ROI Case File No.597: They Championed Store-Level Management, but Had Never Decided How to Group the Stores
They Championed Store-Level Management, but Had Never Decided How to Group the Stores
Chapter 1: They Want Ordering Automated, but Accuracy Won't Improve
"We want to bring in AI demand forecasting and automatic ordering."
Musubu Kubun, head of the sales division at Globex Corporation, said this as he laid out the situation. The company is a food retail chain. "We champion store-level management—each store builds its own floor. But we're short-handed, and ordering eats up the day between manual work and handheld entry. There's no time left for customers."
"Don't you already have automatic ordering?" Claude asked.
"We do. But it's a simple mechanism—replenish when stock falls below a reorder point," Kubun answered. "It can't adjust for season, weather, or promotions. The fine maintenance doesn't get done either, so accuracy never improves. Inventory imbalances between stores aren't visible. I want AI to read demand across all stores with high accuracy."
"Do you intend that forecast to treat every store the same way?" I confirmed.
"...That's how it would end up," Kubun answered. "We say store-level management, but we were thinking of the forecast as uniform. Which stores resemble which, and where to aim first. We've never thought about it in groups."
"Read them uniformly without grouping, and you get a forecast that fits no store," I replied. "Let's break this down with STP."
Chapter 2: STP Asks—Segment, Target, and Position
"This case calls for STP."
Claude wrote "Segment / Target / Position" on the whiteboard.
"STP—Segmentation, Targeting, Positioning—is a framework that divides a population by similar characteristics, narrows to where you'll aim, and defines what you'll offer there," I explained. "The crux is not treating every store as an average. A residential-area store and a station-front store differ in the hours things sell and in what sells. Mix them into one training set and you get a forecast pulled toward the average. Only by dividing does store-level management hold up on the numbers side."
"First, let's measure the current cost," Gemini said, opening ROI Polygraph. The data Kubun had provided went in.
"The monthly cost is out," Gemini read off. "Labor for manual and handheld ordering averages 200 hours a month; at ¥3,500 an hour, that's ¥700,000 a month. Opportunity loss from stockouts caused by poor forecast accuracy averages ¥440,000 a month. Losses from excess inventory, markdowns, and disposal average ¥380,000 a month. Inter-store transfers and missed sales caused by inventory imbalance average ¥320,000 a month. Sales opportunity lost to shrinking customer-facing time, squeezed out by ordering work, averages ¥300,000 a month. Total: ¥2,140,000 a month. Annualized, roughly ¥25,680,000."
Kubun stared at the figures. "I was only looking at the ordering burden. Once you add the sales missed to stockouts and the goods moved at a markdown, it comes to this much."
"Then let's design it with STP," I continued.
[Segment—group stores by similar selling patterns]
"First, segmenting," Claude said. "Location, customer base, floor area, the shape of the peaks by day of week. Bundle the stores whose sales curves resemble each other. Residential type, station-front type, arterial-road type. Not reading every store individually, and not reading them all uniformly, but reading them in bundles. Bundled, accuracy holds even where a single store has little data."
[Target—aim at the bundle where it works most]
"Next, targeting," Gemini continued. "Produce the stockout and disposal figures for each bundle. Start with the bundle carrying the largest loss. The bundle with fast-turning fresh food and heavy weather sensitivity gives the largest return when the forecast lands. Not all stores at once—enter through the bundle where it works."
[Position—change what each bundle is asked to read]
"After targeting comes positioning," I continued. "Time of day and weather matter for the station-front type; day of week and promotions matter for the residential type. Don't feed the same variables to every store. Because the effective inputs change by bundle, store-level management holds on the forecasting side too."
[Return—connect the forecast to ordering and stock transfers]
"Last, returning it," Claude continued. "Even if a forecast comes out, labor doesn't fall unless it connects automatically to ordering. Connect it to the core system, and when imbalance appears within a bundle, propose a transfer. Not ending at reading is what creates customer-facing time."
[Simulating the investment recovery]
"Let's run the numbers with ROI Proposal Generator," Gemini proposed.
