ROI Case File No.451 'The Factory with an Iron Memory'
The Factory with an Iron Memory
Chapter 1: A Memory from Thirty Years Ago
"Build or buy — that's what we can't figure out."
Ryosuke Murakami, IT Promotion Manager at TechNova, spread a single system architecture diagram across the table as he spoke. The paper had yellowed. In the bottom-right corner, a creation date read 1995.
"Our core system was built on an IBM-based platform, developed nearly in-house. It's been running for thirty years."
"Still in active use?" Claude confirmed.
"Yes. We can't shut it down. The production line depends on it. The problem is—" Murakami tapped a point on the diagram. "This data format. It outputs in a proprietary format that modern AI tools can't read directly."
"What specific operations are you hoping to apply AI to?" I asked.
"Four things," Murakami said, folding down his fingers. "Anomaly detection, demand forecasting, similar drawing search, and automated quote generation. Quotes in particular take an average of three hours each. We have four sales staff. In a busy month, we handle over forty quotes. If you do the math—"
"A hundred and twenty hours a month, gone to quotes alone," Gemini said quietly.
Murakami nodded. "That's where we want to apply AI. But our data is in three separate places: CSV output from the core system, CAD data, and PDF drawings from customers. We've been told off-the-shelf products can't handle the integration. So we've been considering custom development — but the cost and time are too much to decide on."
The heart of the case was buried in those words. "Build or buy" was not the real question. The real question was: where do you start?
Chapter 2: The Five Temperatures HEART Asks
"This case calls for HEART."
Claude wrote five letters on the whiteboard: H · E · A · R · T.
"HEART is a user experience evaluation framework proposed by Google," I began. "Happiness, Engagement, Adoption, Retention, Task success — five metrics that determine whether a tool takes root on the ground. Before deciding between build and buy, we need to take the temperature of the people who'll use it."
"Let me start by running the current state through ROI Polygraph," Gemini said, opening a laptop. The operational logs provided by Murakami were entered into ROI Polygraph. The tool returned its results.
"Interesting numbers," Gemini said, studying the screen. "Of the three hours spent on quote generation, AI can replace sixty-eight percent of the routine work. The remaining thirty-two percent involves judgment calls based on the staff member's experience."
"In other words," I picked up, "if you try to replace everything with a custom-built AI, you have to engineer that thirty-two percent of 'judgment' into the system. That's what's driving up development costs."
Murakami raised an eyebrow. "Now that you say it — the part that takes the longest on a quote is reading the specs and pulling up similar past jobs from memory. The actual data entry isn't the bottleneck."
"That's the starting point for HEART," Claude continued.
[H — Happiness: Will the people on the ground want to use it?]
"Murakami-san, how do the four sales staff feel about quote creation?"
"Every one of them says it's a burden," he answered immediately. "Especially at month-end — overtime happens because of quotes."
"The foundation for Happiness on the ground is there. If staff can feel that a tool is cutting their workload, they'll accept it. The question is—"
"Whether they'll actually feel it," Murakami finished.
[E — Engagement: Does it dissolve into daily work?]
"Engagement problems show up in usage frequency," Gemini continued. "Quote creation happens forty-plus times a month. Anomaly detection and demand forecasting, by contrast, risk going unused unless staff actively check them. Quote support is a tool people will use every day. Anomaly detection risks becoming something people say they 'should have' without ever opening."
"Then," I said, drawing together the thread, "the first thing to implement is quote generation support. Start where Engagement is highest."
[A — Adoption: Can staff who aren't tech-savvy use it?]
"The Adoption angle is what drives the build-versus-buy decision," Claude pointed out. "What's the age breakdown of the sales staff?"
"Two are in their thirties, two are in their fifties. The two in their fifties aren't comfortable with systems in general—"
"That's the fork in the road. The key to high adoption is the first experience. Hand over a complex tool first, and the two in their fifties won't keep up. Among existing products, start with the one with the simplest UI."
[R — Retention: Are they still using it three months later?]
