AI tools can boost productivity, but companies need to be wary of overdoing it.
That’s according to a new study published in the Harvard Business Review, which surveyed 1,488 full-time U.S.-based workers in large companies across a variety of industries.
The study found that a small but noticeable number of workers report experiencing mental fog, headaches, and slower decision-making, or what the authors call “AI brain fry.”
“We look at this as kind of the canary in the coal mine,” Matthew Kropp, one of the study’s lead authors and a managing director and senior partner at Boston Consulting Group, told Business Insider. “Those engineers that are the early adopters, that are doing the multi-agent orchestration are experiencing this effect, more and more people are trying to move up to that level.”
Of particular note, the study found that AI tools can boost productivity but only to a certain extent.
“I had one tool helping me weigh technical decisions, another spitting out drafts and summaries, and I kept bouncing between them, double-checking every little thing,” one senior engineer manager told the authors. “But instead of moving faster, my brain just started to feel cluttered. Not physically tired, just… crowded.”
Employees who go from one to two AI tools simultaneously experience “a significant increase in productivity,” they wrote. But the increase is smaller when they go from two to three tools, and then it decreases from three onward, illustrating that multitasking can have its limits.
“We’re not saying we shouldn’t manage multiple agents,” Kropp, who is also chief AI officer of BCG’s X, the firm’s team of roughly 3,000 engineers that build AI solutions, told Business Insider. “It’s the reality. This is how we will be working in a lot of jobs. You will have humans managing agents. I think what’s important is that we are conscious of the fact that it has this impact and that we manage for it.”
Kropp said that he and his coauthors wanted to distinguish between general career burnout and the specific AI-related fatigue some workers are beginning to experience as agentic AI transforms their jobs and their day-to-day routines.
“It’s a very specific effect, because what’s happening is, in order to oversee an agent effectively, this actually a high-congitive load,” he said. “This isn’t something rote. If I’m a software engineer, I’m having it design and write code for me and it really matters what it does, the output has to be good.”
The high stress of managing one agent is then amplified by each additional agent, to the point where workers reach a breaking point. It’s why the authors say businesses should be diligent in crafting their AI policies, including allowing for breaks.
“If I’m 50 times more productive, maybe I should be 20 times more productive, but have better mental health and not want to quit,” Kropp said.
Steve Yegge, a veteran software engineer, recently suggested imposing a 3-hour cap on engineers who use AI-assisted coding.
The occurrence of “AI brain fry” doesn’t mean companies should avoid AI adoption, Kropp said. The study also found that when AI is used to replace “routine or repetitive tasks,” the rate of burnout declines, even if workers’ mental fatigue does not.
“I think part of where it’s coming from is people get how to use these tools, and they can’t stop because it is so empowering,” he said. “And so I think the tools are positive, we just have to manage the negatives.”
The study found that 14% of the workers surveyed said they were experiencing “AI brain fry,” but the percentage of workers in industries varied widely. Employees in marketing (25.9%), human resources (19.3%), operations (17.9%), and software engineering (17.8%) reported a much higher rate of “AI brain fry” than those in legal and compliance roles, which were at roughly 6%.
The differences, Kropp said, are due in part to the adoption rates among industries as a whole. He said that nine months ago, AI agents didn’t have the capabilities they do now, and based on the current rate of change, other industries are likely to experience the upheaval that software engineering and computer programming are going through now.
“If you think about that sort of ladder of adoption, at least in enterprise organizations, you really have less than 5% that are in those top two levels that are even using agents at all, let alone using multiple at a time,” he said. “So, it’s still very early.”
