What does it cost to run an AI calling agent per minute?

Four services bill for every connected minute: carrier telephony, speech-to-text on the incoming audio, the language model per token of conversation, and speech synthesis per character of reply. Telephony and speech synthesis are usually the largest two; the language model is generally the smallest, which is the opposite of most people’s assumption.

Is per-minute or monthly pricing better for an AI receptionist?

Per-minute is cheaper if your call volume is low and steady, because you pay only for what you use. A monthly plan sized by volume is better if your calls are seasonal or spiky, because it moves that variance onto the vendor. The thing to check on any monthly plan is what happens at the ceiling — an overage rate reintroduces per-minute billing above a line you cannot see.

Why is speech synthesis expensive?

It is billed per character and the voices people find acceptable on a phone call are the premium tier, priced several times above the basic ones. Since every reply the system makes has to be synthesised, it scales directly with how much the agent talks — which is one reason a well-designed agent is concise rather than chatty.

How many minutes does a small business actually use?

Multiply your daily call count by average call length. A business taking twenty calls a day at around two and a half minutes each uses roughly 1,500 minutes a month. Most owners estimate high, because the calls they remember are the long difficult ones rather than the thirty-second bookings that make up the bulk.

What an AI calling agent actually costs to run

30 August 2026

Vendors in this category price either per minute or as a monthly plan, and the choice is not arbitrary — it follows from what a minute of call actually costs to produce. Four separate services bill for that minute, all of them by usage, none of them by customer. Understanding the stack is the fastest way to work out whether a quote you have been given is reasonable, and why the cheap-looking option is often the expensive one.

The four meters

Every minute a call is connected, four things are being paid for at once.

Telephony. The carrier minute: the call arriving on a number, and the leg between the carrier and wherever the software runs. Billed per minute, in both directions.
Speech to text. Audio going in, transcribed continuously while the caller is still talking. Billed per minute of audio.
The language model. Every turn of the conversation sends the history plus the business’s own details and gets back a reply. Billed per token, so it scales with how long the call is and how much context the business has.
Speech synthesis. Turning the reply into audio. Billed per character, and the higher-quality voices cost several times the basic ones.

Why the model is not the expensive part

The intuition is that the AI is the cost and everything else is plumbing. It is usually the other way round.

Text generation has become extremely cheap per token, and a phone conversation is short: a three-minute call is a few thousand tokens of dialogue. What is not cheap is the audio on both ends — transcription runs continuously for the whole call whether anyone is speaking or not, and good synthesis is priced like a premium service because it is one.

Telephony is the floor nobody can optimise away. It is a regulated, physical service with a per-minute rate that does not fall because your software got better. For a lot of deployments it is the single largest line.

The costs that do not appear on a per-minute sheet

Three more, and they are the ones that decide whether a vendor is sustainable at the price they quoted you.

Storage and retention. Recordings and transcripts have to be kept, which is cheap per call and not cheap across a year of them. A vendor with no stated retention policy has either not thought about this or is planning to delete your recordings when the bill arrives.

Where the servers are. Every hop between India and a US-hosted model costs real round-trip time, and the fix — running inference in-region — costs more than the cheapest available endpoint. A vendor optimising purely for cost per minute is making a latency decision on your behalf.

The engineering that never ends. Speech models change, carriers change their behaviour, and a system that worked in March fails differently in September. This is a staffing cost dressed up as a software cost, and it is why the market is full of demos and short of deployments.

What this means for how you are billed

Per-minute pricing passes all four meters straight through to you. It looks cheap on a quiet month and is genuinely cheaper if your call volume is low and flat. It also means your bill moves with your busiest week, which is exactly the week you were not thinking about your phone bill.

A monthly plan sized by call volume moves that variance onto the vendor. You pay a predictable number and they carry the risk of a busy month. What you should check is what happens at the ceiling: a plan with an overage rate has quietly reintroduced per-minute billing above a line you cannot see from the pricing page.

The question worth asking any vendor is simply what happens when you exceed the plan. There are three honest answers — it stops, it throttles, or it bills you — and only one of them is a surprise on an invoice.

Doing the arithmetic on your own use

You do not need vendor rate cards for this. Take your average call length, multiply by calls per month, and you have minutes. That single number decides which pricing model is cheaper for you, and it is usually a much smaller number than owners expect — a business taking twenty calls a day at two and a half minutes each is using around 1,500 minutes a month.

Then compare against the thing being replaced rather than against zero. The relevant comparison is not "this costs money and voicemail is free", because voicemail is not free; it costs whatever share of those callers do not ring back.