An AI voice agent is software that answers or places phone calls, holds a real spoken conversation, and takes action on what it hears — booking an appointment, updating a CRM, routing to a human. Unlike an IVR phone menu, it has no fixed script tree: it understands what a caller means and responds accordingly. A production agent costs roughly $1,500–$5,000 to build and $300–$1,000 a month to run, and takes two to three weeks to deploy for a single well-defined job.
This guide covers what is actually inside one, what they are genuinely good at, where they fail, and how to tell whether one is worth building for your business.
What is an AI voice agent?#
Three things have to be true before something deserves the name.
It understands intent rather than keywords. A caller who says "hey, is there any chance you could squeeze me in Thursday afternoon, or Friday if that's easier" is making a single request with a fallback. An IVR hears none of that. An agent handles it as one turn of conversation.
It takes real action. Understanding without action is a transcription service. A real agent reads live calendar availability, writes the booking, updates the record, sends the confirmation — during the call.
It holds context across the conversation. If the caller mentions their name in the first sentence, the agent should not ask for it in the fourth.
Anything that fails one of those three is a phone menu with better marketing.
How does an AI voice agent actually work?#
A live call runs through four stages, on a loop, several times per exchange.
1. Speech to text. Audio streams in and is transcribed in real time. The important word is streaming — the transcript builds as the caller speaks rather than after they stop. This is also where turn detection lives: deciding whether a half-second of silence means "they are thinking" or "they are done". Getting that wrong is the single most common reason an agent feels wrong to talk to.
2. Reasoning. The transcript, the conversation so far, and your instructions go to a language model, which decides what to say and whether to call a tool.
3. Tool calls. "Check availability", "create booking", "look up this customer", "transfer to a human". This is the part that makes the agent useful rather than merely conversational, and it is where most of the build effort goes.
4. Text to speech. The response is synthesised and streamed back, ideally starting before the full sentence has been generated.
The whole loop has to complete in under about two seconds for the conversation to feel normal. That constraint drives nearly every engineering decision in the build — we cover the latency problem in detail here.
What can an AI voice agent actually do well?#
Four jobs, in rough order of how reliably they pay off.
Answering every inbound call#
The most boring use case and the most valuable one. Most small businesses miss a meaningful share of inbound calls — at lunch, after hours, during a rush, when the one person who answers the phone is with a customer. Those callers rarely leave a voicemail. Many of them call a competitor next.
An agent picks up on the first ring at 2am on a Sunday. Even if all it does is capture the enquiry properly and book a callback, the arithmetic usually works out on its own.
Qualifying and booking#
Asking a consistent set of questions, checking the answers against your criteria, and putting qualified people into a real calendar slot. Agents are good at this precisely because they are consistent — they ask question four even when the caller was chatty on question three.
Outbound follow-up#
Calling back a web lead within seconds of the form submission, chasing a no-show, confirming an appointment the day before. Speed matters enormously here, and it is the one thing software does better than any human team — the data on lead response time is stark.
Deflecting repetitive questions#
Opening hours, location, parking, what documents to bring, order status. If a third of your call volume is six questions, that third can be handled instantly.
Where do AI voice agents fail?#
Being straight about this matters more than the sales pitch, because a badly-chosen first deployment sours people on the whole category.
Emotionally charged calls. A complaint, a bereavement, a cancellation someone is upset about. The agent should recognise the tone and hand off fast, not try to resolve it.
Anything with real consequences and no verification path. Medical advice, legal positions, financial commitments. An agent can take the details and book the appointment. It should not give the answer.
Genuinely novel situations. Agents handle the common cases superbly and the long tail poorly. Design for hand-off rather than pretending the tail does not exist.
Poor audio. A caller on a motorway with the window down will defeat any transcription model. Build a graceful "I'm having trouble hearing you, let me get someone" path.
Businesses with very low call volume. If you get four calls a week, the monthly run cost may exceed the value. Be honest about your numbers before you build.
What does an AI voice agent cost?#
Roughly:
| Component | Typical range |
|---|---|
| Build (single-purpose agent) | $1,500 – $5,000 one-time |
| Ongoing platform + telephony | $300 – $1,000 / month |
| Per-minute usage | $0.07 – $0.20 / minute |
The build fee scales with how many distinct call flows you need and how many systems the agent has to write to. Two flows and a Google Calendar integration is the cheap end. Six flows, a CRM, a payment lookup and a bespoke escalation matrix is not.
The full cost breakdown, including the line items vendors tend not to quote, is here.
How long does it take to build?#
Two to three weeks for a single-purpose agent, in four stages:
- Scoping (2–3 days). Which calls, what the agent should do, what it must never do, where it hands off.
- Build (5–8 days). Prompt and flow design, voice selection, integrations, escalation rules.
- Testing (3–5 days). You call it. Repeatedly. In an accent, from a car, while interrupting it. This stage is where the quality actually comes from and it is the one most often cut short.
- Launch and tune (ongoing). Real calls surface things no test script predicted. The first fortnight of transcripts is worth more than the entire testing phase.
Should you build one?#
A reasonable filter, in order:
- Do you miss calls? Check your carrier's stats rather than guessing. If your missed-call rate is under 5%, the case is weaker.
- Is a call worth real money? A dental practice where one new patient is worth £800 has a different equation from an e-commerce store where a call is worth £30.
- Are the calls repetitive? Repetitive is good. It means high coverage from a modest build.
- Do you have somewhere for the agent to write? An agent that cannot see your calendar or CRM is worth a fraction of one that can.
- Can you accept it will not be perfect? It will be more consistent than a rushed human at 5pm on a Friday, and it will occasionally get something wrong. Both are true.
If you answered yes to four of those five, it will almost certainly pay for itself. If you answered yes to two, wait.
Getting started#
The single most useful thing you can do before speaking to anyone is listen to twenty of your own recorded calls and write down what happened in each. Not what you think happens — what actually happened. That document will tell you, and whoever builds your agent, more than any vendor demo.
When you are ready, our AI voice agent service page covers exactly what we build and what it costs, or you can start scoping a project and we will come back with a fixed quote.