The situation
The desk already had a forecasting dashboard. It had been delivered a year earlier, demoed well, and then quietly ignored: schedulers kept planning dispatch from experience and a spreadsheet of rules of thumb. The dashboard was a browser tab nobody opened — a familiar ending for analytics projects that live outside the workflow they are supposed to inform.
The problem underneath the problem
The old forecast failed for two human reasons, not one technical one. It lived in the wrong place — a separate tool the desk had to remember to consult mid-decision. And it overclaimed: a single confident number with no error range, which the desk learned to distrust the first time it was badly wrong. Accuracy alone was never going to fix either.
What we built
The forecast was rebuilt as a component of the scheduling tools the desk already worked in — visible at the moment a dispatch decision is made, not in another tab. Every number ships with its error range displayed honestly, and every forecast is scored against what actually happened, in the open, so trust is earned from a track record rather than asserted in a demo.
The hard part
Earning back trust the first dashboard had spent. The desk was invited to override the forecast freely — and every override was recorded alongside the outcome. Where the schedulers beat the model, the model team studied why and improved it; where the model beat intuition, the record spoke for itself. The override rate has fallen month over month since, for the only reason that lasts: the forecast keeps being right.
Results that held
Forecast error is four points better than the prior method, and — the metric that matters more — the forecast now informs every dispatch decision the desk makes, daily. The previous dashboard was technically alive for a year and used for none of them.