Voice Agents
How Much Does a Voice Agent That Answers Calls Cost?
The published price starts around $0.08 per minute of conversation, but the real invoice is usually several times higher: the advertised rate covers one component, not the whole call. The full price is made up of speech recognition, the language model, speech synthesis and telephony together, plus the one-off setup work.
What actually makes up the price of a voice agent?
A call with an AI agent is not one service, but four running at the same time. Speech recognition turns what the caller says into text. The language model decides what to answer. Speech synthesis turns the answer into sound. Telephony delivers it to the receiver. Each of these four is billed separately.
That is exactly why the per-minute price shown on a marketing page is almost never what you see on the invoice: it is usually the price of just one part. On top of that comes one-off work, the script, integration with your systems, testing on real calls, none of which any rate plan covers.
A practical conclusion before budgeting: it is worth calculating the price not by the cheapest published rate, but by all four parts combined. Otherwise your estimate will differ from the actual bill by several multiples.
What do vendors publish themselves?
Some prices are public, and you can verify them without talking to a salesperson. Below are only the figures that appear on vendors' own pricing pages, with what is not included next to each one.
| Service | Published price | What's NOT included in the price |
|---|---|---|
| ElevenLabs voice agents, additional minutes | $0.08 per minute | Telephony; price rises above the plan limit |
| ElevenLabs Creator plan | $29/month | Minutes above the limit, telephony |
| ElevenLabs Growing Business plan | $390/month | Minutes above the limit, telephony |
| AWS Polly, Neural voices | $16 per 1 million characters | Speech recognition, language model, telephony |
| Google Cloud Text-to-Speech (Wavenet, Neural2, Polyglot) | $16 per 1 million characters | Speech recognition, language model, telephony |
| Telnyx voice agent orchestration with STT and TTS | $0.05 per minute | Language model and telephony |
Why does the bill end up higher than the advertised minute?
Because the advertised rate covers one part. Once you add up all four, recognition, the model, synthesis and telephony, the real minute usually costs several times more than the figure you see on the first screen.
The second layer of cost is work, not a rate. The script, the connection to your CRM or booking system, testing on real calls, and fixes after the first few weeks. This work is one-off, but it is usually what makes up the largest part of the bill in the first month.
Does the Latvian language change the price?
In the rate itself: no. Large vendors do not price by language. Indirectly: yes, and it needs to be planned for.
None of the major vendors publicly measures how well its voice and recognition perform specifically in Latvian and specifically over a phone line. Latvian is on the list of supported languages, but there is no verified quality mark for it. That means you have to test it yourself, on your own scripts and your own conversations, and that time is part of the project cost.
Our own experience recording Latvian voiceover for client videos: the result is most often ruined by numbers the engine reads mechanically, ignored long vowel marks, English abbreviations and overly long sentences. All of this can be fixed in the text, but it takes several passes, which is exactly why the budget needs room for testing, not just for minutes.
What does the law require from a voice agent?
Since 2 August 2026, the EU AI Act's transparency obligations have been in force: a person must be informed that they are interacting with artificial intelligence no later than the moment of first interaction. For a voice agent, that means one sentence right after the greeting, before the questions start.
This also affects the script, and with it the cost: after the disclosure, some callers will immediately ask for a human, so you need to build in a way to transfer to one. We have written about the requirement itself in more detail separately: whether you must disclose that a robot is answering.
How does call volume change the calculation?
The per-minute price multiplies by volume, so the difference between ten calls a day and a thousand is not just a bookkeeping question, it changes the approach too. For low volume, a ready-made subscription solution is usually more cost-effective; for high volume, a system assembled in-house from separate components starts to pay off, because each part of the minute is cheaper there.
The second factor is the complexity of the conversation. An agent that answers the most common questions from a script is cheaper both to set up and to maintain than one that has to reach into a CRM, check order status and log a booking. Price is set not by word count, but by the number of integrations.
When is a voice agent not needed?
