A randomized controlled trial published by METR in July 2025 tested 16 experienced open-source developers across 246 real tasks. The developers predicted AI tools would save them 24% of their time. After each task, they estimated a 20% speedup. The objective measurement: they were 19% slower. That 39-percentage-point gap between perception and reality isn’t a curiosity. It’s the key to understanding why AI is making knowledge workers busier, not freer.
Yesterday, a UC Berkeley eight-month ethnographic study published in Harvard Business Review documented exactly what happens when AI becomes embedded in a workplace. The findings are uncomfortable: AI work intensification is real, measurable, and self-inflicted. Workers voluntarily adopted AI tools and ended up doing more work, not less.
There’s a name for this. In 1865, economist William Stanley Jevons observed that more efficient steam engines didn’t reduce coal consumption — they increased it. Cheaper energy meant more uses for energy. AI is doing the same thing to knowledge work. The tool genuinely makes individual tasks faster. But organizations treat freed capacity the way Victorian factories treated cheap coal: as an invitation to burn more.
The Berkeley study: eight months inside an AI-saturated workplace
The study wasn’t a survey or a self-reported questionnaire. Researchers Aruna Ranganathan and Xingqi Maggie Ye from Berkeley Haas embedded themselves inside a roughly 200-person US tech company for eight months, from April through December 2025. They showed up two days per week, tracked internal Slack channels, and conducted more than 40 interviews across five departments: engineering, product, design, research, and operations.
The critical detail: AI adoption at this company was voluntary. The company offered enterprise subscriptions but didn’t mandate their use. Nobody was forced to adopt or threatened with replacement. Workers chose AI because the tools genuinely helped on individual tasks.
And yet every department experienced the same outcome. AI didn’t create slack in the workday. It created demand for more output. The researchers identified three mechanisms of AI work intensification, each reinforcing the others.
Three ways AI work intensification compounds
- Task expansion. Role boundaries dissolved. Product managers and designers started writing code. Researchers absorbed engineering tasks. Engineers spent extra hours reviewing and correcting their colleagues’ AI-assisted work — a new overhead category that didn’t exist before. The researchers noted that hiring may have been postponed because employees absorbed work that would have justified additional headcount. Workers weren’t doing one job faster. They were doing 1.3 jobs at the same salary.
- Blurred boundaries. AI’s conversational interface made work ambient. Workers sent prompts during lunch breaks, between meetings, during file-loading delays. Some fired off a “quick last prompt” before leaving for the day. The workday lost its natural pauses — the gaps that used to provide micro-recovery. Downtime stopped feeling like downtime because a productive interaction was always one message away.
- Increased multitasking. Workers managed multiple AI threads simultaneously, switching attention across parallel workstreams. Speed expectations rose through normalization, not explicit demands. Nobody told anyone to work faster. People just saw what was possible and adjusted their own baselines upward.
Each mechanism is Jevons’ Paradox in miniature. Efficiency didn’t create leisure. It created demand. And the data isn’t limited to one company. An Upwork survey of 2,500 workers found that 77% of employees using AI reported increased workloads. Microsoft’s own 2025 Work Trend Index found that 90% of users say AI saves time — and 80% of the workforce says they lack the time or energy for their work. Both statements are true simultaneously, which is exactly what Jevons predicted 160 years ago.
The perception gap: feeling faster while falling behind
The METR study quantifies the illusion at the heart of AI work intensification. This wasn’t a survey. It was a randomized controlled trial with 143 hours of screen recordings, manually labeled at roughly 10-second resolution. The repos averaged 22,000+ GitHub stars and over a million lines of code — AI coding tools making experienced developers slower on real projects, not toy benchmarks.
The result is stark. Developers believed AI made them 20% faster. It made them 19% slower. That’s not a miscalibration — it’s a complete inversion of perceived and actual performance. It maps directly onto the Berkeley finding: workers felt more productive but weren’t less busy. Some felt busier.
