ROI Case File No.437 'Before the Red Light Comes On'
Before the Red Light Comes On
Chapter 1: Watching Without Seeing
"Fifty thousand rows of data accumulate every day. Yet breakdowns still happen every month."
The head of production engineering at TransTech Industries displayed a screenshot of the factory's control panel. Temperature, vibration, current readings, operating hours—data automatically collected from equipment flooded the screen like a torrent of numbers.
"We're a global manufacturer of automotive electronic components. We operate five factories in total—three domestic, two overseas. Our main factory houses approximately 280 production machines, each fitted with sensors that continuously collect operational data."
The production engineering director showed the database capacity: approximately 2.4 terabytes over the past three years.
"The problem is how this data is being used."
"What's the current approach?" I asked.
"Every morning, a five-person equipment maintenance team downloads the previous day's data into Excel and manually creates charts. They visually check whether temperature or vibration readings have exceeded thresholds and report any anomalies to the shop floor. This takes about two hours every morning."
"Two hundred eighty machines checked visually by five people," Claude confirmed. "That's fifty-six machines per person. About two minutes of review time per machine."
"Exactly. In two minutes, the best you can do is check whether values exceeded thresholds. There's no time to spot subtle trend changes—for instance, vibration readings that are within threshold but have been gradually rising over the past month."
The production engineering director pulled out another document—a record of equipment failures over the past year.
"Unexpected breakdowns average 4.2 per month. Each breakdown stops the production line for an average of 6.8 hours. The opportunity cost during downtime is approximately 1.8 million yen per hour."
"The monthly loss is," Gemini calculated, "4.2 incidents times 6.8 hours times 1.8 million yen—roughly 51.4 million yen. Annualized, that's approximately 620 million yen."
"Furthermore," the production engineering director continued, "emergency repairs after a breakdown cost about 2.5 times more than planned maintenance. Rush parts procurement, maintenance crew overtime on holidays, premium rates for outside contractors—annual emergency repair costs total approximately 180 million yen."
"There's plenty of data. But the data isn't being used before breakdowns occur," I summarized.
"Exactly. I've heard that AI can enable predictive maintenance. But for a factory like ours—where data exists but isn't being leveraged—will AI really work? Having data and using data are two different things, and I don't know how to bridge that gap."
This wasn't an AI adoption problem. It was a case that required identifying how much waste was hiding in the process between data collection and maintenance decision-making.
Chapter 2: Hunting for Seven Wastes
"Before introducing AI, let's strip the waste from the current process."
Gemini wrote "LEAN" in large letters on the whiteboard and listed seven items beneath it.
"LEAN," I began, "is a philosophy of value creation through the relentless elimination of waste, rooted in the Toyota Production System. At its core is the identification and elimination of seven wastes: overproduction, waiting, transportation, over-processing, inventory, motion, and defects. Viewing the current data utilization process through a LEAN lens reveals room for improvement before AI even enters the picture."
"Do manufacturing waste concepts apply to equipment management processes?" the production engineering director asked.
"They do," Claude answered. "LEAN's seven wastes apply not only to factory production lines but to any process where information flows. Think of the flow from data collection to maintenance decisions as an information production line."
[Waste 1: Overproduction—Data That's Never Used]
"First, the waste of overproduction," I began.
"Fifty thousand rows of data are collected daily from 280 machines," Gemini confirmed. "But how much of that does the maintenance team actually review each morning?"
The production engineering director paused to think. "Honestly—about 15% of the total. Temperature and vibration readings from the top fifty critical machines only. Data from the remaining 230 machines is only looked at retroactively after a breakdown."
"You're collecting 50,000 rows daily, yet 85% simply accumulates unused," Claude pointed out. "On a production line, that's the equivalent of continuously manufacturing products that never sell. Before increasing data collection, you need to define which data is truly necessary for maintenance decisions."
[Waste 2: Waiting—Data That Sleeps Until Morning]
"Second, the waste of waiting," I continued.
"Equipment data is collected in real time," Gemini confirmed. "But humans don't review it until the next morning. That means up to a twenty-four-hour lag between when an anomaly occurs and when someone notices."
"Have breakdowns occurred during that twenty-four-hour window?" I asked.
"Yes," the production engineering director answered immediately. "Of last month's four breakdowns, two involved vibration values spiking overnight during operation. By the time of the morning check, the parts had already failed. The line was already down."
"Data is being generated around the clock, but humans review it only once every eight hours," Claude summarized. "Eliminating this waiting waste alone could prevent roughly half of the breakdowns."
[Waste 3: Over-Processing—The Excel Detour]
"Third, the waste of over-processing," I pointed out.
"Every morning, data is downloaded from the database into Excel, and charts are manually created," Gemini questioned. "Is this step truly necessary to produce the value of a maintenance decision?"
"If you connect a dashboard directly to the database for real-time visualization, the Excel download and chart creation become unnecessary," Claude made concrete. "The two hours of daily data processing can be permanently eliminated with a single dashboard build."
[Wastes 4 Through 7: The Remaining Four]
"Let's briefly organize the remaining four wastes," I suggested.
"Waste 4—inventory waste," Gemini listed. "Spare parts held in anticipation of breakdowns are excessive. If predictive maintenance becomes possible, needed parts can be planned and procured before failure, reducing emergency inventory."
