How to Build an AI Voice Agent for Your Business (2026 Guide)
Build an AI voice agent for your business: pick a platform (Vapi, Retell, Bland, ElevenLabs), write the prompt, connect booking and stay legal.

Table of contents
Difficulty
Intermediate
Time required
4 hours
Tools needed
- A voice agent platform (Retell AI
- Vapi
- Bland AI
- ElevenLabs Agents or Synthflow)
- A phone number (Twilio or platform-provided)
- A booking calendar
- Your FAQ and call scripts
- A test phone
An AI voice agent is software that answers (or makes) phone calls and holds a real conversation: it listens, understands what the caller wants, talks back in a natural voice and can take actions like booking an appointment. For a small business, the most practical first build is an after-hours receptionist that answers common questions, books appointments and hands anything tricky to a human.
You don't need to be a developer to build one in 2026. Platforms like Retell AI, Vapi, Bland AI, ElevenLabs Agents and Synthflow handle the hard parts (speech recognition, the AI model, the voice and the phone line), and you mostly configure a prompt, a phone number and a couple of integrations. Our quick pick: Retell AI for most small businesses that want clear per-minute pricing and a visual builder, Vapi if you have a developer and want to mix and match providers, and Bland AI if you want one all-in rate with no separate model bills.
The harder parts aren't technical. They're choosing a job the agent can actually do well, testing it properly and staying on the right side of calling, recording and privacy laws. This guide covers all three.
How an AI voice agent works
Most voice agents run one of two architectures:
- The pipeline (speech-to-text, LLM, text-to-speech). The caller's audio is transcribed, a large language model decides what to say and what tools to call, and a text-to-speech engine speaks the reply. This is how Vapi and Retell work under the hood, and it's why their pricing pages break costs into transcription, model and voice lines. The upside is control: you can swap in a different voice or model without rebuilding.
- Speech-to-speech (realtime) models. Models like OpenAI's gpt-realtime family take audio in and produce audio out directly, which can cut latency and preserve tone. OpenAI's Realtime API can accept phone calls over SIP from a trunking provider such as Twilio, and it can transfer calls. It's a developer tool, though: you write the webhook and tool code yourself.
Either way, what makes a voice agent useful is tools: small functions the model can call mid-conversation, like "check calendar availability" or "create a lead in the CRM." Without tools, you've built a talking FAQ page.
Good (and bad) first jobs for a voice agent
Start with calls that are frequent, repetitive and low-risk. Good candidates:
- After-hours answering. Anything is better than voicemail, and the stakes are lower when the alternative is no answer at all.
- Appointment booking and rescheduling. Clear goal, structured data, easy to verify.
- FAQs. Opening hours, location, parking, pricing ranges, what to bring.
- Inbound lead qualification. Asking five standard questions and routing hot leads to sales.
Jobs to avoid at first: anything involving emergencies or medical, legal or financial advice; angry customers and complaints; taking card payments over the phone; and cold outbound calling (more on the law below). If a mistake would be expensive or embarrassing, keep a human on it.
Platform options compared (as of September 2026)
All prices below are taken from each vendor's official pricing page in September 2026. Voice pricing changes often, so re-check before you commit.
| Platform | Best for | Pricing model | Starting point |
|---|---|---|---|
| Retell AI | Most small businesses | Per-minute, by component | $10 free credits, pay as you go |
| Vapi | Developers who want flexibility | $0.05/min platform fee plus provider costs | $5 free credits; Core plan $29/month |
| Bland AI | One all-in rate | $0.14/min on Start, $0.12/min on Build | Start plan has no platform fee |
| ElevenLabs Agents | Voice quality, bundled minutes | Monthly plan with included minutes, LLM extra | Free plan with 15 minutes |
| Synthflow | No-code teams | Pay as you go (self-serve) or enterprise contract | Enterprise from $30,000/year |
| OpenAI Realtime API | Custom builds | Per audio token | Developer account |
Retell AI itemizes everything. Its voice infrastructure costs $0.055 per minute, US Twilio telephony $0.015 per minute, most text-to-speech voices $0.015 per minute (ElevenLabs voices $0.040), and the LLM varies, from $0.024 per minute for GPT 5.4 mini to $0.064 for GPT 5.6 Terra or Claude 5 Sonnet. Phone numbers are $2 a month, you get 20 concurrent calls free, and extras like a knowledge base add $0.005 per minute. Retell's homepage says it's compliant with HIPAA, SOC 2 Type II, GDPR and ISO 27001, and it offers a drag-and-drop flow builder.
