How to Build an AI Voice Agent for Your UK Service Business

Every missed call from a potential client is a missed sale. UK consultants, coaches, and agencies miss an average of 23 inbound calls per week — most during the working day, while they are in meetings or delivering client work. An AI voice agent answers every call in under one second, qualifies the caller, books a discovery meeting directly into your calendar, and sends a follow-up SMS automatically. Here is exactly how to build one.

Why Missed Calls Cost UK Service Businesses More Than They Think

The mathematics of missed calls are straightforward but often ignored. A consultancy averaging 23 missed calls per week, with a 20% conversion rate on qualified calls and a £3,500 average project value, is leaving roughly £16,000 per month on the table from unanswered enquiries alone. That is before factoring in the client experience damage — a prospect who called and got no answer will call your competitor next.

The fix has historically been expensive: a receptionist, a virtual PA service, or an outsourced answering service running £400–800 per month. AI voice agents change the economics entirely. A well-configured voice agent running on Vapi costs £30–60 per month in API fees and handles unlimited concurrent calls, 24 hours a day, seven days a week.

The capability has matured significantly in 2026. Latency is now under 800 milliseconds — below the threshold where callers notice they are talking to an AI. Voice quality is indistinguishable from a human in most conversations. And the integration layer — connecting the voice agent to your calendar, CRM, and SMS tools — is now achievable in n8n without custom code.

An AI voice agent is not a phone menu. It is a conversational system that listens, understands context, asks follow-up questions, and takes actions in your business systems — all in real time, while the caller is on the phone.

What Your AI Voice Agent Will Do: The Architecture

Before building, it helps to be clear about what the system does. There are three components:

• The voice layer. Vapi or Retell AI handles telephony — converting speech to text in real time, feeding it to the language model, converting the response back to speech, and streaming it to the caller with sub-second latency. This is the specialist layer. Do not try to build it yourself with raw Twilio; use a voice AI platform that handles the hard parts.
• The intelligence layer. A language model — Claude or GPT-4o — processes the conversation, follows the defined script, asks qualifying questions, handles objections, and decides when to hand off to a booking step. The model receives the conversation transcript as it builds and responds in character as your business.
• The action layer. n8n handles everything that happens after the conversation: writing the call summary to your CRM, booking the discovery meeting via Calendly or Google Calendar, sending a confirmation SMS, and notifying you on Slack with a transcript and booking link.

The three layers connect through webhooks. Vapi sends a webhook to n8n at the end of each call containing the transcript, extracted data, and outcome — booked, not interested, needs follow-up, or transferred to voicemail. n8n handles everything downstream from there.

Step-by-Step Build: AI Voice Agent in Vapi and n8n

Step 1: Set Up Your Vapi Account and Phone Number

Create an account at vapi.ai. In the Vapi dashboard, navigate to Phone Numbers and purchase a UK number. Monthly cost is approximately £3–5. This becomes the number you publish as your business enquiry line, or you forward your existing number to it during out-of-hours periods.

Step 2: Create Your AI Assistant

In the Vapi dashboard, go to Assistants and click Create Assistant. Configure the following fields:

• Name and persona. Give it a human name — Sophie or James — and write a short persona description: "You are Sophie, an assistant at [Your Business Name]. You are warm, professional, and focused on understanding how you can help the caller."
• Model. Select Claude Sonnet or GPT-4o. Claude is preferable for conversations requiring nuance and UK business context.
• Voice. Select a voice from ElevenLabs or Cartesia. For UK professional services, choose a neutral British accent. Listen to at least five options before choosing — voice is the first impression.
• First message. "Good morning/afternoon, thank you for calling [Your Business Name]. I'm Sophie. How can I help you today?"

Step 3: Write the System Prompt

The system prompt is the most important part of the build. It defines how your assistant handles every conversation. A working structure for a UK consultancy or coaching practice:

• Role definition. State clearly who Sophie is, what business she represents, and what her goal is: understanding whether the caller is a good fit and, if so, booking a discovery call.
• Qualification questions. List four to five questions Sophie should ask naturally across the conversation — not as a rigid script, but as guidance. Typically: what brings them to call today; the size and type of their business; their timeline for getting started; whether budget has been allocated. Order matters. Start with the problem they are trying to solve, not the budget.
• Booking trigger. Define the qualification criteria — what combination of answers indicates a good-fit prospect — and instruct Sophie to offer a free discovery call when those criteria are met.
• Hard rules. Include explicit instructions for edge cases: never quote pricing; if asked a question she cannot answer, say the owner will follow up directly; always end the call with a clear next step; if the caller wants to speak to a human or leave a voicemail, transfer immediately without resistance.

