An Egyptian-Arabic AI receptionist for the calls clinics miss

The problem

Most private clinics in Egypt have one receptionist who also handles walk-ins, payments and files. At peak hours the phone rings out, and a patient who cannot get through calls the next doctor on the list. The clinic never learns about the booking it lost.

What MedKlaus does

The clinic keeps its number and its receptionist. Calls that go unanswered for about 15 seconds forward to MedKlaus, which answers in Egyptian Arabic.

• Books, reschedules and cancels appointments.
• Answers questions about fees, location and opening hours.
• Hands the call to staff when it is unsure, and tells callers with urgent symptoms to seek care immediately.
• Writes a structured summary of every call for the clinic: complaint, specialty, doctor, requested slot and outcome.

Engineering for Egyptian Arabic

The voice loop chains Azure AI Speech for recognition and an Egyptian neural voice, and Azure OpenAI for the conversation. We chose the stack by benchmark: seven voice engines on a 41-sentence Egyptian test set, and eleven architectures costed at clinic scale.

• Natural turn-taking. Callers can interrupt mid-sentence, and the reply starts about a second after they stop speaking.
• Local detail. The system reads phone numbers the way Egyptians say them, uses an Arabic pronunciation dictionary and adjusts its grammar to the caller's gender.
• Safe hand-over. Anything outside booking and clinic information goes to a person.

Cost and latency are product metrics

Every call carries a cost ledger priced at the rates in force that day, and an audit dashboard shows latency, tokens and cost for each turn of the conversation. Over ten turns of a test call, the median time from the caller falling silent to the first audio of the reply was 1.09 seconds.

Where it stands

MedKlaus is a working prototype: the full voice loop runs, with cost tracking on every call. Before the first patient call it needs a licensed telephone route, a legal opinion and data-protection licences for health data. We state these gates openly because clinics and regulators should know them.