Systems judged on the shop floor

Southern Kyushu has typhoons most autumns, and Kagoshima receives ash from Sakurajima on many days a year. A store that closes because a data line is down fails its town at the moment the town most needs batteries, water and prescriptions. I therefore design store systems to sell for at least a full day without a connection. Tills hold prices and promotions locally, card payments taken offline are capped at amounts agreed with finance, and each pharmacy keeps its reference data in the store. The current replacement of in-store servers keeps that capability, although a thinner cloud design offered by two suppliers would have cost about ¥380 million less over five years. I test offline trading twice a year in live stores, with the network deliberately disconnected, because a failover tested only in a laboratory has not really been tested.

Our customers pay by cash, credit card, prepaid electronic money, code payments and points, often in one basket that carries both consumption-tax rates. I insist that every payment method runs through the till rather than a separate device, so that a sale, a refund and the tax on each line remain one record. That rule made code payments arrive four months late. It also meant that the qualified invoice system introduced in October 2023, which required our receipts to show our registration number and amounts by tax rate, was a change to one system rather than several. I review payment fees with finance every year, by method and by store type, and I report reconciliation exceptions to our audit and supervisory board members. A payment method the store team cannot reconcile at closing is not cheap, whatever its published fee.

Bread, milk, tofu, eggs and prepared dishes account for about ¥31 billion of our sales and most of our disposal. A forecast can learn weather, weekdays and price changes; it cannot easily know that a school sports day has moved, that a funeral is being held in a small town or that ash is expected over Kagoshima city tomorrow. Our ordering screens propose a quantity, and staff may change it by giving one of nine reasons. Those reasons return to the model and to merchandising each week. An item family becomes fully automatic only after eight consecutive weeks within an agreed error band, and prepared dishes still need a person's confirmation. I measure disposal, evening availability and ordering time together, because improving one by quietly worsening another is easy and of no use to a store manager.

About 2.3 million members used our points card in the past twelve months, and 780,000 use the app each month. Their purchases describe ordinary lives in detail, including medicines and personal products. I keep medicines and 41 sensitive product groups out of supplier analysis and supplier-funded offers, and members must opt in separately before any supplier offer reaches them. We retained the plastic card when we launched the app in 2020, at a running cost of about ¥160 million a year, because many older members in rural stores did not want to use a phone at the till. I work with our legal team on Japan's personal-information law, but a lawful design does not end the question for me. The test I apply is whether a member would be surprised to learn how a purchase had been used.

Labour is now the constraint in almost every one of our towns. We use payment-only self-pay units in 182 stores and full self-checkout lanes in 34 larger stores, and we have added electronic shelf labels in 96. I treat each as a way to move hours rather than simply remove them. Medicines that require a pharmacist or registered seller to be involved cannot be completed at a self-checkout; the lane stops and calls a member of staff. I ask store managers to record where released hours went: to the dispensing counter, the chillers, the shelves or the customer who needs help with the app. Where an automation saves till time but lengthens the wait for advice, I would rather slow its rollout than report a saving that customers experience as poorer service.