The district service manager for Denver had just had his furnace repaired. At a meeting he called, he told a story about the repairman, who carried a portable radio and kept talking on it with other technicians fixing other furnaces elsewhere.
The manager’s own people fixed Xerox photocopiers across Colorado and eastern Wyoming, mostly alone, in other people’s offices. The company’s official position, as the anthropologist Julian Orr and the engineer Norman Crowfoot later wrote, was “that a technician only need know how to read and use hand tools”; the service manual directed the rest. The company tracked calls taken, calls closed, the time each took and parts used.
The people at the table decided to look into radios, so technicians could talk to each other on the road. Then came the catch. A radio would let a manager reach a technician. It would also let a manager listen.
Everyone agreed that monitoring would be a temptation, and that if it was done even once, it would seriously damage what the radios were for. The people in charge of the work had just decided not to watch it.
Where the stopwatch stops
Peter Drucker had been writing about work like theirs for forty years. In a 1999 article he credited Frederick Taylor with the century’s great gain in manual workers’ productivity: look at the work, break it into motions, time each one. But Taylor’s question, Drucker wrote, was how the work should be done. “In manual work, the task is always given.” For knowledge work the first question is different: “What is the task?”
The wheel arriving on the assembly line tells the worker what to do next. In knowledge work, Drucker wrote, “the task does not program the worker”; outside a crisis, a nurse largely decides whether to spend the hour at a bedside or on paperwork. And in most knowledge work, quality is “the essence of the output.”
Work done in the head can’t be run by watching. Start by asking what the task is.
Watch a chess player for ten minutes and you may see her hand move twice. The game was played in between.
Drucker’s method starts with the workers: ask them what their task is, what it should be, what they should contribute and what gets in the way. Knowledge workers, he wrote, “have to manage themselves. They have to have autonomy.” The label was his. In The Atlantic Monthly in 1994 he wrote: “I coined it in a 1959 book, Landmarks of Tomorrow.”
The copier company’s counts recorded the hands. Orr had found the knowledge elsewhere, in “war stories, anecdotes of experience,” the technicians’ shared memory of hard diagnoses.
Last on the dashboard
Imagine you lead six support engineers, and a new tool ranks them by tickets closed. One engineer is last by a wide margin. She is also the one the others message when a ticket makes no sense, and none of those messages is a ticket.
Before answering your director, you ask each person Drucker’s questions. Five name the same obstacle: tickets bounced back for missing fields. Only her answer includes getting whoever is stuck unstuck.
Make her number the target and the messages are the first thing she has reason to drop. It could show up in the team’s reopened tickets, weeks later.
A brief with the thinking done
AI turns half of Drucker’s problem around. You can log every tool call an agent makes, though not necessarily the reasoning behind it. That still doesn’t tell you whether the work was good, or the right work.
In a paper posted in October 2025, OpenAI researchers described GDPval, a test built from real work by professionals averaging 14 years’ experience in 44 occupations. Experts in each field judged quality, as Drucker said knowledge work must be judged: they ranked unlabeled deliverables, a model’s against a professional’s. On a 220-task subset, 47.6 percent of deliverables from Claude Opus 4.1, the top model, were rated as good as or better than the expert’s.
The paper names what it left out. Each task came with its full context, “but in real life it often takes effort to figure out the full context of a task and understand what to work on.” In a separate experiment on an earlier version of the task subset, when prompts were cut to 42 percent of their length on average, GPT-5’s work was rated worse; it “struggled to figure out context.” Tasks needing extensive tacit knowledge or communication between people were out of scope, the kind of knowledge the Denver radios were meant to carry.
With AI, Drucker’s first question moves to whoever writes the brief. Take a piece of work your team finished last month and write the brief twice: the one line you’d normally type, and a full version saying who it’s for, what good looks like and where the data is. Run each a few times. Have the colleague who normally reviews such work rank those outputs and your team’s version, unlabeled, and say why. One ranking is a lead; check the reasons against what the work was for.
Six months, no number
Five work groups, 21 technicians at first, carried radios for a six-month test that ended on December 31, 1991.
The authors reported “no formal results”; the effect couldn’t be measured “with the tools available.” That is the hard edge of Drucker’s idea: stop watching, and you need another way to know. In that article, his evidence is anecdote. Asking his questions and taking action on the answers, he wrote, “usually doubles or triples knowledge-worker productivity, and quite fast”; he cited no data. In a 2001 paper the organization researcher Mats Alvesson took a sceptical view: knowledge is slippery, and where competence and performance are hard to demonstrate, managing rhetoric and image becomes central. The line between hand and head blurs too. Drucker said the manual parts of skilled work, like a surgeon’s, should still be studied by Taylor’s methods.
The technicians called the radios a great success. Some offered to buy them, to make sure they could keep them.