If you’ve ever used an online patient portal to message your doctor in the middle of the night, you won’t be surprised to learn that responding to those messages takes an increasingly big bite out of clinicians’ workdays.

So in recent years, hospitals have begun adopting an AI tool that can draft responses for them. The tool was supposed to make a time-consuming task go more quickly and smoothly, said Philip Barrison, an MD-PhD student at the University of Michigan Medical School who studies AI in healthcare.

Instead, the tool has given doctors and nurses a new to-do list. First they have to read the AI-generated response and decide if it “is actually something that they think they would say,” Barrison said. Humans are suggestible, and looking at something and deciding whether you would have thought of it on your own is a cognitively complex task.

Even if the message looks correct, the clinician still needs to “edit it to the point where they think it’s acceptable” to send to a patient, Barrison said. The AI tool introduces a totally new set of complicated judgment calls into what used to be a relatively straightforward process. As a result, many clinicians have chosen not to use it at all.

They’re fortunate to have the choice. Buoyed by expectations of cost savings and skyrocketing productivity, companies are increasingly asking (and sometimes requiring) employees to use AI to make their work more efficient. Meta, for example, last year instructed some workers to use AI to “go 5X faster by eliminating the frictions that slow us down.” The CEO of Shopify told employees they’d need to prove they “cannot get what they want done using AI” before the company would approve new hires. Some companies are even evaluating or ranking employees based on how much they use AI tools.

Workers in some sectors have found major time savings from AI. But for others, the tools just change the work rather than making it faster. Workers might be spending less time writing patient portal messages, for example, but more time editing the releases the AI tool writes.

At best, this mismatch between employer expectations and employee reality can be an annoyance. In other cases, however, it can result in workers being laid off for failing to meet unrealistic efficiency demands. Some critics say the overzealous adoption of AI in high-stakes settings like healthcare even puts people’s lives at risk. Now workers, unions, and experts are increasingly calling for guardrails to protect employees from inflated expectations around AI — and customers, students, patients, and the general public from mistakes that can happen when managers put AI adoption above all else.

Corporations are increasingly presenting employees with a choice: Use AI to be more productive or “you’re going to be automated out of a job,” said Aiha Nguyen, director of the labor futures program at the research organization Data & Society.

But the effects of AI on productivity aren’t as straightforward as some CEOs have claimed. In one 2025 study, software developers believed AI made them faster, but in fact they took 19 percent longer to complete tasks. (The researchers tried to repeat the experiment this year but had trouble recruiting developers who would agree to work without AI.) And in a recent survey of 5,000 white-collar workers, 40 percent of rank-and-file employees said AI saved them no time at all.

Workers across heavily AI-exposed fields point to hidden timesucks that come with using the technology. Julie, an art teacher, wrote in a response to a Vox reader survey that her school’s administrators routinely suggest using AI for lesson-planning, emails, and progress report comments. She’s tried AI-generated lesson plans, but they don’t account for the fact that kids may work through an activity at different speeds.

“First, I am checking what AI suggests, then I am editing them,” she said. “Why add a step I can accomplish on my own?”

For an employee at an East Coast communications agency, an internal AI tool was supposed to speed up the process of drafting press releases and other documents about the pharmaceutical industry.

“The goal is, I think, to be able to plug and chug into this machine and be able to turn a lot of materials around a lot quicker than we already do,” said the employee, who asked to remain anonymous for fear of career repercussions.

But when the employee tried to use it for basic research, it made too many mistakes. Double-checking its work erased any time savings. When the employee tried using it for communications with clients, its people-pleasing tendencies became a problem, as the tool put a “weird happy spin” even on messages warning of bad news.

“Part of the reason we take a human speed to turn things around is because there is so much nuance behind everything that we do,” the employee told me. “AI is just not going to be able to catch it.”

It’s not just that AI makes errors. With the advent of agentic AI, workers are increasingly being asked to edit and oversee the output of multiple AI tools, a new kind of work that can have unexpected costs.

One recent study of 1,488 workers across industries, for example, found that excessive oversight of AI agents could lead to what the researchers called “AI brain fry,” a kind of cognitive fatigue. “Participants described a ‘buzzing’ feeling or a mental fog with difficulty focusing, slower decision-making, and headaches,” the researchers wrote in Harvard Business Review. Brain fry was also associated with an increased number of errors and an increased desire to quit one’s job.

The researchers also found that while using one or two AI tools increased productivity, adding additional tools produced diminishing returns, and after four tools, productivity actually declined.

Despite such findings, companies continue to pressure employees to use AI, and to cite AI investment as a rationale for layoffs, even as companies that try to link staff reductions to AI adoption tend to struggle on the stock market.

Some workers and organizations, however, are beginning to push back. National Nurses United, the country’s largest nurses’ union, has criticized the use of AI tools in hospitals to estimate staffing needs or to recommend treatment protocols for patients.

There’s no guarantee that these tools will take into account a patient’s individual profile, including underlying medical conditions, the way human clinicians can, Cathy Kennedy, the union’s president, told me. AI is supposed to “help us do our work more efficiently, but at the end of the day, it makes it even more burdensome,” she said.

Hospitals need to evaluate, with nurses at the table, whether AI tools really work as advertised, Kennedy said. “We have to stop — we have to go back and really see if this is truly doing what it needs to do,” she said.

The same is true across industries, Barrison, the healthcare researcher, told me. “Organizations need to be prepared to say when, if they were seeking a return on investment, if they were seeking value in a technology — how do you define what that value is? And if there’s not value there anymore, how do you turn it off?”

Some workers have found ways that AI actually helps them do their work — just not the ones management expected. Julie, the art teacher, likes to use Claude to learn more about topics she’s less familiar with, like kiln-firing ceramics.

Meanwhile, researchers have found that AI can actually reduce employee burnout, if it’s used to complete tasks employees find burdensome. “Everybody in every job has a list of things that they procrastinate on,” said Julie Bedard, a managing director and partner at Boston Consulting Group who led the AI brain fry study. “Those are the places I get, unsurprisingly, a lot of enthusiasm to try AI with.”

But employers won’t find out what those burdensome tasks are unless they listen to rank-and-file employees. “Worker standards and worker rights should continue to be at the heart of all of this,” Nguyen said, “rather than just focusing too much on the AI.”