The quoted number is about a third of the real one
Per-minute inference is the cheapest part. Telephony, transcription, calendar and CRM integration, monitoring and the human review loop together cost more than the model.
Budget roughly three times the quoted per-minute rate for the first six months.
Where the money actually goes
Telephony and carrier fees are fixed and predictable. Model inference scales with call length, which you control through flow design — a tight qualification script keeps average handle time under ninety seconds.
The variable cost that surprises people is quality review. For the first two months someone reads transcripts weekly and tunes prompts. Skip it and accuracy drifts.
When it pays back
The break-even is almost always missed calls. If more than 15% of inbound calls go unanswered, a voice agent pays for itself on recovered volume alone, before any efficiency gain.
Below that threshold, the case rests on speed to lead and agent hours recovered, which are real but slower to bank.
What we would not automate
Anything where a wrong answer is expensive and hard to reverse. Voice agents qualify and book. They should not be quoting bespoke pricing or giving regulated advice.