blog · 13 August 2026 · 6 minutes

How much does an AI voice agent cost — and when you do not need one

We have built voice agents for two companies — a financial services firm and a logistics operator. In both, before the rollout, operators were spending most of their shift on the same handful of questions. Here is how it works on their numbers, what it costs, and most importantly, when you do not need an agent at all.

How an AI agent differs from "press 1"

Everyone knows the phone robot, and its reputation is deservedly bad: "press 1 for your balance", recorded phrases, one step off the path and it is lost while the customer fumes and waits for an operator.

An AI agent is built differently. It understands ordinary, natural speech: the customer talks the way they normally do, interrupts, asks a question in their own words. The agent answers according to your company's policies — the same ones your operators work from — and works out for itself when a question falls outside them and it is time to hand the call to a person. And it hands it over with the context, so the customer does not have to repeat everything.

What the agent does in practice

AI voice agent cost per month, in numbers

Take a typical call centre of three operators. At a salary of $570 each that is around $1,700 a month — every month, plus holidays, sick leave and the cost of finding a replacement when someone leaves.

An AI agent: a one-off build from $1,400, then running costs — at a load of roughly 100 calls a day, about $450–570 a month. The build pays for itself in the first months, and after that the agent costs three times less than a team of operators. Even if it only takes the routine calls and leaves the complex ones to people, over a year the saving runs into thousands of dollars.

An important caveat: these numbers are honest when there is a real volume of calls. What happens when there is not is the next section.

When you do not need a voice agent

It is in our interest to sell agents, but there are three situations where we talk clients out of it ourselves:

• Too few calls. If inbound volume is under 20–30 a day, the economics do not work: running the agent costs more than the time it saves. A messenger bot or a decent answering service is cheaper.
• Every call is a complex consultation. Legal fine print, medical questions, individually negotiated terms — where the cost of an error is high, the agent should at most capture a contact and pass it to a specialist. Building full consultations on AI in those areas is premature.
• Sales rest on personal relationships. If customers have worked with "their account manager" for years, an agent picking up the phone will irritate them. Automate the routine around it — confirmations, reminders, status updates — and leave the live conversation to people.

What a rollout looks like

Based on two projects, the process takes 3–4 weeks: we work through the call scenarios and policies — what the agent takes on and what stays with people; we fix price and timeline in the technical spec; then we work in stages, and each stage ends with a demo of the live agent that you can actually phone and talk to. After handover we fix things free for 90 days.

And one principle we consider mandatory: the agent does not pass itself off as a human. It introduces itself as an automated assistant. This is a matter of trust in your brand — the pretence gets uncovered within seconds and damages the impression far more than talking to a machine ever would.

What to ask a vendor

• Can I phone a demo agent before signing the contract?
• What materials will the agent answer from, and how are they updated?
• What happens during an outage: where do calls go, who finds out first?
• How quickly does a call transfer to a person, and what does the operator see?
• What makes up the monthly cost: minutes, channels, support?
• Where are call recordings stored and who has access to them?