Turning a clinic's phone calls into revenue

A busy dental or medical front desk takes a phone call every few minutes while also checking in patients, taking payments, and calming the waiting room. Every call carries a small promise: we'll call you back, we'll reschedule you, we'll check with the doctor. The promises live in heads and on sticky notes.

The cruel part is invisibility. A practice owner can count no-shows; they cannot count the patient who rang twice, got voicemail, and booked with the clinic across the road. When I sat with the escalation data, missed and mishandled calls were the quiet thread behind an outsized share of churn conversations — practices didn't say 'your software failed', they said 'we're not growing'.

The naive product is a transcription feature: record calls, show text. It demos well and changes nothing — nobody at a front desk has time to read yesterday's phone calls. The unit of value isn't the transcript; it's the action the call should have created.

So the spec inverted: never show the recording first. Listen to every call, decide whether it contains an unkept promise or an unbooked opportunity, and surface only those — as tasks with a draft next step attached. The transcript becomes evidence you can expand, not homework.

Speaker attribution mattered more than I expected. A clinic's calls involve several staff members, and 'who said what' changes both the coaching signal and who the task belongs to. Voice printing each team member — with the team-level settings and assignment logic that implies — is what let the product move from 'this call went badly' to 'this call went badly and here is whose coaching plan it belongs on'.

Then the surface widened in the direction the money was: an unbooked call isn't just a missed task, it's an **opportunity** with a lifecycle — created, staged, worked, converted or lost. Modelling that explicitly turned a support tool into a revenue instrument, and gave practice owners the number they actually repeat to other owners.

Coaching came last and deliberately so. Once calls are transcribed, attributed and scored against call-type scorecards, you can show a practice which behaviours correlate with booked appointments. But scoring people is a trust minefield — it had to arrive after the product had already earned its keep, and it had to grade the call, not the person.

The hard product questions were about trust, not models. What precision does 'the AI says this call needs a callback' have to hit before a stressed receptionist stops double-checking every one? What happens to the task when the patient calls back on their own? Who is allowed to mark a promise kept?

What shipped is a loop, not a feature: calls stream in, the AI transcribes and classifies them, promises and opportunities become cards in an action queue, each card carries a one-tap draft — call back, text, book — and the front desk clears the queue between patients the way they'd clear any inbox.

The interactive mock below rebuilds the experience with invented data. Click a call to see what the AI extracted and what it drafted; toggle to see the same morning without the system.

I came to this product through its angriest users — I ran the escalation queue before I ran the roadmap. That order was a gift: by the time I owned the backlog I had hundreds of hours of listening to what practices actually lose when a call goes wrong. The product decisions that worked were all translations of that listening; the ones that didn't were all clever ideas from planning documents.