The chairman was absolutely certain. In Peter Drucker’s telling, he divided his working time roughly into three parts. A third went to his senior people, a third to his important customers, a third to community activities.
Drucker had a test for claims like this. He liked to ask executives who were proud of their memory to write down how they spent their time, lock the guesses away for weeks or months, and in the meantime have a record kept of what they actually did.
The chairman’s secretary kept his, for six weeks. When she first came in with it, he did not believe her. The record showed almost no time in any of the three areas.
The dispatcher in the corner office
The hours had gone somewhere. Most of them, Drucker wrote, went to acting as a kind of dispatcher: keeping track of orders from customers the chairman personally knew and phoning the plant about them. Most of those orders were going through all right anyhow, and his calls could only delay them.
Memory reports where your time should have gone. Only a record shows where it went.
Drucker blamed memory, not honesty. Senior people, customers and the community were the jobs the chairman knew he ought to be doing, so “memory, obliging as usual, told him that these were the tasks on which he actually had spent his time.” Ask people where their money goes and they describe their budget. The bank statement tells another story.
It took the chairman two or three more logs, Drucker wrote, to accept “that record, rather than memory, has to be trusted when it comes to the use of time.”
What the diaries say
Drucker’s evidence was his own observation of executives, not a study. Survey researchers have since measured one version of the gap. In a 2011 article in the Monthly Labor Review, John Robinson and colleagues compared the hours people estimated they worked with what they wrote in time diaries, using American and Belgian data. Employed people tended to overestimate their work hours by 5 to 10 percent, most of it among those who estimated long weeks. People working under 30 hours tended to underestimate. Diaries aren’t perfect either; the authors suggest treating estimates and diaries as “highly correlated but essentially different.”
Drucker called his chapter “Know Thy Time.” It sits second in The Effective Executive, published in 1967, and it turns the chairman’s lesson into a routine. Effective executives, he wrote, “do not start with their tasks. They start with their time.” The record has to be kept in “real” time, as things happen rather than from memory, for three to four weeks at a stretch, twice a year or so. Then each item faces a blunt question: “What would happen if this were not done at all?” Whatever survives gets gathered into large blocks, because hard work needs a minimum stretch of time. “To spend in one stretch less than this minimum is sheer waste,” he wrote.
He didn’t promise the log would stay clean. Without constant effort, he warned, time drifts back.
Shopify ran a company-sized version in 2023, counting in events and hours the recurring meetings it deleted and later building a calculator that puts a price on any meeting of three or more people. A count of hours removed still says nothing about where those hours went next. Only a log can show that.
Eleven hours of checking
Imagine you lead a support team of five and you’d say your week goes mostly to coaching. You write that guess on a card and put it in a drawer. For two weeks you note what you’re doing every half hour.
The log shows six hours of coaching. It also shows eleven hours of re-reading tickets your team had already closed correctly, because checking felt safer than trusting. That’s the chairman’s dispatcher job, with a help desk instead of a plant. You hand spot checks to a weekly rota and put coaching into two fixed blocks on Tuesday and Thursday mornings.
The speedup nobody logged
In November 2025, Anthropic published an estimate of how much time AI saves people. It sampled 100,000 conversations on Claude.ai and had Claude itself estimate how long each task would take with and without AI’s help. By those estimates, the median saving was about 80 percent, and the savings varied widely by occupation.
Both of those times are estimates, and Anthropic checked its estimator the way Drucker would have: against a record. It used thousands of software tasks from open-source projects, logged as tickets with both a developer’s estimate and the time actually tracked. On 1,000 of them, the developers’ own estimates had a rank correlation of 0.50 with the tracked times, where 1 would mean they put the tasks in exactly the same order as the log. Claude Sonnet 4.5, given only each ticket’s title and description, reached 0.44. That check was about how long tasks take. No one timed the same tickets with and without AI, so it doesn’t test the 80 percent. Its guesses were also squeezed together, too long for short tasks and too short for long ones.
The company named the biggest blind spot itself. The estimates can’t count time people spend on a task outside the conversation, including checking whether Claude’s work is right. That is exactly the time a log would catch.
You can run Drucker’s drawer test on yourself. Write down how much time you think an AI assistant saves you each week, and put the note away. For two weeks, log the start and end of each task, whether you used the assistant, and the minutes spent prompting, waiting and checking its work, including the checking you do after the chat is closed. Then open the drawer. Two weeks is a lead, not a verdict, since the tasks you hand to the assistant may differ from the ones you keep. If the record and the note disagree, keep logging.
The chairman needed two or three more logs before he believed the first one.