ROI Case File No.511: Thirty New Machines, Bought on a Hunch
Thirty new machines, bought on a hunch
Chapter 1: When a Winning Machine Wins, No One Knows Why
"We buy thirty machines at a time, and half of them miss. This has become normal."
Ryosuke Jinbo, Corporate Planning Manager at Quantum Leisure, said this as he opened the procurement approval forms. A pachinko hall operator running fifteen stores. Annual new machine purchases ran seven to eight hundred million yen. "Purchase decisions are left to each store manager. They look at manufacturer pitches, competitor adoption trends, and their own customer base, but in the end it's experience and intuition."
"What's the hit rate?" Claude asked.
"About forty percent of machines exceed the average utilization rate one month after introduction," Jinbo replied. "The remaining sixty percent fall short of expectations. At three to five hundred thousand yen per machine, thirty machines means a twelve-million-yen investment. And this happens multiple times a month."
"Do you have historical purchase data?" I asked.
"We do," Jinbo answered. "Each store's utilization logs, purchase history, competitor activity. The data exists. We just don't have experience analyzing it across stores. The judgment lives entirely inside the store manager's head. The CEO has directed us to 'use AI to improve forecast accuracy,' but we don't know where to start."
"Too many decision axes, no way to prioritize them," I responded. "Let's break this down with MANDALA."
Chapter 2: MANDALA Asks the Center and the Eight Directions
"This case calls for MANDALA."
Claude drew nine cells on the whiteboard. One cell at the center, eight surrounding it.
"MANDALA is a framework that places a central theme at the core and arranges related elements in eight surrounding directions, structuring the whole picture," I explained. "It's known for Shohei Ohtani's use in high school, but it's fundamentally suited to decomposing complex decisions. For multivariate problems like demand forecasting, the choice of input variables determines accuracy. Nine cells let you map them without gaps."
"Let's measure current costs first," Gemini said, opening ROI Polygraph. The procurement data Jinbo provided went in.
"Monthly opportunity-loss costs are out," Gemini read. "Underperformance losses from miss-hit machines average 6.8 million yen monthly—the gap between projected and actual utilization against purchase cost. Early-withdrawal and replacement costs average 1.2 million yen monthly. Store managers' purchase decision labor totals 300 hours monthly across fifteen stores at 4,200 yen/hour, or 1.26 million yen monthly. Manual competitor research labor is 900,000 yen monthly. Opportunity loss from unanalyzed accumulated data is 800,000 yen monthly. Expected-value risk of judgment quality degradation when managers rotate is 600,000 yen monthly. Total: 11.56 million yen monthly. Annualized: roughly 138.7 million yen."
Jinbo stared at the numbers. "I thought it was just the miss-hit losses. Including the judgment labor, the scale is on another level."
"Now let's design with MANDALA," I continued.
[Center cell — Improving demand forecast accuracy]
"The central theme in nine cells is 'shift new machine demand forecasting from intuition to a numerical model,'" Claude said. "Once the center is set, the eight cells around it know what they need to fill. Conversely, when the center is vague, the eight cells scatter."
[Eight directions — Mapping prediction model input variables]
"Around the center, we place the eight factors that determine prediction accuracy," Gemini continued.
"Cell 1: Own-store utilization history—past utilization data by machine type, time slot, and customer segment. Cell 2: Competitor adoption trends—utilization of the same machine types within the trade-area radius. Cell 3: Manufacturer track record—national utilization trends for the same series. Cell 4: Customer-base profile—distributions of age, visit frequency, and dwell time. Cell 5: Seasonal factors—correlations with weekdays, holidays, and paydays. Cell 6: Store characteristics—effects of location, scale, and adjacent facilities. Cell 7: Price and specs—relationship between purchase price and continuation rate. Cell 8: Market trends—popularity cycles in gameplay genres."
"Data exists for all eight elements," I added. "It just isn't used."
[Model design — Integrating the eight cells with weights]
"The design integrates the eight cells into an AI model," Claude continued. "Not a simple average—machine learning calculates each element's predictive contribution. Some stores will weigh own-store history heavily; others will weigh competitor trends heavily. The structure builds a prediction model optimized for each of the fifteen stores."
[Estimating investment recovery]
"Let's run the numbers in ROI Proposal Generator," Gemini proposed.
