How to Build a Production Voice Agent with n8n & Retell AI (2026)

The original approach to building voice agents required stitching together three separate APIs: a speech-to-text provider (like AssemblyAI), an LLM (like OpenAI GPT-4), and a text-to-speech engine (like ElevenLabs). This created a complex orchestration challenge where every millisecond of added latency degraded the conversation experience. Retell AI eliminates this complexity by providing a single managed platform that handles the entire real-time voice conversation loop — speech recognition, LLM reasoning, and voice synthesis — in one optimised pipeline. The result is sub-800ms total response latency, which is below the threshold where humans perceive conversational delay. But a voice agent that can only talk is not useful. It needs to check calendar availability, create CRM records, send notifications, and trigger follow-up workflows. This is where n8n enters the stack. By connecting Retell AI's tool-calling system to n8n webhooks, you get a clean separation between voice infrastructure and business logic. This separation is critical for production environments. Your voice layer scales independently from your business rules. You can update booking logic, add new integrations, or change CRM providers without touching the voice configuration. Each layer is independently testable, deployable, and monitorable.

Understanding the full data flow is essential before writing any configuration. Here is exactly what happens from the moment a caller dials your number to the moment your team receives a Slack notification with the booking confirmation: 1. Inbound Call Arrives: A customer calls your Twilio or Retell-provisioned phone number. Retell AI answers immediately with a configured greeting. 2. Real-Time Conversation: Retell's pipeline streams audio through STT, sends text to the configured LLM (GPT-4o or Claude), and converts the response back to speech. The caller hears a natural voice within 800ms. 3. Tool Call Triggered: When the caller requests an action (e.g. "I'd like to book an appointment for Thursday"), the LLM identifies this as a tool call and sends a structured JSON payload to your n8n webhook URL. 4. n8n Processes the Request: Your n8n workflow receives the webhook, queries Google Calendar for availability, and returns available time slots back to Retell as a structured response. 5. Confirmation Loop: Retell reads the available slots to the caller, the caller confirms, and the LLM triggers a second tool call to finalise the booking. n8n creates the calendar event, updates HubSpot, and sends a confirmation SMS. 6. Post-Call Workflow: When the call ends, Retell fires a post-call webhook to a second n8n workflow. This workflow fetches the full transcript, runs it through GPT-4o for lead scoring and summary extraction, updates the CRM deal stage, and posts a formatted summary to your team's Slack channel.

Start by creating your agent in the Retell AI dashboard. The key decisions at this stage are voice model selection, system prompt engineering, and tool definition. Voice Model Selection: Retell offers multiple voice providers including ElevenLabs and PlayHT. For UK service businesses, select a voice with clear British English pronunciation. Test multiple voices with your actual call scripts before committing — voice quality directly impacts caller trust and completion rates. System Prompt Engineering: Your system prompt defines the agent's personality, boundaries, and conversation flow. Structure it in three sections: - Identity Block: Who the agent is, which company it represents, and its core purpose ("You are Sarah, the AI receptionist for Bright Plumbing Ltd. Your job is to answer customer calls, qualify their plumbing issue, and book appointments.") - Rules Block: Hard boundaries the agent must never cross ("Never quote prices. Never diagnose issues. Always collect the caller's name, phone number, and postcode before attempting to book.") - Flow Block: The expected conversation sequence ("1. Greet the caller. 2. Ask what plumbing issue they need help with. 3. Collect name, phone, and postcode. 4. Check calendar availability. 5. Confirm the booking. 6. Thank the caller and end the call.") Tool Definition: Define the tools your agent can call. Each tool maps to an n8n webhook endpoint. For a standard service business voice agent, you need three tools: - check_availability: Queries Google Calendar for open appointment slots - book_appointment: Creates a confirmed booking with collected details - transfer_call: Escalates to a human operator for emergencies

The n8n webhook is the bridge between Retell's voice layer and your business systems. When Retell's LLM decides to call a tool, it sends a POST request to your n8n webhook URL with a JSON body containing the tool name and arguments. Your n8n workflow must handle three responsibilities: 1. Request Routing: Parse the incoming tool name and route to the correct sub-workflow. Use an n8n Switch node to branch execution based on the tool_name field. 2. Business Logic Execution: For check_availability, query the Google Calendar API with the requested date range and return available 30-minute slots. For book_appointment, create a Google Calendar event, then create or update a HubSpot contact and deal. 3. Response Formatting: Return a structured JSON response that Retell's LLM can interpret and relay to the caller. The response must be concise and conversational — the LLM will read it aloud. Critical implementation details: - Set your webhook to respond synchronously (not async). Retell expects a response within 5 seconds or it will timeout. - Include error handling at every external API call. If Google Calendar is unreachable, return a graceful fallback message like "I'm having trouble checking the calendar right now. Can I take your number and have someone call you back within 10 minutes?" - Log every incoming webhook payload to an n8n Datatable for debugging and audit trails. - Implement HMAC signature verification on the webhook to prevent unauthorised requests.

