Blog

Planning with the average means planning for a case that almost never occurs. Actual demand for an item fluctuates – sometimes strongly, sometimes weakly, often asymmetrically. An average compresses all of that information into a single number and throws away exactly what matters most for replenishment: the risk.

Traditional systems compensate with blanket safety stocks. The result: well-behaved items sit on too much inventory, sporadic items on too little. A probability density, by contrast, describes how likely every possible demand quantity is. From it, the reorder point can be derived directly from the desired service level – individually per item, instead of one buffer for all.

In our projects, precisely this step – from point forecast to distribution – is the biggest lever: 15–30% less inventory at the same or better availability. Not because the model knows the future better, but because it is honest about its own uncertainty.

First: your reorder points haven't been touched in over a year. Demand and lead times change constantly – static parameters are wrong by definition. Second: your planners regularly export to Excel because they don't trust the system. That's not a discipline problem; it's a sign the system makes poor suggestions.

Third: you have overstock and missing parts at the same time. It sounds paradoxical, but it's the typical pattern of blanket safety stocks – too much of the wrong item, too little of the right one. Fourth: nobody can state which service level you are actually targeting. Without a target, no inventory can be optimized. Fifth: inventory coverage is growing faster than revenue.

Each of these symptoms can be quantified – with your own historical data, before anything is changed in your systems. That is exactly what our Proof of Value is for: we calculate what probabilistic replenishment would have saved in your past. Only then do we talk about integration.

A proof of concept answers the question: does it work technically? That is the wrong question. With modern ML methods, almost anything works somehow. The right question is: is it worth it – in euros, on your data, compared to your current process?

That's why our projects start with a Proof of Value: we take your historical data and simulate which decisions the AI system would have made in the past – and what they would have cost or saved. The result is a robust business case with a clearly defined operating point, before a single euro flows into integration.

The side effect is at least as valuable as the number itself: your domain experts can see, transparently, where the model outperforms the status quo and where it doesn't. That transparency later decides adoption in daily operations – and with it, the success of the entire project.