"Waste 5—transportation waste," Claude continued. "When the maintenance team discovers an anomaly, they print the data, bring it to the shop floor, and verbally explain it to the equipment operator. This information transportation becomes unnecessary with tablets or mobile alerts."
"Waste 6—motion waste," I added. "The very act of the maintenance team reviewing data for 280 machines one by one can be replaced by exception management—a system that automatically detects anomalous values and issues alerts."
"Waste 7—defect waste," Gemini concluded. "Emergency repairs after breakdowns are equivalent to quality defects. Redirecting resources from reactive response to prevention eliminates the defect waste at its source."
Chapter 3: Strip the Waste, Then Layer On AI
The production engineering director stared at the whiteboard where seven wastes had been listed.
"I was approaching this as an AI implementation project. But switching from Excel to a dashboard, adding alerts to close the twenty-four-hour lag—these are improvements that come before AI."
"The essence of LEAN," I responded, "is to first remove waste and make the process lean before layering on new technology. If you introduce AI into a waste-filled process, AI merely becomes a waste accelerator. If Excel manual work remains when you add AI, a new waste is born—transcribing AI output back into Excel."
"Execute in three phases," Claude proposed. "Phase one: waste elimination. Build the dashboard, implement automatic threshold alerts, narrow down which data to monitor. None of this requires AI. It can be achieved through configuration changes to existing BI tools and databases. Duration: one month."
"Phase two: AI introduction," Gemini continued. "On top of the lean process with waste removed, layer the AI predictive maintenance model. Analyze trends in vibration and temperature data to issue alerts seventy-two hours before a failure. Start with the thirty machines with the highest breakdown frequency. Duration: three months."
"Phase three: continuous improvement," I added. "Monitor AI prediction accuracy and breakdown trends monthly, analyze cases where predictions were wrong, and keep updating the model. LEAN is not a one-time improvement—it's a perpetual process of seeking out waste."
"Start the pilot with one line at the main factory," Gemini added. "Thirty machines, three months of data. Measure how much effect waste elimination alone produces, without AI. Confirm the improvement from waste removal, then move to phase two. This sequence matters."
"And," I reminded, "record the hours, cost, and time for each of the seven wastes—before and after elimination. Which waste removal produced which effect. This record becomes the blueprint for rolling improvements out to other factories and lines."
The production engineering director stood and bowed deeply. "Thank you. This week, we'll start by discontinuing the Excel data processing and building the dashboard."
Chapter 4: People Move Before the Light Comes On
After he left, Claude said, "LEAN's seven wastes are strongly associated with production lines, but they map beautifully onto data utilization processes."
"Indeed," I answered. "What LEAN teaches is a rigorous perspective: any step that doesn't create value is waste. Data collection, Excel transcription, chart creation—none of these create value in themselves. Value is created only at the moment a maintenance decision is made based on data. Strip waste from every step leading to that moment, one by one. Then, when you layer AI on top, AI can focus entirely on improving the precision of that decision. Get the sequence wrong, and AI becomes a new source of waste."
Gemini added, "And the seven wastes don't disappear once you've eliminated them. When the process changes, new wastes emerge. After AI is introduced, a new waste might appear—responding to false alerts from the AI. Periodically auditing the process, finding new wastes, and eliminating them continuously—that repetition is what reproducibility means in LEAN."
Outside the window, the night lighting of the industrial district faintly colored the sky.
Five months later, a report arrived from TransTech Industries.
Phase one—waste elimination alone—produced dramatic results. The dashboard build and automatic alerts completely eliminated the two hours of daily data processing. The maintenance team now spent their mornings responding to anomaly alerts and conducting visual equipment rounds—in other words, judgment work that only humans can do.
In phase two, an AI predictive maintenance model was deployed on the top thirty machines. Trend analysis of vibration and temperature data enabled alerts an average of forty-eight hours before failure.
Results over five months: unexpected breakdowns dropped from a monthly average of 4.2 to 1.1. Line downtime was reduced from 28.6 hours per month to 5.8 hours. Emergency repair costs were projected to decrease from approximately 180 million yen to about 42 million yen annually.
But the change the production engineering director was most proud of was the maintenance team's mindset. A team that once passively treated data as "something to look at every morning" now proactively audited their own processes, asking "Is there waste here?" One maintenance technician independently analyzed the AI's alert criteria and proposed dynamically adjusting thresholds to account for seasonal temperature fluctuations.
At the end of the report, the production engineering director wrote: "We review LEAN's seven wastes at every monthly team meeting. Last month, the newly discovered waste was 'double-checking equipment data that the AI had already cleared as normal, just to be safe.' Now that AI accuracy has stabilized, this double-check is waiting waste. Every time we eliminate one waste, the next one becomes visible. The fact that this chain never stops—that's the power of our shop floor."
Before the red light comes on, data speaks and people move. The power to keep stripping waste from that flow—that was the factory's reproducibility.
"Having data and having data that creates value are different things. What LEAN asks is: how much waste hides in the process from data collection to decision-making? Identify the seven wastes one by one, eliminate them, and make the process lean. Then layer new technology on top. By respecting this sequence, technology becomes a value amplifier. And waste doesn't disappear once eliminated. Every time the process changes, new waste is born. Instilling in the organization the eyes to keep finding it—that is the essence of reproducibility in LEAN."
Related Files
