AI Catalogue Planning

We ran two experiments that put AI to work inside a Japanese grocery retailer's weekly catalogue planning. Then we turned what both taught us into the blueprint for building it properly.

The Problem

Every week, a major Japanese grocery retailer decides what goes into its printed catalogue. Which products, on which page, at what size, and how much stock to order for each one?. A handful of experienced planners were making hundreds of judgement calls against a hard print deadline, working through a long chain of spreadsheets, and most of what made them good at it lived in their heads.

What We Built

We tried it two ways:

• First we let the AI lead, connected directly to real product and sales figures. It was genuinely good at open questions — _what sold well in this slot last autumn?_ — but it never did things quite the same way twice. That's unnerving when a printer is waiting.

• Then we flipped it round: a screen the planners drive, calling on the AI at six fixed steps from theme to supply forecast. Dependable and easy to follow — but you couldn't go back. Reach step five, spot a problem back at step three, and you started again.

Both versions proved the idea was sound, and both broke in the same place: planning isn't a conveyor belt. So we designed a third version around the way people actually work: start anywhere, change anything, in any order, with the system quietly checking that the whole plan still adds up. The step-by-step route stays available for anyone who prefers it. We wrote it all up as a requirements document.

Where the AI Actually Helps

• The assistant reads live product, sales and supply.
• You can simply ask, "Swap this product and rebalance the page" instead of hunting for the right button.
• Forecasts are kept honest. Each method is scored for accuracy and optional signals like weather sharpen them when available.
• The AI's instructions can be edited by the people who know the trade, the model is a setting rather than a rebuild, and every change is logged, whether a person or the AI proposed it.