- Initial cost: Designing the store classification, preparing sales data, building the AI demand-forecasting model, core-system integration, implementing the auto-ordering logic, and store training—¥5,300,000 in total
- Monthly cost: AI platform usage plus ongoing system operation and maintenance—¥240,000 a month
- Monthly savings: Automation of ordering work = ¥490,000 a month (assuming a 70% reduction), reduced stockout opportunity loss through better forecast accuracy = ¥360,000 a month, reduction of excess inventory and disposal = ¥300,000 a month, resolution of inventory imbalance and recovery of customer-facing time = ¥330,000 a month—¥1,480,000 a month in total
- Net monthly savings: ¥1,480,000 − ¥240,000 = ¥1,240,000 a month
- Payback period: ¥5,300,000 ÷ ¥1,240,000 = approximately 4.3 months
"That's a payback of a little over four months," Gemini summarized. "What works is that you don't read every store uniformly—you divide them into bundles. Mix them into one training set and you get a forecast that misses every store by a little, and the floor stops trusting it. Change the effective inputs by bundle, and the first bundle lands visibly. The investment doesn't swing at air."
Kubun looked over the numbers. "I thought bringing in a high-performance AI would settle it. Without dividing, store-level management doesn't hold up on the numbers side."
"STP is a tool for segmenting, targeting, and positioning," I replied.
Chapter 3: A Deployment Plan That Segments and Targets
"Let me lay out the approach," I said, standing at the whiteboard.
"Month one—prepare the sales data and classify stores by selling pattern. Month two—calculate stockout and disposal figures per bundle and fix which bundle to start with. Months three and four—build and validate the forecasting model for the first bundle. Month five—integrate with the core system and connect to automatic ordering. Month six—verify results and prepare the rollout to the next bundle. Month seven onward—expand to all bundles, revise the classification, and implement stock transfers."
"Since we champion store-level management, shouldn't we read each store individually?" Kubun confirmed.
"Store by store, there isn't enough data," Claude replied. "The units sold of one item at one store vary widely by day, and training directly on that means memorizing coincidence. Bundle similar stores and the tendency becomes clear, and on top of that you can correct for store-level differences. Bundling what resembles each other comes before reading individually."
Kubun took notes as he spoke. "Decide how to group the stores before making it forecast. I can see the sequence now."
Chapter 4: The Day the Bundles Were Formed and the Forecast Landed
Ten months later, a report arrived from Kubun.
In the first bundle, the numbers moved within two months of go-live. "Fresh-food stockouts fell, and at the same time the markdown tags fell. I didn't expect both to improve at once," Kubun wrote.
Ordering work got lighter too. Time spent walking the floor with a handheld shortened. "Store managers who lost two hours a day to ordering can now stand on the floor. The time for customers actually came back," the report said.
The largest change showed up in how store-level management was handled. A state of holding it up as a principle became a state of designing it as a classification. "We kept saying stores differ, without ever organizing how they differ. Once we bundled them, the differences could be explained with numbers," Kubun wrote.
Inventory imbalance also moved toward resolution. Transfers within a bundle reduced missed sales. "What was sitting surplus at the next store was out of stock at this one. Simply making it visible made things move," the report said.
As a side effect, the way promotions were run changed. Different measures worked on different bundles. "We'd been distributing the same flyer to every store. Once we split by bundle, it became clear where it worked and where it didn't, and the way we allocate the spend changed," Kubun wrote.
The final line of Kubun's report read: "I thought the trouble with ordering was forecast accuracy. But the real problem was that we championed store-level management and had never decided how to group the stores. The moment we bundled with STP, both the inputs to read and the order to aim in were decided. Before making it forecast, dividing came first."
The day a company that had never decided how to group its stores became a company that could bundle and take aim, demand forecasting had changed from uniform company-wide learning into a design that groups by similar selling patterns and changes the effective inputs, the report noted.
"Consultations about demand forecasting usually arrive in the form of 'we want a high-accuracy AI.' But there is a question to ask before you make it read. Has it been decided how to divide the population? What STP asks is segmenting and targeting, and then positioning. Bundle, and even sparse data reveals a tendency, and the effective inputs change from store to store. The day a company that treated every store uniformly could decide how to divide, what changed was not the performance of the AI but the very perspective of grouping by similar characteristics."
Related Files
Tools Used
- ROI Polygraph — Visualizing ordering labor, opportunity loss from stockouts, excess inventory and disposal, and the cost of inventory imbalance
- ROI Proposal Generator — Investment-recovery simulation for AI demand forecasting and automatic ordering built from store classification and starting order