Vapi charges a $0.05-per-minute hosting fee and passes model costs through at provider rates. Its pricing page lists Deepgram transcription at about $0.0095 to $0.0099 per minute, OpenAI models at $0.0077 to $0.0452, and ElevenLabs voices at $0.0146 to $0.0238. Telephony through Twilio, Vonage or Telnyx is billed separately. Watch the add-ons: HIPAA compliance is $2,000 a month, and the free tier includes only four concurrent calls and 14 days of data retention.
Bland AI is the simplest to budget. The Start plan charges $0.14 per minute of talk time and $0.05 per minute for transfers, and Bland says the rate covers the LLM, transcription and voice, with no token charges. Start is capped at 100 calls a day and 10 concurrent calls; the Build plan ($299 a month) drops talk time to $0.12 per minute.
ElevenLabs Agents bundles minutes into its plans: 15 on Free, 1,238 on Pro ($99 a month), 3,738 on Scale ($299), with extra minutes at $0.08. LLM usage is billed separately depending on the model, which ElevenLabs doesn't bundle into those figures. Telephony is included at no extra ElevenLabs charge, though your carrier still bills you.
Synthflow is the no-code option, with a visual flow designer and integrations including HubSpot, Salesforce, Zapier and Cal.com. Its homepage says you can start building for free and pay as you go, but the pricing page only lists enterprise contracts starting at $30,000 a year and doesn't publish a self-serve per-minute rate, so get a quote before you plan around it.
For more AI tools that fit a small team's budget, see our pillar guide to the best AI tools for small businesses.
Step-by-step: build your first voice agent
Step 1: Define the job and write the call script
Write down, in plain English, the one job this agent does. For example: "Answer calls outside office hours, answer questions from our FAQ, book or reschedule check-up and cleaning appointments, and take a message for everything else."
Then list:
- The 10 to 20 questions callers ask most (pull these from voicemails, emails and staff)
- The information the agent must collect (name, phone, reason, preferred time)
- What it must never do (quote prices for treatment, give advice, promise anything)
- When it must hand off to a human
This document becomes your system prompt and your test plan. Skip it and you'll spend days chasing weird edge cases.
Step 2: Choose your platform
Use the table above, but the decision usually comes down to two questions. Do you have someone comfortable with APIs and webhooks? If not, lean toward Retell's visual builder or Synthflow. Do you handle health or other sensitive data? Then confirm the vendor will sign the contract you need (see the legal section) before you build anything, and budget for add-ons like Vapi's $2,000-a-month HIPAA tier.
Most platforms offer free credits, so it's reasonable to build the same simple agent on two platforms and compare how they sound on real calls.
Step 3: Get a phone number
You have two routes. The easiest is buying a number inside the platform (Retell lists numbers at $2 a month; Vapi includes one on its free tier). The more flexible route is buying a number from a carrier such as Twilio and connecting it. In September 2026, Twilio lists US local numbers at $1.15 a month and inbound calls at $0.0085 per minute (Twilio US voice pricing).
For an after-hours agent, you usually don't need to change your public number at all. Set your existing phone system to forward calls to the agent's number outside business hours, or when nobody picks up after a few rings.
Step 4: Write the system prompt
The system prompt is the agent's job description. Voice prompts differ from chat prompts in a few ways: replies must be short, the agent should ask one question at a time, and it must say numbers and times the way people speak them. Here's an example for a fictional clinic. Adapt the names, hours and rules to your business.