The honest answer: more often than a salesperson would like. In a Sinch survey conducted in January and February 2026 among 2,527 leaders across ten countries, 74% of companies admitted they had already switched off or rolled back an AI agent they had deployed in customer communication; among those who deployed it with mature governance controls, the figure was even higher, at 81%.
In practice that means: if you get few calls and the questions are typical, a callback form or a simple phone menu tree will work more cheaply and with no risk. If the conversation touches money, health or legal obligations, the cost of a mistake is too high to leave to a system without human review. And if the process itself is broken, an agent will not fix it, it will only speed it up.
What is not publicly known about these costs?
There is no public data on what a voice agent actually costs a Latvian company specifically: vendors publish prices globally, but local implementation costs, the work, the integrations, the testing, are not compiled anywhere. There are also no public measurements of Latvian-language quality over a phone line.
That is why this article contains no "average price in Latvia" and no percentage for time saved. Where there is no data, it is more honest to say so than to fill the gap with a number nobody can later verify.
How we calculate this price
We do not quote a per-minute price over the phone, because without knowing the process, it would not be honest. First we find out how many calls you get per month, which questions repeat, and where the call ends up, a booking, a CRM, or a person's notebook. Only then does the number we give match the actual bill.
The first 90-minute AI consultation is free: we go through your calls and tell you whether you even need a voice agent, and what it would cost. If your process runs over messaging rather than the phone, see AI agent for email.
Sources
- ElevenLabs: voice agent pricing (05.09.2026.)
- AWS Polly pricing (05.09.2026.)
- Google Cloud Text-to-Speech pricing (05.09.2026.)
- OpenAI API pricing (05.09.2026.)
- Telnyx: voice agents (05.09.2026.)
- Sinch "The AI Production Paradox" (2026) (05.09.2026.)
- AI Act Article 50: transparency obligations (05.09.2026.)
Frequently asked questions
How much does one minute of a voice agent cost?
Published rates start around $0.08 per minute, but that is the price of just one component. A full minute includes speech recognition, the language model, speech synthesis and telephony, so on the invoice it is usually several times higher than the advertised figure.
What is included in the price of a voice agent?
Four technical parts, speech recognition, the language model, speech synthesis and telephony, plus one-off work: the script, integration with your systems, and testing on real calls. In the first month, this work is usually the largest part of the bill.
Does Latvian cost more?
Not in the rate itself: large vendors do not price by language. But none of them publicly measures Latvian-language quality, so you have to test it yourself on your own scripts, and that time is part of the project cost.
Do you have to tell the caller that a robot is answering?
Yes. Since 2 August 2026, the EU AI Act's transparency obligations require that a person be informed they are interacting with artificial intelligence no later than the moment of first interaction. For a voice agent, that means one sentence right after the greeting.
When is a voice agent not needed?
When you get few calls and the questions are typical: a callback form or a simple phone menu will work more cheaply. Also when the conversation touches money, health or legal obligations: there, the cost of a mistake is too high to leave to a system without a human.
How often do companies abandon a voice agent?
In a 2026 Sinch survey of 2,527 leaders across ten countries, 74% admitted they had already switched off or rolled back an AI agent they had deployed in customer communication. Among those who deployed it with mature governance controls, the figure was 81%.
Is there a known average price for a voice agent in Latvia?
Not publicly. Vendors publish global rates, but local implementation costs, the work, the integrations, the testing, are not compiled anywhere. So any "average price in Latvia" would be made up.
Which is cheaper: a ready-made solution or a custom-built one?
For low call volume, a ready-made subscription solution is usually more cost-effective, because there is no setup work. For high volume, a system built in-house from separate components starts to pay off, because each part of the minute is cheaper there. The threshold is in your minute count, not in the technology.
Let's work out your numbers
Tell us how many calls you get per month and how they usually end. In a free 90-minute conversation, we'll tell you whether you need a voice agent, and what the real bill would be, before the work starts, not after.