Mecatechnic distributes automotive spare parts across a very long tail catalogue, around one hundred and thirty thousand items on file. The replenishment alert thresholds were maintained by hand in the management system, and so were impossible to keep current at that scale: too cautious on one line, too tight on the next.
The engine
At every export from the management system, a processing chain recalculates three things for each product and supplier pair.
When to order, through a dynamic reorder point based on the last twelve months of demand, its volatility and the supplier lead time. How much to order, through an economic order quantity that balances the cost of placing an order against the cost of holding stock. And what to stop stocking: dormant lines, those with more than twenty-four months of cover or no recent sale, have their threshold brought down to zero, so that capital stops sitting in parts that no longer move.
No black box
The decisive technical choice was to set aside opaque machine learning in favour of classical inventory science, applied rigorously to thirty-six thousand product lines.
A classification crossing value with volatility sets a different target service level for each group, from 97.5% on strategic, stable lines down to 80% on expensive, erratic items, to avoid overstocking. Supplier lead times are measured from the real purchasing history rather than taken from the theoretical lead times quoted, with statistical filtering of outliers. And the real business rules are applied: rounding to pack multiples, carriage-paid levels and minimum order values, exclusions by product range.
Every recommendation can be explained to a buyer, which is the condition for them adopting it. The rollout itself carries little risk: the old manual threshold is kept alongside the new one, so the gap can be measured before switching over.