The situation
A privately-owned group dental practice in the US Midwest had grown to four doctors and fourteen support staff over a decade, acquired a smaller practice two years prior, and absorbed that practice's patient database. The combined active patient file held about 6,200 names. The daily recall report showed 1,800 of those as 'overdue' by more than 14 months.
The practice had a single treatment coordinator handling recall. She was good at it, but there were not enough business hours in the week to make the calls, leave the voicemails, wait for callbacks, and still process the normal intake of new patients and insurance verifications. So the recall list just sat there. Management had accepted that most of those patients were gone.
What we found on a Nano-Pilot was more interesting than that. The lapsed list was not one population. It was at least four, patients who had moved, patients whose insurance had changed and who believed they could not return, patients who had a bad cleaning experience with a specific hygienist who no longer worked there, and patients who simply drifted. The coordinator, when we shadowed her, was intuitively treating all four groups the same way: a friendly voicemail about cleaning schedules.
What we did
We built a workflow that pulled the lapsed list out of the practice management system weekly, classified each patient across the four behavioral categories using their visit history and notes, and drafted a tailored outreach sequence per category. SMS-first with an email fallback, no voicemails in the first pass, explicit unsubscribe honored permanently.
For the 'insurance changed' cohort, the message led with 'we accept most major plans, send us a photo of your current card and we'll tell you your coverage before you book.' For the 'bad experience' cohort, we added a note that the hygienist in question was no longer at the practice. For the 'drifted' cohort, we kept it light, a reminder and a direct-booking link that pre-filled their patient record.
The coordinator approved every message template in writing before it went out. We did not automate the writing of outbound messages; we automated the classification and sequencing. Replies came into a shared inbox that she handled personally. The AI never pretended to be her.
What it was worth
In the first 90 days, 412 appointments were booked from the lapsed list. Of those, 287 showed up, 44 cancelled and rebooked, and the rest were no-shows, roughly in line with the practice's normal new-patient conversion rates. At an average net-of-hygiene revenue per completed appointment of about $180, the first quarter produced just over $74,000 in directly-attributable revenue.
The coordinator spent about eight hours per week reviewing AI classifications and personally handling replies, down from the roughly twelve she had been spending on manual recall calls with far worse conversion. Her total workload went down, her conversion went up, and the four behavioral cohorts kept being processed weekly without backlog buildup.
The practice ran the Implementation Sprint with us, then rolled into a three-month Embedded AI Ops arrangement to extend the approach to new-patient onboarding and insurance verification. They paused the retainer at month four because the backlog was cleared and the maintenance burden was too small to justify the cost, which we told them to do.
What we'd tell another business in this industry
- Classify before you automate. A single 'lapsed patient' outreach sequence was the wrong tool for a list that contained four behaviorally distinct groups. The AI did not write the messages; it just sorted the patients.
- Let the human stay in the loop for replies. Patients noticed and mentioned on review forms that the practice 'remembered them,' which was only true because the coordinator was personally handling every reply. Full automation would have killed the warmth.
- If the retainer should pause, say so. We told them to pause Embedded AI Ops when it stopped paying off, and they came back six months later for a much bigger project. Trust compounds.