- Term: the HBR study by researchers at the consultancy BCG (5 March 2026) calls mental fatigue from excessive use or supervision of AI tools “AI Brain Fry”. It is not a medical diagnosis.
- Prevalence: 14% of AI users in large US companies have experienced it at some point (1,488 respondents, self-reported). That is not 14% of all employees, and the question was about "ever experienced".
- Percentages: the figures on supervision, fatigue, decision fatigue, errors and burnout are relative differences in scale scores, not shares of people. Example: people who supervise a lot of AI report on average 14% more mental effort.
- Context switching: That switching has a cost is well documented in the lab (Monsell 2003, Rubinstein et al. 2001). In the study, the link to the questions on task switching is less clear than the link to information overload.
- Counterpoint: less routine work through AI was linked to 15 % lower burnout scores. A field study found on average 15 % more cases resolved per hour (Brynjolfsson et al. 2025).
- Work rhythm: start, focus, harvest, read notifications only during the harvest, and run only as many runs as you can review. For individual building blocks there is evidence from other fields (Fitz 2019 on bundled notifications, Masicampo and Baumeister 2011); for supervising AI, none has been tested.
On some days I have 5 to 6 AI runs going at the same time. Starting them takes minutes. After that, the results arrive one after another, and each wants to be checked, corrected or approved. I jump back and forth between texts, research and drafts. Whether that is already AI Brain Fry, I can't say, because there is no diagnosis for it.
Anyone who uses AI agents or introduces AI in a business will sooner or later run several runs at once and hit the same question: how many can I still sensibly review? This post separates what the AI Brain Fry study shows, covers the research on context switching, presents the opposing view and describes my own working rhythm.
AI Brain Fry: definition and distinction from burnout
The term “AI Brain Fry” was introduced by Bedard, Kropp, Hsu, Karaman, Hawes and Rosen Kellerman in the HBR article “When Using AI Leads to ‘Brain Fry’” of 5 March 2026. All of them work at the consultancy BCG, and 2 of them are also doing a PhD at the University of California, Riverside. Their definition, loosely translated: mental fatigue from excessive use or supervision of AI tools beyond your own cognitive capacity. No established word has emerged in German. Media write about “AI exhaustion”, “AI burnout” or “AI fatigue”, and a trade portal headlines “fried brain”.
AI brain fry is not the same as burnout. The authors describe burnout as chronic work stress with emotional exhaustion. In their reading, brain fry is an acute overload of attention, working memory and control of one's own thinking. It is not a medical diagnosis.
A note on the source: BCG itself sells consulting on AI adoption. The article recommends putting 70 % of the effort for AI projects into people and processes. That does not make the numbers wrong, but it shows a vested interest in the topic.
The study's figures and how to read them
1,488 full-time employees at large US companies were surveyed, once and by self-report. I did not find a questionnaire or a working paper with the methodology. What matters for every figure is what it actually measures:
| Finding of the study | How to read the number |
|---|---|
| 14% of AI users said they had experienced brain fry at least once | Share of people, but only among AI users in large US companies, and the question was “ever experienced”. It would be wrong to say: “1 in 7 suffers from it.” |
| High instead of low oversight of AI: 14 % more mental effort, 12 % more mental fatigue, 19 % more information overload | Relative differences in self-reported scale scores between 2 groups, not shares of people. |
| With brain fry: 33% more decision fatigue, 11% more minor and 39% more major errors | These are also scale values from self-assessment, not counted errors. “Brain Fry causes 39 % more serious errors” would be wrong. |
| Intention to quit: 25% without, 34% with brain fry | Shares of people compared. I can't recalculate the relative figure HBR cites from the reported values, so it isn't shown here. |
| By function: 26 % of employees in marketing functions and 6 % in legal functions said they had experienced Brain Fry | Functions of large US companies, not industries. |
| AI tools used at the same time: perceived productivity rises clearly from 1 to 2, then more weakly, and drops slightly from the fourth | What was measured is perceived productivity. Whether tools or agents were counted is unclear. No limit for agents follows from it. |
For everyday work, 2 findings matter more than any percentage. The first: what was linked mainly with strain and mental fatigue was the ongoing oversight of AI outputs, plus a workload that had grown through AI. Oversight is at the center of the study. The second concerns context switching: the free-text answers name information overload and task switching as important factors. However, the statistical link with the questions on task switching is less clear than the one with information overload. Context switching does not qualify as the “main cause” in this study.
One more caveat: the authors write that AI use "may be causing" the exhaustion, and that supervision "predicted" it. These are cautious words for good reason. In a cross-sectional design, the opposite direction is also conceivable: people who are overloaded anyway may experience supervision as more taxing.
Context switching and supervision: why they cost so much attention
That switching tasks costs time and accuracy has been researched by psychology for decades, though not on working with AI. The details come from the summaries of the studies, and for Mark from the papers.
