Call Center Economics · 4 August 2026 · 11 min read

AI voice agent cost per minute: what the quote leaves out

The advertised AI voice agent cost per minute is only one line in the bill. What you actually pay is the loaded cost of every call the agent finishes without a human, and a per-minute quote on its own cannot tell you what that is. Here is the arithmetic, built from the rates we publish, so you can run your own numbers through it.

COST PER HANDLED CALLif 85% handled without a humanEUR 0.64if 70% handled without a humanEUR 0.77if 45% handled without a humanEUR 1.20Illustrative, with inputs you set: same 100,000 minutes a month at the sameEUR 0.18 and a three-minute average call, the cost per handled call swingsfrom EUR 0.64 to EUR 1.20 on the share of calls finished without a person.We publish rates, not containment.

You have decided to buy. Someone has sent you a per-minute rate excluding VAT, and you want to know whether it is fair. That question cannot be answered yet, because a per-minute rate on its own does not tell you what a handled call will cost you.

The rate is real. It is also only one of four layers, and not the only one you will pay for. Underneath it sit telephony, speech recognition, the language model, speech synthesis, the build, and the ongoing work of keeping the scenario current. Above it sits the number that decides everything: the share of calls the agent finishes without handing off to a person.

A quote may cover only one layer, so ask which layers yours includes. Below is the arithmetic in the open, at three volumes, built from rates we publish. Every line built from our rates is checkable; the two inputs that are yours, call length and containment, are assumptions, and we flag them where they are used.

The four cost layers behind one minute

  1. Carriage. The phone call itself: the SIP trunk, per-minute termination, the numbers. Your carrier prices this, it changes by country and by direction, and it usually sits outside the vendor's per-minute rate. It is an easy line to leave out of the comparison.
  2. The speech stack. Recognition turns the caller's audio into text, a language model decides what to say, synthesis says it. Three separate costs, billed per second of audio or per token. Where a vendor rents all three from third-party APIs, that cost reaches you inside the rate, so it is worth asking which of the three are actually theirs. Recognition, synthesis and the model all run on our own servers, so call data stays in-region and does not depend on third-party APIs.
  3. The build. Gathering your context, writing and training the scenario, wiring the agent into your CRM and telephony, testing it before it takes a live call. One-time work, and worth confirming in writing whether it is inside the rate you were quoted.
  4. Keeping it right. Products, prices, objections and rules change. Someone has to update the scenario and check what the agent actually said on the line. If nobody owns this, the agent quietly gets worse and nothing in your bill tells you.
The advertised rate is a price for layer two. The bill is all four.

Per minute, per call, or per result

There are three honest shapes for this contract, and they move risk to different places. Per minute, you pay for talk time: easy to audit, and it bills duration rather than outcome. Per call, you pay a fixed amount per conversation whatever its length, which hands duration risk to the vendor, who then prices in a buffer for long calls. Per result, you pay per sale, per qualified lead, or per completed action, and the vendor carries the execution risk. Per minute is what we publish today; pay-for-results is our stated end state.

Whichever shape you sign, the number to compare is the same one, and it is not the rate.

The number that actually matters

Loaded cost per handled call = (all minutes billed x rate, plus any one-time build or maintenance a vendor bills you separately) divided by the number of calls the agent finished without a human. Everything else is an input to this. With us the bracketed extras are empty: billing is talk time only and call analytics is included, so every worked figure below is rate x minutes and nothing else.

Be clear-eyed about the divisor. A failed call still consumes minutes. Someone who hangs up forty seconds in is forty seconds of billed talk time and no result. A call the agent escalates gets paid for twice: once for the agent's minutes, then again for the person who finishes it. Per-minute pricing does not care which of those happened.

AI voice agent cost per minute at three volumes

Take Bene Hotline, our inbound reception and routing agent. Base rate EUR 0.20 a minute excluding VAT, the same nominal figure in USD. Volume discounts run 0, -5, -10, -15, -20 and -25% across the 25k, 50k, 100k, 200k, 400k and 1M minute tiers.

Two inputs are yours, not ours. Assume a three-minute average call and substitute your real figure. And assume the agent finishes 70% of calls without a person. We will not publish a single containment number, because it is set by your call mix and your scenario, not by our software. The 70% just makes the arithmetic visible.

25,000 minutes a month at EUR 0.200%
EUR 5,000 billed; assumed 3-minute call and 70% containment: 8,333 calls, 5,833 without a human
100,000 minutes a month at EUR 0.180%
EUR 18,000 billed; assumed 3-minute call and 70% containment: 33,333 calls, 23,333 without a human
400,000 minutes a month at EUR 0.160%
EUR 64,000 billed; assumed 3-minute call and 70% containment: 133,333 calls, 93,333 without a human

Read the middle row slowly, because it is the whole argument. The raw cost per call attempted is EUR 0.54, which is the number that flatters the quote. The cost per call actually finished is EUR 0.77, and the gap is not a fee, it is the assumed 30% of conversations that consumed minutes and produced a handoff.

Now hold that row still. Same 100,000 minutes, same EUR 0.18, and move only the share the agent finishes alone.

