What a voice AI agent actually costs — five meters, per minute

Five meters, not one

The vendor's own calculator puts 1,000 minutes at $82–$129 a month. The spread is almost entirely the language model: a cheap one lands at the bottom, a capable one near the top.

On top of that sits an optional support package. No package means pure usage billing; Core is $29 a month and Pro is 10% of hosting with a $999 monthly minimum. To start small, take no package and begin with the $5 of free credit.

Is 1,000 minutes a lot?

At three minutes a call, roughly 330 calls — about a month for a site taking fifteen calls a day. Against the cost of a person answering, it is dramatically cheaper. But that figure assumes the AI completes every call.

In practice some calls are handed to a human, and those minutes are still billed on the AI side, so cost does not fall while staff time is added. The decision rests on what share of calls can be completed end to end.

Which calls to automate

Suited

• Routine questions: opening hours, location, whether there is parking
• Taking and changing bookings, where availability can be read from a system
• First-line intake: capturing the request and routing it to a department
• Promising a call back, and recording that promise

Not suited

• First response to complaints, where reading emotional nuance matters
• Confirming amounts or contract terms, where being wrong is expensive
• Callers who need things repeated — the back-and-forth inflates minutes
• Noisy environments, where transcription accuracy drops

Forcing the unsuited categories through the agent lengthens calls, raises cost and lowers satisfaction at the same time. Start with routine enquiries only and transfer everything else immediately, then widen the scope while watching the transfer rate.

Non-English deployments

Transcription and synthesis quality varies far more by engine outside English. Evaluating in English and deploying in another language produces unpleasant surprises — run the evaluation on real recordings in the language you will actually serve.

Latency is the other issue. Transcription, inference and synthesis run in sequence, so there is more of a gap than in human conversation, and on a phone call half a second of silence feels long. Reducing it means choosing a faster model, which trades against capability.

The legal check

The architecture records and transcribes calls, so callers have to be told. An automated notice at the start of the call is the usual approach. Settle where recordings are stored, how long they are kept, and whether they can be used for training — the same questions that apply to meeting AI.

Frequently asked questions