There’s an uncomfortable finding buried in new research from UC Berkeley, and it contradicts nearly everything vendors say about AI productivity. After spending eight months inside a technology company, watching how people actually use generative AI tools, researchers Aruna Ranganathan and Xingqi Maggie Ye found something that should concern every business leader: AI doesn’t reduce work. It intensifies it.
Not because managers demanded more output. Not because the tools failed. Because workers themselves, given faster capabilities, voluntarily expanded what they did, when they did it, and how much they juggled at once. The productivity gains were real. The cost was real too.
What the Berkeley researchers actually found
The study, published in Harvard Business Review, followed approximately 200 employees at a US technology company from April to December. This wasn’t a survey or a self-reported diary study. Ranganathan and Ye conducted twice-weekly in-person observations, tracked communication channels, and carried out more than 40 in-depth interviews across engineering, product, design, research, and operations.
Critical Context: This is ethnographic research—direct observation over months—not the kind of two-week pilot study that typically dominates AI productivity claims. The methodology matters because it captures behaviour people don’t self-report.
The core finding is disarmingly simple. One participant put it plainly: “You had thought that maybe…you save some time, you can work less. But then really, you don’t work less.”
Three distinct patterns of intensification emerged, each feeding into the others.
Three ways AI makes people work more
Task expansion
When AI tools made certain tasks faster, workers didn’t pocket the time savings. They absorbed work that would previously have been outsourced, deferred, or left undone. Product managers started writing code. Researchers took on engineering tasks. The boundaries between roles blurred—not through deliberate reorganisation, but through individual choice.
This is the bit that should make leaders uncomfortable. Nobody asked these people to do more. The tools made it possible, so they did it. The psychology here is straightforward: when something becomes easy enough to attempt, people attempt it. But “easy to attempt” and “easy to do well” are different things, and the cognitive load of managing unfamiliar responsibilities doesn’t disappear just because an AI helped draft the first version.
Strategic Insight: Task expansion looks like productivity from above—more output per person. It looks like burnout from below. The gap between these two perspectives is where organisational risk accumulates.
Blurred work boundaries
AI tools reduced the friction of starting tasks. That sounds like a benefit, and in isolation it is. But reduced friction also means reduced barriers. Workers slipped work into lunch breaks, commute times, and evening hours. Not because deadlines demanded it, but because starting felt effortless.
This pattern will be familiar to anyone who lived through the smartphone revolution. When you can check email anywhere, you check email everywhere. AI tools are doing the same thing to actual productive work, not just communication. The difference is that producing a draft, running an analysis, or reviewing code feels more like accomplishment than reading emails does. It’s harder to recognise as overwork when it feels productive.
Reality Check: “I can do this quickly” becomes “I should do this now” becomes “I’m always doing this.” The progression is gradual enough that most workers don’t notice until exhaustion sets in.
Cognitive overload through multitasking
With AI handling routine elements of multiple projects, workers took on more concurrent work streams. The study found increased multitasking across the board—people managing three, four, five parallel initiatives because AI assistants made each individual stream feel manageable.
The research literature on multitasking is unambiguous: humans are bad at it. We don’t actually do multiple things simultaneously. We context-switch, and every switch costs time, attention, and quality. AI tools masked this cost by making each individual task feel lighter, whilst the aggregate cognitive load grew heavier.
Hidden Cost: Individual task completion times dropped. Total time spent working increased. That’s the paradox in a sentence.
Why workers do this to themselves
The obvious question is why rational adults voluntarily take on unsustainable workloads when nobody is making them do so. The Berkeley research suggests several reinforcing mechanisms.
| Driver | Mechanism | Organisational effect |
|---|---|---|
| Capability expansion | ”I can do this now” mindset | Role boundaries dissolve |
| Reduced friction | Low barrier to task initiation | Work fills all available time |
| Perceived productivity | Each task feels efficient | Total workload grows unnoticed |
| Social comparison | Colleagues doing more too | Informal expectations escalate |
| Achievement satisfaction | More output feels good | Burnout lags behind by months |
There’s also a competitive dynamic at play. When your colleague uses AI to take on additional responsibilities, standing still feels like falling behind. Nobody formally raises expectations, but informal norms shift quickly. Within months, doing what AI makes possible becomes what’s expected.
Strategic Reality: Organisations that deploy AI tools without managing this dynamic aren’t getting free productivity. They’re borrowing against their workforce’s sustainability.
