Vapi.ai Tutorial: Build Voice AI Agents from Scratch (2026)

What is Vapi.ai?

Vapi.ai is a real-time voice AI platform that lets you build AI phone agents without managing the underlying infrastructure. It handles the complete voice pipeline: receiving phone calls, transcribing speech in real-time (Deepgram), passing to the language model (GPT-4o, Claude, or others), and converting the response back to speech (ElevenLabs, OpenAI TTS, or Cartesia).

All of this happens in under 500ms, creating the natural conversation flow that makes voice AI actually usable for real business applications.

Use Cases That Generate Real Revenue

Understanding the business applications helps you choose the right configuration:

• AI Receptionist: Answers calls, qualifies inquiries, books appointments in Google Calendar — 24/7, no sick days
• Outbound Sales Agent: Calls leads from a list, qualifies them, and books demos with human reps
• Customer Support: Handles Tier-1 support (tracking, FAQs, basic troubleshooting)
• Appointment Reminders: Calls patients/clients before appointments to confirm or reschedule
• Survey and Feedback Collection: Calls customers after service delivery to collect NPS scores

Revenue potential: A dental practice paying €800/month for an AI receptionist that handles scheduling, reminders, and FAQs is a straightforward sale when the alternative is a part-time receptionist at €1,500–€2,000/month. Our Voice Agent course (€39) covers exactly how to build and sell this.

Account Setup and First Call

Getting started with Vapi takes about 10 minutes:

1. Go to vapi.ai and create an account
2. Add a payment method (you get $10 in free credits to start)
3. Go to Dashboard → Assistants → Create Assistant
4. Configure the basic settings (name, model, voice)
5. Click Test Call to speak to your assistant in the browser

Configuring Your Assistant

The assistant configuration is where you define how your voice agent behaves. Here are the key settings:

System Prompt (First Messages)

This is the most important configuration. The system prompt defines the agent's personality, knowledge, and behavior. Here's an example for a dental receptionist:

You are Sarah, a friendly receptionist for Smile Dental Clinic.
Your job is to:
1. Greet callers warmly and ask how you can help
2. Book, reschedule, or cancel appointments
3. Answer common questions about services and pricing
4. Collect patient information for new patients

Clinic information:
- Hours: Monday-Friday 9am-6pm, Saturday 10am-3pm
- Address: 123 Main Street, Dublin
- Services: General dentistry, whitening, implants, orthodontics
- New patient exam: €85

Keep responses concise — this is a phone call, not a text message.
Always confirm appointment details before ending the call.

Voice Selection

Vapi supports multiple TTS providers. Our recommendations:

• ElevenLabs voices — highest quality, most natural. Use for client-facing agents.
• Cartesia Sonic — ultra-low latency (<100ms). Best for fast, responsive conversations.
• OpenAI TTS — good quality, fast, and cost-effective. Good middle ground.
• PlayHT — good quality with many voice options and languages.

Language Model Settings

• Model: GPT-4o-mini for most cases (fast, cheap). GPT-4o for complex reasoning.
• Temperature: 0.3–0.5 for consistent, professional responses. 0.7+ for more creative agents.
• Max tokens: 150–250 for phone conversations. Longer responses feel unnatural on calls.

Handling Real-Time Interruptions

Natural conversations require barge-in support — the ability for callers to interrupt the agent while it's speaking. Vapi handles this through its transcription layer.

// Vapi assistant configuration (JSON)
{
  "model": {
    "provider": "openai",
    "model": "gpt-4o-mini",
    "temperature": 0.4,
    "maxTokens": 200
  },
  "voice": {
    "provider": "elevenlabs",
    "voiceId": "21m00Tcm4TlvDq8ikWAM"
  },
  "transcriber": {
    "provider": "deepgram",
    "language": "en",
    "model": "nova-2"
  },
  "conversation": {
    "interruptionsEnabled": true,
    "numWordsToInterruptAssistant": 2,
    "silenceTimeoutSeconds": 30
  }
}

Tool Calls: Connecting to External Systems

Tool calls (function calling) let your voice agent access real data during the conversation. When the agent needs to check availability or book an appointment, it calls a function which returns real data from your backend.

Example: Booking an Appointment

// Tool definition in Vapi
{
  "name": "check_availability",
  "description": "Check available appointment slots",
  "parameters": {
    "type": "object",
    "properties": {
      "date": {
        "type": "string",
        "description": "Date to check, format YYYY-MM-DD"
      },
      "service_type": {
        "type": "string",
        "description": "Type of dental service needed"
      }
    },
    "required": ["date"]
  },
  "server": {
    "url": "https://your-n8n.com/webhook/check-availability"
  }
}

The server.url points to an n8n webhook. When the AI agent calls this tool, n8n receives the parameters, queries Google Calendar or your booking system, and returns available slots as JSON. The agent then speaks the results naturally.

Integrating with n8n for CRM Updates

After a call, Vapi sends a webhook with the full call summary, transcript, and any data collected during tool calls. Use n8n to process this data:

1. In Vapi Dashboard, go to Settings → Webhooks
2. Add your n8n webhook URL as the End of Call Report webhook
3. In n8n, create a workflow with a Webhook trigger
4. Parse the call data (customer name, phone, appointment booked, etc.)
5. Create or update contact in HubSpot/Salesforce/Airtable
6. Send confirmation SMS or email if needed

Transcription and Call Recording

Vapi automatically transcribes and records every call (if enabled). The end-of-call report includes:

• Full transcript — every word spoken by agent and caller
• Call summary — AI-generated summary of what was discussed
• Structured data — information extracted via tool calls
• Sentiment analysis — caller satisfaction indicators
• Duration and cost — for billing and analytics

Cost Breakdown

Understanding costs is essential for pricing client projects correctly:

• Vapi platform: ~$0.05/minute base cost
• LLM (GPT-4o-mini): ~$0.01–$0.02/minute
• Voice (ElevenLabs): ~$0.03–$0.05/minute
• Transcription (Deepgram): ~$0.005/minute
• Phone number: ~$1.15/month per number + $0.015/minute

Total: approximately $0.10–$0.15 per minute for a production voice agent with premium voice quality. A 3-minute call costs $0.30–$0.45.

Deploying with a Real Phone Number

To receive actual phone calls, you need to provision a phone number:

1. Go to Phone Numbers in Vapi Dashboard
2. Click Buy Phone Number — choose your country and area code
3. Assign the number to your assistant
4. Test by calling the number from your phone

For outbound calling campaigns, use Vapi's Calls API to programmatically initiate calls from a list — perfect for appointment reminders or follow-up sequences triggered by n8n.

Next step: If you want to build voice AI agents professionally and charge €1,500–€4,000 per project, our complete Voice Agent course (€39) covers everything from this tutorial plus advanced topics: multi-language support, sentiment escalation, and client delivery frameworks.