Overflow — AI Receptionist

The whole product fits in one call record. It discloses what it is, it recognises a medical concern, and it stops — rather than trying to help.

The business keeps its own phone system. Overflow is a forwarding rule inside it. The staff line rings first, for a set number of seconds. If a person answers, the agent never joins the call at all. If nobody does, the call forwards and the agent picks up, recording which condition caused it.

In production the phone system has usually already decided why it is forwarding and passes the reason as call metadata; the router derives it itself only when it owns the ring leg. If the agent’s leg is unreachable, the call falls through to voicemail rather than dead air.

Most AI phone agents handle escalation inside the prompt — transfer the call if the caller seems upset. That is a suggestion to a model. A persistent caller can talk it out of firing, and it behaves differently from one call to the next.

Here the escalation layer runs before the model is asked for anything. Five rules, evaluated in order, first match wins. A second check runs against the generated reply and can also hand off. The call transfers immediately, with no attempt to resolve, when the caller asks for a person in any wording, mentions a medical concern, disputes a charge, is misunderstood twice in a row, or the call passes its time cap.

If no human is free to take the transfer, the agent takes a callback request and states when someone will call back. It does not leave a caller in a loop.

The failure mode that matters here is an agent confirming a booking that never happened. If the calendar provider times out or errors mid-call, this agent says a person will confirm, and records a callback.

That is structural rather than a matter of careful wording. A successful-booking result is unconstructable without a reference returned by the provider, and the deferred path scans its own outgoing sentence against 35 confirmation phrases before speaking it. Booking writes are never retried, because a retried write that already landed is a double booking.

Vapi handles speech-to-text and text-to-speech only. It is configured as a custom LLM pointed at a FastAPI orchestrator, so no reply is generated inside the voice platform. Every sentence the caller hears comes out of application code, after the escalation rules have already run on that turn.

That arrangement is what makes the behaviour inspectable. When the agent does something wrong, there is a specific line that decided.

Prices, hours, services, policies and the agent’s answers are edited through the console by staff. Escalation detection logic is read-only there, because it is code rather than configuration.

Every account — voice platform, telephony, the phone number, storage — would be registered to the business. The design assumes I am a collaborator who can be removed without the system losing access, data or function.