AI Didn't Give You Time Back. It Widened the Job.
Two things people said about AI, in the same building, in the same month.
Asked about a particular task, they described momentum: prompting, iterating, getting somewhere. Asked about their work in general, the tone changed. They said they felt busier, more stretched, and less able to switch off.
Neither group was wrong, because it was the same group. Researchers at Berkeley spent eight months watching how that happens.
They watched instead of asking
Most of what we know about AI at work comes from surveys, which ask people to summarize their own experience in a dropdown. This was an Research done by embedding with the people you are studying and observing what they actually do, rather than asking them to report it afterwards. It trades scale for fidelity: you learn about far fewer people, but you catch the things nobody would think to mention on a form, like the fact that they now send prompts during lunch. , which is a different instrument.
Aruna Ranganathan and Xingqi Maggie Ye of Berkeley Haas spent eight months inside a 200-person technology company. Ye ran on-site observation, sat in meetings, and conducted more than forty interviews across engineering, product, design, research and operations.
That design matters for how much weight to put on it. One company, one sector, no control group, so it cannot tell you how common this is. What it can do is show the mechanism in detail, which a survey never will.
The job got wider in three directions
They found AI did not free up time. It produced More effort packed into the same hours, rather than more hours. It is a term labor researchers have used for decades about other technologies, and it is worth knowing because it does not show up in the numbers a company usually watches. Headcount is flat, hours on the timesheet are flat, output is up, and the thing that changed is how densely the day is packed. , and it did it in three distinct ways.
- Scope expanded. People started doing work that would previously have belonged to someone else, or that nobody would have attempted. In the researchers’ phrase, “the scope of what counted as ‘my job’ widened.”
- Boundaries dissolved. Work moved into the gaps that used to be pauses. People sent prompts during lunch, before meetings, or in the evening when something occurred to them.
- Parallel processing. Rather than doing one thing, people kept several threads alive at once, running AI in the background while reviewing code or sitting in a meeting. Some ran multiple agents at the same time.
None of this was asked for. That is what makes it more than a complaint about overwork. No manager set a higher target. Capability arrived, and the definition of a reasonable day quietly moved to match it.
Why it feels good and adds up badly
Ye’s observation about the two registers is the most useful thing in the study. In the small moments, people reported momentum. Zoomed out, they reported strain.
That is not people being confused about their own lives. Both readings are accurate descriptions of different things. Each individual act of using AI genuinely is faster and often genuinely is satisfying. The aggregate of a hundred such acts is a wider job done at a higher tempo with fewer gaps in it, and nobody experiences that as a decision they made.
It also explains why this is hard to notice from inside. You do not get a moment where the work visibly increases. You get a series of moments where something is easier than it used to be.
How this fits with “the savings vanish into slack”
This conflicts with an earlier post, and I think resolving it makes the findings more useful.
The AI productivity paradox argues that individual time savings are real but get absorbed as slack, distributed five minutes here and twenty there, never pooled into anything measurable. This study says the savings get absorbed as more work. Those sound opposed.
They are the same phenomenon meeting different constraints. The saving has to go somewhere, and where it goes depends on what limits your output.
- It becomes slack when the pace is set by something outside you: a review queue, an approval, a weekly meeting. Finishing your part sooner just means waiting longer at the same place.
- It becomes scope when the work has no natural stopping point, which is most knowledge work. Finishing sooner means doing more.
Neither turns into time off by default. That only happens if somebody decides it should, which is the one thing neither the tool nor the process will do on its own.
What to do with that
The researchers recommend that you build a deliberate practice around AI rather than letting it set the pace. Their suggestions, and what they look like for one person:
- Decide what you are not taking on. Expanded scope is the mechanism, so the counter is an explicit boundary on it. If AI lets you absorb a task that used to be someone else’s, that is a choice worth making on purpose rather than by drift.
- Batch instead of reacting. They recommend grouping non-urgent updates. The tool makes an instant response possible, which is not the same as making it a good idea.
- Protect a window. A block with no prompting in it, for the kind of thinking that does not survive being interrupted every four minutes.
- Watch for the lunch prompt. It is the clearest single tell. Not because one prompt matters, but because it marks the moment a pause stopped being a pause.
For managers, this is the permission problem inverted
The same failure shows up twice, in opposite directions, and both are worth recognizing.
In the permission problem, the team wanted to use AI and could not get time for it, so the work happened after hours or not at all. Here, nobody withheld anything. People simply took on more, filled the gaps, and ran three things at once, and no one ever said that was the expectation.
Those look like opposites and they are the same omission: nobody said out loud what the job now is. A team with no stated priority invents one, and what it invents is more.
The takeaway
AI did not hand these people spare time. It widened what they attempted, dissolved the gaps between tasks, and let them run several things at once, and none of it was requested by anybody.
The saving is real. Where it lands is a decision. If nobody makes that decision, it goes into scope, and the only visible symptom is that a job that used to end now does not.
Sources: the study is by Aruna Ranganathan and Xingqi Maggie Ye of Berkeley Haas, described in AI promised to free up workers’ time. UC Berkeley Haas researchers found the opposite (opens in a new tab) and written up by the authors in AI Doesn’t Reduce Work, It Intensifies It (opens in a new tab) (Harvard Business Review, February 2026). Eight months, one 200-person technology company, more than forty interviews. The quoted phrases and the three forms of intensification come from Berkeley’s own write-up; the HBR article is paywalled and I have not read it. Worth holding the scope in mind: a single company in one sector, so this shows a mechanism rather than measuring how widespread it is. The reconciliation with the productivity paradox is mine, not something either piece of research proposes.
Related: AI Productivity Paradox: Your Team Feels Faster and the Company Can’t Tell is the other half of where the time goes. AI Anxiety at Work: You’re Stressed and More Productive. Both Are True. is what this feels like from the inside.