The problem
A two-chair dental clinic in Cairo misses most calls between 6 and 10pm — exactly when working patients call to book. Hiring a full-time receptionist costs more than the missed bookings feel worth, so the phone just rings out and the patient books the clinic across the street.
The build
Two people assemble it, they don’t invent it: a SIP number for the clinic’s published phone line, a WhatsApp Business number for voice notes, a speech-to-speech agent primed with the doctor’s slots and price list, and a shared calendar as the source of truth. The agent answers in Egyptian Arabic, negotiates a time, sends a WhatsApp confirmation, and flags uncertain bookings for the doctor to glance at each morning. Nothing is trained — it is prompting, call routing, and calendar glue.
Why now
Three years ago this needed a custom Arabic-dialect ASR pipeline, a dialogue manager, and a TTS voice nobody had — a research project. The Realtime API’s public beta (Oct 2024, GA Aug 2025) collapsed that into one API call, and Azure now ships Egyptian Arabic neural voices (ar-EG) off the shelf, so the dialect half is bought, not built.
First customer
Private dental and derma clinics in Greater Cairo: visible missed-call pain, a named payer (the doctor-owner), and a fee smaller than one receptionist salary. Offer: we answer every call for a flat monthly fee, starting with one pilot clinic that prepays month one.
Challenges
Dialect transcription of proper names under street noise will be wrong sometimes — the confirmation loop is the product, not a detail. Latency on real mobile calls can make the agent feel drunk; the pilot must be measured on real lines, not demos. And trust: one double-booked root canal undoes ten perfect weeks, so every booking stays human-glanceable until the error rate earns autonomy.