Think of the person on your team who took to AI fastest. The one with a fleet of assistants running at once, prompts saved in a document, an answer to every "can AI do this?" question in the building. A year ago they looked like the future. Lately they look tired. Not lazy-tired or overworked-tired in the ordinary way, but a particular kind of frayed — the sort where they have to get up and walk away from the screen to think a single clear thought.

There's a reason for that, and it has nothing to do with the person. The tools that were supposed to take work off our plates have, for the people using them hardest, quietly put a different kind of work on. We were sold AI as a way to do less. What it actually did was change the shape of the job — from doing the work to supervising a machine that does a rough version of it — and supervising turns out to be more tiring than doing, not less.

This is the part of the AI story almost nobody is pricing in. The output looks like a clean productivity win, so the cost stays invisible. The report got written, the code got shipped, the emails got cleared. What you can't see on the dashboard is what it took out of the person to get there: the constant switching between tools, the low-grade vigilance of checking work that arrives confident and fluent and sometimes wrong, the mental effort of holding three half-finished AI conversations in your head at once. There's a name for where that ends up now.

Brain fry is a real thing, and it has a shape

The consultants at Boston Consulting Group have started calling it "AI brain fry" — mental exhaustion from pushing the use and supervision of AI tools past our cognitive limits. In their survey of 1,488 full-time US workers, 14% said they'd experienced it. The people reporting it weren't dabblers poking at a chatbot once a week. They were the heavy adopters, the ones running small armies of agents that all need watching, drafting long prompts, reading through more machine-generated output in a day than a person can properly hold.

The symptoms they described are worth repeating, because they're physical: brain fog, trouble focusing, headaches, decisions that come slower than they used to. Some said they had to physically step away from the computer to reset. And the roles hit hardest weren't the ones you'd guess — marketers led the list, ahead of HR, operations, engineering and finance. What links them is the sheer volume of AI they're each expected to wrangle, whatever the job on the door.

Whether this is a passing phase — the rough early days of tools we haven't learned to live with — or a permanent feature of the work, I don't know. What's not in doubt is that it's real now, and it's hitting exactly the people a business can least afford to wear out.

If you've felt this yourself, the useful thing to understand is why supervising drains you faster than doing did. When you write something from scratch, there's friction, but it's one continuous task and your attention has somewhere to settle. When you orchestrate AI, you're doing something closer to managing a room full of fast, eager, slightly unreliable juniors. You frame the task, wait, read what comes back, work out whether it's any good, decide what to keep, and start again — across several tools at once, each with its own quirks. Every loop is a small act of judgement, and that's the most expensive thing your brain does. Do enough of those in a day and you're spent, even though the tab count makes it look like you were barely working.

There's a quieter reason it wears you down, too. The machines are built to feel trustworthy. A recent Stanford study found that AI models endorse the user roughly 49% more often than another person would — they're agreeable by design, because agreeableness is what keeps people coming back. That's lovely until you remember your job is to catch the times they're wrong. You're being asked to stay sceptical of something engineered to sound reassuring. Vigilance against a thing that's actively trying to please you is exhausting in a way that vigilance against an ordinary error isn't. And the harder an organisation pushes people to lean on these tools, the more of that draining checking it's asking for — usually with less time to do it.

The mandate makes it worse

Now layer on the bit that turns an individual strain into an organisational one. Employers have moved from suggesting AI to requiring it. JPMorgan now tracks how its 65,000 engineers use AI tools, sorting them on a dashboard into light, heavy and non-users, and folding that into performance reviews. The message is plain enough: use the tools, and be seen to use them.

I understand the instinct — leaders who've spent on AI want to know it's actually being used. But measuring usage rather than outcomes rewards exactly the behaviour that fries people. An engineer at a bank is now asked to do two things that pull against each other: lean on AI enough to register as a proper adopter, and check every line of its output thoroughly enough to meet banking-grade standards. Under deadline, those two demands don't sit comfortably together, and the person in the middle absorbs the tension. I've written before about how to handle an employer who mandates AI without telling you how — but the harder fix sits with the people setting the targets, who are counting logins when they should be counting results.

This is a different problem from AI making us worse at thinking — that's the opposite failure, where you hand the thinking over and slowly lose the muscle. The cognitive tax is what happens when you keep the judgement and apply it relentlessly, across more decisions than a human day was built for. One is the cost of thinking too little; the other is the cost of supervising too much. The same tool can hand you either bill, and the heaviest users are the ones most exposed to the second.

Treat your attention as the scarce resource

Using less AI isn't the fix. Used well it expands what a small team can do, and stepping back from it just hands the advantage to whoever doesn't. What helps is to stop treating your own attention as infinite and the AI's as the thing to optimise. It's the other way round. The tool is abundant and cheap; your capacity to make good calls in a day is finite and the actual asset. Protect it deliberately.

That means setting a few boundaries, and none of them are clever:

  1. Cap how many tools run at once. Pick the two or three that earn their place and close the rest. Most of the fry comes from switching, not from any single tool. A smaller, deliberate kit beats a sprawling one every time.
  2. Batch the AI work, don't thread it through everything. Set aside blocks for the prompt-and-review loop rather than dipping in every few minutes. Constant context-switching is where the energy leaks out, and interleaving AI with everything else maximises the switching.
  3. Decide what "good enough" looks like before you prompt. Name the standard the output has to clear up front. Without that line, you'll either over-check everything or trust too readily — and the over-checking is what burns the hours and the patience.
  4. Keep a daily block that's AI-free. Thinking without a machine in the loop is restful, and it's still where your best original work comes from. The reps keep your judgement sharp for the moments the AI gets it wrong.
  5. If you manage people, measure output, not usage. A login count tells you nothing about value and rewards the exhausting behaviour. Ask what got better, not how many prompts got sent.

The promise of AI was more capacity. For the people using it most, it's delivered the opposite — busier, more drained, and wondering why a technology that does so much of the work has left them with less in the tank. That's not the tool's verdict, it's a design choice we keep making by default. Decide how AI fits your day before it decides for you, and treat the limited, irreplaceable resource — your own clear thinking — as the thing worth protecting.