Forecasting demand without a crystal ball: AI to support planning
On a shop floor, an operator shows me his Monday morning planning file. The two-week order forecast lands in front of him, he looks at it, he sighs, and he takes an average. Nobody really trusts it, so we guess. It’s a scene I see again in nearly every plant I visit, and it’s exactly the kind of pain where AI can help, as long as you don’t ask it for the impossible.
The real problem: planning blind
The one or two week forecast is the most strained horizon. Too far out to know, too close to ignore. When the firm order hasn’t dropped yet, the operator improvises: a bit of history in his head, a bit of instinct, and an average to cover himself. The result is safety stock inflated by caution, last-minute changeovers, and a plan redone three times in a week. It isn’t a lack of skill: it’s that we ask a human to hold dozens of products, seasonalities and weak signals in his head at once. Nobody does that well by hand.
What AI does, and what it doesn’t
AI is good at this precise task: crossing the sales and order history with the available signals (seasonality, trends, the behavior of a customer or a product family) to rough out a forecast. It sees regularities a human brain can’t hold across that many references at once. But beware the trap: it doesn’t deliver a truth. It proposes a starting number, more honest than an average, and that’s all. The planner stays in control. He knows a customer will place a big order because he had him on the phone, he knows a line is down for maintenance next week. None of that is anywhere in the history. AI proposes, the human adjusts with field knowledge. It’s a starting point, not a verdict.
The prerequisite: usable sales data
Before any of this, you need a foundation. A forecast is only worth as much as the history it leans on. If your sales and order data is scattered, incomplete or full of holes, AI will learn on noise and produce forecasts that are confident but wrong. This isn’t a technical detail: it’s the prerequisite. The good news is that this data almost always exists somewhere, in the ERP or in spreadsheets. You just have to make it usable, and that’s often a short job.
Where AI has no business
Let’s be clear about the limit. AI is built for the regular, the repetitive, what resembles the past. It has no business on the exceptional event: a customer doubling his order for a reason it can’t know, a supplier shortage, an unheard-of commercial spike. In those situations, human judgment decides, full stop. Handing the exceptional to the model is the surest way to get it wrong with confidence. AI roughs out the repetitive work to free up brain time for what truly matters: the decisions only a human who knows his ground can make. AI with us, not in our place.
For the full picture, read the guide AI in industry. See also: The data that sleeps. Wondering where to start? Gauge your AI maturity in 2 minutes, or let’s talk for 20 minutes.