I have been using AI coding tools daily for over a year now. Claude Code, Cursor, Perplexity, ChatGPT for brainstorming, various MCP servers, automation scripts. At some point in the last few months, I noticed something I could not explain.
I was shipping more code than ever. My output was objectively higher. But by 3 PM most days, my brain felt like someone had microwaved it. Not tired in the normal “long day of coding” way. A different kind of exhaustion. A fog that made it hard to make even simple decisions, like what to eat for dinner or whether a variable name was good enough.
I thought it was just me being dramatic. Then a BCG study dropped in March 2026, and it gave this feeling a name: AI brain fry.
Turns out, I was not the only one.
The BCG Study That Changed the Conversation
Boston Consulting Group surveyed 1,488 full-time workers in the US and published the results in Harvard Business Review. The findings are not subtle.
Workers using three or fewer AI tools reported genuine productivity gains. No surprise there. But once people crossed the threshold of four or more AI tools, productivity plummeted. Not stayed flat. Plummeted.
The cognitive costs were measured precisely:
- 14 percent more mental effort expended at work when heavily overseeing AI output
- 12 percent greater mental fatigue
- 19 percent greater information overload
- 34 percent of workers experiencing brain fry actively intended to quit their jobs, compared to 25 percent without the condition
Julie Bedard, the BCG managing director behind the study, put it simply: “People were using the tool and getting a lot more done, but also feeling like they were reaching the limits of their brain power.”
That sentence captures exactly what I have been experiencing. More output. Less capacity. A strange combination that does not feel sustainable.
What AI Brain Fry Actually Feels Like
Before the BCG study gave it a clinical name, developers were already describing it in forums, Twitter threads, and Slack channels. The descriptions are remarkably consistent.
Participants in the study described a “buzzing” feeling or mental fog. Difficulty focusing. Slower decision-making. Headaches. If you have spent a full day reviewing AI-generated code across multiple tools, context-switching between agents and editors and chat interfaces, you probably know exactly what they are talking about.
It is not the same as being physically tired from a long coding session. Traditional coding fatigue comes from sustained deep focus on a single problem. You are tired because you went deep. AI brain fry comes from going wide. You are tired because you were monitoring, evaluating, context-switching, and making micro-decisions across multiple AI interactions simultaneously.
The mental model is closer to air traffic control than carpentry.
The Productivity Paradox Nobody Warned Us About
Here is the part that genuinely surprised me.
An eight-month observational study at a US technology company with around 200 employees tracked what actually happens when a team adopts AI tools. The researchers identified three specific patterns:
Task expansion. When AI made unfamiliar work feel accessible, people started doing things outside their role. Product managers started writing code. Researchers handled engineering tasks. Engineers found themselves coaching colleagues who were now using AI to venture into technical territory. Everyone’s job description quietly expanded.
Blurred work-life boundaries. The conversational nature of AI prompting made it feel less like work and more like chatting. People started prompting during lunch breaks, during meetings, while waiting for files to load. One researcher noted that work “naturally extended into evenings without deliberate intention.” Because prompting an AI does not feel like sitting down to code, the psychological barrier to working off-hours disappeared.
Increased multitasking. Workers managed multiple concurrent AI threads simultaneously. One agent building a feature. Another reviewing a PR. A third answering a research question. The study described it as “a sense of always juggling, even as the work felt productive.”
The conclusion from the researchers is worth quoting directly. Workers thought AI would enable reduced hours, but “really, you do not work less. You just work the same amount or even more.”
Ninety-six percent of C-suite executives expected AI to improve productivity. Seventy-seven percent of employees reported it actually increased their workload.
Those two numbers sitting next to each other tell you everything about the disconnect between how leadership thinks AI works and how people experience it in practice.
Why Developers Get Hit the Hardest
The general workforce data is concerning enough. But developers face a specific version of this problem that I think is worth calling out separately.
When a marketing team uses AI to draft emails faster, the cognitive overhead is relatively low. The output is text. You read it, tweak it, send it. The feedback loop is simple.
When a developer uses AI to generate code, the cognitive overhead is massive. You are not just reading output. You are:
- Evaluating whether the logic is correct
- Checking if it follows your existing patterns
- Verifying it does not introduce security vulnerabilities
- Confirming it handles edge cases the AI did not think about
- Tracing how it interacts with the rest of your codebase
- Deciding whether the approach is architecturally sound
Every single AI-generated code block requires a mini code review in your head. And unlike reviewing a human colleague’s PR, where you can trust that they tested it and thought about it, you know the AI might have confidently generated something that looks correct but is subtly wrong.
This is why the METR study from mid-2025 found that experienced developers believed AI made them 20 percent faster, but objective measurements showed they were actually 19 percent slower. The cognitive cost of verification was eating the productivity gains.
Pull requests containing AI-generated code had roughly 1.7 times more issues than human-written code alone. So you are reviewing more code, catching more problems, and using more mental energy per line. The math does not add up the way the productivity dashboards suggest.
The Three-Tool Threshold
The BCG finding about three tools being the tipping point resonates with my own experience. Let me explain why I think three is the magic number.
With one or two AI tools, you build a mental model of what each tool is good at. You develop intuition for when to trust the output and when to double-check. Your brain learns the tool’s patterns, and reviewing its output becomes semi-automatic.
At three tools, you can still manage. You have your editor AI, your chat AI, and maybe a research tool. Each has a different interface and different strengths. But you can hold all three mental models simultaneously.
