Sales Call Summary Agent

An AI-powered agent that automatically summarizes sales calls using Vercel's Sandbox architecture. The agent analyzes call transcripts and generates structured summaries with objections, action items, and insights.

Template Note: This template uses Gong as a starting example for call transcript integration. You can adapt it to work with other call recording platforms (Zoom, Google Meet, etc.) by modifying the webhook handler and transcript fetching logic.

Extensible: This template can be extended to integrate with Salesforce (or another CRM of your choice), Slack (to post call summaries), and other services. The demo files in /demo-files/context/ demonstrate how CRM data and other context can be provided to the agent.

Features

• Structured Summaries - AI-generated summaries with tasks, objections, and key insights
• Sandbox Agent - Uses Vercel Sandbox for secure code execution and file exploration
• bash-tool - Generic bash tool for AI agents, compatible with AI SDK
• Demo Mode - Works out of the box with mock data (no Gong credentials needed)
• Objection Tracking - Identifies and scores how well objections were handled
• Durable Workflows - Built with Vercel Workflow DevKit for reliability

One-Click Deploy (Demo Mode)

You can deploy to Vercel and try it immediately with demo data with one click:

Demo mode is enabled by default - no Gong credentials required to test!

Local Development (Demo Mode)

1. Clone and Install

1
git clone https://github.com/vercel-labs/call-summary-agent
2
cd call-summary-agent
3
pnpm install

2. Link to Vercel (creates .env.local automatically)

1
vercel link
2
vercel env pull

3. Run

1
pnpm dev

4. Test

1
# Trigger the agent with demo data
2
curl -X POST http://localhost:3000/api/gong-webhook \
3
  -H "Content-Type: application/json" \
4
  -d '{}'

Production Setup (Real Gong Data)

To use real Gong API data instead of demo data:

1. Set Environment Variables

Add these to your Vercel project settings or .env.local:

1
GONG_ACCESS_KEY=your_gong_access_key
2
GONG_SECRET_KEY=your_gong_secret_key

Note: Demo mode is automatically enabled when Gong credentials are missing.

2. Configure Gong Webhook

1. Go to your Gong settings > Integrations > Webhooks
2. Create a new webhook with:
   • URL: https://your-app.vercel.app/api/gong-webhook
   • Events: Select "Call completed"
3. Save and test the webhook

Architecture

Environment Variables

Demo Mode Details

Demo mode uses realistic mock data including a sample 20-minute product demo call. The demo files are organized in the /demo-files folder:

1
demo-files/
2
├── webhook-data.json              # Mock Gong webhook payload
3
├── transcript.json                # 20-minute call transcript
4
└── context/
5
    ├── gong-calls/previous/       # Historical calls
6
    │   ├── demo-call-000-discovery-call.md
7
    │   └── demo-call-intro-initial-call.md
8
    ├── research/                  # Background info
9
    │   ├── company-research.md
10
    │   └── competitive-intel.md
11
    └── playbooks/
12
        └── sales-playbook.md

These files are loaded into the sandbox for the agent to explore.

Configuration

Custom System Prompt

Override the default system prompt:

1
AGENT_SYSTEM_PROMPT="You are a sales call analyst..."

How It Works

1. Webhook Received: Gong sends call data when a call completes (or mock data in demo mode)
2. Workflow Started: The durable workflow begins processing
3. Transcript Fetched: Call transcript is retrieved from Gong API (or mock data)
4. Sandbox Created: A secure sandbox is created with call files
5. Agent Runs: The AI agent explores transcripts using bash-tool
6. Summary Generated: Structured output with tasks, objections, insights

Workflow Steps

Each step uses the "use step" directive for:

• Automatic retries on failure
• State persistence
• Observability in Vercel dashboard

Sandbox Tools

The agent uses bash-tool for exploring call transcripts via shell commands:

1
# List call files
2
ls gong-calls/
3

4
# Search for pricing discussions
5
grep -r "pricing" gong-calls/
6

7
# View call metadata
8
cat gong-calls/metadata.json
9

10
# Find objections
11
grep -i "concern\|issue\|problem" gong-calls/*.md

All bash commands are logged for observability:

1
[bash-tool] INFO: Bash command starting { command: 'grep -r "pricing" gong-calls/' }
2
[bash-tool] INFO: Bash command completed { command: '...', exitCode: 0, stdoutLength: 605 }

Output Schema

The agent generates structured output:

1
{
2
  summary: string,           // Comprehensive call summary
3
  tasks: [{
4
    taskDescription: string,
5
    taskOwner: string,
6
    ownerCompany: 'internal' | 'customer' | 'partner'
7
  }],
8
  objections: [{
9
    description: string,
10
    quote: string,
11
    speaker: string,
12
    speakerCompany: string,
13
    handled: boolean,
14
    handledAnswer: string,
15
    handledScore: number,    // 0-100
16
    handledBy: string
17
  }]
18
}

Customization

Custom Playbooks

Add playbook detection by configuring config.playbooks in lib/config.ts.

Development

1
# Install dependencies
2
pnpm install
3

4
# Run development server
5
pnpm dev
6

7
# Build for production
8
pnpm build

Project Structure

1
sales-call-summary-agent/
2
├── app/
3
│   ├── api/gong-webhook/    # Webhook endpoint
4
│   ├── layout.tsx
5
│   └── page.tsx             # Status page
6
├── demo-files/              # Demo mode files
7
│   ├── webhook-data.json
8
│   ├── transcript.json
9
│   └── context/             # Additional context files
10
├── lib/
11
│   ├── agent.ts             # ToolLoopAgent configuration
12
│   ├── config.ts            # Centralized configuration
13
│   ├── gong-client.ts       # Gong API helpers
14
│   ├── mock-data.ts         # Demo mode loader
15
│   ├── sandbox-context.ts   # File generation for sandbox
16
│   ├── tools.ts             # Agent tools (bash-tool)
17
│   ├── types.ts             # TypeScript types
18
│   └── logger.ts            # Logging utility
19
└── workflows/
20
    └── gong-summary/
21
        ├── index.ts         # Main workflow
22
        └── steps.ts         # Workflow steps

API Reference

POST /api/gong-webhook

Receives Gong webhook payloads and triggers the summary workflow.

Request Body: Gong webhook payload or empty {} in demo mode.

Response:

1
{
2
  "message": "Workflow triggered",
3
  "callId": "1234567890"
4
}

Output

The function returns a string output which can be configured into an outputSchema of your choice.

Contributing

Contributions are welcome! Please open an issue or submit a pull request.