Use case assessment
Is AI demand forecasting worth building?
Forecasting is decades old and mostly solved. The AI label is new; the method probably shouldn't be.
This is a well-established statistical problem. Classical methods frequently match or beat machine learning on business time series, and they are far cheaper to run.
What it usually means
Predicting future demand for products or services to drive purchasing, production or staffing.
Is it genuinely AI?
It is genuinely statistics, and it has been since long before the current labelling. Exponential smoothing, ARIMA and seasonal decomposition are mature, interpretable, and cheap.
The honest framing: machine learning is one family of methods for a problem that already had good ones. It sometimes wins, usually where you have many related series and rich external signals. It frequently does not, and the literature on forecasting competitions has been uncomfortable reading for the ML case for years.
What the simpler version looks like
Establish a naive baseline first: next period equals last period, or equals the same period last year. Then seasonal exponential smoothing. Measure both against your current process.
That baseline is the assessment. Any proposal for a sophisticated approach should have to beat it by a margin worth the run cost, and a surprising number cannot.
What it costs to run
A statistical model runs on a schedule and costs almost nothing. A machine learning pipeline needs feature engineering, retraining, monitoring for drift and someone who understands it when it behaves oddly.
The number that decides it: the accuracy improvement, translated into inventory or staffing cost, against the annual run cost of the more complex approach. If nobody has done that arithmetic, the project is not ready.
Where it sits under the EU AI Act
Demand forecasting is not an Annex III high-risk use. Where forecasts drive workforce scheduling, be aware that worker management is an Annex III area — forecasting demand is not, but automatically allocating shifts on the basis of it moves closer to that boundary.
When it is worth building
When you have many related series, genuine external drivers to incorporate, and a baseline that you have measured and found wanting. When forecast accuracy translates into a large enough cost to justify the operational overhead.
Not because the current spreadsheet feels unsophisticated. Sophistication is not the objective; the inventory cost is.
Common questions
Is machine learning better than statistical forecasting?
Not reliably. Classical methods like exponential smoothing and ARIMA frequently match or beat machine learning on business time series, and they are interpretable and far cheaper to run. ML tends to win where there are many related series and rich external signals.
What should you do before building a forecasting model?
Establish a naive baseline: next period equals last period, or equals the same period last year. Then seasonal exponential smoothing. Any sophisticated proposal should have to beat that by a margin worth its run cost.
Is demand forecasting high-risk under the EU AI Act?
No. But if forecasts feed automated workforce scheduling, worker management is an Annex III high-risk area, and the allocation step sits closer to that boundary than the forecast does.
Scoring 40 AI use cases in a regulated enterprise — the baseline test that removed roughly one proposal in ten before any budget moved.