The client had years of patient data and an intuition that they were losing people they should have kept. The task was to turn that intuition into a ranked list a finite team could actually work through on a Monday morning.

Defining churn correctly took longer than building the first model, and it mattered more. A dental patient does not cancel anything — they simply stop returning — so churn had to be expressed against recall behaviour, measured relative to each patient's own interval and history rather than a fixed window that would misclassify half the base.

The signals that carried weight were not always the obvious ones. Time since last visit relative to expected recall mattered, as anticipated, but so did provider continuity and patterns of broken and rescheduled appointments — patients who had lost their usual clinician were at materially higher risk than the raw interval suggested. Declined treatment and payment friction added further signal.

We deliberately ranked on expected value rather than probability alone. A patient at eighty per cent risk worth very little is a worse call than one at forty-five per cent worth a great deal, and the output had to be a worklist for a team with limited hours rather than an academic ranking.

Delivery into the workflow was the final requirement. Scores landed in the tools staff already used, each with the reason attached, so the person making the call knew why the patient was on the list. We monitor drift and retrain on a schedule — patient behaviour shifts, and a model left alone becomes confidently wrong, which is worse than having no model at all.