Carrefour: a tomato cannot wait for the perfect forecast

Fresh food has an inconvenient habit: it ignores corporate strategy and deteriorates on its own schedule. We can spend a long time debating forecast accuracy, but the salad left at closing time will be quite tangible. This is why I find AI in grocery operations interesting. The decision has a deadline that cannot attend another meeting.

Carrefour publicly identifies combatting food waste as an area of work. That provides the context for my independent proposal. I am not describing Carrefour's internal algorithms or suggesting a model will automatically solve waste. Sometimes the underlying problem begins with delivery terms rather than demand prediction.

Food safety, dates and temperature constraints belong in explicit rules that a generated recommendation cannot override. Personal customer data is largely unnecessary here. A mature solution produces a fairly ordinary result: fewer losses, a reasonable evening selection and an associate who understands why a particular action was suggested. That is enough ambition for a tomato.

Look at the time remaining

A forecast needs to concern a batch with a particular selling window, not merely an item over a day. It needs sales, stock, arrival times and storage constraints. AI suggests replenishment quantities and presentation order. If batch dates are missing, it should report uncertainty rather than manufacture freshness mathematically.

Then comes the very practical question of when to change the price. Markdown recommendations must stay within approved rules. Staff can see the stock, remaining time and expected demand behind a suggestion. They confirm or reject it with a reason. The retailer then gains a history of decisions, not just a record of money written off after the useful decision-making window has closed.

Even a good markdown can collide with online order fulfilment. Promoting a batch already reserved for orders is unhelpful. Training customers to wait for the final discount can also damage ordinary sales. We therefore need category margin and regular demand, not simply a heroic chart showing how quickly one unfortunate packet finally found a buyer.

Waste is only part of the picture

A limited assortment with reliable batch identification offers a manageable place to test this. Compare similar days while accounting for weather, promotions and deliveries. The category manager should define markdown limits and mandatory human decisions beforehand. Otherwise the experiment tests staff members' ability to guess policy, which is an unnecessarily difficult way to introduce technology.

The department needs to compare waste by value, category gross profit after markdowns and key-item availability towards the end of the day. Empty shelves can reduce waste while disappointing shoppers. Aggressive discounts can reduce write-offs while hurting economics. Discussing all three together makes the result less convenient to present and much more useful.