What Is an AI Voice Agent and When Does It Pay Off?

An AI voice agent is software that talks with your website visitors out loud: the visitor asks a question by speaking, the agent answers in speech, in real time. Before anything else, one clarification that saves a lot of confusion. Our voice bot works as a microphone button inside the chat window on your website. It has no phone number, it does not take calls and it does not dial out. Phone calls are coming soon, but they do not exist today, so if a vendor promises you an agent that answers your phone line, ask for a live demo.

Here is what the conversation actually looks like, what the agent can and cannot do by voice, why Serbian is a demanding language for voice agents, who benefits, and what it costs.

What an AI voice agent is, and what it is not

It is the same bot that already answers in your chat, given ears and a voice. Same content, same tools, same knowledge. The only difference is that the visitor speaks instead of typing.

It is:

• a microphone button in the chat widget on your site, provided the site runs over a secure connection (https);
• a real-time conversation with 10 voices to choose from and adjustable speaking speed;
• help for a visitor browsing from a phone, driving, or simply not in the mood to type;
• a way to clear repetitive questions outside business hours.

It is not:

• a phone system. No number, no inbound or outbound calls, and calls are coming soon;
• a replacement for text chat. Part of the work still happens by typing, see below;
• available on WhatsApp, Messenger, Instagram or Viber. On those channels the bot stays text-only;
• audio passing through our servers. The sound goes straight from the visitor's browser to OpenAI, which handles speech recognition and speech output, and our server never sees that audio.

What a conversation sounds like

The visitor opens the chat and taps the microphone, and the browser asks for permission. The agent greets them with a line you define, and the conversation runs. A session lasts up to one hour, and if the visitor stays silent past the inactivity limit you set (30, 45 or 60 seconds), the session closes on its own so it stops burning minutes. In your dashboard you later see the full conversation, the questions the bot could not answer, and the answer confidence score over time.

An example of a booking conversation:

Visitor: Hi, do you have anything open this week?

Agent: Hello. We have Wednesday at nine thirty and Thursday at five in the afternoon. Which one works better?

Visitor: Thursday, please.

Agent: Great. The appointment takes thirty minutes. Tell me your name, and type your phone number into the chat so no digits get lost.

Visitor: Marko Jovanovic.

Agent: Noted. Thursday at five in the afternoon. Add the phone number and the appointment is booked.

Notice the phone number is never dictated. Spoken digits, email addresses and street names are the easiest place for a conversation to go wrong, so it is better for the visitor to type them than for the agent to ask "was that a six or a seven" three times.

What the voice bot can and cannot do

By voice it can:

• search your company's content and answer from it;
• check open slots and book an appointment;
• read out your list of services with prices;
• explain details from a specific page of your site.

By voice it cannot, these stay in text chat:

• taking an order;
• checking order status or cancelling an order;
• producing a payment link;
• checking stock;
• cancelling or rescheduling an existing appointment.

So treat voice as an addition, not a swap. For online stores it handles "are you open on Saturdays" and "how long does delivery take", while everything touching the cart stays in text, as described on the AI for e-commerce page.

Why the bot does not make things up

A spoken mistake sounds more convincing than a typed one, so this matters more with voice. Your content (website, PDFs, spreadsheets, question and answer pairs) is cut into smaller pieces and stored in a separate database that belongs to your company alone. For every question the bot searches that content first and answers from what it found. Your own instructions sit at the very end of its rules, because when instructions conflict the model follows the one closest to the end, which means your wording has the last word. When it finds nothing, it says so and offers a contact or a callback instead of guessing.

There is also a nightly document check that warns you when a file was read badly, for example a scanned PDF or a catalogue whose font is missing Serbian letters. The bot can barely find anything inside such a file, and in conversation that shows up as "I do not have that information" even though the information exists.

Why Serbian is hard for voice agents

If you are running an AI that speaks Serbian, the language itself is a bigger obstacle than the technology. Three reasons stand out.

Cases

Serbian nouns change endings constantly, so one service name shows up in several forms within a single conversation, and city names do the same. The agent has to understand that two differently spelled forms are the same thing, otherwise it will tell a visitor you do not cover a city you actually cover.

Prices and numbers

A price written as 4,500 RSD has to be spoken as words, not read digit by digit. The same goes for times, dates and percentages. Get this wrong and the agent sounds like a parking machine.

English words inside Serbian sentences

Product names, brands and model codes are usually English and land in the middle of a Serbian sentence. Both recognition and pronunciation slip there, so the names that matter most should be written clearly in your content. Details on the product itself are on the AI voice agent page.

Who it pays off for, and who should wait

It pays off when you have decent website traffic and a stream of repetitive questions that end in a booking or a visit: clinics, salons, service shops, hotels, real estate agencies. "Do you have anything free on Thursday" is a perfect voice question. For one industry in detail, see the AI for clinics page.

Wait if your main channel is WhatsApp or Instagram, since voice does not run there. Wait if what you really need is someone answering the phone, because that is still ahead of us. Wait if your traffic is small enough that five questions a month are handled fine in text. And wait if your content is messy, because voice will only deliver empty answers faster.

What it costs

Voice minutes come with the Pro and Business plans, and messages are consumed in parallel: every visitor turn in a voice conversation counts as a message, just like a chat question, a message from a social channel, or a submitted chat form.

Annual billing saves 12 percent, and Enterprise is negotiated. Appointment booking comes with Pro and Business, or as an add-on at 79 EUR per month. Full details are on the pricing page.

If you run out of minutes, a 200 minute voice pack costs 50 EUR as a one-off, and each extra minute beyond that is 0.25 EUR. Extra messages are 0.05 EUR each (50 for 2.50 EUR, 100 for 5 EUR, 200 for 10 EUR, 500 for 25 EUR).

Two things to know in advance. A top-up creates an invoice with a 7 day term and is paid by bank transfer, and the quota is added to your account only once the payment is confirmed, so plan top-ups before you run dry. And when a quota is used up, the bot stops answering new questions until it renews or is topped up. There is no automatic charging past your quota, so the bill cannot surprise you.

Phone calls: coming soon, not today

The question comes up in every conversation: can this answer the phone instead of the front desk? Phone calls are coming soon. Today the agent has no number and touches no telephony at all, it lives in the website chat. We will announce it when calls are ready, and until then keep the phone line staffed.

Trying it

The trial runs for 7 days and needs no card. Installation is one line of code on your site, and the content is pulled from your website, PDFs and spreadsheets. Practical advice: run the text bot for a week first and read the list of questions it could not answer. That list tells you whether voice is worth turning on at all, and exactly what it will be working on when you do.