Retail has made real progress in forecasting. Machine learning can process more signals, recognize patterns that conventional systems miss, and produce forecasts at a level of detail that would have been impossible to achieve a decade ago.
Still, the inventory problem remains enormous. IHL Group estimates that inventory distortion, the combined impact of stockouts and overstocks, costs global retailers $1.73 trillion annually.
What does that number tell us? While forecasting is important, it’s only part of the job. Retailers must still decide which purchase orders to release, where to position inventory, whether to redirect stock between distribution centers, and how to respond to unpredictable changes.
This is where the real gap appears. A model generates a better prediction, but the surrounding business process continues as before. Buyers still spend hours working through spreadsheets. Recommendations arrive outside their normal workflow. Users see a number without understanding what influenced it. Without making an impact on day-to-day decisions, any model, however technically successful it is, isn’t enough.
Rules-based forecasting has served retailers well for decades. It has always been effective when applied to stable and limited demand drivers. And it still is. However, as the industry grows, forecasting is becoming more and more complicated. Planners nowadays need to manage thousands of SKUs across multiple locations, with long supplier lead times, regional differences, promotional activity, port conditions, changing inventory positions, and, not least, geopolitical events. Static forecasting rules weren't built for such complexity.
The problem can be solved with machine learning. It can identify non-linear relationships, incorporate new signals and incrementally improve as more data becomes available. But soon it becomes apparent that building an algorithm or a model isn't the hardest part. Often the key difficulty is data. To make an algorithm work, you need to feed it with lots of data, including inventory positions that are sometimes inconsistent; purchase-order data that may sit in a legacy enterprise resource planning system; or key additional data such as promotional information that is frequently siloed. And after gathering distributed data, it’s an altogether different challenge to encourage the teams to actually adopt a new way of working, casting aside a process they may well have created.
A pragmatic AI forecasting capability needs three things working together: a better predictive model that produces forecasts at the right SKU, location and time; incrementally combining the signals that genuinely affect demand; and dovetailing into existing business processes and creating sound recommendations for expert teams to approve.
That final component is where all projects should start. If the recommendation being made doesn’t generate a return on investment, look elsewhere.
For a supply team to act on a model’s recommendations, it needs to understand what they're based on. Say, for instance, that a system recommends releasing one container before another. An experienced buyer would always ask which SKUs are driving the decision, how demand has changed, what inventory is already available, and which open orders were considered. This is because buyers are the ones who are responsible for inventory and customer experience, and they won’t act before seeing the reasoning.
That is why explainability should be designed in from the start. This level of visibility changes the conversation. Users start by checking the model against their own judgment. When the recommendations are sound, they begin to trust it. As time goes by and the recommendations are consistently good, teams become more confident in accepting its decisions.
With that said, the human role remains critical. A model can't account for information it has never received. It grounds its recommendations on available data, and it might not know that a particular supplier is experiencing production difficulties or that some competitor has just launched a successful promotion campaign. This is where human experts still give retailers competitive advantage. Buyers or planners always have a better understanding of business context and thus make better decisions. However, with a fire hose of signals and decisions to make in a very short space of time, an artificial intelligence-enhanced recommendation system will make a great team even better.
So how do you make an effective AI model that backs up your decisions? Firstly, there's no need to apply AI to the entire supply chain at once. Start by picking one decision where you have available data, measurable value,and manageable scope. Use four questions to assess that opportunity:
Signal quality: Do we have enough data? A model needs data to train and to run, and that data should be available, accessible and reasonably clean. A first use case shouldn’t need six months of integration work to get there. Take historical sales information, inventory positions or open purchase orders to teach your model.
Value clarity: what outcomes do we improve? Decide this before you pick the model. For instance, it could be stockout frequency, buyer processing time or logistics cost. If the outcome can't be named and tracked, the use case needs more work.
Explainability requirement: what do users need to see before they act? As mentioned before, users need to see the reasoning behind the model’s recommendation. Before picking a model, think about what explanation it needs to provide to encourage users to trust its suggestions and act upon them.
Proof of concept followed by a pilot feasibility: can we test the concept? Better forecasting is the cornerstone for making a sound recommendation system. Proving this can be done in four weeks.
Before scaling your concept, try testing it small. Build a pilot that covers a known group of SKUs, one distribution center, or one purchasing decision. Its job is to answer a narrow question: can the model perform well enough to justify the next investment? This pilot should take no longer than three months, and have a demonstrable ROI to build trust in the process.
The SVEP exercise prevents teams from choosing a use case because it looks impressive in a demonstration. It keeps the focus on whether that use case can become a viable operating capability.
Once an opportunity passes the SVEP test, adoption should progress in evidence-led stages:
Validate the ROI. Start with known SKUs, accessible data and established models. Avoid building a full platform at this point. The immediate goal is to establish whether machine learning can meet or exceed the relevant forecasting benchmark for the selected category.
Deliver to real users. Next, the capability should reach the people who will use it — buyers or planners. Build an MVP around one specific process, show users why the system recommends what it recommends, and track both the technical and the business metrics. This is where a forecasting model becomes useful. The important question changes from “Is the model accurate?” to “Does it improve the way people make decisions?”
Scale with evidence. Expand only when the proof is there:
- The forecast meets the benchmark;
- People actually act on the recommendations;
- Business numbers such as stockout frequency, transfer volume, or processing time are getting better, and
- The architecture is stable enough to handle a wider rollout.
After that, retailers can add more SKUs, distribution centers and use cases, without over-engineering something that hasn’t yet proved its value.
Forecast accuracy isn’t going anywhere. Promotion planning, replenishment, logistics, inventory allocation: all of it depends on having a credible view of future demand.
But the bigger opportunity today is to make that forecast operational, so that it changes what people actually do.
So where to begin? With one decision that matters. You can test it against four things: how strong the signal is, how clear the value is, whether the results can be explained, and whether a pilot is realistic. If it passes, you can prove that the forecast works, put it in front of real users, and scale it.
That’s how retailers get from more accurate predictions to supply chain decision intelligence — by building a habit that can be repeated, and turning complex signals into timely action that people trust.
Stuart Hogg is client director, retail at Intellias.
