Machine learning transformed supply chain forecasting. The next transformation is turning those forecasts into decisions.
Over the past decade, machine learning has quietly rewritten what is possible in supply chain forecasting. Models that once leaned on simple historical averages now learn from seasonality, promotions, lead time variability, weather and live demand signals, anticipating shifts with a speed and accuracy that would have seemed unreachable not long ago. Forecasting, genuinely, has been redefined.
And yet many organizations that have invested heavily in better forecasts find the same problems persisting: shortages that still surprise them, inventory still in the wrong place, decisions that still arrive late. The forecasts got sharper. The outcomes did not move as much as expected.
The reason is subtle but fundamental. A forecast is a prediction, and a prediction is not a decision.
Two different questions
It is worth being precise about what machine learning is doing when it forecasts. Predictive analytics answers one question: what is likely to happen? Given the data, it estimates a future quantity, such as demand next month, the probability a shipment is late, or when a stock level will run down.
That is genuinely valuable. But it stops one question short of the one that actually changes an outcome: what should we do about it? Knowing a shortage is likely does not reorder anything. Knowing a lane will be delayed does not reroute anything. The step from a prediction to an action, choosing the right move given constraints, costs and trade-offs, and then taking it, is a separate capability. In analytics terms, it is the leap from predictive to prescriptive.
Most supply chains have not made that leap. According to APQC’s 2024 research, fewer than half of supply chain organisations have adopted prescriptive analytics, even as 65 percent name advanced analytics the trend they expect to matter most over the coming years. Enormous effort goes into predicting, and far less into deciding. As one industry analysis put it, the investment exists, but the decision layer does not.
Why better forecasts hit a ceiling
There is a technical reason this gap matters, and it is worth understanding rather than glossing over.
Prediction and decision are different mathematical problems. Prediction is estimation: infer the most likely value of something uncertain. Decision is optimisation: given that uncertain estimate, and real constraints like budget, capacity, shelf life and supplier lead times, choose the best action available. A model can be superb at the first and contribute nothing to the second.
Researchers Bertsimas and Kallus, in foundational work on the shift from predictive to prescriptive analytics, describe exactly this tension: you cannot simply forecast first and decide later as though the forecast were certain, because the best decision depends on the uncertainty in the prediction itself. Deciding well is its own discipline.
This is why, past a point, chasing forecast accuracy delivers diminishing returns. The bottleneck stops being how well can we predict the shortage and becomes whether anything decides what to do about it, in time. A forecast that is two percent more accurate but still lands in a spreadsheet no one acts on changes nothing.
The next transformation
If the last decade of supply chain AI was about prediction, the next is about the decision that follows it, and the encouraging part is that closing the gap rarely requires a bigger model or a dedicated data science team.
It requires connecting the signals an organization already has and building recommendations around its recurring decisions, the reorder, the substitution, the reallocation, so a forecast becomes a specific, proposed action rather than a chart. And it requires governance, which becomes essential the moment a system starts recommending or executing: every recommendation bounded, and traceable back to the data and logic behind it. Automation without traceability is not intelligence; it is a black box with authority. The goal was never to take people out of the decision. It is to hand them a decision worth approving.
Machine learning gave supply chains foresight. The organizations pulling ahead now are the ones turning that foresight into action, reliably and traceably.
From prediction to decision
The forecast was never the finish line. It was the setup. The value has always lived in the decision that follows, made early enough, and confidently enough, to change the outcome.
Data, then intelligence, then action. Machine learning has more than proven itself on the first two. The frontier, and the real prize, is the third.
Turning prediction into a recommended, traceable action is exactly what we built Optivian to do.
Sources and further reading
- Descriptive, predictive and prescriptive analytics stages, and predictive versus prescriptive definitions. IBM; e2open; KNIME.
- Fewer than half of supply chain organizations have adopted prescriptive analytics; 65 percent rank advanced analytics as the highest impact trend (APQC 2024); “the investment exists but the decision layer does not.” Supply and Demand Chain Executive.
- Analytics exists to “drive decisions and actions.” Davenport and Harris, Competing on Analytics (2007).
- The predictive to prescriptive problem, and optimal actions under prediction uncertainty. Bertsimas and Kallus, From Predictive to Prescriptive Analytics.
- Automated decision systems must be governed and traceable to their data and logic. KNIME.