AI voice agents for business calls: how they work and where they fit

The phone rings while the receptionist is with a patient. The property manager’s office is closed, but a tenant has water dripping through the ceiling. Half the calls a car service takes are “can I move my appointment?”. Callers who get no answer rarely try again. They call someone else.

For years the choice was a phone menu (“press 1 for bookings”) or more people. AI voice agents add a third option: software that answers in normal speech, works out what the caller wants, checks your calendar or order system, does the routine part and passes the rest to a person. This guide explains how they work, where they fit, what to keep human and how to run a pilot without testing your customers’ patience.

How a voice agent works

What happens on a call

1. 01 Call comes in Your number, forwarded by your phone provider
2. 02 Speech becomes text Transcribed while the caller speaks
3. 03 The model decides Follows your script and rules
4. 04 Your systems answer Calendar, CRM, order status
5. 05 Text becomes speech Spoken back in a synthetic voice
6. 06 Call logged Transcript, summary, next step

A call passes through five parts, and each one adds a little delay:

• Telephony. You keep your number. Calls are forwarded to the agent through a phone provider: always, only after hours, or when nobody picks up within a few rings.
• Speech-to-text. The caller’s voice is turned into text while they speak.
• The language model and its tools. The model follows your script and rules, works out what the caller wants and uses tools (defined connections to your calendar, CRM or order system) to check and act. How that works is explained in how AI agents use tools.
• Text-to-speech. The answer is spoken in a synthetic voice. It’s streamed, so the agent starts talking before the whole sentence is ready.
• Logging. The transcript, a summary, the outcome and the next step go to your CRM or helpdesk.

There’s also a newer option: speech-to-speech models, which take audio in and produce audio out without separate text steps. They respond faster and sound more natural, especially when people interrupt. The pipeline above is easier to inspect, because you can read the text at every step, and it lets you pick the best speech provider for your language. Both are workable today. We choose per project, after testing with real calls in the caller’s language.

Speed and interruptions

In normal conversation people reply within a fraction of a second. A silence of two or three seconds on the phone feels as if the line went dead. That shapes the whole design:

• Every step streams, so the agent starts speaking before it has finished thinking.
• Slower lookups get a short filler: “Let me check the calendar.”
• The caller’s record is fetched from the caller ID before they finish saying hello.
• The conversation runs on a fast model. Heavier work, such as the summary for your CRM, happens after the call.

Callers also interrupt, and the agent has to stop talking when they do. The opposite problem matters just as much: people pause mid-sentence while they look for a date in their diary. Telling when someone has actually finished speaking is one of the things that separates a good voice agent from an irritating one.

Slovak and other languages

Slovak works well with the main speech providers today, as do English and most European languages. The hard parts are similar in every language: names, street names, product codes, numbers and dates (“on the twenty-second”), and callers who mix languages. Test your own vocabulary during the pilot. And don’t make the agent take down an email address letter by letter over the phone. Send a text message with a link instead.

Where voice agents fit

Good candidates share three traits: the call type repeats, the answer or the action lives in one of your systems, and a mistake is easy to put right. Voice is one support channel among several; what to automate across email and chat as well is covered in AI in customer support.

• Booking and rescheduling for clinics, salons, car services and property viewings.
• After-hours and overflow answering, so calls that would ring out get a real answer, or at least a structured message.
• Order and delivery status for a distributor or an e-shop, read straight from the order system.
• Fault reports for a property manager. The agent takes the address and the problem, sends an SMS asking for a photo, creates a ticket, and sends real emergencies (a gas smell, a burst pipe) straight to the person on call.
• Qualifying inbound sales calls. A few questions about what the caller needs, then a booked call with the right salesperson. More in AI for sales teams.
• Reminders and confirmations to existing customers, which cut no-shows. Cold sales calls are a different matter: consent rules apply, and most people hang up on them anyway.

Here’s what a routine call looks like when the agent is connected to the booking calendar:

Moving an appointment by phone

1. AI: Good morning, you’re speaking with the clinic’s virtual assistant. How can I help?
2. You: Hi, I need to move my appointment on Thursday. Something came up at work.
3. Tool call: bookings.find(phone: caller, day: Thursday)
4. Result: Thu 9:30 · dental check-up · 30 min
5. AI: I can see your check-up on Thursday at 9:30. I can offer Friday at 8:00 or Monday at 14:30. Which suits you better?
6. You: Monday afternoon is fine.
7. Tool call: bookings.move(to: Mon 14:30)
8. AI: Done: Monday at 14:30. You’ll get a text message with the confirmation. Anything else?

