The AI Intensity Trap: When Productivity Gains Become Workload Creep

A study is making the rounds this week with a headline designed to travel: “Harvard study proves AI makes workers burn out.” Within 48 hours, it was on LinkedIn, Reddit, and a dozen tech blogs. The framing was irresistible. People who distrust AI shared it as vindication. People who sell AI dismissed it as fear-mongering. Almost nobody read what the researchers actually said.

Here is what happened. Two UC Berkeley Haas researchers --- Aruna Ranganathan and Xingqi Maggie Ye --- published findings in Harvard Business Review from an eight-month ethnographic study of 200 employees at a US technology company. The study ran from April through December 2025. HBR is a practitioner magazine, not an academic journal. The research is described as “in-progress.” It has not been peer-reviewed.

None of that makes it wrong. It makes it preliminary. The distinction matters because the finding itself is important enough to deserve precise language, not headline inflation.

What the Researchers Actually Found

The study documents three mechanisms through which AI tools intensify work rather than reduce it.

Task expansion. AI fills knowledge gaps, enabling employees to take on work outside their defined roles. Product managers started writing code. Researchers began doing engineering tasks. The friction that once kept people within their competencies --- “I don’t know how to do that” --- evaporated. AI made everything feel adjacent.

Blurred boundaries. When starting a task requires nothing more than typing a prompt, the activation energy drops to near zero. Employees in the study began prompting AI during lunch breaks, in meetings, between tasks. The boundary between “working” and “not working” dissolved --- not because anyone mandated it, but because the tool made it frictionless.

Increased multitasking. Workers managed multiple AI threads simultaneously. Writing code in one window while AI generates alternatives in another. Running parallel agents on adjacent problems. The cognitive load didn’t decrease. It redistributed across more concurrent streams.

These three mechanisms create a self-reinforcing cycle: AI accelerates individual tasks, which raises expectations for output, which deepens reliance on the tool, which expands the scope of what each person is expected to do, which increases overall intensity.

The researchers warned about the risk of burnout. They did not claim AI causes burnout. That is an important distinction the headlines erased.

The Study That Wasn’t From Harvard

Let me be precise about what this research is and is not.

It is not a Harvard study. Ranganathan and Ye are Berkeley faculty who published in Harvard Business Review --- a magazine that accepts practitioner-oriented writing, not a peer-reviewed journal. Calling this “a Harvard study” is like calling a New York Times op-ed “a New York Times investigation.” The platform is not the institution.

It is a single-company study. Two hundred employees at one US technology firm. The findings may reflect that company’s culture, management style, or tool deployment patterns. We cannot generalize from n=1, no matter how carefully the ethnography was conducted.

It is in-progress research. The authors describe it as such. The final, peer-reviewed version may include additional analysis, caveats, or revised conclusions.

This matters because the media ecosystem doesn’t do nuance. A Berkeley ethnographic study published in a practitioner magazine becomes “Harvard proves AI bad” in the time it takes to write a LinkedIn post. And once that framing takes hold, it shapes policy conversations, vendor negotiations, and organizational decisions that affect real people.

The pattern is familiar. We saw the same thing with the METR randomized controlled trial last year, where “experienced developers were 19% slower with AI” became the headline --- stripped of the context that these were 16 developers working on mature codebases they had contributed to for years, using early-2025 tools. The finding was real and important. The universalization of it was not.

The Perception Gap Is the Real Story

Here is what connects the Berkeley study to the broader evidence base. Across multiple studies, with different methodologies and different populations, one finding keeps replicating: people systematically misjudge their own productivity with AI tools.

The METR trial found developers believed they were 24% faster. They were 19% slower. The Berkeley study found employees experienced AI as reducing their workload. Their observable behavior showed the opposite --- more tasks, more hours, more concurrent streams.

Google’s DORA 2024 report adds another dimension: 75% of developers report feeling more productive with AI. But each 25% increase in AI code adoption correlates with a 1.5% dip in delivery speed and a 7.2% drop in system stability.