Prompt: after-hours receptionist for a fictional dental clinic You are Maya, the virtual receptionist for Brightside Dental, a family dental clinic at 214 Elm Street, Springfield. You answer calls when the front desk is closed. You are an AI assistant, and you say so at the start of every call. OPENING (say this first, word for word): "Thanks for calling Brightside Dental. I'm Maya, the clinic's AI assistant. This call may be recorded to help us serve you. How can I help?" YOUR JOBS 1. Answer questions using only the FAQ below. 2. Book, reschedule or cancel check-ups and cleanings using the calendar tools. 3. Take a message for anything else. STYLE - Keep replies to one or two short sentences. This is a phone call. - Ask one question at a time. Confirm names by spelling them back. - Say times naturally ("Tuesday at two thirty"), never as digits. - If the caller interrupts, stop and listen. - Friendly, calm, plain language. No medical jargon. BOOKING - Collect: full name, date of birth, phone number, new or existing patient, and appointment type (check-up, cleaning, or check-up and cleaning). - Call check_availability, offer at most two slots, then call book_appointment. Read back the date and time before confirming. - New patients: book a "new patient exam" (60 minutes) only. FAQ - Hours: Monday to Friday 8am to 6pm, Saturday 9am to 1pm. - Parking: free lot behind the building. - Insurance: we accept most major PPO plans. For specific plan questions, take a message for the billing team. - Cancellations: please give 24 hours' notice. HARD RULES - Never give medical or dental advice, diagnoses or treatment prices. - If the caller mentions severe pain, swelling, bleeding that won't stop, trouble breathing or swallowing, or a knocked-out tooth, say: "That sounds urgent. If you have trouble breathing or swallowing, please hang up and call 911. Otherwise I'll connect you to our on-call dentist now." Then call transfer_to_on_call. - If the caller asks for a person, is upset, or you don't understand them after two tries, call transfer_to_on_call during on-call hours or take a message. - Never make up information. If it isn't in this prompt, say you'll pass the question to the team and take a message. - Don't discuss other patients or confirm whether someone is a patient. ENDING Summarize what happened ("You're booked for Thursday at ten"), ask if there's anything else, then thank them and end the call.For more on why structure like this works, see our guide on how to write better prompts.
Step 5: Connect tools: calendar, CRM and messages
Tools are how the agent does things. On most platforms you define a function with a name, a description and the fields it needs (for example,
check_availabilitywith a date and appointment type). When the model calls it, the platform sends a webhook to a URL you control, and your server or automation replies with the result.Common setups:
- Calendar booking. Some platforms have native booking actions (Retell lists "check availability and book," and Synthflow integrates with Cal.com). Otherwise, point the webhook at your booking system's API.
- CRM and lead capture. Retell lists HubSpot, GoHighLevel, Make and n8n among its integrations. A no-code automation tool can catch the webhook, create the contact and post a summary to your team chat. Our guide to automating your business with n8n and AI walks through that pattern, and Zapier vs Make vs n8n helps you pick the tool.
- Post-call summaries. Most platforms send an end-of-call webhook with the transcript. Save it, and email the front desk a summary each morning.
Keep tools narrow. "book_appointment" is safer than "update_anything_in_the_database."
Step 6: Set up the human handoff
Every agent needs an exit. Decide who gets transferred calls, during which hours, and what happens when nobody answers (usually: take a message and promise a callback).
Vapi's
transferCalltool supports both blind transfers and warm transfers, where the agent speaks a summary to the staff member before connecting the caller (Vapi docs). Warm transfers are worth the extra setup, because nobody likes repeating their story. Vapi's docs also note the prompt should tell the assistant to acknowledge the request and call the transfer tool in the same response. Otherwise it may say it's transferring without actually doing it.Step 7: Test with a real call checklist
Test on real phone calls, not just the browser test button, and get people other than the builder to call. Run through this checklist:
- Latency. Does the pause after you stop speaking feel natural, or awkward?
- Interruptions. Talk over the agent mid-sentence. Does it stop and respond to what you said?
- Accents and noise. Call from a car, a busy street and a speakerphone. Ask colleagues with different accents to call.
- Names, numbers and dates. Spell an unusual surname. Give a phone number quickly. Say "next Tuesday" and "the 3rd."
- Tool calls. Does the booking actually appear in the calendar? Does a double-booked slot get refused?
- Off-script questions. Ask about something not in the FAQ. It should take a message, not invent an answer.
- Emergencies and transfers. Say the emergency trigger words. Does it give the right instruction and transfer?
- Adversarial callers. Ask it to ignore its instructions or reveal the prompt.
Log every failure, fix the prompt or tool, and re-run the whole list. Small prompt changes can break things that worked yesterday.
Step 8: Launch small and monitor
Go live on a narrow slice first, like weekday evenings only. For the first few weeks, read transcripts daily. Look for calls that ended abruptly, repeated questions, failed tool calls and callers asking for a human. Those are your prompt fixes.
Also watch the bill. Per-minute costs add up if calls run long, so check average call length and set concurrency and spending limits where your platform allows it.
