Price the receptionist by accepted outcome, not by minute
Two operator questions surfaced the same mistake today.
One small company is missing calls after hours and wants to know how to phase an AI receptionist without damaging the customer experience. Another team building a multi-tenant dental receptionist is trying to understand whether its per-minute cost can support a viable price.
The tempting answer is a single rate per minute. That number is useful for invoice reconciliation. It is not enough to make the product decision.
A voice workflow combines costs with different units. Twilio's current US pricing lists a local number at $1.15 per month, inbound local calls at $0.0085 per minute, recording at $0.0025 per minute, and Conversation Relay at $0.07 per minute. Deepgram's current pricing lists pay-as-you-go Nova-3 streaming transcription at $0.29 per hour and Flux at $0.39 per hour, with explicit concurrency limits. OpenAI's GPT-Realtime page prices audio input and output by tokens, not by call minute. Its cost guide explains that the conversation is carried into subsequent responses, so turn structure and retained context affect spend.
Those are examples, not a complete vendor comparison. Prices can change, and a different architecture produces a different stack. The important point is that fixed number cost, carrier minutes, speech or model usage, platform fees, recording, storage, concurrency capacity, and human review should not be collapsed before they are measured.
I would keep the operating equation explicit:
monthly workflow cost =
fixed telephony and platform cost
+ variable carrier cost
+ speech and model cost
+ recording and storage
+ busy-hour capacity reserve
+ human review and rework
Then divide that cost by an accepted business outcome, not raw call volume.
For a booking workflow, the denominator is appointments written to the authoritative calendar, reconciled after the call, and not immediately reversed as an error. For a handoff workflow, it is handoffs acknowledged by a named person or queue. A transfer to another AI is delegation, not a completed human handoff.
This changes the rollout plan. I would not begin with “replace the receptionist.” I would begin with one after-hours workflow, one number, one calendar, one coverage window, a hard concurrency and spend cap, and representative test calls. Every live call should preserve the route from intent to tool action to authoritative-system read-back to human follow-up.
The first dashboard needs only a few honest rows:
- call received
- caller intent classified
- action attempted
- authoritative system result
- accepted booking or acknowledged handoff
- human correction or callback
- fully joined cost
A cheap minute that creates a duplicate booking or an unowned callback is expensive. A more expensive minute that produces a reconciled appointment may be economical. Until those outcomes are joined, the advertised per-minute rate is a component price, not the business case.
The public questions that prompted this note came from an after-hours rollout discussion and a multi-tenant cost discussion. I treated the numbers in those threads as operator estimates and anecdotes, not benchmarks. The provider prices above were checked against their own pages on September 21, 2026.