An eight-month ethnographic study published in Harvard Business Review tracked workers across engineering, product, design, research, and operations roles. The finding: AI adoption increases workload intensity rather than reducing hours.
Not exactly shocking on its own. But stack it against our coverage of METR's controlled experiments, and something interesting emerges.
Efficiency Gains, Captured Before You See Them
METR found experienced developers were 19% slower with AI coding assistants, despite believing they were 20% faster. That's a perception gap: workers feel more productive while actually producing less.
The HBR study captures the flip side. Even when productivity gains are real, they don't translate into reduced workload. The time "saved" gets immediately reabsorbed through three mechanisms:
Scope creep: Each efficiency gain reveals "10 more smaller tasks." Accessible capability invites endless expansion of responsibility.
Cognitive overload: AI eliminates routine work, leaving humans handling only high-intensity problem-solving without the mental breaks that simpler tasks once provided. One participant described supervising multiple agents as creating exhausting context-switching demands.
Jevons paradox: As output increases, expectations scale proportionally. Workers voluntarily take on longer hours because AI makes ambitious projects feel achievable.
"Are people leveraging LLMs making more money while working the same hours? Are they working fewer hours for the same pay? If neither is true, LLMs haven't improved worker welfare."
That's from the HN discussion, and it's a useful test. By that measure, the productivity gains are going somewhere. Just not to the workers producing them.
The convergence between these studies matters because the methods differ so sharply. METR used controlled experiments with stopwatches. The HBR researchers used ethnographic observation over eight months. Different methods, different contexts, same conclusion: AI productivity benefits are systematically captured by expanding job scope rather than worker wellbeing.
METR explained why developers overestimate speed gains (context switching, debugging AI suggestions, integration difficulty). HBR explains where the real gains go when they do exist: absorbed by organizational expectations that expand to consume whatever efficiency becomes available.
Our read: these aren't separate findings. They're two views of the same phenomenon. The perception gap and the intensity gap are symptoms of the same structural problem. AI tools change the shape of work, but power dynamics determine who benefits from that change.
It Doesn't Get Easier, You Just Go Faster
Multiple commenters invoked the Greg LeMond cycling quote: "It doesn't get easier, you just go faster."
The parallel is apt. Every efficiency technology in history has faced this dynamic. Email didn't give us more leisure time; it gave us more email. But there's a difference with AI. The intensity increase happens at the cognitive level, not just the volume level. As one study participant noted, supervising agents requires sustained high-level attention without the mental buffer that routine work once provided.
The floor of cognitive demand rises.
This matters for sustainability. You can increase email volume without fundamentally changing the difficulty of reading emails. You cannot strip out all the easy cognitive tasks and expect the same hours of output without consequences.
So what do you do with this? The productivity optimists aren't wrong. AI genuinely makes certain tasks faster. The METR study itself noted that 50 hours of experience may be insufficient, and that less familiar codebases might show real gains. But the organizational response is predictable. Faster task completion doesn't mean earlier departures; it means more tasks. The efficiency gains are real, but the beneficiary isn't the worker. It's the expanding job description.
For individuals, the implication is tactical: if you're using AI tools to work faster, you're funding scope creep for your employer. Want to capture the gains yourself? Set boundaries, or work independently.
For organizations, the research suggests that AI adoption increases intensity even as it increases output. Staff for it, or watch burnout follow the productivity bump.
The tools work. The distribution of benefits is the problem.