Would a local AI phone receptionist pay off for a clinic?
For small digital businesses that build or resell AI automations. Put in a client's numbers and your own costs, see whether a receptionist that answers missed calls from a box in the client's office pays for itself, then copy a starter build prompt. About two minutes; every default is an example to replace.
The stack (Sherpa-ONNX speech-to-text, SetFit intent routing, Kokoro text-to-speech) follows the open-source axiom-voice-agent project. Its speed claims are the developer's own, listed under Sources.
EXAMPLE ASSUMPTIONS · CLIENT SIDE: replace with the client's real numbers
EXAMPLE ASSUMPTIONS · YOUR OFFER: assumed pricing, not a market rate
EXAMPLE ASSUMPTIONS · COST RATES: replace with your own quotes
EXAMPLE RESULTS · computed from the assumptions on this page. Illustrative, not a forecast or a measured result.
How the example math works: recovered revenue = weekly calls × 4.33 weeks × booking share × booking value. Client profit = recovered revenue × the client's margin − your monthly fee. Payback = setup fee ÷ daily client profit. Your margin = (monthly fee − box running cost) ÷ monthly fee.
What this leaves out: callers who would have called back anyway, calls the system handles wrong, the box itself (hardware), your setup and support time, taxes and churn. Count the client's real missed calls from their phone logs before you pitch.
Starter build prompt (paste into Claude, Cursor or a similar tool)
Sources & what we checked
- axiom-voice-agent (GitHub repo): the repo description (the GitHub About box) claims "<400ms latency" on 4 GB VRAM and "optimized for GTX 1650"; the README prerequisites list 2–3.6 GB VRAM for GPU acceleration. These are the developer's claims, not our measurements, and the README text gives no per-component millisecond figures.
- Kokoro-82M model card: the usage examples write audio at 24 kHz (24000 Hz).
- ITU-T G.711: telephony PCM is sampled at 8000 samples per second (8 kHz). Kokoro's 24 kHz output must be resampled to 8 kHz before it goes down a G.711 SIP leg.
- Agent-R2 (arXiv 2611.08219, Zhejiang University and Baidu): Agent-R2-4B scores 51.0% on GDPval vs 48.0% for Qwen3-235B-A22B-Thinking-2507, and 14.8% vs 13.8% on OfficeQA. Some frontier models score higher on GDPval in the paper's own table (for example Claude-Opus-4.5 at 72.6%).
- Everything else on this page (call volumes, conversion, tickets, offers, per-minute rates, power) is an example assumption. No vendor's pricing is quoted.