What is an AI voice agent? A guide for service businesses

Say “voicebot” and most people picture a bank’s phone line: press 1, press 3, listen to another menu, then give up. An AI voice agent has one thing in common with that machine, which is that neither of them is a person. We’ve run 1,700,000+ AI calls over 3.5 years since 2022, so what follows comes from real phone lines rather than vendor demos.

The short version: the agent listens, understands what the caller wants and acts on it, within limits you set. The rest of this guide explains how that works, where it fits a small service business and where it doesn’t.

Voicebot vs IVR vs AI voice agent: what’s the difference?

People use “voicebot” for quite different things. It’s easier to split them into three generations.

IVR: “Press 1 for bookings”

IVR (interactive voice response) is the classic phone menu. It plays a recorded message and waits for a key press. It can’t do anything that nobody recorded into its menu tree in advance. A caller who needs something outside the menu ends up in a queue, or hangs up.

For a small service business, IVR rarely makes sense. A garage or a salon doesn’t have ten departments to route calls between.

Keyword voicebots

The second generation listened to the caller’s voice, but it only recognised a handful of preset words and phrases (“booking”, “complaint”, “opening hours”). If the caller said the same thing a different way, the bot didn’t follow and asked again. Most of the bad reputation voicebots still have comes from this generation.

AI voice agents

A current voice agent runs on a large language model (LLM), the same kind of technology behind chat assistants. The caller speaks normally, and the agent works out what they mean and answers in its own words. It stays within the script and the information the business gives it.

	IVR	Keyword voicebot	AI voice agent
How the caller answers	key presses	a few set words	normal speech
What the system says	recorded prompts	recorded or synthetic prompts	its own replies, within a script
When the caller goes off-topic	the menu restarts	“Sorry, I didn’t catch that”	it responds, then steers back
Typical use	routing calls in large organisations	simple queries	reception, bookings, outbound calls

How does an AI voice agent work?

Each call runs as a loop of three steps, repeated after every sentence the caller says:

1. Speech to text. The caller’s voice is transcribed while they speak.
2. Language model. The model receives the transcript along with the script, your business information and the conversation so far. It decides what to say and, where needed, triggers an action, such as checking the calendar for a free slot.
3. Text to speech. The reply is spoken in a synthetic voice and played down the line.

Why response time matters

People expect very short pauses in conversation. A study of turn-taking across ten languages measured an average gap of around 200 milliseconds between one speaker finishing and the next starting (Stivers et al., PNAS 2009). Each of the three steps above takes time, so an agent always answers a little more slowly than a person would.

If the pause runs too long, callers assume the line has dropped and start talking again. That’s why so much work on voice agents goes into speed: a shorter first sentence in each reply, faster models, transcribing while the caller is still speaking. When you compare providers, listen to a real phone call, not a demo recorded into a studio microphone.

Interruptions

The second thing that separates a usable agent from a frustrating one is what it does when the caller cuts in. A good agent stops talking and answers what the caller just said. A poor one finishes its prepared sentence, and the caller feels they’re back with a keyword bot.

This is harder than it sounds. The agent has to tell whether the caller wants to say something, is only saying “yeah, yeah”, or is talking to someone else in the room.

Inbound calls: an AI receptionist for small businesses

Most small firms start here, with calls that nobody would otherwise pick up.

Calls you can’t get to

A typical picture: the roofer is up a ladder, the dentist has a patient in the chair, the estate agent is halfway through a viewing. The phone rings and nobody answers. Some callers leave a message or ring back. The others try the next business in the search results.

The agent can take the call after a few rings. It finds out who is calling and what they need, so instead of a voicemail you get a logged request with contact details.

Evenings and weekends

Plenty of people sort out personal admin after work. They want to book the car in or arrange a salon appointment at 8pm, after you’ve closed. An AI receptionist takes those calls as well, and in the morning your CRM shows who rang and why.

Bookings straight into the calendar

Once it’s connected to your calendar, the agent offers free slots, the caller picks one and the booking is saved. A confirmation can go out by SMS or email. This is the biggest difference from voicemail: the call ends with an appointment, not a promise to ring back. The calendar side is covered on our booking calendar page.

Answering common questions

Opening hours, whether you work on a particular car make, where to park, whether a referral is needed. These questions come up again and again, and the agent answers them from the material you provide. When it doesn’t have an answer, it notes the question and arranges a callback.

Outbound calls: when the agent rings out

The agent can also make calls. Outbound calling needs more care, partly for legal reasons and partly because of how people see your business.

Sensible outbound uses:

• Fast follow-up on enquiries. A customer fills in a form on your website and ticks the box agreeing to a call. The agent rings within minutes, asks the follow-up questions and books a visit or a quote.
• Appointment reminders. It calls the day before to check the customer is still coming, and offers another slot if they aren’t. You can then give the freed-up slot to someone else.
• Bringing past customers back. It calls people who haven’t booked for a while, or who enquired and didn’t go ahead. Read the PECR section below before you run this one.

What I wouldn’t do is buy a list of numbers and have the agent cold-call strangers with an offer. In the UK, that breaks PECR unless every person on the list specifically agreed to automated calls from you, and a bought list won’t give you that. Even where it is legal, a cold call from a robot does nothing for your reputation.

Where an AI voice agent works and where it doesn’t

The agent does well on calls that repeat and have a clear goal. The more a call looks like a form (who, what, when, where), the better it copes.

