How to Build an AI Receptionist With Vapi: A Step-by-Step Setup for Small Businesses

How to build an AI receptionist with Vapi

A two-chair salon, a plumbing office and a law practice all share one problem: the phone rings while the person who should answer it has their hands full. Off-the-shelf AI receptionists fix that for $49 to $300 a month, but they come with someone else's rules. Building on Vapi costs roughly 8 to 13 cents a minute, lets you decide exactly what the agent says, and takes an afternoon if you follow an order. This guide gives you that order.

Vapi is a platform that handles the phone side of a voice agent: it answers the call, converts speech to text, passes it to a language model, speaks the reply, and calls your tools when the agent needs to do something. You choose each part and write the instructions.

Is building it yourself the right call?

Build when you want custom rules, want to own the setup, or sell this to clients. A pest control operator documented on this site (builtwithagents.ai/strategy/pest-control-ai-voice-receptionist-vapi-n8n) built a Vapi agent with n8n that answers emergency calls, qualifies the infestation, checks real calendar availability and books with technician notes attached. A builder featured at builtwithagents.ai/strategy/salon-ai-receptionist-3000-vapi-make sells a Vapi and Make salon receptionist as a $3,000 package, a self-reported price.

Buy instead when you will not maintain it. Goodcall, Rosie and similar tools cost more per minute but give you a working agent in an hour with someone else fixing the plumbing. Our AI receptionist cost guide compares both on the same bill.

What you need before you start

You need a Vapi account, a business phone line you can forward, a calendar if the agent will book (Google Calendar is the common choice), and an automation tool such as Make or n8n if you want booking or text confirmations. You also need one page of written content: your hours, services, prices you are willing to state, service area and the questions you want asked on every call.

What it costs

As of October 2026, Vapi's pricing page lists hosting at $0.05 per minute, with $5 in free credit for new accounts. Provider costs pass through at cost: Deepgram transcription is about $0.01 a minute, OpenAI language model costs range from about $0.008 to $0.045, and ElevenLabs voice costs about $0.015 to $0.024. Vapi estimates $82 to $129 a month for 1,000 minutes, so at an average three-minute call that is about 330 calls for roughly $100.

Carrier fees are extra. Twilio inbound is $0.008 a minute. The Core success package is $29 a month and raises you to 10 concurrent calls and 5 phone numbers; most small businesses do not need it at the start. The free plan allows 4 concurrent calls.

Add your tools. Make and n8n both have free tiers that cover low volume, and a paid plan if you outgrow them. The real cost is your time: plan on three to six hours to build and test your first agent properly.

Step by step

Step 1: Write the call script on paper

Before you open Vapi, write what a perfect call sounds like for your three most common call types. For a plumber that might be an emergency leak, a quote request and a question about an existing job. For each, list the information you need and what the caller leaves with. Keep it to six questions or fewer per type.

Step 2: Create the assistant

In the Vapi dashboard, create a new assistant from the blank template. Set a first message that identifies the business and the assistant, for example: "Thanks for calling Ridgeline Plumbing, this is the automated assistant. How can I help?" Saying it is automated up front keeps the call honest and reduces hangups later.

Step 3: Pick the model, voice and transcriber

Start with a fast, inexpensive model for the language step, since speed matters more than cleverness on a phone call. Pick a natural voice from the ElevenLabs list and listen to it read your greeting. Leave the transcriber on its default English setting unless you serve a second language, in which case set the language explicitly and test it.

Step 4: Write the system prompt

This is the work that matters. Put these blocks in order: who the assistant is and for which business; the call types and what to collect for each; the facts it may state (hours, service area, prices you allow); a list of things it must never do; and when to transfer to a person. Tell it to ask one question at a time, keep replies to one or two sentences, and never read a list. Add rules such as: do not give legal, medical or pricing advice, do not promise arrival times, and if unsure say a team member will call back.

Step 5: Add a transfer tool for emergencies

In the assistant's tools, add a transfer call tool pointing at the owner's cell. In the prompt, define exactly what triggers it: an active emergency, a caller who asks for a person twice, or a caller who sounds upset. A reliable escape hatch is the difference between a useful agent and an angry customer.

Step 6: Connect booking through a webhook

Create a function tool in Vapi that calls a webhook URL. Build the other end in Make or n8n: receive the caller's name, number and requested time, check your calendar for a free slot, create the event, and return a confirmation sentence for the agent to read out. Have the same workflow text the caller a confirmation and email you the summary. Keep the scenario simple at first: one calendar, one appointment type.

Step 7: Attach a phone number and forward your line

Create or import a phone number in Vapi (importing from Twilio is common), assign it to the assistant, and then set your business line to forward to it, either always, after a few rings or after hours. Forwarding after three or four rings lets a human answer when they can and sends the overflow to the agent.

Step 8: Test with 20 real calls

Call it yourself, and ask friends to call. Try a normal booking, an emergency, a caller who changes their mind, a caller with a noisy background, one who interrupts constantly and one who asks for something outside the script. Read the transcripts in Vapi's call logs. Every bad answer becomes a new sentence in the prompt. Expect three or four rounds.

Step 9: Go live and review weekly

Turn on forwarding and read all calls for the first two weeks. After that, skim the summaries weekly and check any call longer than five minutes, because those usually mean the agent got stuck.

A worked example: a three-person plumbing office

Say a plumbing office takes 200 calls a month averaging two and a half minutes, so about 500 minutes. At Vapi's estimated $0.08 to $0.13 a minute that is roughly $40 to $65 in platform and provider costs, plus a few dollars of carrier fees. A call comes in at 6:40 pm: the assistant greets the caller, learns a water heater is leaking into a garage, asks for the address and whether the shutoff valve is reachable, and tells the caller how to close it. It then triggers the transfer tool because the rule says active water means a person, and the owner's phone rings with the caller already on the line. A second call at 7:10 pm is a quote request, so the agent collects the details, books the next free estimate slot through the calendar webhook and texts a confirmation. These figures are illustrative, but the pattern is the point: routine calls get booked without you, and urgent ones reach you fast.

Mistakes that make it sound broken

Long prompts full of exceptions make the agent hesitate. Cut anything you do not need. Letting the agent read long lists, like every service you offer, wastes time and bores the caller. Skipping the human handoff turns a small failure into a lost customer. And forgetting call recording laws can cause trouble: some US states require consent from everyone on a call, so put a line in the greeting if you record, and check your state's rules.

When not to build it

If you take fewer than about 50 calls a month, a voicemail greeting and a text-back may be enough. If your calls are long, emotional or complicated, a person is better. And if nobody at your business will spend an hour a week reading transcripts, buy a managed tool instead.

Frequently Asked Questions