Answering the calls that were going to voicemail at peak hours

Reception staff at nine practices were handling calls alongside patients standing in front of them. At peak — Monday mornings, and the hour after each practice opened — a large share of calls went unanswered. Call-back logs showed most callers never tried again; they called another practice.

The overwhelming majority of those calls were routine: booking, rescheduling, opening hours, whether a particular practice took a particular insurance plan. High volume, low complexity, and consistently displacing the patient at the desk.

A previous phone-menu system had been removed after patient complaints. The group's position going in was firm: anything that felt like a menu tree, or that made a patient repeat themselves to a human afterwards, would not be deployed.

We scoped hard. The agent handles four intents — book, reschedule, cancel, answer a practice-information question — and everything else escalates immediately. Refusing to broaden that list was the most important decision in the project, and we pushed back on it twice during delivery.

Clinical questions escalate unconditionally. Not by classifier confidence, not by threshold: any utterance touching symptoms, pain, medication or treatment advice routes to a human, and the agent says so plainly. In healthcare the asymmetry of a wrong answer makes this the only defensible design.

Latency was treated as a product requirement, not an optimisation. Past a certain response delay, callers start talking over the agent and the conversation degrades badly. We budgeted the speech pipeline end to end and used a smaller, faster model for turn-taking decisions, keeping the larger one for the substantive turns.

The rate of unanswered calls at peak dropped substantially once the agent could pick up the routine bookings, rescheduling and information requests that used to go to voicemail, with genuine overflow still routing to the human queue.

Slot fill from inbound calls improved, largely by recovering callers who previously hung up and rang a competitor — that effect was most of the business case.

Reception staff, initially and reasonably wary, became the strongest advocates within a month — not because the agent was impressive, but because Monday mornings stopped being unmanageable.