Automated Cold Calling for B2B Leads with Vapi + OpenAI: A Step-by-Step Build

You can spin up a fully hands-free outbound dialer that greets prospects, answers objections, and writes the outcome straight into your CRM - all without hiring additional SDRs. In this guide I'll show you how to connect Vapi's voice-AI gateway to OpenAI's conversational model, enrich the call with ElevenLabs TTS, and push results through a webhook to a CRM of your choice. By the end you'll have a production-ready pipeline that can run 50-100 calls per hour and feed qualified leads directly into your funnel.

Voice AI agent is a software service that receives a phone call, turns the caller's speech into text, runs that text through a language model, and replies with synthesized speech, all in real time.

What you need

\*Pricing notes are deliberately vague; always verify the latest rates on the provider's website before committing budget.

Rough build time: 8-12 hours for a developer comfortable with Docker and REST APIs.

Build it

1. Spin up the infrastructure

Create a Docker-Compose file that launches n8n, PostgreSQL, and a tiny Reverse Proxy (Caddy) to expose the webhook endpoint securely.

# docker-compose.yml
version: "3.8"
services:
 db:
 image: postgres:15
 environment:
 POSTGRES_USER: aab_user
 POSTGRES_PASSWORD: aab_pass
 POSTGRES_DB: aab_calls
 volumes:
 - db_data:/var/lib/postgresql/data

 n8n:
 image: n8nio/n8n:latest
 ports:
 - "5678:5678"
 environment:
 - DB_TYPE=postgresdb
 - DB_POSTGRESDB_HOST=db
 - DB_POSTGRESDB_PORT=5432
 - DB_POSTGRESDB_DATABASE=aab_calls
 - DB_POSTGRESDB_USER=aab_user
 - DB_POSTGRESDB_PASSWORD=aab_pass
 depends_on:
 - db

 caddy:
 image: caddy:2
 ports:
 - "443:443"
 volumes:
 - ./Caddyfile:/etc/caddy/Caddyfile
 depends_on:
 - n8n

volumes:
 db_data:

This file creates three containers: a PostgreSQL instance for persisting call data, n8n as the orchestrator, and Caddy to terminate TLS for the public webhook URL.

Run docker compose up -d. Verify each service is reachable (docker ps should list three containers).

2. Register a Twilio phone number

Log into the Twilio console, buy a local number, and point its Voice → Webhooks URL to the Caddy-exposed endpoint, e.g.:

https://yourdomain.com/webhook/vapi

Set the request method to POST and enable Webhook for both Incoming Calls and Call Status.

3. Connect Twilio ↔ Vapi

Create a Vapi Outbound Flow that uses the Twilio number as the source. In the Vapi dashboard go to Flows → New Flow → Outbound and set:

• Source Number: the Twilio number you just bought
• Destination: dynamic; will be supplied by n8n when the workflow triggers
• Callback URL: https://yourdomain.com/webhook/vapi (the same endpoint you configured in Twilio)

Vapi will POST a JSON payload for each call event (ringing, answered, speech, etc.) to that URL.

4. Design the n8n workflow

1. Webhook Node - listens on /webhook/vapi. Set Response Mode to "Immediately respond 200 OK" so Twilio/Vapi don't time out.
2. IF Node - checks event.type field. If event.type === "answer" proceed; otherwise ignore.
3. HTTP Request Node (Speech-to-Text) - call Vapi's /speech-to-text endpoint, passing event.recording_url. Store result in {{ $json["transcript"] }}.
4. OpenAI Node - prompt the latest transcript with a system prompt that encodes your sales script. Example prompt (see code block below).
5. ElevenLabs TTS Node - feed OpenAI's reply text, receive an audio URL, and return it to Vapi via a HTTP Request PUT /call/{callId}/audio.
6. PostgreSQL Node - INSERT call metadata (caller, transcript, AI reply, confidence score) into calls table.
7. CRM HTTP Node - POST qualified lead JSON to your CRM's REST endpoint (HubSpot contacts, for example).

#### OpenAI Prompt (Step 4)

{
 "model": "gpt-4o-mini",
 "messages": [
 {
 "role": "system",
 "content": "You are a friendly B2B sales rep for a SaaS company that offers a data-enrichment platform. Your goal is to qualify the prospect, capture company name, size, and pain point, and schedule a 30-minute demo if interest is shown. Keep the conversation under 90 seconds."
 },
 {
 "role": "assistant",
 "content": "Hi, this is Alex from DataBoost. I'm calling because we help companies like {{company_name}} improve lead quality. Do you have a minute to see if we might help?"
 },
 {
 "role": "user",
 "content": "{{transcript}}"
 }
 ],
 "temperature": 0.2,
 "max_tokens": 200
}

The JSON is sent directly to https://api.openai.com/v1/chat/completions. The response's choices[0].message.content becomes the text that ElevenLabs will speak back to the prospect.

5. Deploy the webhook in Caddy

Create a Caddyfile that terminates TLS with Let's Encrypt and forwards /webhook/vapi to the n8n container:

yourdomain.com {
 encode gzip
 reverse_proxy /webhook/vapi n8n:5678
}

Reload Caddy (docker exec caddy caddy reload). Verify the endpoint works with a simple curl -X POST https://yourdomain.com/webhook/vapi -d '{}'. You should receive 200 OK.

6. Test end-to-end

1. Use Twilio's Call Simulator (or dial the number from a real phone).
2. When the call connects, Vapi streams inbound audio to the webhook.
3. n8n converts speech → text → OpenAI → ElevenLabs → Vapi, completing a round-trip conversation.
4. After the call, check PostgreSQL (SELECT * FROM calls ORDER BY created_at DESC LIMIT 1;) and your CRM dashboard for a new lead entry.

If everything logs correctly, you have a functional automated cold calling for b2b leads pipeline.

7. Add lead scoring

Create a second IF Node after the OpenAI response that looks for keywords like "interested", "demo", "budget". Set a numeric score (e.g., 80) when those appear; otherwise assign a lower score (20). Store the score in the PostgreSQL row and send it to the CRM as a custom field. This enables your sales team to prioritize follow-ups automatically.

Where this breaks

Key warning: The whole chain is only as reliable as its slowest link. In practice, the OpenAI request adds ~500 ms of latency; budget extra buffer in your call flow to avoid breaking the natural conversation rhythm.

For a deeper technical reference, see n8n's documentation.