ROI Case File No.531: No One Had Ever Measured the 'Slight Burden'
No One Had Ever Measured the "Slight Burden"
Chapter 1: The Burden Was Only Ever Spoken of as a Feeling
"Our workforce is aging. The load is concentrating on the younger staff, and at this rate they'll burn out. We want to fix it somehow with AI agents."
Soichi Ukai, CEO of TechFuture, opened with that, calling it an urgent request. The company had about fifty employees. "Fifteen at headquarters, and the rest scattered one or two at a time across regional field offices nationwide. The floor is full of manual checking work, where mistakes and lost time happen constantly."
"Specifically, which burden weighs on you the most?" Claude asked.
"That's just it—I can't say clearly," Ukai answered. "When I ask the floor, I get impressions back: 'a little hard,' 'pretty hard.' But no one can put a number on what's eating how much time. We have no in-house knowledge of AI deployment, and there are few examples in our industry. I can't even see where to begin."
"Have you ever measured the size of the burden?" I confirmed.
"Never," Ukai answered. "All we share is the felt sense that we're busy."
"If the burden is only ever spoken of as a feeling, the first thing we have to do is measure it," I responded. "Let's break it down with BOM."
Chapter 2: BOM Asks for the Parts List of Work-as-Product
"This case needs BOM."
Claude wrote three letters on the whiteboard: "BOM."
"BOM—a Bill of Materials—is a table that lists every part making up a single product, with quantities and unit costs," I explained. "Applied to work, it's powerful. You take a vague lump called 'checking work' and break it into a list of constituent tasks, giving each a weight in labor hours. Then 'a slight burden' and 'a heavy burden' can be laid side by side—not as impressions, but in the same unit of labor hours. It's a tool for measuring what can't be measured."
"Let's measure the current cost first," Gemini said, opening ROI Polygraph and entering the data Ukai had provided.
"The monthly burden cost is in," Gemini read out. "Manual checking work averages 220 hours a month; at ¥3,600 an hour, that's ¥792,000 a month. Rework and losses from checking errors average ¥650,000 a month. Effort to confirm and reconcile information dispersed across the regional offices averages ¥500,000 a month. The expected value of turnover and recruitment risk from load concentrating on young engineers averages ¥700,000 a month. Opportunity loss from initiatives stalling for lack of AI know-how averages ¥400,000 a month. The expected value of key-person risk from work depending on individuals averages ¥550,000 a month. The total is ¥3,592,000 a month—roughly ¥43.1 million a year."
Ukai stared at the figures. "What I could only call 'a little hard' comes to over forty million yen. Turning an impression into money makes that much difference."
"Then let's design it with BOM," I continued.
[Turning Work into a Parts List—Breaking the Burden into Constituent Tasks]
"First, we write out a day on the floor as a parts list," Claude said. "We break the single word 'checking work' into its constituent tasks—document review, figure reconciliation, transcription, report writing, communication between field offices. The same way you split a product into parts, you split the work down to its smallest units. As a lump it can't be touched, but split into parts, each can be evaluated one at a time."
[Weighting Each Part—Measuring 'Slight' and 'Heavy' in Labor Hours]
"Next, we assign each part a weight in labor hours," Gemini continued. "Measure how many hours a month each task takes, and you find that the figure reconciliation the floor called 'a little hard' is actually the heaviest part. Conversely, the report writing that felt 'pretty hard' sits mid-rank in labor hours. Measure the weights, and the gap between impression and reality becomes visible."
[Judging AI Substitutability—Which Parts to Hand to Agents]
"We sort the weighted parts by AI substitutability," I continued. "The routine parts with little room for judgment—figure reconciliation, transcription, standard reporting—can be handed to AI agents. Parts that require human judgment stay with people. Move the heavy parts to agents first, and the investment doesn't swing and miss."
[Reassembly—A New Process Chart for People and AI]
"Finally, we reassemble the work between people and AI," Claude continued. "We redesign the process chart for after the parts handed to AI are removed. Young staff can focus on judgment work, and even a one-person field office gets an agent to handle first response. Because there's a parts list, who carries how much of what becomes visible as a process."
[Estimating the Payback]
"Let's run the numbers with ROI Proposal Generator," Gemini proposed.
