Skip to content

Item-level forecasts that start from the planner's rules

Demand forecasting is often done in a spreadsheet, and safety stock reflects the planner's experience more than the order history.

How demand and stock are planned today

The planner usually extracts customer orders and sales history from the ERP system, takes them into a spreadsheet and adjusts them with what they know, for example the promotion announced by sales, or the customer who orders only at the end of the quarter. As long as the items are few, the method holds. When the codes become many, the review concentrates on the best sellers and the rest follow safety stock levels set years earlier, so stock-outs come to light when the customer's order has already arrived.

The model proposes forecasts and stock, the planner approves the plan

In the example, the Data Intelligence forecasting models read the order history, the promotions in the calendar and suppliers' lead times and estimate the demand for each item. On that basis they propose a stock level and reorder point, applying the rules the planner already uses, such as a supplier's minimum order quantity. The planner looks first at the items where the proposal departs from the current plan, corrects where they know what the data does not say and decides the orders.

  1. The models read orders, promotions and supplier lead times
  2. They propose a forecast and stock level for each item, following the rules in use
  3. The planner corrects the variances and approves the plan
  4. The planner's corrections feed into the next cycle

Illustrative example

What data-driven stock planning needs

ItemIn this case
DataIn this caseOrder and sales history by item, stock on hand, suppliers' lead times and minimum order quantities, promotions, the stock rules in use
ModulesIn this caseData Foundation, Data Intelligence, Workspace
Who decidesIn this caseThe planner on plan, orders and priorities; the head of planning on the rules the models apply
What we measureIn this caseThe gap between forecast and actual demand, stock-outs, idle stock and the planner's time, against today's method
AI Act risk classIn this caseThe models estimate demand for items and do not assess people, so the case stays well away from the high-risk uses in Annex III that we come across most often, such as the creditworthiness of individuals; we still set the classification with risk and compliance before development
Where to startIn this caseWith Deployment on one family of items, when the order history is in the ERP system and the planner has the time to write down the rules they apply

Without a clean history the forecast does not hold

If items change code at every revision, or if returns and cancelled orders stay mixed in with sales, the model learns a demand that never existed, and the first job is on the data, with Data Foundation. For made-to-order products there is no history at all and the planner's judgement remains the best source. With few items and stable demand, finally, the spreadsheet costs less than a model.

The case comes from the Industrials page.

Let's start from an item that ran out of stock

Taking an item that went out of stock in recent months, or sat idle in the warehouse, we look with the planner at which data would have given the warning in time.