- Initial cost: Data infrastructure, predictive model development, store-by-store tuning, and manager operational training: 12.8 million yen total
- Monthly cost: Model operation and ongoing data collection infrastructure: 380,000 yen/month combined
- Monthly savings: Reduced miss-hit losses = 4.2 million yen/month (hit rate improving from 40% to 60%), reduced early withdrawals = 700,000 yen/month, reduced manager decision labor = 760,000 yen/month, reduced competitor research labor = 540,000 yen/month. Total: 6.2 million yen/month
- Net monthly savings: 6.2 million yen − 380,000 yen = 5.82 million yen/month
- Payback period: 12.8 million yen ÷ 5.82 million yen = approximately 2.2 months
"Just over two months to recover," Gemini summarized. "The big lever is miss-hit loss reduction. Lifting the hit rate from 40% to 60% alone produces an impact of around 4 million yen monthly."
Jinbo checked the numbers. "Seeing the AI system's introduction cost, I assumed it would take two years to recover. The moment we organized the elements into nine cells, the view shifted."
"MANDALA is a tool for seeing the whole," I responded.
Chapter 3: An AI Forecasting Rollout Designed in Nine Cells
"Let's organize the rollout," I said, standing at the whiteboard.
"Months 1–2: Clean three years of utilization and purchase history across all fifteen stores; build the data foundation. Month 3: Define input variables for the eight cells; set up competitor data collection pathways. Months 4–5: Develop the prediction model; tune individually for each of the fifteen stores. Month 6: Pilot operation; compare AI predictions against store managers' judgments. Month 7: Production launch as decision-support for managers. Month 8 onward: Monthly learning from prediction-versus-actual deltas to continuously improve model accuracy."
"Will store managers still make decisions?" Jinbo asked.
"Final judgment stays with the manager," Claude responded. "The AI provides decision material. Experience and intuition don't become worthless; the basis for the intuition gets backed by numbers. The manager's role shifts from 'decision maker' to 'judge who uses the AI well.'"
Jinbo took notes. "The field's concern that AI would replace managers should ease with this framing."
Chapter 4: The Day Intuition Was Reinforced by Data
Nine months later, Jinbo's report arrived.
The new-machine hit rate, three months after the AI prediction model went live, had risen from 40% to 63%. Miss-hit losses dropped by 4.5 million yen monthly on average. "The model is picking up variables managers weren't seeing. Interactions between competitor utilization patterns and our customer base—factors humans can't fully calculate in their heads—are now reflected in predictions," Jinbo wrote.
Manager decision labor was also substantially reduced. With AI prediction scores and judgment rationale displayed on screen, candidate machine shortlisting now took less than half the time. "We used to spend days drafting an approval document. Now a fully-supported document goes out in half a day," the report noted.
The most unexpected shift showed up in inter-manager information sharing. With each of the fifteen stores' models individually optimized, the differences between stores became quantitatively visible, and discussions about why a machine hit at one store but missed at another became routine. "Manager-to-manager exchanges used to be casual chat. Now they're structured around numbers," Jinbo wrote.
Judgment-quality risk during manager rotations was also mitigated. With the AI prediction model already trained on each store's characteristics, judgment quality held even when managers changed. "First-month hit rate after rotation used to drop to the twenties. Now it stays around 60%," the report said.
A side effect was a shift in negotiating power with manufacturers. Because the AI prediction could flag machines unsuited to the store's customer base in advance, the basis for declining unreasonable pitches became explicit. "Manufacturers also stopped pushing unreasonable proposals. The way they treat a data-equipped customer changes," Jinbo wrote.
The prediction model also keeps improving. Monthly prediction-versus-actual deltas feed back as training data, and accuracy gained four points over nine months. "It's a structure where the more you use it, the better it gets," the report noted.
At the end of the report, Jinbo wrote: "The reason AI deployments fail is that no one decides what to feed into them. The moment we filled in the nine cells of MANDALA, the design started moving. Once the center is set, the periphery follows. Experience and intuition aren't things to discard. Broken into nine cells and put into words, intuition reveals itself as learnable variables."
The mornings of writing approval documents while staring at piles of miss-hit machines had turned into mornings illuminated by data, he wrote.
"Requests for AI demand forecasting come in often. They fail because what to predict stays vague and what to feed in stays unorganized. MANDALA asks about the center and the eight directions. Set the center and the periphery is determined; align the periphery and the model can be assembled. Experience and intuition aren't to be thrown out. Break them into nine cells, put them into language, and intuition reveals itself as learnable variables. At a hall buying thirty machines at a time and missing half, on the day a manager's mental model was unfolded into nine cells, what changed wasn't prediction accuracy—it was the structure of judgment itself."
Related Files
Tools Used
- ROI Polygraph — Visualizing miss-hit losses, decision labor, and manager rotation risk
- ROI Proposal Generator — Investment recovery simulation for AI demand forecasting infrastructure
