AI promised to save you 7 hours a week: 37% vanish into rework

Every AI tool you added was supposed to buy back time. Collectively, they did the opposite.

An eight-month ethnographic study from UC Berkeley Haas tracked 200 employees at a U.S. tech company and found a pattern nobody predicted: workers using AI didn't slow down. They sped up, took on tasks they'd never attempted, and stretched their workdays into evenings and lunch breaks, often without being asked. The researchers called it "work intensification," and it showed up in three forms: expanded task scope, dissolved time boundaries, and parallel processing of multiple AI workflows at once.

The assumption was simple. Faster tools, less work. The data said otherwise.

37% of your AI time savings vanish before you use them

A Workday and AlixPartners survey of 3,200 workers across North America, Asia-Pacific, and EMEA found that 85% of respondents saved one to seven hours per week using AI. But 37% of those saved hours disappeared into rework: correcting errors, rewriting AI-generated content, and verifying outputs that looked right but weren't.

That means a worker saving five hours per week actually reclaims about three. The other two go to babysitting the tool that was supposed to help.

"I got the eight hours to two hours, but now I can get 20 hours of work," admitted one CTO at Dun & Bradstreet, capturing the trap perfectly. AI didn't reduce the workload. It raised the ceiling on what management considered possible.

The cognitive cost nobody budgeted for

Boston Consulting Group and UC Riverside surveyed 1,488 full-time U.S. workers and coined a term for what happens next: "AI brain fry." Workers with high AI oversight responsibilities reported 14% more mental effort, 12% greater mental fatigue, and 19% more information overload than those with low oversight.

Here is what makes this worse: productivity actually increased when workers used one to three AI tools. The moment they added a fourth, self-reported productivity collapsed. More tools didn't mean more output. They meant more context-switching, more oversight, and more decisions per hour than the human brain can sustainably manage.

Fourteen percent of AI-using workers reported experiencing brain fry. Among marketing teams, that figure hit 26%. Those affected made 39% more major errors and showed 34% of them actively planning to quit.

The real problem is not the tools. It is what fills the gap.

The Berkeley researchers observed something the productivity data alone couldn't explain. When AI compressed a six-hour task into 40 minutes, employees didn't log off. They stayed at their desks and filled those reclaimed hours with more tasks, more threads, more parallel workflows. One worker described it bluntly: "You had thought that maybe you could work less. But then really, you don't work less."

This is what cognitive scientists call the efficiency paradox: making a process faster doesn't reduce total effort if the system responds by adding more processes. The same research linking mind-wandering to divergent thinking (correlations of r = .14 to .16 across 865 participants) suggests the hours being consumed by AI busywork are exactly the hours your brain needs for creative incubation. When you fill every gap with another AI-assisted task, you eliminate the unstructured mental space where original ideas form.

Google reports that 50% of its code is now AI-generated, producing "well over a 10% velocity gain" across tens of thousands of engineers. But velocity and value are not the same thing. KPMG cut meeting prep time by 75%. AES compressed a 14-day audit process into one hour. In every case, the freed time was immediately reallocated to more work, not to thinking.

What the 6% who profit from AI actually do differently

Only 6% of companies generate real returns from AI, according to McKinsey and Goldman Sachs data. The Berkeley researchers proposed a framework for joining them: developing an "AI practice," meaning intentional norms around how, when, and how much to use AI tools.

Their three specific prescriptions: structured pauses between AI-assisted tasks (so speed doesn't crowd out reflection), sequential task batching instead of parallel processing (reducing the context-switching that drains cognition), and protected focus time with no AI tools running.

The irony is sharp. The companies extracting the most value from AI are the ones deliberately limiting its presence in their employees' workdays. Not because the tools are bad, but because human cognition has bandwidth limits that no software update will fix.

Your next AI-saved hour has two possible futures. You can fill it with another task. Or you can protect it as the space where your best thinking actually happens. The Berkeley data is clear on which one pays off.