Every AI voice platform prices the same way: cents per minute, big friendly number, calculator widget on the pricing page. And every team that takes one to production discovers the same thing — the per-minute meter is the smallest line in the real budget.
I run an AI voice agent in production: a cold-calling agent that dials newly registered FMCSA broker authorities for a US carrier’s sales team, handles objections, and drives to a booked demo. This is the cost structure I wish someone had written down before I built it.
What the meter is actually metering
A voice agent is four metered services in a trench coat:
- Telephony — the actual phone call (Twilio, Telnyx, or a platform’s bundled minutes)
- Speech-to-text — transcribing the human in real time
- The LLM — deciding what to say next, every conversational turn
- Text-to-speech — the voice itself, which ranges from cheap-and-robotic to expensive-and-eerily-good
You can buy them bundled (a managed voice-AI platform) or assemble them yourself. As of August 2026, the market prices out roughly like this:
Two honest notes on that chart. First, the assembled stack’s discount is not free — you are now the one wiring latency, interruption handling, and failover, which is real engineering. Second, model choice dominates: a premium ultra-natural voice can cost more per minute than the LLM and telephony combined, and for B2B cold calls a mid-tier voice converts essentially the same in my experience.
The napkin math for a cold-calling agent
Usage costs scale with talk time, and most dials never become talk time. The formula:
Monthly usage ≈ dials × connect rate × avg talk minutes × cost per minute (plus a small per-dial telephony charge for the calls nobody answers)
Worked example at the volumes a small sales operation actually runs — 2,000 dials a month, a 25% connect rate, and an average conversation around 1.5 minutes:
| Managed platform (~$0.15/min) | Assembled stack (~$0.06/min) | |
|---|---|---|
| Conversations | 500 | 500 |
| Talk minutes | ~750 | ~750 |
| Talk cost | ~$113 | ~$45 |
| Unanswered-dial telephony | ~$20 | ~$20 |
| Usage total | ~$135/mo | ~$65/mo |
Sit with that for a second: five hundred cold conversations a month for the price of a team lunch. This is why “the calls are too expensive” is never the real objection to AI calling. The real costs live elsewhere.
The budget nobody puts on the pricing page
Where those non-obvious lines come from:
Numbers and registration. US carriers now aggressively filter unregistered outbound traffic. Real campaigns need registered numbers, sensible rotation, and monitoring for “Spam Likely” labeling — a small but permanent line item, and a project the first time you do it.
Data and suppression. The agent is only as good as its list. My agent runs on a daily feed of newly registered FMCSA authorities; yours will have its own source, plus DNC scrubbing against the federal registry and your internal suppression list. The compliance side deserves its own article — I wrote it here — because getting it wrong is priced per call, by statute.
Human attention. This is the line that decides success, and it never appears on a platform pricing page. Someone has to listen to a sample of calls every week, notice that the agent fumbles a specific objection, fix the prompt or flow, and verify the fix. Someone has to handle the edge cases the agent escalates. Budget real hours for it — roughly eight a month in my example, more in the first two months, and staff it with someone who has authority to change the script. An unwatched voice agent does not stay good; it drifts, and it drifts in front of your prospects.
So when does it pencil out?
Compare against the human alternative honestly. A part-time SDR making the same 2,000 dials costs thousands of dollars a month, works one time zone, and quits. The agent’s fully loaded ~$1,000/month wins that comparison easily — if the calls are the kind an agent can do.
It pencils out when:
- The call is high-volume, repeatable, and survivable when imperfect — top-of-funnel qualification, appointment setting, check-ins
- The list refreshes itself (new registrations, inbound leads, renewals) so the agent never starves
- There is a clean handoff — a booked demo, a captured email, a transfer to a human who closes
It does not pencil out when volume is low (a human is cheaper than the setup cost), when every prospect is a named account where one bad call burns real money, or when nobody on your team will own the weekly review. That last one kills more deployments than any technical limit.
The build-versus-platform choice follows the same logic as build-versus-buy for a TMS: platforms to validate, assembled stack when volume makes the per-minute delta pay for the engineering.
If you are trying to work out whether a voice agent makes sense for your operation — or you have one that demos well and drifts in production — that is a conversation I do for a living, and the case study of mine that runs daily is the proof I bring to it.