Step 4: Add the Booking Tool

In the Vapi assistant configuration, navigate to Tools and add a function tool called book_appointment. This connects the conversation to your calendar. The function accepts the caller's name, email address, phone number, and preferred time slot, then triggers a webhook to n8n where the booking is processed.

Define the function parameters: caller name as a required string, caller email as required, phone number as optional, preferred time as a free-text string the caller states ("tomorrow afternoon", "next Tuesday morning"). n8n handles parsing the preference against actual calendar availability. Keep the function description short and action-oriented — Vapi uses it to decide when to call the tool during conversation.

Step 5: Build the n8n Post-Call Workflow

In n8n, create a workflow triggered by the Vapi end-of-call webhook. The webhook URL goes into your Vapi assistant's server configuration. The workflow handles four outcomes:

1. Booking made. Parse the caller name, email, and preferred time from the webhook payload. Check your Google Calendar or Calendly for the next available slot matching the preference. Create the booking, send a confirmation email to the caller, and post a Slack notification to you with the transcript and booking details.
2. Interested but not booked. The caller expressed interest but did not commit to a time. Create a follow-up task in your CRM assigned to you for the next business day, with the call transcript attached.
3. Not a fit. Log the call summary to your CRM with the qualification outcome. No follow-up action needed.
4. Voicemail or transfer. Transcribe the voicemail audio using Whisper and send the transcript to Slack for same-day review.

Step 6: Configure Call Routing

Decide how and when the AI agent answers calls. Three setups work well for different business types:

• After-hours only. Forward your business number to the Vapi number between 6pm and 9am and on weekends. You answer during business hours; Sophie covers everything else. Best starting point for sole traders and small teams who want to test before committing to full-time routing.
• Overflow routing. Configure your business number to ring you first. If unanswered after three rings, forward to Sophie. The most popular setup — you handle the calls you can; Sophie catches everything you miss.
• Always-on intake. All inbound enquiry calls go to Sophie first. She qualifies and books; only confirmed discovery meetings reach your calendar. Useful when you receive high inbound volume and want to protect your time from unqualified conversations.

Testing, Going Live, and What to Monitor

Before publishing the number, test the agent by calling it yourself. Run through at least ten scenarios: a well-qualified prospect, a vague enquiry, someone asking about pricing, a wrong number, a caller who wants to speak to a human, and several edge cases your system prompt does not anticipate. Update the prompt after each gap until all scenarios are handled correctly. Record your test calls and listen back — what sounds fine on a transcript often sounds abrupt or stilted in the actual audio.

Once live, monitor three metrics weekly through Vapi's dashboard and your n8n logs:

• Conversion rate. The percentage of inbound calls that result in a booked discovery meeting. Benchmark: 15–25% in the first month, rising to 30–40% as the prompt is refined.
• Escalation rate. The percentage of calls where Sophie could not handle the conversation and transferred to voicemail or requested a callback. Target: under 10%.
• Call duration. Average call length tells you whether the conversation is flowing naturally. Under two minutes suggests Sophie is cutting conversations short; over six minutes suggests the qualification questions are too numerous or too formal.

The most common refinement after going live is the system prompt. Callers ask things you did not anticipate. When they do, update the prompt to handle those cases. Most voice agents reach a stable, high-performing state after two to three weeks of live operation and prompt iteration. This connects directly to the observability framework — the same monitoring principles that apply to text-based agents apply to voice.

Common Mistakes to Avoid

Three mistakes account for most early failures:

Overcomplicated scripts. The system prompt tries to cover every edge case in the first version. The result is an agent that sounds stilted and formulaic. Start with five core qualification questions and one outcome — book or do not book. Add complexity after you have three weeks of live call data showing where callers go off-script.

Skipping the GDPR disclosure. UK GDPR requires that callers are informed when their call is being recorded or processed by an automated system. Add a disclosure to Sophie's opening message: "Just to let you know, this call may be recorded and processed by our AI assistant for booking and follow-up purposes." Short, clear, and legally necessary for UK businesses. If you work with FCA-regulated clients, check your specific Consumer Duty obligations before going live.

No fallback to voicemail. Some callers refuse to engage with an AI agent and simply want to leave a message. Always configure a fallback: if the caller says "I'd rather speak to a human" or "can I leave a voicemail," Sophie should acknowledge, thank them, and transfer immediately. Never trap a caller in a conversation they do not want to have. The human-in-the-loop principles apply here too — your voice agent should know when to hand off, not try to handle everything itself.

A voice agent built correctly handles the intake work that previously fell through the cracks after hours, during meetings, and on weekends. It recovers revenue that was already being lost without anyone noticing — and it does it for roughly the same monthly cost as a single business lunch.

If you want help building a voice agent for your specific business — or integrating it with your existing CRM and calendar setup — book a free 30-minute call. We will walk through your call volume, your qualification criteria, and the exact configuration that fits your setup.