At four or more, something breaks. You are now context-switching between fundamentally different AI interaction patterns. Each tool has different strengths, different failure modes, different output formats. Your brain cannot maintain reliable intuition for all of them at once. So instead of semi-automatic review, everything requires full conscious evaluation.
That is where the 14 percent increase in mental effort comes from. Your autopilot stops working. Everything becomes manual.
I wrote about how I use AI tools in my daily workflow and one of the points I made was that I intentionally keep my tool stack small. At the time, I framed it as a preference. Now I realize it was self-preservation.
The Workload Creep Problem
There is a second dynamic that makes AI brain fry worse over time, and it is more insidious than the tool-switching problem.
When you start shipping faster with AI, your output goes up. Your manager notices. Your team notices. The natural response is to give you more work, because clearly you have the capacity. Or you give yourself more work, because you feel like you should be doing more now that you have these powerful tools.
The TechCrunch piece on this nailed it: “The employees’ to-do lists expanded to fill every hour that AI freed up, and then kept going.”
This is not a technology problem. It is a management and self-management problem. The time AI saves gets immediately reinvested into more work, not into rest, thinking, or deeper focus on fewer things. The output ceiling rises, but so does the expectation, and the human brain does not scale the same way the tools do.
I have caught myself doing this. Finishing a feature in half the time, then immediately starting the next one instead of taking the break I would have taken if I had coded it manually. The AI made the work faster, but it did not make me less human. I still need the same recovery time between intense cognitive tasks.
What the Research Says Actually Helps
The BCG study did not just identify the problem. It tested interventions. Here is what actually reduced brain fry:
Training on AI tool usage. Workers who received proper training on how and when to use AI tools experienced significantly less cognitive overload. Not “here is how the tool works” training, but “here is when to use it and when to step away” training. Understanding the tool’s limitations prevented the constant vigilance that drives fatigue.
Batching AI work. Instead of interleaving AI interactions throughout the day, scheduling dedicated blocks for AI-assisted work and dedicated blocks for non-AI work reduced the context-switching cost. This mirrors the Pomodoro approach but applied specifically to AI usage.
Breaks before demanding tasks. Scheduling deliberate breaks before tasks requiring deep judgment or decision-making helped counteract the cumulative fatigue from AI oversight. The brain needs transition time between “monitoring AI output” mode and “making important decisions” mode.
What I Changed in My Own Workflow
After reading the research and honestly reflecting on my own patterns, I made some concrete changes.
I consolidated to fewer tools. I stopped trying every new AI tool that launched. I use Claude Code for most development work and Perplexity for research. That is my core stack. Everything else gets evaluated against a high bar: does this genuinely solve a problem my current tools cannot handle, or is it just a different interface for the same capability?
I stopped prompting during breaks. This was the hardest habit to break. The conversational nature of AI tools makes it feel harmless to fire off a quick prompt while eating lunch. But those micro-sessions were preventing genuine mental recovery. Lunch is now a no-screen break for me, and the afternoon brain fog has noticeably decreased.
I time-box AI sessions. I work in 90-minute focused blocks with AI tools, then take a 15-minute break doing something that does not involve evaluating AI output. Reading, walking, or just staring at a wall. The key is giving the “evaluation” part of my brain a genuine rest.
I review in batches, not in real-time. Instead of reviewing AI output line by line as it generates, I let the agent complete a task fully, then review the entire diff at once. This is one focused review session instead of continuous monitoring. It uses less mental energy for the same amount of oversight.
I stopped measuring myself by output volume. This is the cultural shift that matters most. When AI multiplies your output capacity, judging yourself by how much you ship is a trap. The question is not “how many features did I complete today” but “did I make good decisions today.” Those are different metrics, and only one of them correlates with burnout.
The Uncomfortable Truth for the Industry
Here is what I think the industry needs to hear, even though it is not what most AI companies want to say.
AI tools are genuinely useful. I write about them because I believe that. I use them every day because they make me better at my job. But the narrative that AI just makes everything faster and better, with no tradeoffs, is not honest.
The tradeoff is cognitive. The more AI output you supervise, the more mental energy you spend on evaluation rather than creation. At some point, you cross a line where the evaluation cost exceeds the creation savings. That is brain fry.
The junior developer crisis is partly about companies thinking AI can replace human judgment at scale. AI brain fry is about individuals learning that supervising AI at scale is its own form of exhausting work. Both problems come from the same root: underestimating the human cognitive cost of working with AI systems.
The companies that figure out how to give developers the right amount of AI assistance, not the maximum amount, are going to have healthier, more sustainable teams. The developers who figure out how to use AI intentionally, with clear boundaries and genuine rest, are going to outlast the ones who try to run at maximum AI-augmented capacity indefinitely.
The tools will keep getting better. But your brain is not getting an upgrade anytime soon. Treat it accordingly.
What This Means Going Forward
AI brain fry is not going to go away because models improve. Better models mean higher quality output, which is great, but they also mean more ambitious tasks, more output to review, and more pressure to use AI for everything.
The solution is not less AI. It is smarter AI usage. Set boundaries. Consolidate your tools. Take real breaks. Stop measuring productivity by volume alone. And push back when your workload expands just because AI made it theoretically possible to do more.
The developers who thrive in this era will not be the ones who use the most AI tools. They will be the ones who use AI tools the most deliberately.
That is a different kind of skill. And it is one worth developing right now.