ROI Case File No.463 'A Thousand Needles, Two Maps'
A Thousand Needles, Two Maps
Chapter 1: I Know What AI Can Do. I Don't Know Where to Start.
"I'm starting to understand what AI can do. But I don't know where to begin."
Kosuke Hamada, Quality Assurance Manager at TechNova, opened a notebook as he spoke. The pages were dense with small writing, almost no margins left. Personal logs of experiments with ChatGPT and Copilot sat side by side with operational pain points collected from across the organization.
"Some staff are already using it individually," I confirmed.
"Yes," Hamada nodded. "But it's scattered. Someone uses it for document creation, someone else for translation, others barely at all. There's no company-wide strategy. From management, we've been told to start small and build a success story. The problem is, we can't agree on where to place that small start."
"What operations are you considering?" Claude asked.
"Two of them," Hamada answered. "The first is automated sample selection. In our quality inspection process, deciding which samples to pull from which lots is currently done by experienced staff using intuition. Three people spend two hours on this, three times a week. We think we could train AI on the decision criteria and automate it."
"And the second?" Gemini prompted.
"Quality assurance proposal writing. Four proposals per month, six hours each. The data is scattered everywhere—just gathering it consumes half the time. Two people handle this. We're wondering if AI could integrate the data and auto-generate proposal drafts."
"You're stuck on which one to do first," I confirmed.
"That's right," Hamada said. "Both feel important intuitively. I want to make the decision with data."
"Let's get the data," Gemini said quietly.
Chapter 2: ROI Analysis and the Question of Priority
"This case calls for ROI analysis."
Claude drew two vertical lines on the whiteboard. Sample selection on the left. Proposal writing on the right.
"ROI stands for Return on Investment," I explained. "When deciding which operation to invest in first, making that call based on gut feel or the loudest voice causes small starts to fail. ROI analysis lines up investment, savings, and payback period in numbers, and asks: which one comes first? Running two operations simultaneously isn't a small start—it's risk dispersion disguised as focus dispersion."
"Let's measure current labor with ROI Polygraph," Gemini said, opening his laptop and entering the operational logs and interview results Hamada had provided.
The numbers came back.
"Here are the current costs for sample selection," Gemini read aloud. "Three staff, three times per week, two hours each. Monthly labor: 72 hours. At ¥3,200/hour, that's ¥230,400/month. Annualized: ¥2,764,800."
Hamada raised an eyebrow. "That much for one operation."
"For proposal writing," Claude continued: "Two staff, four proposals per month, six hours each. Monthly labor: 48 hours. Same hourly rate: ¥153,600/month. Annualized: ¥1,843,200."
"Combined, these two operations alone are consuming ¥4,608,000 per year," Gemini summarized.
"Now let's project the post-AI numbers," I said, opening ROI Proposal Generator.
[ROI Projection: Sample Selection Automation]
"AI automation of sample selection requires historical data as a prerequisite," Claude explained. "The precision of the model depends on how much past sampling records and inspection results have been accumulated. How complete is the historical data?"
"We have five years' worth," Hamada answered. "But the format varies by staff member."
"One month for data cleansing, two months for model training and validation—three months of preparation total," Gemini projected. "After that, based on comparable cases, we'll estimate a 65% reduction rate. Monthly savings: ¥149,760. With an initial investment of ¥800,000, payback is 5.3 months. Add the 3-month prep period and the real-world payback is 8.3 months."
[ROI Projection: Proposal Writing Automation]
"The core challenge with proposal writing is the scattered data," Claude continued. "Before AI can generate drafts, there needs to be a foundation that consolidates data in one place. Where is it scattered right now?"
"Across servers in three departments, individual PCs, and emails from external testing organizations," Hamada answered.
"Two months to build the data integration layer, one month to implement and tune AI draft generation—three months of preparation," Gemini projected. "Reduction rate from data collection automation alone: 50%. Monthly savings: ¥76,800. Initial investment ¥600,000. Payback: 7.8 months. Including prep: 10.8 months."