What organisations need to do
Ranganathan and Ye propose a framework they call “AI Practice”—deliberate organisational habits that counteract the natural tendency toward intensification. The framework has three components, and all three require active management rather than passive guidelines.
Intentional pauses
Structured reflection intervals that protect against unchecked acceleration. This means building actual stops into workflows—not suggesting that people “take breaks when they need to,” because the research shows they won’t. Scheduled review points where teams assess whether AI-enabled expansion has pushed work beyond sustainable boundaries.
For UK organisations, this connects directly to the duty of care under health and safety legislation. If AI tools systematically drive overwork, and the organisation knows about the research, failing to implement safeguards creates legal as well as ethical exposure.
Sequencing
Deliberate work pacing and notification batching. Rather than letting AI’s instant availability drive constant task-switching, organisations should design workflows that channel AI assistance into focused blocks. This is harder than it sounds because it means overriding the tools’ default behaviour, which is optimised for responsiveness rather than sustainable pace.
Implementation Note: Notification batching alone won’t solve this. The sequencing needs to be structural—built into project management processes, not left to individual discipline. People don’t self-regulate well when the tool makes it feel effortless to keep going.
Human grounding
Protected time for dialogue and collaborative perspective-building. The research found that AI-assisted individual work was crowding out the informal conversations, mentoring, and collaborative thinking that organisations depend on. When everyone is heads-down producing AI-assisted output, nobody is building the shared understanding that makes that output useful.
Four challenges most organisations will miss
The measurement problem. Standard productivity metrics will show improvement—more output per person, faster completion times, higher throughput. These metrics won’t capture the unsustainability of the pace or the erosion of collaboration. Organisations that manage by dashboard will see success right up until they see resignations.
The training gap. Workers taught how to use AI tools aren’t taught how to manage the psychological effects of those tools. There’s no onboarding module for “your new copilot will make you feel like you should be working constantly.” This gap between technical training and behavioural awareness is where most AI rollouts create harm.
Warning: ⚠️ If your AI training programme covers prompting techniques but not workload management, you’re equipping people to overwork more efficiently.
The manager blind spot. Middle managers face a genuine dilemma. Their teams are producing more, morale seems fine in the short term, and raising concerns about overwork when output is up feels counterintuitive. The Berkeley study suggests that the visible signals of intensification lag behind the actual experience by months. By the time managers notice problems, the damage is already significant.
The equity dimension. Not every worker responds to AI tools with equal intensification. Some will expand aggressively, others will maintain boundaries. Without explicit norms, the expanders set the pace and the boundary-maintainers look like underperformers. This creates a toxic dynamic where sustainable working practices become career-limiting.
The bottom line for UK businesses
This research arrived at a moment when UK businesses are under pressure to demonstrate AI returns. The temptation is to celebrate the productivity numbers and ignore the sustainability question. That would be a mistake.
The three patterns identified by Ranganathan and Ye—task expansion, blurred boundaries, and cognitive overload through multitasking—are not bugs in how people use AI. They’re predictable consequences of deploying powerful tools without managing the human response to them. The solutions exist, but they require active organisational intervention rather than individual willpower.
Here’s what matters most:
- Measure sustainability, not just output. Track hours worked, boundary maintenance, and wellbeing alongside productivity metrics.
- Design workflows around human limits. Build pauses, sequencing, and collaboration time into AI-assisted processes rather than hoping people will self-regulate.
- Train for behaviour, not just technique. Include workload management and boundary-setting in every AI training programme.
- Set explicit norms early. Don’t let informal escalation dynamics establish your organisation’s working culture by default.
Take Action: Before your next AI tool deployment, audit your current measurement systems. If they only track output and speed, they’ll confirm the productivity story while missing the intensification story entirely.
The Berkeley research doesn’t say AI tools are harmful. It says they change behaviour in ways that organisations must actively manage. That’s a more nuanced message than “AI boosts productivity” or “AI causes burnout,” and it requires a more nuanced response than either enthusiasm or resistance. The organisations that get this right will keep the productivity gains without paying for them in workforce sustainability.
Source: Ranganathan, A. and Ye, X.M. (2026) “AI Doesn’t Reduce Work—It Intensifies It,” Harvard Business Review, 9 February. Available at: hbr.org
Analysis by Resultsense — making sense of AI in the UK.