The perception gap is not a curiosity. It is a governance problem. If the people using AI tools cannot accurately assess whether those tools are helping, then organizational decisions based on user sentiment are unreliable. And right now, most organizations measure AI impact through surveys and self-reports --- the exact instruments the evidence says are misleading.

You cannot manage what your people misperceive.

Why This Isn’t Really About AI

Strip away the technology and the Berkeley study describes a pattern as old as knowledge work itself.

Give people a tool that makes tasks easier. Watch them take on more tasks. Marvel at how busy everyone is. Wonder why no one feels less stressed.

Email did this. Smartphones did this. Slack did this. Every tool that reduces the friction of one task creates the capacity --- and eventually the expectation --- to do more tasks. The tool is never the problem. The absence of organizational norms governing how the tool is used is the problem.

This is what makes the AI version more consequential. Previous tools reduced friction incrementally. AI reduces it categorically. When a product manager can go from “I don’t know Python” to “here’s a working script” in three minutes, the friction reduction isn’t 10% or 20%. It approaches zero. And when activation energy approaches zero, the only thing standing between an employee and an ever-expanding workload is organizational discipline.

Most organizations don’t have that discipline. They have AI tools.

The Governance Gap

The Berkeley researchers describe what happened to 200 employees. They don’t prescribe what organizations should do about it. That’s appropriate for researchers. But it leaves a vacuum that vendors and pundits are happy to fill --- vendors with “AI wellness” products, pundits with “AI is bad” takes. Neither is useful.

The useful question is: who decides how AI is used in your organization?

Not who decides whether to adopt AI. That question is settled for most companies. The question is who sets the norms for how it’s used. Who defines which tasks are appropriate for AI assistance and which require direct human engagement? Who monitors whether AI is expanding scope beyond what was intended? Who notices when “AI helps me do my job” becomes “AI changed my job without anyone deciding it should”?

In most organizations, the answer is: nobody. AI adoption is happening bottom-up, driven by individual employees finding tools useful. This is how every productivity technology has been adopted. And it is how every productivity technology has eventually created problems that required organizational intervention.

The difference is speed. Email took a decade to reshape work norms. Smartphones took five years. AI is doing it in months. The window between “this is helpful” and “this has changed everything and we didn’t decide it should” is compressing faster than organizations can respond.

What “AI Practice” Looks Like

We use the term “AI Practice” to describe what most organizations are missing: a deliberate set of organizational norms governing how AI tools are used, by whom, for what purposes, and with what boundaries.

AI Practice is not AI policy. Policy is a document that sits in a shared drive. Practice is behavior that is reinforced through management, measurement, and culture.

Concretely, AI Practice addresses the three mechanisms the Berkeley study identified:

For task expansion: Define role boundaries that AI doesn’t override. If a product manager’s job is product management, their AI use should enhance product management --- not transform them into an amateur engineer. The fact that AI can fill a knowledge gap doesn’t mean it should. Some knowledge gaps exist for good reasons: specialization, accountability, quality control.

For blurred boundaries: Establish norms around when AI is and isn’t appropriate to use. This sounds paternalistic until you recognize that every healthy organization already does this for other tools. You don’t take sales calls during board meetings. You don’t write code during customer dinners. The norm isn’t about restricting freedom. It’s about preserving the distinction between focused work and always-on availability.

For increased multitasking: Measure outcomes, not activity. If a team is running six concurrent AI threads and producing better results, that’s fine. If they’re running six concurrent AI threads and producing more volume at lower quality, that’s a problem. The measurement system needs to distinguish between the two --- and right now, most don’t.

The Evidence in Full

A responsible assessment of AI’s impact on productivity requires looking at the full evidence base, not cherry-picking studies that support a predetermined conclusion.