What it costs: a worked example
Say your agent handles 500 calls a month averaging 3 minutes, or 1,500 minutes. Using only the published rates above (September 2026):
| Platform setup | Per-minute math | Monthly estimate |
|---|---|---|
| Retell AI: platform voice, GPT 5.4 mini, Twilio telephony | $0.055 + $0.015 + $0.024 + $0.015 = $0.109 | About $164 + $2 number = $166 |
| Vapi: Deepgram, OpenAI, ElevenLabs voice, Twilio inbound | $0.05 platform + provider costs = $0.0903 to $0.1374 | About $135 to $206 + $1.15 number |
| Bland AI Start (all-in, no transfers) | $0.14 | $210 |
| ElevenLabs Agents Pro | $99 includes 1,238 min; 262 extra at $0.08 | About $120, plus LLM usage and carrier fees |
A few caveats. Vapi's range depends on which model and voice you choose, and the Twilio figure is the local inbound rate. Transfers, knowledge bases and add-ons cost extra on most platforms. And none of these include your own time building and maintaining the agent, which is the real cost for most small businesses.
Legal must-haves before you go live
This isn't legal advice, and rules vary by country and state. But a few requirements apply to almost every business.
Outbound calls need consent
In February 2024, the FCC issued a declaratory ruling (adopted February 2, released February 8) confirming that the Telephone Consumer Protection Act's restrictions on "artificial or prerecorded voice" cover AI technologies that generate human voices. In the FCC's words, callers "must obtain prior express consent from the called party before making a call that utilizes artificial or prerecorded voice simulated or generated through AI technology." If the call is telemarketing, the rules require prior express written consent. The ruling also points to the FCC rule that such messages must clearly state the identity of the business at the start.
In practice: inbound agents are the safe starting point. Don't point an AI agent at a list of phone numbers unless you have documented consent for each one.
Call recording consent
US federal law allows recording when one party to the call consents (18 U.S.C. § 2511(2)(d)), but several states require everyone's consent. California, for example, makes it an offense to record a confidential communication "without the consent of all parties" (Penal Code § 632). Since you can't know where every caller is, the simple fix is a recording notice at the start of every call, as in the sample prompt.
Tell callers they're talking to AI
Disclose it upfront. It's good practice everywhere, and in the EU it's becoming a legal duty: Article 50 of the EU AI Act says AI systems that interact directly with people must be designed so those people "are informed that they are interacting with an AI system," unless that's obvious from context. According to the Act's text, Article 50 applies from 2 August 2026.
Privacy: GDPR and HIPAA
If you take calls from people in the EU or UK, voice recordings and transcripts are personal data. GDPR's Article 13 requires telling people who you are, why you're processing their data, the legal basis and how long you'll keep it, at the point you collect it. A short line in the greeting plus a link in your follow-up text or privacy policy is a common approach. Set sensible retention periods on your platform and check where it stores data.
If you're a US healthcare provider, the vendor handling patient calls is likely a business associate under HIPAA. The rules say you may share protected health information with one only if you get "satisfactory assurance" it will safeguard the information, documented in a written agreement (45 CFR § 164.502(e)). In short: no signed business associate agreement, no patient calls. Retell and Synthflow both advertise HIPAA compliance, and Vapi sells it as a paid add-on, but confirm the agreement in writing.
Frequently asked questions
How much does an AI voice agent cost per minute?
As of September 2026, typical all-in costs on the platforms above run roughly $0.09 to $0.14 per minute before add-ons. Bland AI charges $0.14 per minute on its Start plan, Retell's itemized rates add up to about $0.11 with a small model, and Vapi ranges from about $0.09 to $0.14 depending on providers.
Can I build an AI voice agent without coding?
Yes, for simple jobs. Retell and Synthflow offer visual flow builders, and platform-provided phone numbers remove most setup. You'll still need some technical comfort for calendar or CRM connections, though no-code automation tools can handle the webhooks.
Do I have to tell callers they're talking to an AI?
It's strongly recommended everywhere and increasingly required. The EU AI Act's Article 50 requires it in most cases, and for outbound calls in the US the FCC treats AI voices as "artificial" under the TCPA, which brings consent and identification rules. A one-line disclosure in your greeting covers you.
Is it legal to use an AI voice agent for outbound sales calls?
Only with the right consent. Under the FCC's 2024 ruling, AI-voice calls need prior express consent, and telemarketing calls need prior express written consent. Most small businesses should start with inbound calls.
Will callers hang up on an AI receptionist?
Some will, especially if the agent is slow or doesn't understand them. Short replies, a clear disclosure, fast answers to common questions and an easy path to a human reduce that. Review transcripts regularly to find where callers drop off.

Written by
Panoptix Editorial Team
Our editors test AI tools hands-on for weeks before we publish a word. We pay for our own subscriptions and never accept payment for rankings.
How we test AI tools →

