Latest / The Operations Podcast with Fexingo: Process, People, and Profit in Modern Business / How a Japanese Convenience Store Chain Reduced Spoilage by 60 Percent
Transcript
- Lucas: You walk into a convenience store in Tokyo any evening around seven p.m., and you see shelves of bento boxes, onigiri, sandwiches — all fresh, all made that morning. And you know that whatever doesn't sell by midnight gets thrown out. Luna: Right, that's the dark side of convenience. Fresh food has a shelf life measured in hours, not days. Lucas: Exactly. And the industry average for spoilage at Japanese convenience stores — what they call 'waste rate' — has historically hovered around eight percent of all fresh items. That's billions of yen in lost product and a massive sustainability problem. Luna: So how do you fix that without running out of stock and disappointing customers? Lucas: That's the operational tension. FamilyMart, one of the big three chains here with over sixteen thousand stores, tackled it with a demand forecasting AI. And the result? They cut spoilage by sixty percent. Luna: Sixty percent. That's huge. From eight percent down to what, about three percent? Lucas: Three point two percent, according to their fiscal 2025 reports. That translates to roughly thirty billion yen saved annually — about two hundred million dollars — just from not throwing away unsold food. Luna: Okay, I want to understand how the AI actually works. It's not just looking at historical sales, right? Lucas: No, and that's the key. The system ingests three categories of data. First, point of sale data from every store, updated in near real-time. Second, weather forecasts — temperature, rain probability, even humidity, because that affects what people want to eat. Third, local event calendars: school holidays, sports games, festivals. Luna: So a store near a stadium orders more on game days. That makes sense. Lucas: Exactly. But the granularity is what's impressive. The AI forecasts demand at the SKU level — each bento variety — for each store, for each hour of the day. And it updates the forecast continuously as new data comes in. Luna: So the store manager isn't manually ordering every morning anymore? Lucas: That's the biggest operational shift. Before, store staff would spend hours each day checking shelves and placing orders based on gut feel and simple rules. Now the system generates an optimal order quantity for each item, and the staff's job becomes exception-based: they override the AI only when they have local knowledge the system doesn't. Luna: But does that create any friction? I imagine some managers felt like their expertise was being sidelined. Lucas: There was definitely skepticism early on. FamilyMart rolled it out gradually, starting with a pilot in five hundred stores. They found that when the AI made a bad call — say, under-ordering a popular item — managers would lose confidence quickly. So they built in a safety net. Luna: What kind of safety net? Lucas: A backup inventory buffer. For each high-turnover item, the system orders about five percent more than the forecast suggests, just in case. That buffer adds a small amount of waste, but it prevents stockouts, which are more damaging to customer trust. Luna: So they optimized for total profit, not just waste minimization. A stockout loses a sale and maybe a customer. A little extra waste is cheaper. Lucas: Exactly. And this is the nuance a lot of ai inventory stories miss. The goal isn't zero waste — it's the economic optimum. FamilyMart's internal models showed that eliminating all waste would require such conservative ordering that they'd lose more in forgone sales than they'd save. Luna: That's a really important point. The operations problem is always a trade-off. Lucas: Speaking of trade-offs, I want to mention something that makes this show possible. We don't run ads — never have — and we want to keep it that way. A small number of listeners already chip in monthly through buy me a coffee dot com slash fexingo, and that literally funds the research time and production for episodes like this one. Luna: Yeah, it's a straightforward model. If the show has helped you think differently about a process at work, or given you a concrete example to bring to a meeting, that's the reason we exist. No pressure, just putting it out there. Lucas: So back to FamilyMart. One other dimension worth noting is how the system handles new products. When a new bento flavor launches, the AI has zero historical data. How does it forecast? Luna: I was wondering about that. It must use some kind of similarity matching. Lucas: Exactly. The system classifies every product by attributes — protein type, sauce, price point, package size — and then finds similar past products. So a new chicken teriyaki bento would be compared to previous chicken bentos, adjusted for seasonal factors. Luna: That's smart. And over time, as actual sales come in, the forecast gets more accurate. Lucas: Right. FamilyMart reported that forecast accuracy for new items reaches mature-item levels within about two weeks. That's critical because new product launches are when waste risk is highest — stores tend to over-order to avoid disappointing customers. Luna: Two weeks to converge on a stable forecast? That's fast. How do they measure accuracy? Lucas: They use something called weighted absolute percentage error, or WAPE. The metric gives more weight to high-volume items because a ten percent error on a best-selling bento matters more than the same error on a niche salad. Their target is under fifteen percent WAPE at the store sku level. Before the AI, they weren't even measuring it systematically. Luna: So part of the improvement comes from just having a metric to manage against. Lucas: Absolutely. Good operations starts with measurement. And FamilyMart didn't stop at ordering. They also linked the forecast system to their supply chain — production schedules at the commissaries that make the bentos, truck routing, even the timing of deliveries to each store. Luna: So the entire chain from factory to shelf is coordinated around the same demand signal. Lucas: Yes. Previously, each link in the chain had its own forecast, and mismatches accumulated. Commissaries would overproduce because they didn't trust the stores' orders. Stores would over-order because they didn't trust the commissaries. The shared forecast broke that cycle. Luna: That's a textbook example of the bullwhip effect in reverse. Instead of amplifying variability, they dampened it. Lucas: Exactly right. And the results speak for themselves. FamilyMart's waste rate is now the lowest among Japanese convenience store chains. Competitors like Seven-Eleven Japan and Lawson are catching up, but FamilyMart had a two-year head start with this system. Luna: So what's next for them? Is there another layer of improvement? Lucas: They're working on dynamic pricing for items nearing expiration. If a bento has four hours left on its shelf life, the system could suggest a discount — say, twenty percent off — to encourage purchase rather than disposal. They piloted it in two hundred stores last year and saw a further twelve percent reduction in waste on discounted items. Luna: That seems like a natural extension. But it adds complexity — you need digital shelf labels, real-time price updates, and staff training. Lucas: It does. And it raises a question: at what point does the operational complexity outweigh the waste savings? FamilyMart's data suggests the marginal cost of dynamic pricing is worth it, but they're being careful about rollout. Luna: That's the kind of decision operations people face every day. Not just 'can we do this?' but 'should we do this, given everything else?' Lucas: Right. And that's why the FamilyMart case is such a good study. It's not a story about technology magically solving a problem. It's about a disciplined process of measurement, iteration, and trade-off management. Luna: And a sixty percent reduction in spoilage is a pretty good reward for that discipline. Lucas: It is. And it leaves me wondering: what other industries with fresh-food waste could apply the same playbook? Grocery stores, fast-casual restaurants, even hospital cafeterias. Luna: Maybe that's a future episode. For now, I'll think twice before grabbing that last bento at eleven p.m. Lucas: Actually, you should grab it. That's the whole point — the system made sure it was there for you.