# Take a real call without picking up: a 90-minute AI voice agent build

Source: https://getstartedwithai.net/tutorials/build-ai-voice-agent-basic
Site: Get Started With AI
Tagline: Build AI automations that actually run your business.
Audience: solo professionals, freelancers, consultants and AI implementers in training
Contact: hello@getstartedwithai.co.uk

This Markdown companion is generated from the prerendered page for AI crawlers.
# Take a real call without picking up: a 90-minute AI voice agent build

A working AI voice agent now answers a real phone number for the cost of a Pret coffee per hour of calls. Retell and Vapi both run at £0.10 to £0.25 per minute all-in (model, speech-to-text, text-to-speech, telephony). This 90-minute build produces a phone number you can call from your mobile, that answers two FAQs, books a discovery slot via Calendly and transfers anything else to your line. By the end you have call transcripts to review and a £20 daily cost cap that prevents a runaway weekend bill.

## Pick a platform

- Retell AI: cleanest pricing at about £0.06/min orchestration plus telephony and model costs. Best for first builds because the all-in cost is transparent.
- Vapi: more flexible architecture but multi-part pricing means real cost runs £0.12 to £0.27/min. Better for advanced builds with custom routing.
- Bland AI: pricing tied to a plan, from £240/month. Worth it if you run more than 3,000 minutes a month.
- For this tutorial: Retell. The setup is the simplest and the cost is the easiest to forecast.

## What you need

- A Retell account at retellai.com.
- A Calendly or Cal.com link with a 30-minute meeting type.
- A short FAQ list of two real questions you get on calls.
- About £5 of telephony credit on Retell to test with.
- 90 minutes uninterrupted.

## Step 1: Define the agents job (10 minutes)

A bad voice agent tries to handle every call. A good one has a tight scope. Decide before you build.

- Job: answer inbound calls during working hours, handle two FAQ questions, book a discovery slot, escalate anything else to your mobile.
- Out of scope: detailed scoping, pricing negotiation, refunds, anything sensitive.
- Hand-off rule: if the caller says speak to a human, says complaint, or asks anything outside the FAQ, transfer to your mobile.

## Step 2: Write the agent prompt (15 minutes)

1. In Retell create a new agent. Name it after the job.
2. In the agent prompt, write a clear role brief: who the agent is, who they work for, what they handle, what they escalate. 200 to 400 words.
3. Voice: pick the closest match to your brand. ElevenLabs voices are integrated. Preview a few before picking.
4. First message: Thanks for calling [your business]. Im an AI assistant. I can answer common questions or book you a call with the team. How can I help?
5. Tone instructions: short sentences, friendly, ask one question at a time, never invent information.

## Step 3: Load the FAQ entries (10 minutes)

1. Open Knowledge Base in Retell. Create one knowledge entry per FAQ.
2. Each entry: question (as a real caller would phrase it), answer (under 30 seconds when read aloud), source (where this is canonical, e.g. your pricing page).
3. Two real FAQs to start: How much do you charge? and Do you work with [common sector]? Test both by typing them into the simulator and listening to the agent reply.

## Step 4: Add the calendar booking tool (20 minutes)

1. In Retell open the Tools tab. Add a Custom Function called book_meeting.
2. Schema: input is caller_name (string), caller_email (string), preferred_day (string).
3. Webhook: a Zapier webhook URL. The Zap receives the call data and calls Calendly to create a booking. (See the Zapier first automation tutorial for the webhook setup.)
4. In the agent prompt, instruct: If the caller wants to book a call, ask for their name, email, and preferred day. Then call the book_meeting function with those values. Confirm the booking once the function returns.

## Step 5: Configure escalation (10 minutes)

1. In Retell add a Transfer Call action to your mobile number.
2. In the agent prompt: If the caller asks for a human, says complaint, or asks something outside your knowledge base, say I will transfer you now and call the transfer function.
3. Test in the simulator: type I want to speak to a human. Confirm the agent transfers correctly.

## Step 6: Get a phone number and test (15 minutes)

1. In Retell buy a phone number (about £1 a month plus per-minute usage).
2. Attach the number to your agent.
3. Call the number from your mobile. Run three test calls: an FAQ, a booking, and a speak-to-a-human.
4. If anything misfires, watch the call transcript in Retell. The agent prompt is usually where the fix lives.

## Step 7: Decide if it is ready for production (10 minutes)

A voice agent that handles 70 percent of calls cleanly is worth deploying. Below that, fix before launching.

- Run 10 real test calls with friends or colleagues acting as callers.
- Score each call: clean handle, escalated correctly, broke and sounded weird.
- If 7 of 10 are clean, deploy and monitor. If less, iterate the prompt and the FAQ first.
- Set a cost cap on the account so a runaway loop cannot burn through more than £20 a day.

## What you ship

A phone number that answers calls, handles two FAQs, books discovery slots and transfers everything else to your mobile. Charge clients £4,000 to £15,000 for the same build, plus a separate £400 to £900 monthly retainer for monitoring. Run it on yourself first, gather two weeks of call transcripts as proof, then sell the case study.

## Common mistakes

- Trying to make the agent handle every call from day one. Pick three things it must do, escalate everything else.
- Skipping the cost cap. A runaway agent on Vapi has been known to burn £200 in a single bad weekend.
- Voice that does not match the brand. The voice is the brand on a phone call. Spend 20 minutes picking the right one.
- No human-takeover path. Always include a transfer rule. Callers will ask, and the answer cannot be no.

## What to build next

Pair this with the AI privacy and security tutorial for the disclosure and consent obligations on AI voice calls. The EU AI Act and UK ICO guidance both expect you to disclose AI involvement in the first message.

## Related builds

More tutorials in Automation. Each one runs in 60 to 90 minutes and ships a finished asset.

### Capture every lead automatically: a 30-minute Zapier build

30 to 60 minutes. Solo Pro.

[Open tutorial](/tutorials/first-zapier-automation-30-min)

### Decide between native AI Steps and a Custom GPT in 10 minutes

60 to 90 minutes. Solo Pro.

[Open tutorial](/tutorials/zapier-ai-features-vs-custom-gpt)

### Run automations Zapier cannot: the first Make scenario

60 to 90 minutes. Solo Pro.

[Open tutorial](/tutorials/first-make-scenario-visual)

### Self-host the automation stack: when n8n is worth the effort

60 to 90 minutes. Solo Pro / Implementer.

[Open tutorial](/tutorials/n8n-self-hosted-when-worth-it)

## Questions people ask before they start

### Do I need to disclose that the caller is talking to AI?

In the UK and EU: yes, in most contexts. The first message should say I am an AI assistant or equivalent. Skipping the disclosure is both unethical and a regulatory risk under the EU AI Act.

### How accurate is the speech-to-text?

Excellent for clear UK English. Drops noticeably for strong regional accents, multi-speaker calls or noisy backgrounds. Test with real speakers before deploying.

### What does it actually cost?

On Retell, expect about £0.12 to £0.18 per minute all-in for a UK call (model plus STT plus TTS plus telephony). A small business with 200 minutes a month pays about £30. Most pay back the build cost within 3 to 4 months.

### Can the agent take payment?

Possible, but the legal exposure makes it a bad idea. PCI compliance on voice channels is complex and the audit trail is harder to defend. The cleaner pattern: agent qualifies and books a call, payment happens on a Stripe link in a follow-up email.

### How does the agent handle long callers who ramble?

Set a soft turn-taking timeout in Retell at 8 to 12 seconds. If the caller has not finished, the agent waits. If silence stretches past the timeout, the agent prompts politely.
