AI Demand Forecasting in Manufacturing: A Worked Scenario

AI demand forecasting in manufacturing works better as support for planners than as a crystal ball. The scenario below is a hypothetical, typical example: a plastics injection molder uses its order history to suggest how much to produce, and the production planner reviews before deciding.

The scenario: a plastic packaging plant

Picture a mid-sized company with six injection molding machines, making tubs and lids for food, cleaning and cosmetics customers. Today, the planner builds the month in a spreadsheet based on experience and open orders.

The symptoms are familiar: some months the best-selling lids run out and customers wait; other months slow movers pile up, tying up inventory and cash. Frequent mold changes make it worse, since each one costs hours of idle machine time.

The problem to solve

• Big customers place orders only days ahead, but resin and molds need planning earlier.
• Demand shifts with the seasons: beverage and summer packaging rise in hot months, with peaks before holidays.
• Two or three customers carry much of the volume, and one of them changing course moves the whole plant.

The solution: statistical forecasting with human review

1. Organize the history

Pull two to three years of orders by item, customer and month from the ERP. Clean out what distorts the picture: one-off project orders, returns, strike or shutdown months.

2. Build the model

The system combines trend, seasonality and the effect of variables such as the calendar and customer promotions. For a portfolio like this, classic statistical models or simple machine learning are usually enough. You don't need a giant model.

3. Deliver the forecast as a range

Instead of a single number, the forecast gives an expected value and a range, for example 80,000 to 110,000 units. The range tells the planner how far to trust each item.

4. Planner review

The planner compares the forecast with what they know: a customer opening a new line, a stalled construction project, a competitor that lost supply. They adjust the numbers, and the adjustment feeds back into the system as learning.

Expected result

In scenarios like this, the typical gain is fewer stockouts on fast movers, less idle inventory on slow ones and fewer last-minute mold changes. The improvement depends on data quality and the starting point, so a realistic goal is to cut planning error gradually, not eliminate it.

Track a few indicators: average forecast error by product family, service level (orders delivered on time) and days of inventory. Compare the quarter before the pilot with the one after.

Lessons from the scenario

1. Clean data matters more than the algorithm.
2. Start with a few items, the ones that weigh most on revenue.
3. New items and one-off orders have no history: handle them separately with human judgment.
4. The forecast doesn't replace the planner; it gives back time to think about exceptions.
5. Review the model every quarter, because customers and markets change.

Projects like this are part of our automation and AI work for manufacturers. If you want to know whether your order history is sufficient, start a conversation through the contact page.

Frequently asked questions

How much history do I need to forecast demand?

Ideally two to three years, to capture seasonality. With less you can still forecast, but with a wider margin of error.

Does AI forecasting replace the production planner?

No. It suggests numbers and ranges, and the planner applies market and customer knowledge that never shows up in the data.

Does it work for new products?

Generally not well, since there is no history. For launches, use analogy with similar products and human judgment, and let the model take over after a few months of sales.