The strongest peer-reviewed evidence comes from Brynjolfsson, Li, and Raymond’s 2023 study, published in the Quarterly Journal of Economics. They found a 14% productivity increase in customer support, with a 34% increase for novice workers. This is rigorous, peer-reviewed, and specific to a domain where AI’s advantages are clear: structured interactions, measurable outcomes, established best practices.

Bick and colleagues at NBER found 5.4% time savings for AI users, translating to 1.4% in aggregate across the workforce. Not the revolution vendors promise, but not nothing.

The METR trial showed a 19% slowdown for experienced developers on familiar codebases --- but with only 16 participants using early-2025 tools. The sample is too small and the tools too dated to draw universal conclusions. What it does demonstrate convincingly is the perception gap: developers thought they were faster when they weren’t.

The Berkeley ethnographic study adds qualitative depth to what the quantitative studies suggest: even when AI is helpful at the task level, organizational dynamics can convert those gains into intensified workloads.

Taken together, the evidence says something more interesting than either “AI works” or “AI doesn’t work.” It says: AI changes work in ways that are difficult to perceive from inside the experience, and organizations that don’t measure and govern those changes will be surprised by the outcomes.

What To Do With This

If you’re leading an organization that uses AI tools --- which, at this point, means most organizations --- here is what the evidence suggests.

Stop relying on how people feel about AI. The perception gap is replicated across studies. Self-reported productivity gains are unreliable. Instrument your systems. Measure actual throughput, cycle time, quality, and scope creep. Compare equivalent work with and without AI assistance.

Define AI boundaries before the tool does it for you. The Berkeley study shows that without explicit boundaries, AI expands the scope of every role it touches. This may or may not be desirable. But it should be a decision, not an accident.

Measure intensity, not just output. More output is only valuable if it doesn’t come at the cost of sustainability. Track hours worked, concurrent task counts, and boundary violations alongside traditional productivity metrics. If your team is producing 30% more while working 40% harder, you haven’t gained productivity. You’ve borrowed from the future.

Treat AI norms as a management responsibility. Individual employees will optimize for their own productivity. That’s rational behavior. But individual optimization doesn’t guarantee organizational health. Someone needs to watch the system-level effects --- and that someone is management.

Read the studies, not the headlines. The Berkeley research is worth reading. So is the METR trial, the Brynjolfsson study, and the DORA report. They tell a more nuanced and more useful story than any headline. Leaders who make decisions based on “Harvard says AI causes burnout” will make different --- and worse --- decisions than those who understand what was actually studied, how, and with what limitations.

The Trap

The intensity trap is not that AI makes you work harder. It’s that AI makes working harder feel like working smarter. The subjective experience is one of capability and flow. The objective reality --- measured in hours, scope, and sustainability --- may tell a different story.

The organizations that avoid this trap won’t be those that reject AI. They’ll be those that govern it: setting boundaries, measuring reality instead of perception, and treating AI adoption as an organizational design challenge rather than a tool deployment exercise.

The Berkeley study didn’t prove that AI is bad for workers. It documented what happens when a powerful tool meets an organization without the norms to manage it. That’s not an AI problem. It’s a leadership problem.

And leadership problems have leadership solutions.

Sources

Aruna Ranganathan and Xingqi Maggie Ye. “Research: AI Tools Make Work Harder, Not Easier.” Harvard Business Review, February 2026.
METR. “Measuring the Impact of Early-2025 AI Models on Experienced Open-Source Developer Productivity.” July 2025.
Erik Brynjolfsson, Danielle Li, and Lindsey R. Raymond. “Generative AI at Work.” Quarterly Journal of Economics, 2023.
Alexander Bick, Adam Blandin, and David Deming. “The Rapid Adoption of Generative AI.” NBER Working Paper, 2024.
Google DORA. “State of DevOps Report 2024: The AI Impact.” Google Cloud, 2024.

Victorino Group helps organizations build the governance infrastructure that turns AI from an intensity trap into a sustainable advantage. If your AI adoption is outpacing your organizational norms, let’s talk.