[What the Two Projections Say]
"Here's the summary from ROI Proposal Generator," I said, writing the numbers on the whiteboard.
- Sample selection: Monthly savings ¥149,760, investment ¥800K, payback 8.3 months
- Proposal writing: Monthly savings ¥76,800, investment ¥600K, payback 10.8 months
- Combined monthly savings (once both are complete): ¥226,560 / approx. ¥2.71M/year
"Both savings impact and payback speed favor sample selection first," Gemini said quietly. "The first move of the small start is sample selection automation. The numbers have determined the order."
Hamada exhaled deeply. "Intuitively, proposal writing felt more visibly painful, so I assumed it should come first."
"Pain and ROI are different questions," I replied. "Pain is an emotional question. ROI is an investment question. Starting with the higher number is also the language that works for management."
Chapter 3: How to Use Two Maps
"With the sequence decided, let's design the execution plan for each," I said, stepping to the whiteboard.
"Sample selection automation moves in three phases," Claude said.
"Phase 1 (1 month): Data cleansing. Convert five years of historical data to a unified format. Absorbing the differences between individual staff formats is the core work of this phase. Cut corners here, and AI judgment accuracy suffers."
"Phase 2 (2 months): Model training and validation. Train AI on the sampling criteria and have staff run in parallel on real cases to verify accuracy. During this period, both AI and humans operate simultaneously. Trust AI only when its decisions align with staff judgment at 90% or above."
"Phase 3 (post-launch): Redefine staff roles," Gemini added. "Decide in advance what the three staff members will do after automation. If the 72 freed-up hours don't flow toward new value creation, the savings don't translate into organizational benefit."
Hamada took notes and said: "One of the three has been saying for a long time she wants to focus on anomaly analysis. This creates that time."
"That is the real outcome of a small start," I said. "Not reducing time, but designing what to do with the time you reduce. That is where ROI completes itself."
"For proposal writing," Claude continued, "we start three months after sample selection is on track. Build the data integration layer, then implement draft generation on top. Not running both simultaneously is the actual meaning of the phrase 'small start.'"
Chapter 4: The Day the First Needle Was Threaded
As Hamada prepared to leave, he closed his notebook and said:
"I came here not knowing where to start. Today, the order was determined with numbers. ROI analysis is a framework for deciding priority—not which is right, but which comes first. That's what I understand now."
"The phrase 'small start,'" I replied, "doesn't mean starting small. It means starting in the right order. When there are a thousand needles, deciding which one to thread first is what 'small start' actually means. ROI analysis is the tool that finds that first needle."
Outside the window, white steam rose quietly from the factory chimney.
Seven months later, a report arrived from Hamada.
The sample selection AI reached 92% accuracy in its fourth month of operation and graduated from parallel running with staff. Monthly labor dropped from 72 to 24 hours. Using the freed-up time, one staff member began dedicating herself to anomaly pattern analysis—and discovered a equipment deterioration signal that had gone undetected for three years. Early intervention based on that discovery avoided an estimated ¥4M in product scrapping risk, the report noted.
Approval for proposal writing automation came through after the success story from sample selection was presented to management, with work beginning in month six. Hamada wrote: "Without the first success story, the budget for the second would never have been approved. The sequence opened the next door."
Of a thousand needles, the first one had been threaded.
"A small start is not starting small. It is starting in the right order. ROI doesn't ask which is right—it asks which comes first. Running two operations simultaneously is a dispersion of focus, and the seed of failure. Thread the first needle and the path to the second becomes visible. Needles must be threaded one at a time—however many there are, the order is always one."
Related Files
Tools Used
- ROI Polygraph — Operational labor and cost measurement, savings visualization
- ROI Proposal Generator — Dual-operation ROI comparison and payback simulation
