A regional retailer’s buyers were ordering replenishment stock the way they always had: last year’s number, this week’s gut feel, and a spreadsheet that only one person fully understood. It worked well enough for the fast-moving lines they knew by heart and badly for the long tail — slow items that quietly tied up cash on the shelf, and others that ran out just as demand picked up. Nobody could look at thousands of products each week, so most of them were ordered on autopilot.
What mattered most
- The buyers had real knowledge a model could not — a promotion coming, a supplier issue, a local event — and the tool had to respect that, not fight it.
- A forecast nobody trusts gets ignored, so it had to be explainable and tied to something the buyer recognised.
- The output had to reach the order, not sit in a separate report that needed re-keying.
- Demand shifts with season and trend, so a forecast set once and forgotten would drift out of date.
How we approached it
We began with the data that already existed — sales history, current stock, lead times — and got it clean and connected before any forecasting. The model produces a suggested order quantity per item, but it never places an order. It surfaces the items worth a second look — the ones where the suggestion and the recent pattern diverge most — so the buyer’s attention goes where it earns its keep, instead of being spread evenly across a catalogue nobody can read in full. For the steady lines, the suggestion is usually right and waved through; the judgement is saved for where it matters.
The buyer can always override, and we made a point of recording when they do and how it turns out — so over time both the team and the model learn where human knowledge beat the number and where it did not. The suggestion appears inside the ordering workflow the buyers already used. We measured against availability and against stock held — the two things the team is judged on — rather than against forecast accuracy in the abstract, which can look impressive while the shelves are still wrong.
Where it stands
Buyers now start from a considered number rather than a blank cell, and spend their time on the items the tool flags rather than scrolling through everything. They still place every order, and they still override the model when they know something it does not — which is exactly the relationship we wanted. The long tail gets attention it never used to, and the spreadsheet only one person understood is no longer a single point of failure.