Two hundred customers stopped ordering and nobody noticed.

An industrial distributor with roughly four thousand trade accounts, most of them buying on a predictable rhythm: a workshop that orders every five weeks, a plant that orders quarterly, a contractor that orders when a project starts. That rhythm is sitting in the order history in plain sight.

The inside sales team called accounts in territory order, working through a list. So an account that had quietly stopped ordering four months ago would be called when its turn came around, not when it went quiet, and by then it had usually found another supplier.

The simplest possible build

This engagement is included here because it was almost embarrassingly simple, and it produced the second-highest return of anything in this list. There was no model doing anything clever. There was a reorder interval computed per account from its own history, a due date derived from it, and a queue sorted by how overdue an account was rather than by territory.

The hard part, as always, was that the same customer existed several times: once under the workshop name, once under the proprietor's name, and sometimes once more under a GST registration entered differently. Until those collapsed into one account, no rhythm could be computed at all.

The result

A hundred and eighteen dormant accounts placed an order within two quarters, from the same number of calls made in a different order. The sales manager's observation was that his team had not worked harder, they had simply stopped calling people who were fine and started calling people who had left.