"Retention is directly tied to the data integration wall," Gemini continued. "With CSVs, CAD, and PDFs scattered in separate places, staff have to manually assemble data every time they use the tool. That friction drives drop-off at the three-month mark. Before bringing in any AI tool, the priority is building a data foundation that brings those three types into one place."
"So that's — the part that needs custom development," Murakami realized.
"To be precise," Claude corrected, "build the data integration layer in-house, and use an off-the-shelf product for the AI functionality that runs on top. You don't need to build everything from scratch."
[T — Task Success: Does the job actually get done?]
"The final metric, Task Success, is the simplest question," Gemini concluded. "Does the quote get finished faster, at the same quality? ROI Polygraph estimates that with a data integration foundation in place and an off-the-shelf AI on top, quote creation time can be reduced from three hours to around fifty minutes per job. That's roughly eighty hours a month in workforce savings across forty quotes."
Chapter 3: The Map at the Fork
"To summarize," I said, standing at the whiteboard. "This decision isn't all-custom or all-off-the-shelf. HEART's five metrics have drawn a clear line between what to build and what to buy."
Claude sketched a diagram.
- Build in-house: The data integration layer that unifies three data types
- Use off-the-shelf: Quote support AI running on top (one function to start)
- Subsequent phases: Anomaly detection, demand forecasting, similar drawing search
"Let's model the investment plan in ROI Proposal Generator," I suggested. Data integration build cost, off-the-shelf AI implementation cost, workforce reduction impact — the estimate came back. Payback period: eleven months. Compared to a full custom build, the payback period was less than half.
"Why is the gap that large?" Murakami asked.
"With a full custom build, there's a dead period before completion where you receive no operational benefit," Claude answered. "With off-the-shelf plus a custom foundation, value begins the moment the data integration goes live. The value of time is what's showing up in the numbers."
Murakami nodded slowly. "Understood. We start by building the data integration foundation, then pilot one quote support AI on top of it."
"Set the pilot period at three months," Gemini added. "Record the five HEART metrics weekly, and let the numbers make the decision at three months. Use those results to determine expansion into anomaly detection and demand forecasting. That sequence is the safest path forward for a factory that's been running on a thirty-year-old core system."
Chapter 4: The Morning the Iron Memory Stirs
Murakami stood and folded the 1995 system diagram with care.
"For thirty years, that system ran without a fault. That's exactly why no one wanted to touch it. This time too — build the foundation before layering AI on top. Same thinking. Don't break what's underneath. Add to it."
"The essence of HEART," I replied, "is not the performance of the tool. It's designing the experience of the people who use it. No matter how capable an AI, if the staff stop using it, it's worth nothing. The five metrics need to read normal from the very beginning. That question comes before build or buy."
Outside the window, morning light over the industrial district was already catching on metal rooftops.
Five months later, a report arrived from Murakami.
Three weeks after the data integration foundation was built, pilot operation of the quote support AI began. Over the three-month pilot, all five HEART metrics exceeded baseline. The most significant change: the two staff in their fifties began voluntarily increasing their own usage — saying, "This actually works."
Quote creation time per job dropped from three hours to forty-seven minutes. Monthly workforce savings reached roughly eighty-five hours. The estimated payback period was eleven months; the projection was revised to nine months ahead of schedule.
In his report, Murakami wrote: "When we decided to build the data foundation first, there was internal pushback — voices that wanted to start using AI right away. But I believe no one would have used a tool without the foundation in place. HEART was the framework that showed us the right order of operations."
The day a factory with an iron memory began learning new words.
"Build or buy is not the real question. The real question is whether the experience of the people using it has been designed. The five metrics HEART asks — Happiness, Engagement, Adoption, Retention, Task Success — are none of them about the tool. They're all about people. The reason a thirty-year-old system kept running is that the people on the floor trusted it and kept using it. A new AI should be designed from the same question."
Related Files
Tools Used
- ROI Polygraph — Operational log analysis and visualization of automatable workload
- ROI Proposal Generator — Investment payback simulation: custom build vs. off-the-shelf
