AI doesn't reduce work. It intensifies it.
An 8-month ethnography embedded inside a 200-person US tech company found that workers using AI did more, faster, for longer hours, and nobody asked them to. Here is why that is a harder problem than it sounds.
The productivity optimists have a reliable story, and it goes like this. AI tools arrive. Knowledge workers use them. Time gets saved. People leave the office at a reasonable hour, reclaim their evenings, and finally start the side project they have been deferring for three years.
It is a tidy story. It is also, on the evidence, the wrong one. An 8-month ethnography published in Harvard Business Review in February 2026 watched what people actually did, rather than what they said they did, and found close to the opposite.
How did the researchers actually study this?
Aruna Ranganathan (Associate Professor at UC Berkeley Haas) and Xingqi Maggie Ye did something this field rarely bothers with: they showed up. Instead of surveying employees about their AI use, which mostly measures what people think they are supposed to say, they embedded inside a real 200-person US tech company. Ranganathan spent two days a week on-site from April to December 2025, and the team ran more than 40 interviews across engineering, product, design, research, and operations.
That distinction matters more than it sounds. Ask someone “has AI made you more productive?” and they will say yes, partly because they feel it has and partly because yes is the expected answer. An observer watching you still typing at dinner is a much harder thing to talk your way around. This is an ethnography in the proper social-science sense, not a panel experiment and not a dashboard scrape.
What did the study actually find?
The core finding is stated plainly in the paper: “We found that employees worked at a faster pace, took on a broader scope of tasks, and extended work into more hours of the day, often without being asked to do so.”
Read that last clause again, because it is the whole study. Nobody raised the targets. Nobody moved the deadlines. The extra work was voluntary, and that is exactly what makes it interesting rather than just sad. The researchers’ explanation is mechanical and convincing. Generative AI removes the blank page, and the blank page is one of the most effective procrastination devices ever invented. Take it away and the activation energy for hard tasks collapses, so people attempt things they used to defer. They also reach sideways, into adjacent roles, because AI narrows the gap between “can write code” and “can write some code.” Work expands in volume and in scope at the same time. And then it follows everyone home.
Why did workers take on more without being asked?
Here is the part worth sitting with. The extra effort was experienced, by the workers themselves, as enjoyable experimentation. That is not cynicism talking. AI tools genuinely are engaging: fast feedback, surprising output, a real sensation of your own capability expanding. People were not grinding. They felt like they were exploring, which is precisely what made the additional load invisible, including to the people carrying it.
“Because the extra effort is voluntary and often framed as enjoyable experimentation,” the researchers wrote, “it is easy for leaders to overlook how much additional load workers are carrying.” That is a clean description of a blind spot that organisations have every incentive to leave un-fixed. If your staff are happily taking on more, logging the hours by choice, and rating the experience positively, what exactly is supposed to trigger an intervention? There is also a structural cost the study surfaces: experienced engineers ended up checking the output of juniors who were using AI to write code above their level. AI levels beginners up faster than it levels up experts, and the resulting quality-assurance burden landed on senior staff on top of their existing work, not instead of it.
What does voluntary work intensification actually cost?
The paper is not alarmist, but it is clear-eyed. “What looks like higher productivity in the short run,” the researchers wrote, “can mask silent workload creep and growing cognitive strain as employees juggle multiple AI-enabled workflows.” The danger is not that AI makes people tired. It is that the pattern conceals itself. Output rises, hours rise, visible stress stays low because the work feels chosen and positive, and cognitive load accumulates quietly until it shows up as slower decisions, weaker output, and eventually people leaving.
“For workers, the cumulative effect is fatigue, burnout, and a growing sense that work is harder to step away from, especially as organizational expectations for speed and responsiveness rise.” That final phrase is the trap. Even if the intensification starts entirely worker-driven, it resets the baseline. Once an organisation has seen what its teams can produce with AI, “without being asked” quietly becomes “it would be strange if you stopped.”
What do the researchers recommend?
They are careful not to cast AI as the villain. The problem is that neither organisations nor workers have built the habits to absorb a step-change in individual capability. They propose three responses. Intentional pauses: before a major decision is finalized, require “one counterargument and one explicit link to organizational goals,” not to slow things for its own sake but to widen the attention field enough to catch the drift fast workflows induce. Sequencing over responsiveness: “by regulating the order and timing of work (rather than demanding continuous responsiveness) sequencing can help organizations preserve attention, reduce cognitive overload, and support more thoughtful decision-making.” Let the team lead the AI, not the reverse. Human connection as a circuit-breaker: “as AI enables more solo, self-contained work, organizations can benefit from protecting time and space for listening and human connection.” None of this requires limiting AI adoption. It requires managing it, which is the harder and less glamorous job.
Who captures the value when AI boosts output?
There is a question underneath all of this that nobody enjoys asking out loud: if AI makes knowledge workers more productive, who keeps the value? The Anthropic Economic Index, published in April 2026, found 10% of surveyed AI users said their employer, not they, was capturing the productivity gains. Different study, different sample, but the direction of travel rhymes. AI produces measurable output increases; whether those land in worker pay, smaller headcount, or wider margins is a distribution question, not a productivity one, and it is mostly unanswered. What this ethnography adds is the texture survey data cannot: the intensification happens voluntarily, with genuine enthusiasm, in a way that looks from the outside like exactly the company you would want to run. Which is why it stays invisible until it has been happening for a while.
Why this study deserves attention
The instinctive pushback is to ask whether more work is even bad. Maybe it is fulfilling. Maybe the output justifies the hours. Those are fair questions, and the ethnography cannot answer them; it does not tell us whether these workers ended the eight months happier or unhappier. What it can tell us is that they worked harder, took on more, did it by choice, and that the cumulative effects of that pattern, sustained, point toward the outcomes the researchers describe. The in-depth reading is not “AI makes work worse.” It is this: AI changes the structure of work in ways productivity metrics do not capture, and those structural changes carry costs that tend to arrive later than the benefits. If you are making decisions about AI adoption at an organisational level, deploying the tools, setting the expectations, designing the workflows, that is the sentence to keep on your desk.
Read the study: AI Doesn’t Reduce Work, It Intensifies It - Harvard Business Review, February 9, 2026.