AI for Store Replenishment: What the Algorithm Should Decide, What the Buyer Should, and How to Tell If It Helps

Automated store replenishment is not new. What is new is calling it an AI employee and asking it to take more of the work: not only to compute a quantity, but to prepare the order, explain it and, eventually, send it. Before deciding how much of that to hand over, it is worth looking at what happens in chains that already order this way, because the people using those systems keep changing their output, and recent field studies have measured what those changes do.

1. What an order is made of, and who should decide each part

A store order is a small calculation wrapped in several judgements. The calculation is mechanical: expected sales until the next delivery arrives and is on the shelf, plus a buffer for variability, minus stock on hand and stock already on order. The judgements are not: which supplier, whether a promotion or a local event will lift demand, whether a new product deserves room, whether stock in the back room is real. Splitting the work along that line gives each side what it does well:

2. What the evidence says about overrides

The most detailed recent study follows an upmarket European supermarket chain that had used automated replenishment for more than three years, in two perishable groups, packaged meat and ready meals¹. Store staff could accept or change each proposal. Their cuts were not random: after a product had recently run out, they reduced the proposed order markedly more often, reading the low recorded sales of the empty days as low demand. Those cuts then made a new stock-out of the same product much more likely. The authors call it a self-inflicted stock-out.

Overrides in the other direction can be right. In a convenience-store chain that introduced a replenishment recommender in some stores and not in others, managers raised or brought forward orders on popular items ahead of demand the system had not yet detected, and held back on slow movers². The two kinds of override side by side:

The convenience-chain rollout also showed what the recommender itself did: reorder points rose, orders became smaller and more frequent, service levels improved by 2.9% without extra inventory or lost sales, and managers spent 36.5% less time ordering each day². The lesson from both studies is the same. The question is not whether people should override, but which overrides carry information the system lacks.

3. What helps people use recommendations well

Adherence is rarely high even when the tool is good. At Zara, whose clearance-pricing tool had raised revenue in a pilot, country managers followed its price recommendations in 61.0% of cases in their own stores and in 46.4% of cases for franchises; where the tool recommended no markdown at all, managers marked down by a median of 19.9%³. That study is about pricing, not ordering, but the pattern is familiar to anyone who has deployed a replenishment system.

Three findings point to what helps. Adding a reference point to Zara's screen, so managers could see what following the tool would mean, raised franchise adherence by about nine and a half percentage points³. In online experiments, people who blend their own forecast with an algorithm's at a fixed weight make errors 20% to 61% larger than needed, and showing which inputs the algorithm uses, without opening the model, cut errors by a quarter⁴. In two field experiments on inventory ordering, showing managers how the algorithm had performed improved their orders, and showing where it had performed badly helped more than showing where it had done well⁵.

For a replenishment tool that suggests three design rules: show the numbers each line is built on, show the tool's own record including its misses, and flag the overrides that the evidence says go wrong. The supermarket study estimated that simply blocking cuts made right after a stock-out would have removed a third of the self-inflicted stock-outs and raised profit by about 1%, at a holding cost increase of the same order, and suggested an alert might achieve much of it¹.

4. Granting authority by evidence

The same logic sets how much an AI replenishment employee should be allowed to do. It should start by advising, so its proposals can be compared with what buyers actually ordered. It can move to preparing drafts for approval once its advice is mostly accepted, and to executing after a single confirmation once people rarely change its drafts. Each step is granted per type of action, a store order separately from a supplier order, and only after a probation with written thresholds. The guide to AI agents sets out the full ladder. Some orders should stay with a person at every level:

5. Measuring the effect

Retailers' own reports show why this needs care. Ahold Delhaize describes using "AI-driven order algorithms" to balance availability with minimising waste⁶, and reports its food waste falling, but attributes no figure to the algorithms; closures, donations and better shrink control moved the same number. Group totals cannot separate an ordering tool from everything else that changed that year.

The comparison that can is store against store. Keep a share of comparable stores where the AI only advises, measure shelf availability in both groups over the same weeks, and read the effect as the change in the gap between them. Measure stock value at the same time, because availability bought with more inventory is not an improvement; the out-of-stocks guide and the dead-stock guide cover reading both sides store by store.

6. A note on regulation

In the EU, a replenishment recommender used by a chain's own buyers is generally not one of the systems the AI Act's transparency article targets; that article concerns AI that interacts directly with people, such as a chat assistant, which must make clear it is an AI. Two other provisions are closer to home: deployers must ensure their staff have sufficient AI literacy, and systems used to monitor and evaluate workers' performance are classed as high-risk⁷. Ranking individual buyers by how their overrides turned out would come close to the second; measuring the tool against control stores does not.

7. Limitations

The supermarket study covers five stores, two perishable groups and no promotions, and its profit figures come from a simulation of policies that were not run live. Two of the experiments are known here from their published abstracts only, which give no sample sizes. The Zara data are from 2010 to 2013 and concern pricing. Taken together, the studies agree on direction more firmly than on size.

How marql's AI supply employee works

marql's AI supply employee follows this design. Each morning from six o'clock it reads the order baskets of marql's replenishment workspace, where quantities come from a buffer engine rather than the language model. As an adviser it proposes; when it is allowed to prepare, it writes a draft the buyer edits, approves or rejects like any other, labelled "Prepared by marql's AI supply employee". A draft nobody touched is withdrawn the next morning and rebuilt from fresh data. A basket that needs a person stays with the buyer, with the reason: too large, a supplier known only from invoices, or not the supplier's order day.

Every proposal opens on "Why this was proposed", with the cover now and after the delivery and the lead time behind each line. While it only advises, it is compared with what people ordered on their own. Its KPI is shelf availability in its stores against control stores where it only advises, with stock value as the limit it may not exceed to get there. To execute orders after confirmation it has to pass an exam on a retail simulator first; executing within limits is not open to any customer.