Stopping the calls we did not need was the proof. The model was the point. The three intent tiers had done their job, but they were built for a smaller operation than the one we had become. Two questions sorting every lead into three buckets was blunt where the scale now needed precision. A decile model scores each lead on its own expected conversion, so call strategy, nurture and media spend can follow the evidence rather than the bucket it fell into.
I ran alignment sessions with six stakeholders across admissions, marketing, paid media and data engineering, synthesised where they converged, and took the model to senior leaders, who endorsed it. Then the groundwork: I audited the signals we actually had, wrote the analysis query myself, around 600k leads with point-in-time features, and set an interpretability requirement so the team could trust the scores.
The model is designed, documented end to end, and lands in the next couple of months. When it does, the share of call effort we can safely cut should climb from the roughly 27% the tiers reached toward half. The whole workstream is written to run without me as the dependency.