- Initial cost: AI agent build, turning work into a parts list, automating checking work, the inter-office coordination platform, and on-site training—¥7.4 million total
- Monthly cost: agent operation and ongoing model updates combined—¥260,000 a month
- Monthly savings: checking-work hours cut = ¥630,000 a month (assuming 80% reduction); rework loss cut = ¥520,000 a month; field-office confirmation effort cut = ¥380,000 a month; retention effect from easing the load on young staff = ¥450,000 a month—¥1,980,000 a month total
- Net monthly savings: ¥1,980,000 − ¥260,000 = ¥1,720,000 a month
- Payback period: ¥7.4 million ÷ ¥1,720,000 = about 4.3 months
"Payback in just over four months," Gemini summarized. "What makes it work is handing the heavy parts to agents in order. Rather than trying to automate all checking work at once, you prioritize the parts that are heavy in labor hours. Because there's a parts list, where to aim the investment is decided by numbers."
Ukai said as he checked the figures, "I left 'we're busy' as an impression the whole time. Split it into parts and assign labor hours, and you can see where to start."
"BOM is a tool for turning an impression into a parts list," I responded.
Chapter 3: A Rollout Plan That Hands Over the Heavy Parts First
"Let me lay out the approach," I said, standing at the whiteboard.
"Months one and two—interview a day on the floor, turn checking work into a parts list, weight each task in labor hours. Month three—judge AI substitutability and fix priorities. Months four and five—build AI agents starting with the heavy parts, automating figure reconciliation and transcription. Month six—pilot at headquarters and verify the people-and-AI process chart. Month seven—roll out to field offices, with agents handling first response. Month eight onward—expand the substitution range while updating the parts list, and accumulate individually held know-how into the agents."
"Will it work even at a one-person field office?" Ukai confirmed.
"It will," Claude responded. "A one-person office is painful because there's no one nearby to ask. If an agent handles routine tasks and first response, even one person can follow the same procedure as headquarters. Because the parts list spells out 'which parts the agent holds,' differences between sites are unlikely to appear."
Ukai said as he took notes, "I never thought of measuring the burden. Measure it, and the order in which to reduce it gets decided."
Chapter 4: The Day the Measured Burden Grew Lighter
Nine months later, a report arrived from Ukai.
Three months after the AI agents went live, manual checking work was down 80% versus before. "Figure reconciliation and transcription moved almost entirely to the agents. The work that ate most of a young staffer's day vanished," Ukai wrote.
Rework from checking errors dropped sharply too. Oversights stemming from manual reconciliation fell, and the floor's trust recovered. "The habit of double-checking on the assumption of errors disappeared. The checking-of-the-checking is gone," the report said.
The most unexpected change showed up in the young staff's intent to quit. With the load eased and time freed for judgment work, the mood on the floor changed. "The young people who said they were 'about to burn out' can now spend time on design and proposals. The voices saying they wanted to quit stopped," Ukai wrote.
The one-person field offices changed as well. With agents handling first response, isolated sites could now operate by the same procedures as headquarters. "The quality variance we'd resigned ourselves to with 'it can't be helped, it's just one person' got leveled out by the agent," the report said.
As a side effect, the burden became a shared language. With the burden made visible as a parts list and labor hours, the floor's discussions changed. "Meetings that stalled at the impression 'it's hard' turned into the concrete: 'this part is heavy, so we tackle it first,'" Ukai wrote.
At the end of Ukai's report, he had written this: "A burden can't be reduced as long as it stays a feeling. The moment we broke it into parts with BOM and assigned labor hours, where to lighten it from was decided. What can't be measured can't be improved."
The day a company where no one could measure the "slight burden" became one that could measure burden as a parts list, business efficiency had turned from a rallying cry into the work of removing parts one at a time, the report read.
"Companies where the floor's burden never lightens share something in common: the burden is left as a 'feeling.' 'A little hard,' 'pretty hard'—the felt sense is spoken, yet no one measures it in labor hours. What BOM asks for is the parts list of work-as-product. Split the vague lump into constituent tasks and give each a weight in labor hours. Hand the heavy parts to AI in order, and the investment doesn't swing and miss. The day a company that had never measured the 'slight burden' could measure it as a parts list, what changed was not the performance of the AI agents but the very perspective that converts burden into a unit you can measure."
Related Files
Tools Used
- ROI Polygraph — Visualizing checking-work hours, rework loss, and load-concentration risk on young staff
- ROI Proposal Generator — Payback simulation for AI agent deployment starting from a parts-list view of work
