On January 19, 2010, Japan Airlines filed for bankruptcy protection. By the count on Kazuo Inamori’s official site, its debt came to about 2.3 trillion yen. That day its shares fell as low as 3 yen. “Everyone believed secondary bankruptcy was inevitable,” Inamori said later.
He was the man the government had asked to save it. Inamori had founded Kyocera and KDDI, would turn 78 at the end of the month, and said he would do the job without pay. His official archive describes the staff he found as lacking a sense of crisis, or any feeling that the company was theirs.
Hajime Tozaki, an aviation expert at Waseda University, described how the old JAL had kept score. It looked at income and spending across the whole network and paid little attention to whether individual routes made money.
Inamori wanted to know exactly. He rearranged the accounting system he had built at Kyocera to fit an airline. By his account, the profit of every route and every flight became available the following day.
The question was who would read those numbers first.
The report goes to the unit
Peter Drucker had asked that question in 1954. In The Practice of Management, his chapter on management by objectives gave it a partner: self-control through measurement. Each manager, he wrote, “should have the information he needs to measure his own performance and should receive it soon enough to make any changes necessary for the desired results.”
Then he added: “this information should go to the manager himself, and not to his superior. It should be the means of self-control, not a tool of control from above.”
Give people a goal and their own numbers in time to act, and they can steer themselves.
A fuel gauge lets a driver change how she drives while the tank is still half full. A fuel report sent to the fleet manager a month later can only judge the trip.
Drucker set two companies side by side. At General Electric, traveling auditors studied every managerial unit at least once a year, “but their report goes to the manager of the unit studied.” At a large chemical company, a control section audited every unit too, but the results went only to the president, who then called in the managers. The managers had a nickname for the control section, he reported: “the president’s Gestapo.” More and more of them, he wrote, were running their units for the best showing on the audits rather than the best performance.
Three days a month
JAL did both. Inamori’s next-day figures gave employees, in his words, “real-time awareness of their results.” The airline’s own report says a divisional profitability system was put in place from April 2011 “to raise the profit consciousness of each department.”
The figures also went up. Each month, departments presented them over a three-day meeting while Inamori, by his account, told their leaders “the next steps to take.” The same numbers worked as a mirror and as a review.
Whether the mirror is what turned JAL around, the case can’t tell you. Banks forgave 521.5 billion yen of debt, a government-backed fund put in 350 billion, and the plan called for about 15,600 job cuts. Inamori himself didn’t credit the accounting alone. In 2012 he told more than 4,000 business managers that “the revival of Japan Airlines did not result from my actions.”
The number that arrives too late
Imagine you lead the night shift at a warehouse. Your shift’s picking errors reach you once a month, as one line on your manager’s slide, three weeks after the fact.
You ask for the count by aisle, sent to you the morning after each shift. In the second week, the errors bunch up on the two nights a new temp covers one aisle, where two products share a shelf. You relabel the shelf and walk him through it. The monthly line still goes to your manager, and now it arrives with a reason next to it.
Asked to check its own work
AI systems are often told to measure themselves. A common prompt: review your answer and fix any mistakes.
In a paper at ICLR 2024, researchers from Google DeepMind and the University of Illinois tested that on reasoning benchmarks, with models available in 2023. With no outside feedback, the models scored lower after reviewing their own answers. GPT-3.5 fell from 75.8 to 38.1 percent on a commonsense quiz after one round. On grade-school math it mostly kept its answers, and when it changed one, it was more likely to turn a right answer wrong than a wrong one right. Earlier reports of gains, the authors found, had used the correct answers to decide when to stop. If you already have the ground truth, they wrote, there seems to be little reason to use the model at all.
Drucker’s version needed a measurement, and a second look isn’t one. For an agent, that means a check that doesn’t depend on its own opinion, such as tests it can run or a handful of cases with known answers. Ask it to report the check’s raw output, then run the same check yourself on a copy before you believe the summary.
At JAL, the figure for each flight arrived the day after it flew.