How Much Does an AI Voice Agent Cost in 2026?

An AI voice agent costs between roughly $0.06 and $0.20 per minute in platform and model fees, plus a build cost that usually lands between $8,000 and $60,000 depending on how many systems it has to reach. The per-minute number is the one vendors quote. The build cost is the one that decides whether the project is worth doing.

How much does an AI voice agent cost?

Three costs, and they behave differently:

• Per-minute running cost — $0.06 to $0.20. Speech-to-text, the language model, text-to-speech and telephony. Scales linearly with call volume, and falls as models get cheaper.
• Build cost — $8,000 to $60,000. One-time. Driven almost entirely by how many of your systems the agent must read from and write to.
• Running maintenance — 10–20% of build cost per year. Your policies change, your systems change, and the agent has to change with them.

A useful sanity check: at 2,000 calls a month averaging four minutes, per-minute costs land around $480–$1,600 a month. Against that, one full-time agent handling the same volume costs substantially more in most markets. The running cost is rarely what makes or breaks the case — the build cost is.

What actually drives the build cost

Vendors quote per-minute because it is small and comparable. It is also the part you have least reason to care about. Here is what the larger number is made of.

Number of systems (the dominant factor)

An agent that only answers questions from a script is cheap and nearly useless. One that reads your booking system, checks your CRM, writes back a call summary and triggers a confirmation email is doing four integrations, each with its own authentication, error handling and edge cases. Cost scales with integrations far more than with call volume.

Whether it can act, or only inform

Reading is straightforward. Writing — booking, cancelling, refunding, updating — needs validation, permission scoping and an approval path for anything irreversible. Expect the ability to act to roughly double the build against a read-only agent.

Verification requirements

If the agent discusses anything account-specific, it must establish who it is speaking to. Verification against a system of record is a workflow in itself, and in regulated contexts it is the longest part of the build.

Languages

A second language is not a setting. It is a second set of prompts, a second voice, a second evaluation set and a second round of failure discovery. Budget close to a further half of the original build per additional language you genuinely support.

Call complexity

"What time do you close" is one turn. "I need to move my booking, but only if the new date has availability and the price difference is under fifty pounds" is a negotiation with branching state. Handle time, model cost and build cost all rise with turn count.

What the per-minute price does not include

Read a quote carefully for these, because they are commonly outside it:

• Telephony. Numbers, inbound and outbound minutes, often billed separately.
• Failed and abandoned calls. Usually still metered.
• Transfers to humans. The AI minutes before the transfer are charged, and the human minutes after are your existing cost — a transferred call costs more than either alone.
• Silence and hold. Some platforms meter wall-clock, not speech.
• Evaluation runs. Testing against recorded calls consumes the same model budget.

When a voice agent does not pay

Worth saying plainly, because the answer is sometimes no:

• Under about 500 calls a month. The build cost has too little volume to amortise against. Fix routing and hours first.
• Consultative calls. If the value of the call is the judgement of the person taking it, automating it removes the product.
• No system to read. An agent without live data is an expensive phone tree.
• Highly variable calls. If no three calls resemble each other, there is no pattern to automate and you will spend the build discovering that.

A realistic first project

The pilots that work are narrow: one call type, one language, read-only or single-action, with a hard rule that anything outside the pattern transfers to a person. That is typically $8,000–$20,000, runs six to eight weeks, and produces a real number — deflection rate, containment rate, cost per resolved call — that tells you whether the wider programme is worth funding.

The projects that fail are the opposite: every call type, three languages, full booking authority, no measured baseline. They fail on integration and edge cases, not on the AI.

How to compare quotes

Ask every vendor the same five questions and the range narrows quickly:

1. What is the all-in cost per minute including telephony, and what is metered — speech or wall-clock?
2. Which of my systems will it read from and write to, and is that in the quoted build?
3. What happens on a transfer to a human, and what does the customer experience at that moment?
4. What is the confidence threshold, and what does the agent do below it?
5. What does year two cost if nothing changes, and what if my policies change?

The fifth question separates a platform subscription from an engineering relationship, and it is where most of the disappointment in this category comes from.

Frequently asked questions

Is an AI voice agent cheaper than a call centre?

Per handled call, usually yes at volume, once the build is amortised. But the comparison is rarely like for like: an agent handles the repeatable calls well and the unusual ones badly, so what you are really comparing is a smaller human team plus an agent against a larger human team. Model it that way and the numbers are more honest.

How long until it pays for itself?

At meaningful volume, typically six to fifteen months from go-live. Below about 500 calls a month, often never. The build cost is fixed and the saving is per call, so payback is almost entirely a volume question.

Do customers accept talking to an AI?

They accept it when it resolves the call faster than waiting on hold, and resent it when it traps them. Acceptance correlates with how quickly and gracefully the system hands to a person, not with how human it sounds.

What about accents and noisy lines?

Recognition has improved considerably but is not uniform across accents, and a poor line still degrades it. Design for it: confirm critical details back, and treat repeated recognition failure as a transfer trigger rather than something to push through.

Can it make outbound calls?

Technically yes — collecting missing information, confirming appointments, chasing documents. Legally, outbound automated calling is regulated differently in every market, and consent requirements vary. Check the rules for the country you are calling into before you build the campaign, not after.