AI Voice
The AI front desk that actually answers
Most clinics do not have a phone problem. They have a coverage problem that shows up on the phone. Here is how we scope voice agents for healthcare intake.
A clinic calls us because patients are complaining about hold times. Almost every time, the phone is the symptom. The real number is how many people hung up before anyone picked up — and how many of those never called back.
Start with the call log, not the demo
Before writing a line of code we pull four weeks of call data and sort it by intent. In most practices the top five intents cover well over seventy percent of volume: booking, rescheduling, prescription refills, billing questions, and directions. None of those need judgment. All of them need to happen right now.
That distribution is the scope. Everything else routes to a person, and it routes with the transcript attached so the person does not start from zero.
Latency is the product
A voice agent that answers correctly in four seconds feels broken. One that answers adequately in eight hundred milliseconds feels like a good receptionist. We budget end-to-end latency before choosing any model, then work backwards: speech recognition, retrieval, generation, synthesis. If the budget does not close, the scope shrinks — not the target.
Escalation is a feature, not a failure
The agent should hand off early and cleanly. Any hint of clinical judgment, any frustration in the caller's voice, any question the knowledge base does not cover: transfer, with context. The goal was never to eliminate the front desk. It was to stop the front desk from spending its whole morning reading back appointment slots.
What to do next
Pull your last month of call records and count the abandons. If that number surprises you, the conversation is worth having.