In 1978, Masaru Ibuka, Sony’s co-founder and a serious music lover, packed the company’s newest portable on a business trip abroad. The TC-D5 could record and play in stereo. It was also the size of a textbook and weighed about 1.7 kilograms. On the plane, with a pair of headphones, Ibuka enjoyed the music. He found the machine far too heavy.
Before his next trip to the United States, he went to Norio Ohga with a request. “I can do without recording, so could you just install the circuits for stereo sound into this Pressman?”
The Pressman was a mono tape recorder small enough to sit in the palm of a hand. Engineers took the recording function out and put stereo in its place. When Akio Morita, the other co-founder, heard it, he said: “This might be big.” He wanted it in shops in four months.
Not everyone was convinced. “No one wants a ‘tape player’ that can only play tapes. If it can’t record anything, it’ll never sell,” some at Sony objected.
A recorder with something missing
The objection made sense. Judged as a recorder, the new machine did less than the one it came from.
Ibuka had never asked for a better recorder. He wanted stereo music on a long flight without carrying 1.7 kilograms. Judged against that, recording was something he had already said he could do without.
People hire a product to get a job done, and the job may not be the one the maker thinks. That is the core of Jobs to Be Done, an idea the Harvard Business School professor Clayton Christensen had been honing with colleagues for more than a decade when his school wrote it up in 2011. In his account, a job comes up in someone’s life and they “hire” something to do it. Being 18 to 35 with a college degree, he said, “may be correlated with the decision, but it doesn’t cause it.” A teaching note he co-wrote quotes the marketing professor Theodore Levitt: “People don’t want to buy a quarter-inch drill. They want a quarter-inch hole!”
The Walkman went on sale on July 1, 1979. By the end of August, the first 30,000 had sold out.
The milkshake nobody can check
His best-known illustration was a milkshake. In the version his school published, a fast-food chain asked people who fit its profile of a milkshake drinker what would make the shake better, and made the changes. Sales did not improve.
One of Christensen’s fellow researchers then spent a full day in a restaurant, noting who bought milkshakes and when. Forty percent were bought first thing in the morning, by commuters who ordered them to go. In Christensen’s words, they “faced a long, boring commute and needed something to keep that extra hand busy.”
It is a good story, and a hard one to check. The teaching note says “the product and the company in this example have been disguised.” The versions cited here give no sales figures for what came after.
Nor do the theory’s champions agree on what it found. In 2013 Tony Ulwick, founder of the consultancy Strategyn and creator of its own jobs-to-be-done process, argued that a boring commute is “a context within which you execute a job.” The job, in his reading, was to “get breakfast on the go.” To his knowledge, no restaurant had used the insight to launch a breakfast milkshake with skyrocketing sales.
Two people working from the same theory read the same story and wrote down different jobs. The theory tells you where to look. What you find there depends partly on who is looking.
The question the logs can’t ask
The researcher’s sharpest tool, as the teaching note describes it, was a question asked at the door: the last time you had this job and didn’t buy a milkshake, what did you hire instead?
AI chatbots come with plenty of logs and no door. In September 2025, economists and OpenAI researchers published a working paper on how people use ChatGPT. They sampled about 1.1 million conversations from consumer accounts (Free, Plus and Pro), drew one message from each, sent between May 2024 and July 2025, and had automated classifiers label them, with no person reading them. Writing was the most common use at work, about 40 percent of work-related messages in June 2025. And about two-thirds of all writing requests asked ChatGPT to modify text the user had provided, by editing, critiquing or translating it, rather than to write something new. Much of the time, people brought their own draft.
It is a working paper, not peer reviewed, and six of its seven authors list OpenAI as an affiliation. It leaves out business and enterprise accounts. Its labels describe what a message is about. The category labels themselves can’t tell you the situation behind it: that someone pasted in a paragraph and asked for it shorter says nothing, by itself, about who the paragraph was for.
The door question still works. Ask five people on your team about the last time they used your AI tool: what were they trying to get done, and the last time they had that job without the tool, what did they use? Set the answers beside the reason the tool was bought. A gap is a lead to follow up, not a verdict.
Everyone asked for more charts
Imagine you run the analytics team at a retail chain. Your sales dashboard has a feedback button, and the requests agree: more charts, more filters, a dark mode. You ship all three. Weekly visits don’t move.
So you sit beside one regional manager while she uses it. She filters to her twelve stores, exports the table, rounds the numbers and copies four of them into an email to her director. The dashboard is one step on the way. Her job is the email.
You walk back to your desk. In the feedback queue, a new request is waiting: someone would like a fifth chart.