Notice the first sentence. The caller knows straight away they’re talking to an assistant, which is both a legal requirement (more below) and good manners.

How does this compare with what you may use today?

Phone menu, call centre or AI agent?

	Phone menu (IVR)	Outsourced call centre	AI voice agent
Answers at 2 a.m.	Yes	Partly	Yes
Understands “I need to move Thursday”	No	Yes	Yes
Books and updates records itself	No	Partly	Yes
Calms an upset caller	No	Yes	Partly
Cost per extra call	Very low	High	Low
Time to change the script	Days	Weeks	Hours

Each option has its place. A phone menu is cheap and predictable for a simple split between departments. A call centre handles anything a person can, at a person’s price. An AI agent sits in between: it understands free speech and acts in your systems at a low cost per call, but it doesn’t replace judgement.

What to keep human

• Complaints and upset callers. The agent should recognise them and transfer them early.
• Medical symptoms and other emergencies. The agent tells the caller what to do (in Europe, call 112) and alerts a person. It never tries to assess the situation itself.
• Payments and card details. Send a secure payment link by SMS instead of taking numbers by voice.
• Anything outside the rules you defined, and anyone who asks for a person.

The hand-off is part of the product. The agent says it’s passing the caller to a colleague, transfers the call and puts a summary on the colleague’s screen, so nobody has to repeat themselves. If nobody is available, it takes a message and tells the caller when to expect a call back. Then someone has to actually call back. How to design these checkpoints is covered in human in the loop.

Disclosure, recordings and GDPR

This isn’t legal advice, but a few points apply to almost every voice project in the EU:

• Say it’s an AI. Since August 2026 the EU AI Act requires that people are told when they’re dealing with an AI system, unless it’s obvious. A first sentence like “You’re speaking with the clinic’s virtual assistant” covers it and sets expectations.
• Tell callers if you record or transcribe the call, and why. Under GDPR you need a lawful basis for both.
• Keep only what you need, for as long as you need it. Decide how long recordings and transcripts are stored and who can read them.
• Know where the audio goes. Speech, model and voice providers all process the call. Choose ones that offer processing in the EU and don’t train on your data, and sign their data processing agreements.

The wider picture is in our guide to AI, GDPR and the EU AI Act.

What it costs to run

Running costs are mostly per minute of conversation: speech recognition, the language model and the voice, plus the phone line. For routine calls that’s usually a fraction of what the same minutes cost in a person’s time. Estimate it from your real call volumes and average call length before you build anything.

Then compare it with the cost of the calls you miss today: the booking that went to a competitor, the tenant who complained to the owner, the order question that turned into an email thread. On the build side, a focused agent connected to one or two systems falls under “one process” on our pricing (€5–10k). More call types and integrations move it into the next tier.

How to run a pilot

A pilot checklist

• Listen to 50–100 real calls Or read the call notes, and count the call types
• Pick 3–5 call types The ones that repeat and end in an action in your systems
• Write the script and the limits What the agent may do, and what it must hand over
• Connect one or two systems Usually the calendar and the CRM
• Test it by phone every week With awkward callers, background noise and mumbled dates
• Go live in a narrow slot After hours, or for one call type only
• Review every call at first Fix the script from real transcripts
• Compare with the baseline Missed calls, bookings, hand-offs, complaints

What to measure during the pilot:

• Share of calls the agent completes without a person
• How often it hands over, and why
• Bookings made or changed by the agent
• Callers who hang up in the first 15 seconds, a sign that the opening or the speed is off
• Complaints about the agent
• Missed calls before and after

There’s a bonus most companies don’t expect. Once every call is written down, you can see why people call, which questions your website doesn’t answer and where bookings get lost. A month of transcripts is often enough to fix the FAQ, a confusing form or the opening-hours page.

What to do next

1. Count last month’s calls by type, including the ones nobody answered.
2. Pick the three to five types that repeat and end in a booking, an update or a status answer.
3. Write down what the agent may do and what always goes to a person.
4. Check what your calendar, CRM or order system lets you connect to.
5. Run a narrow pilot with every call reviewed, then decide on the numbers.

If you’d like to know which of your calls an agent could take, have a look at how we build AI voice agents, or send us a short description of your typical calls.