adjuster

Entity Table Node

Maintain a static CSV for reference data lookups:

state_code,state_name,timezone
CA,California,America/Los_Angeles
TX,Texas,America/Chicago
FL,Florida,America/New_York

Use for mapping codes to names, looking up regional settings, or maintaining reference data that doesn’t change often.

Platform Architecture

Understanding how the pieces connect:

Organization (Your Company)
    └── Project (Claims Automation)
            └── Agent (Claims Voice AI)
                    └── Tool (FNOL Workflow)
                            └── Workflow (slug: fnol-to-snapsheet)

Concept	Description
Organization	Your company’s Liberate instance
Project	Container for related agents and workflows
Agent	A configured Voice/Email/SMS AI
Tool	Links an agent to a workflow
Workflow Slug	Unique ID in the URL that identifies the workflow

When you create a workflow, a unique slug is generated in the URL (e.g., fnol-to-snapsheet-a1b2c3). This slug tells the AI which workflow to execute when a tool is invoked.

Next Steps

Workflow Quickstart

Build your first workflow step-by-step

Node Reference

Detailed docs for each node type

Connections Guide

Set up integrations with your systems

FNOL Example

Complete claims intake workflow

Getting Started

Voice AI

Email & SMS AI

Use Cases

Integrations

Workflow Builder

Dashboard & Analytics

Support

Voice AI Workflows

How Voice Transcripts Work

Example: FNOL to Snapsheet

AI Prompt Node

Parsing a Transcript

Mapping Node

Example: Format for Snapsheet API

Task Node

Configuring a Connection

Task Node Configuration

Handling the Response

Returning Data to Voice AI

Error Handling

Branching on Errors

Common Error Scenarios

Additional Node Types

Code Node

Decision Table Node

Entity Table Node

Platform Architecture

Next Steps

Workflow Quickstart

Node Reference

Connections Guide

FNOL Example

​How Voice Transcripts Work

​Example: FNOL to Snapsheet

​AI Prompt Node

​Parsing a Transcript

​Mapping Node

​Example: Format for Snapsheet API

​Task Node

​Configuring a Connection

​Task Node Configuration

​Handling the Response

​Returning Data to Voice AI

​Error Handling

​Branching on Errors

​Common Error Scenarios

​Additional Node Types

​Code Node

​Decision Table Node

​Entity Table Node

​Platform Architecture

​Next Steps

Workflow Quickstart

Node Reference

Connections Guide

FNOL Example

How Voice Transcripts Work

Example: FNOL to Snapsheet

AI Prompt Node

Parsing a Transcript

Mapping Node

Example: Format for Snapsheet API

Task Node

Configuring a Connection

Task Node Configuration

Handling the Response

Returning Data to Voice AI

Error Handling

Branching on Errors

Common Error Scenarios

Additional Node Types

Code Node

Decision Table Node

Entity Table Node

Platform Architecture

Next Steps