Works well:

• booking appointments, including moving or cancelling them
• questions about opening hours, availability and prices, if you want to share them
• taking enquiry details before an engineer or salesperson gets in touch
• confirmations and reminders
• triage: what’s urgent and what can wait until morning

Best left to a person:

• Complex complaints. A customer whose repair went wrong needs someone who can make a decision. The agent can log the complaint and pass it on, but a person should deal with it.
• Emotional situations. An angry or frightened caller, a health problem, a bereavement. An automated voice is the wrong response here, even if it could technically reply.
• Price negotiation and large jobs. The agent can book the meeting. A person should close the deal.
• Anything where wrong information would cost something. Professional advice, diagnoses, legal questions. The agent shouldn’t answer anything that isn’t in its material.

Handing over to a person

Decide in advance when the agent passes a call on. Common triggers:

• the caller asks for a person
• the topic is outside the script
• the caller is upset
• the call is a complaint

The agent then transfers the call or arranges a callback. You get a notification with a summary of what was said, so the customer doesn’t have to repeat everything.

An agent with no route through to a human is worse than having no agent. That’s the first rule I’d set.

Accents, names and languages

Current speech recognition handles everyday conversation well, including people correcting themselves mid-sentence. Synthetic voices also sound far more natural than they did a few years ago. These are the weak spots worth testing:

• Names and street names. An unusual name can be transcribed wrongly. Having the agent read key details back helps (“So that’s Ms Oyelaran, 12 Larch Road, is that right?”).
• Numbers. Phone numbers, registration plates, order numbers. Reading them back helps here too.
• Strong regional accents. Usually fine, but not always. Test with colleagues or friends who sound like your actual callers.
• Noise and poor signal. Calls from a car or a building site are harder to transcribe, for software and for people.

Our agents work in English and Czech, and they switch between the two if the caller does.

What the law says in the UK and EU

This is a short overview, not legal advice. If you plan outbound marketing calls, check your setup with an adviser first.

Telling callers they’re talking to an AI (EU AI Act)

Article 50(1) of the EU AI Act requires AI systems that interact directly with people to be designed so that those people are informed they’re dealing with an AI system, unless that’s obvious to a reasonably well-informed person. Article 50(5) says the information must be given clearly, at the latest at the first interaction. Under Article 113, these rules apply from 2 August 2026.

In practice, the agent introduces itself as your business’s AI assistant in its first sentence. If you’re a UK firm with callers in the EU, assume the rule applies to you. If all your callers are in the UK, the AI Act isn’t what binds you, but we set every agent up the same way anyway. A caller who works it out halfway through the call has good reason to feel misled.

Recording and transcripts (GDPR and UK GDPR)

A call almost always contains personal data, and agent calls are usually recorded and transcribed. The ICO’s guidance says you must give people privacy information at the time you collect their data. That information covers why you process it, your lawful basis, who receives it, any transfers abroad and how long you keep it. The EU GDPR sets the same requirement in Article 13.

On a call, that usually means a short line at the start (“this call is recorded and transcribed”) and a fuller privacy notice the caller can find on your website. Decide where recordings are stored and how long you keep them before you go live. If your CRM stores data outside the UK or EU, that transfer belongs in the notice too. HighLevel, which our CRM runs on, stores data in the US under the EU-US Data Privacy Framework.

Automated marketing calls (UK PECR)

Regulation 19 of PECR bans direct marketing calls made with an automated calling system unless the person you’re calling has told you beforehand that they consent. The regulation’s definition covers systems that transmit sounds that aren’t live speech. A synthetic AI voice isn’t live speech, so we treat AI marketing calls as automated calls.

The ICO’s guidance adds three points:

• The consent must specifically cover automated calls. General marketing consent, or consent to live calls, is not enough.
• Every automated call must give your name and a contact address or freephone number.
• You must let your number, or another number people can contact you on, display to the person you’re calling.

A reminder about an appointment the customer booked themselves is a different situation from calling to sell something. If a call mixes the two, treat it as marketing.

What you need before you launch

The technology is now the easy part. Preparing the content takes more work. This is the order I’d follow:

1. Pick one or two scenarios. For example, “inbound calls outside opening hours” and “appointment reminders”. Starting with everything at once rarely pays off.
2. Write down how the call should go. How the agent introduces itself, what it asks and in what order, when it ends the call and when it hands over to a person.
3. Collect common questions and answers. Go through what people ask you on the phone. Twenty questions with short answers will cover most calls. Also write down what the agent must not answer, such as exact repair prices.
4. Connect the calendar and CRM. Without a calendar the agent can’t book anything. Without a CRM, call outcomes get lost in email. Ideally the transcript, summary and booking save to the contact record automatically.
5. Decide which number the agent answers. It can take calls on your existing number (after a few unanswered rings, for example) or get its own number. Our phone system page explains how we handle numbers and call forwarding.
6. Test with practice calls. Ring it yourself. Be awkward, interrupt, ask odd questions. Get colleagues to do the same.
7. Listen after launch. For the first few weeks, read the transcripts and adjust the script wherever calls get stuck.

How to tell whether it’s paying off

Without measurement, you can’t tell whether the agent is helping or just picking up the phone. I’d track these:

• Calls answered by the agent, especially outside opening hours. This shows how many calls would otherwise have gone nowhere.
• Calls that ended with a result: a booking, a logged enquiry, an answered question.
• Share of calls handed to a person. If the agent passes on almost everything, the script or the material behind it isn’t good enough.
• Hang-ups in the first few seconds. If callers often hang up right after the introduction, try a different opening line or voice.
• No-shows before and after you start reminder calls.
• Response time on web enquiries, meaning the time between a form being sent and the first call.

Compare these figures with how things were before launch. So for a few weeks beforehand, write down how many calls you miss and how many enquiries are waiting for a callback.

To see what one of these calls looks like step by step, including what gets written to the CRM along the way, go to our AI voice agents page.