“We can either go slow and deliberate, or fast and chaotic. We are going fast and chaotic.”
Kaz Nejatian, Shopify’s chief operating officer, sent that line to staff in early January 2023. Then a bot went into everyone’s calendar and deleted every recurring meeting with three or more people: nearly 10,000 events, more than 76,500 hours, by the company’s estimate. Nobody could add one back for two weeks. Wednesdays would be meeting-free again.
A month later, Nejatian told NPR what one engineer had told him: for the first time in a very long time, they got to do what they were hired for, write code all day.
The deleted meetings were easy to count. What the cleared Wednesdays turned into was harder to see.
An empty workbench
Free hours aren’t deep work until you spend them on something hard, undistracted.
Cal Newport, a computer science professor and author, named the idea on his blog in November 2012. Deep work, he wrote, means “cognitively demanding activities that leverage our training to generate rare and valuable results, and that push our abilities to continually improve.” Its opposite, shallow work, is the tasks “almost anyone, with a minimum of training, could accomplish”: email replies, logistical planning, tinkering with social media. Shallow work tempts, he argued, because it’s easy, which makes it feel productive, and because people enjoy the back-and-forth.
In his 2016 book Deep Work, Newport stretched that into a larger claim, summed up on his own website: the ability to focus without distraction is becoming increasingly valuable.
A cleared calendar is an empty workbench. It’s necessary, and it builds nothing by itself. Newport’s advice is about what happens next. Deep work takes energy, he wrote, so don’t expect to slide into it straight from shallow work; build a ritual that carries you across. His own, in that 2012 post: shut down the computer, switch off the overhead lights and leave on only the desk lamp, brew a cup of coffee, and put a “do not disturb” sign on the door.
Protected time can also drift. In a study published in 1999, the researcher Leslie Perlow helped a software team protect three mornings a week. A year later, with the team’s rewards unchanged, the practice had faded, as Maker’s Schedule, Manager’s Schedule describes.
An argument, not a finding
Newport’s case is an argument, made in essays and in books for a broad audience. The research under it is narrower than the claim.
Switching between tasks does have a measured cost; Sophie Leroy’s experiments on attention residue found people did worse on a task after leaving another unfinished. But the cost doesn’t always show up as worse output. In a 2008 lab study, Gloria Mark and colleagues had 48 people, most of them German university students, answer a pile of emails as a human resources manager while a “supervisor” interrupted them about every two minutes. Leaving out the time the interruptions themselves took, the interrupted participants spent less time on the emails, and the study found no significant difference in quality. They paid in stress: after only 20 minutes, they reported more stress, frustration, time pressure and effort.
Routine email is shallow work by Newport’s definition, so this doesn’t test his claim directly. It does show that on routine tasks, interruptions can be absorbed at a price people feel rather than see. Whether deep work is gaining value across the economy is a bigger claim, one a lab study like this can’t settle.
Shopify’s own numbers have the same gap. Bloomberg reported in July 2023 that average meeting time per worker had fallen 14 percent over the first five months of the year; Nejatian said that helped toward a projected 18 percent rise in finished projects. That is a projection, from the company that made the change.
The Wednesday that filled itself
Imagine you’re an analyst whose team has just adopted meeting-free Wednesdays. By noon on the first one, you’ve cleared forty messages and tidied a dashboard. Nothing interrupted you. Nothing hard got done either.
The next Tuesday evening, you write down the one problem for Wednesday morning: why the churn model overpredicts for annual customers. You close chat until noon. By lunch you have a page of notes and a test that fails for the right reason. The messages are still there at 12:15.
When the assistant does the hard part
Newport’s definition has a second half that’s easy to skip: work that pushes your abilities to keep improving. A 2026 study from Anthropic tested what an AI assistant does to that part.
Fifty-two software engineers, mostly junior, who used Python weekly but had never used a library called Trio, built two features with it. Some, assigned at random, had an AI assistant in the sidebar that could produce correct code on request. Then everyone took a quiz on the concepts they had just used.
The AI group finished about two minutes faster, a difference that wasn’t statistically significant. On the quiz, it averaged 50 percent; the hand-coding group averaged 67. The widest gap was on debugging. The researchers suspect the hand-coders, who hit more errors and fixed them alone, learned from the struggle. Participants who mostly asked the assistant conceptual questions or for explanations scored better than those who handed it the coding, though the authors say this pattern shows an association, not a cause.
The limits matter: the sample was small, the quiz came right after the task, and the setup differed from agentic tools like Claude Code, where the authors expect any effect on learning to be larger.
Next time you learn a tool with an assistant, ask it for explanations rather than finished code on one piece, then write the next piece without it and see whether it runs and passes its tests. Try it on a few tasks before you decide; one task tells you little.
Somewhere at Shopify, early in 2023, an engineer spent a whole day writing code. The calendar showed the hours were free. Only the work could show what they were for.