AI voice agents are software that answers or makes phone calls for a business in a natural voice, understands what the caller says, and acts on it: answering questions from the business’s own information, booking into the real calendar, taking details for the CRM and passing the call to a person when it should. Today they handle routine calls well. They still need clear rules, good information and someone looking after them.
This page brings together what we cover across our Voice AI guides. Platform prices below come from each vendor’s own pricing page as at November 2026; check their page for current prices. CentriWeb builds and looks after voice agents for clients, on platforms of this kind, so weigh our view with that in mind.
What AI voice agents are
An AI voice agent is a phone line that listens, thinks and talks. Unlike a phone menu that asks callers to press 1 for sales, it holds a conversation: the caller says what they need in their own words and the agent responds. Our guide to IVR vs conversational AI sets out that difference in detail.
Most businesses meet voice agents in one of two jobs:
- Inbound: an AI receptionist that answers when staff cannot, after hours, at lunch or on a busy morning.
- Outbound: calls the business makes, such as callbacks people asked for and appointment reminders. These have their own consent rules, covered in our guide to AI outbound calling.
What a voice agent can and cannot do on a business phone line today
What it does well
- Answer every call, at any hour, including several at once.
- Answer common questions from your own information: hours, service areas, what you do and do not do, how pricing works.
- Take the caller’s name, number, address and the details of the job, and write them to your CRM.
- Book, move or cancel appointments in your calendar, following your rules.
- Send a text confirmation or a link after the call.
- Transfer to a person, or take a message, when the call needs one.
Where it still struggles
- Judgement calls. A complaint, a distressed caller or a complicated quote is better with a person.
- Information it was never given. It cannot answer well from a website that is out of date.
- Noisy lines and strong accents can cause errors in what it hears, which is why testing with real callers matters. Our guide on AI voices and the Australian accent covers this for one market.
- Long pauses and interruptions. Delays of a second or two feel odd on the phone, and how well an agent copes with being talked over varies.
Our guide to the pros and cons of an AI receptionist weighs these against a human receptionist.
What happens on a typical call
As an example, a caller rings a plumbing business at 7 p.m. about a leaking hot water system. The agent answers with the business name, says it is an AI assistant, and asks what the problem is. It checks the suburb against the service area, explains that after-hours call-outs carry an extra fee, and offers the first morning slot. The caller takes it. The agent books the job in the calendar, creates the contact in the CRM with a note about the leak, and texts a confirmation. Had the caller said water was coming through the ceiling, the agent would have followed the emergency rule and transferred to the on-call plumber.
Every step in that call came from something the business decided in advance: the greeting, the service area, the fee, the booking rule and the emergency rule. The agent is only as good as those decisions.
How AI voice agents are built: the parts
Every voice agent, whichever platform it runs on, is made of the same five parts. Our guide to how an AI voice agent works explains each in plain English.
| Part | What it does | What to ask about it |
|---|---|---|
| Speech to text | Turns the caller’s words into text | How well it hears your callers, on mobiles and in noise |
| Language model | Decides what to say and which action to take, following your instructions | Speed, cost per minute, and how it is told what it must never say |
| Text to speech | Turns the reply into a voice | Whether the voice suits your business and your callers’ accent |
| Telephony | Connects the agent to a real phone number | Whether you can use a number in your country and keep your existing one |
| Calendar and CRM connection | Books jobs and records the caller’s details where staff will see them | Which calendar and CRM it connects to, and what happens if the connection fails |
The first four are bought from the platform or its providers. The fifth is where most of the work lies, because it is specific to your business. Our guide to connecting a voice agent to your CRM covers it, and our guide to writing an AI receptionist script covers the instructions that sit on top.
What AI voice agents cost to run
Most platforms charge by the minute, built from those parts. Two examples from their own pages:
- Retell AI’s pricing page lists voice agents at $0.07 to $0.31 a minute, depending on the voice and model. Its parts include $0.055 a minute for its voice engine, language models from $0.0016 to $0.08 a minute in the examples shown, standard voices at $0.015 a minute, and telephony at $0.015 a minute in listed countries. New accounts get $10 in credit and 20 concurrent calls are included. Calls are billed to the second, and silence and hold time are billed.
- Vapi’s pricing page lists $0.05 a minute for hosting, with speech to text, model and voice passed through at the provider’s cost. Its calculator estimates $82 to $129 a month for 1,000 minutes, and it lists four concurrent calls without a package.
Telephony may be a separate account. Twilio’s United States voice pricing lists a local number at $1.15 a month, $0.0085 a minute to receive calls and $0.014 a minute to make them. Rates differ by country.
As an example with your own numbers: 600 calls a month averaging two minutes is 1,200 minutes. At a combined $0.11 a minute, which is Retell’s own worked example before telephony and add-ons, that is about $132 a month in platform usage, before the number, the build and the upkeep. Those figures are shown in dollars on pages written for a United States audience, so confirm the currency.
The usage bill is usually the smaller part. Building the agent, connecting it and looking after it each month cost more, whether in your own time or someone else’s. Our guide to what an AI receptionist costs sets out the full picture, and our Retell AI pricing and Vapi pricing guides work through each platform’s bill.
How to judge an AI voice agent before buying
- Ring it yourself, many times. Call from a mobile, in the car, with background noise. Interrupt it. Ask something it should not answer.
- Ask what it does when it does not know. It should say so and take a message or transfer, never guess.
- Watch a booking go through. Check the appointment lands in your real calendar with the right details, and the caller gets a confirmation.
- Check the handover. Who gets transferred calls, and what happens if nobody answers?
- Ask where the recordings and transcripts are kept, who can see them and for how long. Recording rules differ by country and state, so check yours; our guide to call recording laws in Australia covers one example.
- Ask who changes it. When your prices or hours change, who updates the agent, and how quickly?
- Ask for the monthly bill on your volume, including telephony and concurrency.
Our comparison of the main AI voice agent platforms groups the options by type, and Retell AI vs Vapi compares two developer platforms side by side.
Keeping a voice agent working after launch
The first week is the easy part. Over the following months, prices change, staff leave, the calendar tool updates its connection, and callers ask questions nobody planned for. Someone needs to:
- Read a sample of transcripts every week and fix answers that were wrong or unclear.
- Update the agent’s information whenever prices, hours or services change.
- Check that bookings and CRM records are still landing correctly.
- Watch the monthly minutes and bill against what was expected.
Name that person before you start.
When to build an AI voice agent and when to buy one
There are three broad ways to get one:
- A ready-made AI receptionist product. Set up through a form, priced per month. Quick to start, limited in what it connects to.
- Build it yourself on a platform such as Retell AI or Vapi. Full control, but someone on your side has to write the instructions, connect the calendar and CRM, test it and maintain it.
- Have it built and looked after. Someone else builds it on a platform and stays responsible for it.
Buy a ready-made product when calls are simple, volumes are modest and it works with the calendar you already use. Build it yourself when you have someone technical with time to look after it every month. Have it built when missed calls cost real money and the agent must book into a live calendar and write to the CRM every time. Our guide on whether to build a voice agent yourself or hire someone sets out both paths in more detail.
Where to start with AI voice agents
- Count your missed calls for two weeks, and note when they happen.
- List the five questions callers ask most and the answers you want given.
- Decide what the agent may do on its own and when it must hand over.
- Choose how you will get it: product, platform or built for you.
- Start with one job, such as after-hours calls, and widen it once it works.
If you would rather have it built and looked after, our AI phone and chat agents service covers that work, and every Voice AI guide sits on our Voice AI topic page.