An AI voice agent answers your business phone line, turns what the caller says into text, lets a language model work out what they want, and speaks a reply back, then either handles the job itself (booking, quoting, answering a question) or passes it to a human with notes attached. For a small business it works well on repetitive calls like appointment booking and basic enquiries, and badly on anything emotional, unusual or legally sensitive, which is the bit most sales pages skip over.
What an AI voice agent is
Strip away the marketing and an AI voice agent is three pieces of technology stitched together and answering your phone. Speech-to-text turns the caller's voice into written words. A language model (the same family of tech behind ChatGPT) reads those words, works out what the caller wants, and decides what to say back. Text-to-speech turns the reply into a voice that comes down the line. How quickly all of this happens varies by provider and by the model and voice you pick, and there's no published standard figure, so judge the pause on a test call: if you notice the gap before each reply, your customers will too.
It is not a phone tree with pre-recorded options. A real agent can hold a loose conversation, cope with "erm, can you push that to Thursday instead", and update a calendar or CRM in real time. That's the difference between this and the IVR systems everyone already hates.
The steps that happen on every single call
Here's the actual sequence, stripped down:
- Call comes in, voice agent picks up, usually within one or two rings.
- Caller speaks, speech-to-text converts it to text in real time.
- The language model checks that text against a script and your business data: opening hours, prices, availability, FAQs.
- It generates a response, which gets voiced back through text-to-speech.
- If the task is simple (book a slot, take a message, answer a price question), it completes the action directly in your calendar or booking system.
- If the task is complex or the caller sounds upset, confused, or wants something outside the script, it flags the call for a human and either transfers live or sends a summary and recording for someone to follow up within an agreed window.
That last step is where most first set-ups go wrong. A typical example: the agent cheerfully offers Thursday at 2pm, a slot a colleague booked by hand an hour earlier. The fix is to give it live read access to the calendar and a rule never to offer a slot it can't see. No provider or regulator publishes a standard rate for how many calls should go to a human; it depends on how many of your calls are routine, so measure your own hand-off rate for the first month and expect it to sit well above zero. The handoff itself comes in three forms, and you can pick one per call type: a warm transfer, where the agent rings your mobile and reads you a one-line summary before connecting the caller; a summary with the recording, emailed or dropped into your CRM for a call back; or an SMS alert to whoever's on call with the caller's name, number and reason. When you're ready to shop around, voice agent platforms compared here make the shortlist easier.
A worked example: a plumbing business in Leeds
Say you run a three-van plumbing business and you currently miss around 30 percent of inbound calls because your team is up a ladder or already on another job. Each missed call is a potential job worth £150 to £400. You set up an AI voice agent connected to your booking calendar and your pricing list.
A caller rings about a leaking tap. The agent asks three questions: postcode, whether the leak is active or a slow drip, and whether anyone's home today. Based on the answers it offers the next available 90-minute slot, books it straight into the calendar, texts the customer a confirmation, and sends your team a one-line summary: "Leaking tap, active, Leeds LS6, booked Thursday 2pm, customer home all day." No human touched that call.
Now a different caller rings with a burst pipe and water coming through the ceiling. The agent recognises urgency language, skips the standard booking flow, and either transfers live to whoever's on call or texts an emergency alert immediately rather than queuing it for the next admin check. That split, routine job automated, urgent job escalated, is what a built system does. If your agent can't make that distinction, it's not ready to answer your phone unsupervised.
What would that cost on usage pricing? Say the business takes 200 calls a month at three minutes each, so 600 minutes. Retell AI's pricing page uses $0.11 a minute as its worked example, which makes $66 a month; its published pay-as-you-go range of $0.07 to $0.31 a minute puts the same 600 minutes at $42 to $186. That's one provider, in US dollars, and phone number rental and call charges can sit on top, so check the current pricing page before you budget.
What it costs and what you get
Pricing in this space runs one of two ways: a flat monthly plan, where the price depends on the provider and your call volume (check each one's current pricing page, as plans change often), or usage pricing per minute of conversation. Usage pricing is easier to compare because some providers publish the parts. Retell AI's pricing page lists pay-as-you-go at $0.07 to $0.31 a minute in US dollars, and breaks its own $0.11 example down like this:
- Language model: $0.04 a minute
- Voice infrastructure: $0.055 a minute
- Text-to-speech: $0.015 a minute
Vapi's pricing page shows component prices the same way, for example transcription at around $0.0095 to $0.0099 a minute and voice at $0.0146 to $0.0238 a minute, and lists $5 of free credit to start. So your total depends on which model, transcription and voice you choose, not just the platform. A business taking 200 calls a month at three minutes each uses 600 minutes, which at Retell's example rate is about $66 a month before setup and integration costs.
Setup is five jobs, in this order: forward your number to the agent, connect your calendar, load your price list, write a script that sounds like your business rather than a generic template, and set the escalation rules. How long that takes depends on how tidy your calendar and price list already are. If you bring in someone to build and tune it, costs vary widely by builder, so ask for a fixed quote that includes the test calls and the first month of script changes. If you compare a few options before committing, our breakdown of the best AI voice agents for small business covers what nine of the main providers charge and what's included.
A question readers often ask next: Which AI agents are free to try?
Where it falls apart
This is the part most articles on this topic avoid saying out loud: an AI voice agent does not save you as much money as the sales pitch implies, because it does not remove the need for human oversight, it just moves it. Someone still needs to check transcripts weekly, listen to a sample of calls, catch the ones where the agent confidently gave a wrong price or booked the wrong slot, and retrain the script. Skip that and within a month you'll have a trail of small errors feeding into your calendar and your customer data that nobody notices until a customer complains.
