An AI-driven demand forecasting system for a distribution business

The challenge

Replenishment decisions were made from last month's sales and individual buyer judgement. Fast-moving lines ran out during peak periods while slow lines tied up capital in stock that would not move for months. Nobody could explain why a purchase order had been raised, which made the whole process impossible to improve.

What we did

We worked through this in stages, starting with what had to be true before anything else could be built.

1. Consolidated sales, stock and seasonality data from the source systems into one clean pipeline before modelling anything.
2. Built forecasting models per product category rather than one model across the whole catalogue, because demand behaviour differs by line.
3. Made every recommendation explainable — each suggestion shows the drivers behind it so buyers can challenge it.
4. Put the models under monitoring from day one, with drift alerts and a scheduled retraining cycle.

The result

Measured against the baseline captured during discovery, not against a target chosen afterwards.

---