AI Voice Agents: How They Work and When They Make Sense
How an AI voice agent answers calls, which jobs it handles well and badly, how per-minute costs add up, and the UK call-handling rules to follow.

An AI voice agent is software that answers or makes phone calls and holds a spoken conversation: it converts the caller’s speech to text, decides what to say or do using a large language model connected to your information and systems, and speaks the reply. It works well for predictable calls such as bookings, opening hours, message taking and routing. It works badly for complaints, sensitive conversations and anything that needs judgement or regulated advice.
Voice agents have improved quickly, and the better ones now sound natural. That makes the design decisions more important, not less: what the agent may do, when it hands over, and what callers are told.
How does an AI voice agent work?
A call passes through four layers, usually in well under a second per turn.
- Telephony. A phone number, or a connection to your existing phone system, routes the call to the agent. This might be a new number, an overflow rule on your current line, or an out-of-hours divert.
- Speech to text. The caller’s words are transcribed as they speak. The system also detects when they have finished, and when they interrupt.
- The model. A large language model reads the transcript with its instructions: who it represents, what it may and may not do, your prices and policies. It can call tools, such as checking a calendar, creating a booking or logging a message in the CRM.
- Text to speech. The reply is spoken in a chosen voice.
Some newer systems use speech-to-speech models that handle audio directly, without a separate transcription step, which cuts the delay. OpenAI’s Realtime models are one example. Either way, the same design questions apply.
Behind the call, an automation layer, often a tool such as n8n, does the follow-up: sending a confirmation text, emailing a summary to the right person, updating records. Our comparison of n8n, Zapier and Make covers that layer.
When does an AI voice agent make sense?
The best uses share three features: a high volume of similar calls, a clear right answer, and a cost to missing the call.
Good fits:
- Missed and out-of-hours calls. Taking a proper message, with name, number and reason, rather than voicemail nobody returns.
- Bookings and rescheduling. Appointments, consultations, tee times, viewings, where the agent can check a live calendar.
- Routine questions. Opening hours, parking, what to bring, membership fees, how to renew.
- Routing. Working out what the caller needs and transferring them to the right person or team, with a summary.
- Overflow. Answering when every line is busy, instead of leaving callers in a queue.
Poor fits:
- Complaints and distressed callers, who need a person quickly.
- Regulated advice in law, finance or health, beyond general information.
- Calls where the caller’s identity must be verified before discussing an account, unless the verification is designed carefully.
- Taking card details by voice.
- Low call volumes, where a good voicemail and a fast callback cost less.
A worked example, with illustrative numbers: a membership body’s phone line rings most at lunchtime, when two of the three staff are on break, and roughly a fifth of calls go to voicemail. The voice agent answers only when the line is busy or out of hours. It handles renewal questions and event bookings, takes messages for everything else and texts the caller a confirmation. Complaints and anything about a member’s payment go straight to a person, or to a callback list if nobody is free. The measure is calls answered, and callbacks completed the same day.
How much does an AI voice agent cost?
Voice agents are usually priced per minute of call time, built up from the parts above: the voice platform, the language model, speech generation and telephony. At the time of writing (late 2026), Retell AI’s published pay-as-you-go pricing, for example, ranges from $0.07 to $0.31 a minute depending on the model and voice chosen. Other platforms use similar per-minute models, sometimes with a monthly fee.
To estimate running costs, multiply your expected minutes by the per-minute rate. A practice that diverts 300 calls a month averaging three minutes would use 900 minutes. At a rate of $0.15 a minute that is $135 a month in usage. These figures are illustrative; get a quote for your own call pattern and chosen voice.
The build is usually the larger cost: writing the call flows, connecting the calendar and CRM, setting up transfers and testing with real calls. A voice agent with bookings takes more work than a chatbot, because spoken conversations go off-script more often. Our post on what an AI chatbot costs in the UK explains the build-versus-running split in more detail.
What rules apply to AI call handling in the UK?
This is general information, not legal advice. Check your own position, particularly if you work in a regulated sector.
Inbound calls
- Tell callers. Say in the greeting that they are speaking to an AI assistant, and how to reach a person.
- Explain recording and transcription. If calls are recorded or transcribed, your privacy information must say so, why, and for how long you keep them. Add the voice platform and model provider to your records as processors, with data processing terms in place.
- Keep retention short. Decide how long recordings and transcripts are kept and delete them on schedule.
- Offer a person. A caller who asks for a human should be transferred or promised a callback.
Outbound calls
Outbound marketing calls are stricter. Under the Privacy and Electronic Communications Regulations, the ICO’s telephone marketing guidance says automated marketing calls need the person’s specific prior consent; general marketing consent, or consent to live calls, is not enough. The guidance describes automated calls as recorded messages and does not yet address AI-generated conversations directly, so the cautious approach is to treat an AI marketing call the same way. Service calls, such as appointment reminders to existing clients, are a different category, but still need care.
How to design one that callers do not hate
- Start narrow. One job, such as out-of-hours messages, before bookings and routing.
- Write the call flows. The top ten reasons people call, what the agent says, and when it transfers.
- Make handover easy. “Speak to a person” should always work.
- Keep answers short. Spoken replies of more than two sentences lose callers.
- Test with real calls. Have staff call it with awkward requests, accents and background noise before the public does.
- Review transcripts weekly at first. Fix the misunderstandings, then move to monthly.
Where we fit
As a UK AI agency, we build voice agents that answer, book and route calls, connected to your calendar and CRM and tested with your team, as part of our AI integrations work. For membership bodies, the phone line is often one of the first things we automate; the membership organisations page shows how it fits with the rest of the work.
Questions about AI Voice Agents: How They Work and When They Make Sense
Should callers be told they are speaking to an AI?
Yes. Say so in the greeting. It is honest, it sets expectations, and if the call is recorded or transcribed you need to tell callers how their information is used under UK GDPR anyway. Callers also tend to speak more clearly once they know.
Can an AI voice agent take card payments?
It is best not to. Reading card numbers aloud to a system that records and transcribes calls creates serious security and compliance problems. Send a secure payment link by text or email instead, or transfer to a payment line built for the purpose.
Will an AI voice agent replace our receptionist?
It usually covers the calls a receptionist cannot take: lunchtime, out of hours, busy periods and simple repeat questions. Most organisations keep a person for complex, sensitive and high-value calls, with the agent transferring those straight through.
Related reading
Need help with any of this?
Send the brief and we reply the same day with a price and a date.


