What AI voice agents actually cost to run

Nobody can quote you a single number, and anyone who does is either bundling heavily or guessing. What we can do is name every component so you can price it against your own call volume.

We are deliberately not printing per-minute rates here. They change frequently across every vendor involved, and a stale number in a blog post is worse than no number — check the live pricing pages linked below.

The four metered components

1. Telephony. The actual phone call — inbound or outbound minutes, plus a monthly cost per number. Billed per minute, scales linearly with talk time. See Twilio’s voice pricing for the shape of this.

2. Speech-to-text. Converting the caller’s audio into text, billed per minute of audio. Also linear with talk time.

3. The language model. Generating the agent’s replies, billed per token. This is the component people expect to dominate and it usually does not — a two-minute call is a small number of tokens unless the agent is loaded with a very large context on every turn.

4. Text-to-speech. Converting the reply back to audio, billed per character or per second. Higher-quality voices cost more, and the difference is audible.

Plus: the platform. Vapi, Retell and similar orchestrate all four and add their own fee — sometimes a per-minute margin, sometimes a subscription, sometimes both. Their pricing pages (Vapi, Retell) are the authority.

What actually drives the bill

Total talk time, more than call count. Three components are per-minute. A 90-second qualification call and a nine-minute meandering one differ by a factor of six on most of the bill.

Whether the agent is concise. This is a prompt-design problem with a direct cost consequence. An agent that asks one clear question and waits is cheaper and better than one that narrates.

Silence handling. Agents that keep the line open through long pauses bill for that silence. Endpointing configuration is a real cost lever.

Failed and abandoned calls. They still cost telephony and often speech. A high abandon rate is paying for calls that produce nothing.

Retries on outbound. Outbound campaigns multiply everything by the retry policy. Three attempts per contact is three times the cost per contact, not per connection.

Where the cost is not

Build cost is one-off; running cost is forever. Businesses routinely negotiate hard on the build and never look at the running cost again, which is backwards for anything past low volume.

Integration is usually the bigger line. The agent talking is the easy part. The agent reading a live calendar, writing to the right CRM fields, routing correctly and handing over cleanly is where the engineering time goes — and where the value is.

The number worth calculating instead

Cost per booked appointment, not cost per minute.

An agent that costs more per minute but qualifies properly and books reliably can be far cheaper per outcome than a cheap one that produces transcripts nobody acts on.

To calculate it you need the agent writing structured outcomes into the CRM, which is another reason the integration work matters more than the voice.

A cheaper option that is often better

If you get a handful of calls a week, a voice agent is the wrong tool. Missed-call text-back captures the same enquiries for a fraction of the cost and complexity. We covered the ladder of options in stopping lost leads after hours.

Volume is what makes a voice agent make sense. Below it, simpler automation wins on every axis.