Retailers do not put forecasts on shelves. They put products there. That sounds obvious until you see an AI project celebrate a lower prediction error while the store still runs out of a popular item on Friday and receives too much of it on Tuesday.

Demand forecasting matters, especially when a chain has thousands of products and stores. But the useful question is what happens after the number appears. Can a buyer change an order? Is the supplier still accepting it? Does the stock record reflect what is actually in the back room? If the answer to any of these is no, the most elegant prediction in the world has found a very expensive way to be ignored.

What the Cainz case does and does not prove

Japanese home-improvement retailer Cainz gives us a useful public example. In a Google Cloud customer case, the company describes replacing a replenishment approach based on recent moving averages with AI-assisted demand forecasts. The account says Cainz first tested the system on a limited set of products at one store in 2022, then expanded it. By August 2024 it had deployed the forecasting solution in 209 stores. Google Cloud also reports that data preprocessing across the network took about 50 minutes.

That is a supplier-published case, not an independent evaluation. The 50-minute figure concerns data processing; it is not a claim that shelves became perfectly stocked, or that a rival retailer would see the same result. The case is interesting precisely because the team had to solve mundane operational problems as it scaled. A model was only one component. Data had to move through the pipeline fast enough to be useful, and its output had to feed the retailer’s planning process.

The first pilot was small. I wish more AI projects copied that detail rather than the vendor logo. Starting with a handful of products gives a team room to discover why its inventory file says there are six units when the store can find only two. That is not a model problem. But if nobody notices it, the model gets blamed later.

Name the decision before you pick the model

“Improve forecasting” is too broad to manage. A better starting question is: which decision should change, for whom and by when?

A garden-products buyer may need a weekly store-level order recommendation before a supplier cutoff. A store manager may need a daily warning that a promoted item is likely to sell out. A distribution planner may need to move stock between locations. Those are related decisions, but they have different time horizons, data needs and owners.

Write down the actual decision path. The forecast arrives on Monday morning. A planner checks it against the latest promotion, known supplier delays and local demand. The planner approves or changes the order by noon. The supplier confirms quantities by late afternoon. If the forecast arrives after the ordering window closes, it is a beautifully formatted history lesson.

Give the person making the decision a visible reason for the recommendation and a way to override it. Keep the override reason. A repeated correction may reveal a missing promotion feed, an inaccurate stock count or a supplier constraint that the model never sees. Overrides are not proof that employees “resist change.” Sometimes they are the most useful dataset in the project.

Measure shelf availability alongside forecast accuracy

Forecast accuracy is worth tracking, but it is a diagnostic measure. Retail customers care whether the item is available. Finance cares what it cost to keep it available. Store teams care whether replenishment creates another hour of avoidable work.

For a pilot, define a small group of comparable stores or categories and agree on the baseline before deployment. Then watch on-shelf availability, stockouts, excess inventory, markdowns, emergency transfers and staff time. Track the share of suggested orders that people change, and why. A forecast that becomes more accurate while markdowns rise or staff spend longer repairing orders has not earned a rollout.

Be careful with the denominator. If an item was not on the shelf, recorded sales cannot show how much customers would have bought. A model trained only on historical sales may learn the retailer’s past stockouts as if they were weak demand. It is worth checking shelf availability and lost-sales signals before congratulating the model for predicting what the old process allowed customers to buy.

The test also needs a credible comparison. A seasonal category can look brilliant for a month simply because the weather behaved. A promotion can make a weak model look clever. Compare like with like where possible, record major changes during the pilot and explain the result in operational terms. If a simpler ordering rule performs as well, keep the simpler rule. Retail has enough software subscriptions already.

Design the moment of disagreement

The most revealing screen in a forecasting tool is often the one where a human disagrees with it. What can the planner see? Which source data are stale? Can they change the quantity without opening three other systems? Does the system learn the reason, or merely record that somebody clicked a different number?

Agree on what the AI may recommend and what it must never approve alone. A short shelf-life product, a high-value seasonal item and a basic line with reliable supply should not all share the same tolerance for error. Where supplier capacity is tight, the order may be capped regardless of predicted demand. Where customer promises are already made, a potential stockout deserves earlier human attention.

This is where AI adoption becomes a management decision. People need to know who owns a bad order, who can stop an automated recommendation and when a pilot should pause. Those are not glamorous questions. They are the difference between a system the team trusts and a dashboard everyone politely ignores.

The operating result to look for

Cainz’s public case shows that demand forecasting can scale across a large store network when the data-processing architecture and planning workflow scale with it. It does not establish a universal ROI. The lesson for another retailer is to test its own decision chain, not borrow somebody else’s performance claim.

Start with one category, one ordering decision and a buyer who is willing to tell the project team when its suggestion makes no sense. If the pilot improves availability without shifting the cost into waste, markdowns or extra work, expand. If it does not, you have learned something useful while the experiment was still small enough to change.

Source: Google Cloud, “Cainz: Using Vertex AI for demand forecasting to optimize sales-planning predictions” (customer case, accessed 25 October 2026). Reported Cainz deployment and processing figures are attributed to this vendor-published account.

About the author: Kirill Naumov is a consultant and entrepreneur working across retail, customer experience and applied AI. His background includes 11 years with Apple, seven years teaching MBA participants and 16 years of cross-sector work. See Kirill Naumov’s website and his live online AI workshop for retail, sales and learning teams