Strong regional accents, background noise and callers who talk over the agent all make mistakes more likely, and how much varies by provider, which is why the test calls below include each of them. Emotional calls, complaints, and anything with legal or financial consequence (refunds, cancellations within a contract, accusations of poor work) should always route to a person. An agent that tries to "handle" an angry customer with a scripted apology usually makes things worse, not better. Businesses that get the best results treat the agent as a filter for the routine calls, not a replacement for the judgement calls, and you'll only know what share of your calls is routine by measuring it.
Before it goes anywhere near your real number, make 20 test calls yourself and log each one as pass or fail in a simple sheet. Cover a normal booking, a reschedule, an interruption mid-sentence, a call from a noisy van with the radio on, a friend with a strong accent, a price question that isn't on your list and an angry caller asking for a manager. A pass means it got the facts right and either finished the job or handed over cleanly; a fail is anything it guessed. Fix the script, rerun only the failures, and go live once the whole set passes.
There's more on the best VoIP for small business in a separate post.
How it fits with the rest of your admin
A voice agent on its own is one piece. The real value shows up when it talks to the rest of your systems. A booked job should trigger an invoice, and a job that's overdue for payment should trigger a reminder without you having to remember to chase it, which is why pairing a voice agent with something like cloud based invoicing closes the loop from "phone rings" to "money in the bank." If chasing late payers is already eating your time, our guide to the best invoice reminder and late payment tools is worth reading alongside this one, because a voice agent that books jobs but leaves you manually chasing invoices only solves half the problem.
There's a data protection side too. If the agent records calls or stores names and numbers, you're processing personal data, so tell callers at the start ("this call is handled by an automated assistant and may be recorded"), keep recordings no longer than you need them, and make sure your privacy notice mentions it. The ICO's guidance on AI and data protection has a chapter on transparency in AI that covers exactly this. Treat it as general guidance, not legal advice.
If you'd rather not build and tune this yourself, working with someone who does hands-on AI implementation for small business will usually get you a working, tested agent faster than trial and error, because the mistakes (wrong routing rules, scripts that don't match how customers speak) are the same ones most businesses make on their first attempt.
When you're ready to compare, look at the best AI voice cloning tools.
Want AI doing the heavy lifting in your marketing?
I build the systems that handle the boring 80 percent, so you get your week back. Done properly, with the human kept in.
Getting started without wasting money
Start with one call type, not your whole phone line. Pick the most repetitive request, booking a standard appointment, answering your three most common questions, taking a message outside hours, and build the agent for that alone. Run it alongside your human team for the first two to four weeks rather than replacing anyone, check the transcripts every few days, and only widen what it handles once it's getting the simple cases right consistently. Businesses that try to automate everything on day one are the ones who end up switching the whole system off within a month. Here's a starting script for that first call type, using the plumbing example:
- Opening line: "Thanks for calling [business name], I'm the automated assistant. I can book a visit or take a message."
- Three questions, in this order: postcode, how urgent it is, and the preferred day.
- Rule one: never quote a price that isn't in the price list.
- Rule two: never promise a time that isn't free in the calendar.
- Rule three: transfer at once on the words "burst", "flooding", "gas smell", "complaint" or "solicitor".
- Close: read the booking back to the caller and confirm it by text.
If you want to keep an eye on what's changing in this space, it moves fast, our weekly AI news roundup for small business tracks the tools and updates worth knowing about.
Related: ai consultant contract what should be in it for small businesses.
That's a phone agent for customers, and for your own use there's a guide to talking with ChatGPT out loud.
I run AI training and workshops for teams, and help businesses with AI enablement and workflow automation, so the busywork runs itself and people get time back. Most of that work is with SaaS marketing teams and small businesses. Here's how we can work together.
Frequently asked questions
Can an AI voice agent completely replace a receptionist?
For most small businesses, no. It can handle the repetitive calls, bookings, FAQs and basic quotes, but complaints, emotional calls, and anything unusual still need a human, so most businesses keep someone checking in on it rather than switching the phone over entirely.
How much does an AI voice agent cost per minute?
Published pay-as-you-go prices vary by provider: Retell AI lists $0.07 to $0.31 per minute in US dollars, with its worked example at $0.11 per minute (language model $0.04, voice infrastructure $0.055, text-to-speech $0.015). Check the current pricing page before you budget, since prices change, and add any number rental or call charges.
Do I have to tell callers it's an AI voice agent?
Plan to. The ICO's AI and data protection guidance has a transparency chapter for systems that process personal data, so say at the start that an automated assistant is handling the call and whether it's recorded. That's general guidance, not legal advice, but it also saves you from the caller who feels tricked halfway through.
What happens if the AI voice agent gets something wrong?
It can book the wrong slot, misquote a price, or misunderstand a strong accent, which is why checking a sample of call transcripts weekly and keeping a clear escalation path to a human is part of running one, not an optional extra.
Can I try an AI voice agent for free?
Some platforms give starter credit: Vapi's pricing page lists $5 of free credits to get started. Spend it on the 20-call test plan before you connect your real number, not on a live line.
Sources worth reading
Let's get your marketing running on AI.
I'm Lilach. 23 years in marketing. Currently obsessed with AI.
My day job is making your lead generation and ads run on AI. SEO and content that compound. Automation, systems, and processes quiet enough to leave alone. Forty hours a week reclaimed.
Done for you, consulting, coaching, or training.
Run a business, a brand or a blog? I feature tools, products and expert guest posts for an audience of 15,000. See how it works