The Harvard/BCG “jagged frontier” study adds the essential nuance. When 758 BCG consultants used AI on tasks within its capability zone, they completed work 25% faster at 40% higher quality. But on tasks outside AI’s frontier — tasks requiring judgment, synthesis, or business problem-solving — consultants were 19 percentage points less likely to produce correct solutions. Below-average consultants improved 43%; top performers improved only 17%.
This explains why workers feel productive. They are productive on tasks within AI’s sweet spot. But organizations respond by piling on more tasks, including ones outside that frontier. The perception gap is the psychological lubricant that keeps the Jevons cycle spinning. You feel fast, so you take on more. The more you take on, the more lands outside AI’s zone. And you don’t notice the drag until you’re already overcommitted.
Who pays the tab: burnout by seniority
The cost of AI work intensification doesn’t fall evenly. A separate DHR Global survey of 1,500 professionals paints a clear gradient: 62% of associates and 61% of entry-level workers report that burnout has reduced their engagement. For managers, it’s 48%. Directors, 44%. C-suite executives, 38%. The pattern is obvious. Junior workers absorb the most task expansion and pay the highest price.
And within companies, the view varies sharply by rank. The people closest to the work see AI’s limits. The people furthest from it see the savings.
Here’s the cruelest irony: the most enthusiastic AI adopters are burning out fastest. The power users — the workers companies point to as proof AI is working — are the canaries. Their to-do lists expanded to fill every hour AI freed up, and kept going.
The end state is already visible. Amazon cut 30,000 corporate jobs since October 2025 as part of broader restructuring — while posting $21 billion in quarterly profit and committing $100 billion in capex. The sequence is becoming a pattern: AI intensifies work for existing employees, the gap between AI hype and enterprise reality narrows enough to justify headcount cuts, and the survivors absorb even more. In 2025 alone, companies directly cited AI in 55,000 job cuts — a 12x increase from two years prior.
AI practice: cognitive ergonomics, not corporate wellness theater
The Berkeley researchers don’t stop at diagnosis. They propose a framework they call “AI practice” — and it’s worth taking seriously because it’s grounded in eight months of observation, not a whiteboard brainstorm. The framework has three pillars:
- Intentional pauses. Before acting on AI-generated output, require one counterargument and an explicit link to organizational goals. This isn’t “mindfulness.” It’s quality control for decision velocity.
- Sequencing. Batch non-urgent notifications. Protect focus windows from interruptions. Advance work in coherent phases rather than continuous AI-mediated multitasking.
- Human grounding. Protect time for brief check-ins, shared reflection, and structured dialogue. AI-mediated work is fast and isolating. Human grounding counters both effects.
Think of it as cognitive ergonomics. In the 1990s, standing desks and screen breaks became standard responses to the physical toll of computer work. AI practice is the equivalent response to the cognitive toll of AI-mediated work. Developer Simon Willison, reflecting on the study, put it plainly: we’ve disrupted decades of existing intuition about sustainable working practices, and it’s going to take discipline to find a new balance.
The urgency is real. Only 11% of organizations have deployed AI agents in production, according to Deloitte. When agents move from pilot to production — when AI doesn’t just assist tasks but initiates them — the ambient work problem compounds. BCG’s data shows 50% of companies are redesigning workflows for AI, while the other 50% bolt it onto existing processes. That second group is where intensification hits hardest.
The efficiency trap we’ve seen before
We’ve known since 1865 that efficiency doesn’t reduce consumption. Jevons watched it happen with coal. We watched it happen with email, smartphones, and Slack. AI isn’t an exception to this pattern. It’s the most powerful demonstration of it yet.
The tool works. The problem is what we do with the surplus. If voluntary adoption at a 200-person tech company produced this level of intensification in eight months, what happens when mandatory AI agent deployments hit Fortune 500 organizations with 200,000 employees who never asked for the change? The tipping point will come when a major employer implements a structured AI practice framework and publishes the results — proving that intentional constraints on AI use produce better outcomes than unconstrained adoption. Until someone runs that experiment, the Jevons cycle keeps spinning. And we keep mistaking the acceleration for progress.
Get the Daily Pulse
Sharp analysis on what's actually moving in AI. No hype, no filler, no weekly digest.