EUR 0.64
if 85% finish without a human (3-minute call assumed)
EUR 0.77
if 70% finish without a human (3-minute call assumed)
EUR 1.20
if 45% finish without a human (3-minute call assumed)

Going from 25,000 minutes to 400,000 minutes moves the rate by a factor of 1.25. Across the illustrative 85%-to-45% range above, containment moves the cost per handled call by a factor of about 1.9 without the rate changing at all. Which is why negotiating two cents off the rate is close to a waste of a meeting.

Negotiate the rate and you win pennies. Fix containment and you win the invoice.

What to make a quote answer in writing

Take any quote, ours included, and make the vendor answer these in writing. Where a vendor will not, that is your answer.

When per-minute pricing is the wrong deal for you

We sell per minute, so read this as a warning against us where it applies. There are two situations where a flat per-call or per-result price is genuinely better for the buyer, and no volume discount fixes either.

The first is when your calls are long by nature and each one is worth little. Compliance scripts your own industry requires you to read out, long verification reads, a base that likes to talk: if you cannot compress the duration and the value of the conversation does not grow with its length, per-minute pricing charges you for exactly the thing you do not benefit from. Ask for a per-call price and let the vendor own the duration.

The second is when the outcome is countable and you would rather not carry the execution risk. Under per-minute pricing, a scenario that escalates half its calls bills the same as one that closes them, and the difference lands entirely on you. If what you want can be counted, a sale, a qualified lead, a booked appointment, then ask to be priced on that. It is the harder deal for a vendor to sign, which is precisely why it tells you how confident they are.

Two conditions, stated plainly

Long calls worth little each, or a countable outcome you do not want to carry the risk on. In both cases a flat fee beats per minute, and you should say so.

Our rates in full

Base per-minute rates, excluding VAT. The same nominal figure applies in EUR and USD; that is regional parity, not a currency conversion. Latin America is priced at roughly half these rates.

Locator, AI speech analytics0%
Per minute, excluding VAT. Reviews and scores 100% of the calls your human team makes
Bene Hotline0%
Per minute, excluding VAT. Inbound reception and routing
Bene Qualifier0%
Per minute, excluding VAT. Lead qualification
Bene Sales0%
Per minute, excluding VAT. Full-cycle selling

Volume tiers are 25k, 50k, 100k, 200k, 400k and 1M minutes a month. The three agents discount 0, -5, -10, -15, -20 and -25%. Locator runs a steeper curve of its own: 0, -12.5, -25, -37.5, -50 and -62.5%.

Two things people expect and will not find. There is no subscription: billing is for talk time only, with call analytics included rather than sold on top. And the trial is priced 20% above base, not below, because trial volume sits below the first discount tier. If a trial is priced under the base rate, it is worth asking what was built into the base rate to make that possible.

If you want your own numbers rather than ours, the cost calculator is public, and it works from your call volume, your region and the product you are pricing; the pricing block on the homepage lists the rates. The lowest rate on that list is Locator at 0.10 a minute, on the calls your human team already makes: it reviews and scores 100% of them, where a manual QA team gets through 3-5%, which is our own published claim rather than an industry study.

Frequently asked questions

How much does an AI voice agent cost per minute?
Ours: Locator speech analytics at 0.10, Bene Hotline at 0.20, Bene Qualifier at 0.25, Bene Sales at 0.30, excluding VAT, the same nominal figure in EUR and USD. Volume discounts reach -25% for the agents and -62.5% for Locator across tiers from 25,000 to 1M minutes a month. Latin America is priced at roughly half those rates. A trial runs 20% above the base rate.
Is the per-minute rate the whole cost of a voice agent?
No. There are four layers: telephony carriage, the speech stack itself, the one-time build, and ongoing scenario maintenance. Then divide the total by the number of calls the agent finished without a human. Run it with your own inputs: at 100,000 minutes a month the Bene Hotline rate is EUR 0.18, so an assumed three-minute average call is EUR 0.54 of talk time, and if 70% of calls finish without a person the cost per handled call is EUR 0.77. Both the three minutes and the 70% are illustrative inputs you should replace with your own; containment is not a figure we publish, because it depends on your call mix and your scenario.
Do I pay for a call the agent fails to complete?
Under per-minute pricing, yes. Talk time is talk time. Someone who hangs up at forty seconds is billed for forty seconds, and a call the agent escalates is paid for twice, once for the agent and once for the person who finishes it. That is the arithmetic reason containment matters more than the headline rate.
Is there a subscription, or do I only pay for talk time?
No subscription. Billing is for talk time only, and call analytics is included rather than sold separately. The trial is priced 20% above the base rate.
When is per-minute pricing worse for me than a flat fee?
Two cases. When your calls are long by nature and each one is worth little, so you are paying for duration you cannot compress. And when the outcome is countable and you would rather the vendor carried the execution risk. In both, ask for a per-call or per-result price instead.
What is the cheapest way to find my real cost per call?
Work the formula with your own two inputs rather than the illustrative ones in this article: your average call length, and the share of calls finished without a person. For the rate side of it, the cost calculator is public and works from your call volume, your region and the product you are pricing. The cheapest rate we publish is Locator at 0.10 a minute on the calls your human team already makes: it reviews and scores 100% of them, where a manual QA team gets through 3-5%, which is our own published claim rather than an industry study.