The post-call workflow transforms raw call data into actionable business intelligence. Retell fires a webhook when each call ends, sending the full transcript, call duration, caller number, and call recording URL. Your n8n post-call workflow should execute these steps: 1. Transcript Parsing with GPT-4o: Send the full transcript to OpenAI's GPT-4o with a structured extraction prompt. Extract: caller name, phone number, postcode, job type, urgency level (1-5), whether an appointment was booked, and a 2-sentence summary. 2. Lead Scoring: Based on the extracted data, assign a lead score. High-intent signals include: explicitly requesting a booking, mentioning urgency ("emergency", "flooding", "no hot water"), and providing full contact details without prompting. 3. CRM Deal Update: Update the HubSpot deal with the extracted summary, lead score, call recording URL, and transcript link. Move the deal to the appropriate pipeline stage (New Lead, Qualified, Appointment Booked, or Needs Follow-Up). 4. Team Notification: Post a formatted summary to your Slack channel with key details: caller name, job type, urgency, whether booked, and a direct link to the CRM record and call recording. 5. Follow-Up Trigger: If the caller did not book an appointment, trigger an automated SMS follow-up sequence via Twilio. Send a personalised message within 5 minutes referencing the conversation topic. This post-call pipeline ensures that no call data is lost and that your team has full context before ever speaking to the customer.

The final step connects your voice agent to real phone lines and validates the entire system end-to-end. Phone Number Provisioning: You have two options. Retell AI offers built-in phone number provisioning — you can purchase a UK number directly in the dashboard and it is automatically connected to your agent. Alternatively, use Twilio SIP trunking to connect your existing business number to Retell via SIP credentials. Testing Protocol: Before going live, run these validation checks: - Happy Path Test: Call the agent, provide all required details, and verify the booking appears in Google Calendar, the CRM deal is created, and Slack receives the notification. - Edge Case Tests: Test with background noise, heavy accents, interruptions mid-sentence, and rapid topic changes. - Failure Mode Tests: Temporarily disable Google Calendar access and verify the agent gracefully handles the error without hanging up. - Concurrent Call Test: Run 3-5 simultaneous test calls to verify your n8n webhook handles concurrent requests without race conditions. Monitoring & Alerting: Configure n8n's error workflow to send immediate Slack alerts when any webhook fails. Set up a daily health check workflow that verifies Google Calendar API connectivity, Retell agent status, and HubSpot API authentication. Go-Live Checklist: - Voice agent greeting and scripts approved by business owner - All tool-calling webhooks verified with test data - Post-call workflow producing accurate CRM updates - Error handling tested for all external API failure modes - Phone number ported or provisioned and routing confirmed - Team trained on reading Slack notifications and CRM records

Based on production deployments across UK service businesses, here are the performance benchmarks you should target: Response Latency: Sub-800ms from end of caller speech to start of agent response. Retell AI achieves this natively. If your n8n webhook adds more than 2 seconds, optimise by pre-caching calendar data or moving non-critical operations to async sub-workflows. Speech Recognition Accuracy: 95%+ on standard UK English with moderate background noise. Retell's built-in models handle most accents well. For industry-specific terminology (e.g. "combi boiler", "soil stack", "consumer unit"), add these terms to the agent's system prompt context so the LLM recognises them in conversation. Booking Completion Rate: 60-75% of qualified callers who engage with the booking flow complete a booking. The primary drop-off reason is unavailable time slots — ensure your calendar has sufficient availability during peak calling hours. Cost Per Call: Retell AI charges per minute of conversation. Combined with n8n hosting, OpenAI API costs for post-call processing, and Twilio telephony, expect £0.15-0.30 per minute of total call cost. For a typical 3-minute service booking call, that is £0.45-0.90 — significantly cheaper than a human receptionist at £12-15/hour.