- Switching costs: After a task switch, people in the lab respond more slowly and usually make more errors. Preparation lowers the costs but does not remove them (Monsell 2003, review of simple tasks). They rise with the complexity of the rules, fall with cues and are higher for unfamiliar tasks (Rubinstein, Meyer, Evans 2001).
- Attention residue: Attention has trouble letting go of unfinished tasks, and performance on the next task suffers. The technical term for this is “attention residue” (Leroy 2009, 2 experiments).
- Interruptions: Among 24 employees of one company (each observed for 3.5 days, study duration 13 months), 57 % of work sessions were interrupted. They returned to the interrupted work on the same day after 25 min 26 s on average. The spread is huge (standard deviation 54 min 48 s), and 77 % were resumed at all on the same day (Mark, Gonzalez, Harris 2005). In the lab, interrupted people worked faster, but with more stress and effort (Mark, Gudith, Klocke 2008, 48 participants, mostly students).
- Oversight of automation: People who monitor an automated system check less under multiple demands (complacency) and follow suggestions too quickly even as experts (automation bias). Training does not prevent this (Parasuraman and Manzey 2010, review article).
My take: if you run 5 threads, you switch between 5 tasks and constantly review other people's results at the same time. Both together fit what the study and this research describe. It is not proven for working with AI. The lab and office studies date from before AI agents, and the interruption studies have small samples.
The other side: handing off routine brings relief
The same study has a second message that is easily lost next to the strain figures. Employees who had less routine work thanks to AI reported 15% lower burnout scores. The scores for mental fatigue were not lower. It is again an association from self-report, but it argues against reading AI use across the board as a burden. Less fatigue was also associated with leadership answering questions about AI (15% less) and with employees feeling that their company values work-life balance (28% less).
A field study points in the same direction. In “Generative AI at Work” (published 2025), Brynjolfsson, Li and Raymond analyzed data from 5,172 support employees, among whom an AI assistant was introduced in stages in 2020 and 2021. On average they resolved 15 % more cases per hour. Less experienced and less qualified staff benefited more; among the most experienced, speed gains and quality losses were small. Employee well-being was not measured directly, and support with a 2020 assistant is something different from managing several agents.
Strain and relief sit side by side in the data. My reading: it depends on how much oversight fits within your own capacity. Less AI does not follow from that.
Work rhythm: start, focus, harvest
From here on I describe my way of working. It is the experience of a single solo entrepreneur, not a tested method. Where there is evidence from other fields for a building block, it is noted there. Whether these building blocks work the same way when supervising AI is an open question.
The basic idea: I limit my touchpoints with the AI and leave the duration of the runs open. They may take a long time, as long as I only look in at fixed times. This gives 3 blocks:
| Block | What happens | Duration |
|---|---|---|
| Start | Send off all tasks one after another, each with a clear result | 15 to 30 min |
| Focus | A task of its own that only I can solve. No reading results, no watching runs | 90 to 120 min |
| Harvest | Review results, approve, start corrections as new tasks (which is also the next start) | 30 to 45 min |
Start: batch similar jobs
At the start, a clean task is what counts. I send off similar tasks as a batch, 5 to 6 in parallel. What doesn't fit in the batch waits for the next start. Preparation and cues lower switching costs (Rubinstein, Monsell). That is why it's worth adding a sentence to the task about how you will recognize the finished result and what you plan to do with it. This is a transfer from the lab.
Focus: read notifications only at harvest time
In the focus block I read no results and look at no runs. Notifications may arrive, but they are only read at harvest time. That is less than bundling: they still arrive one by one, I just leave them unread.
Bundling has a smartphone field experiment. Fitz and colleagues (2019) had 237 people get notifications as usual, bundled 3 times a day, or not at all. The bundled group reported more attention, better mood and less stress. The group with none gained little, but felt more anxiety and fear of missing out. Bundling beat switching off. To support the focus block, have notifications delivered in batches. Whether just leaving them unread works equally well, the study doesn't show. iPhone and Pixel settings are in the post on constant availability.
Harvest: review, approve, note changes
At harvest time I review the results one after another, approve and send off corrections as new tasks. At the same time the next start begins. Anyone who switches between 2 results or back to their own task can note in 1 sentence where and how they will continue. In 2011, Masicampo and Baumeister showed in lab experiments that a concrete plan for when and how a goal will be completed ends the disruptive effect of an unfinished goal. Whether this applies to AI results left lying around, I found no study on.
After the harvest and the new start, before the next focus block begins, a real break helps. A meta-analysis (22 samples, 2,335 people) found that micro-breaks raise vitality and lower fatigue, both to a small degree (Albulescu et al. 2022). For performance on demanding work, the effect is not established; the recommendation rests on well-being.
As many threads as you can review
The bottleneck is reviewing, not starting. A day is full when the harvest no longer fits into working hours. This matches experience reports. Addy Osmani ties the number of agents to his review rate and writes that for most people it is a low single-digit number (The Orchestration Tax, 2026). Simon Willison says he can only concentrate on reviewing and accepting one larger change at a time (2025).
The HBR study's threshold of 3 tools is no good as a yardstick for runs, because it is unclear what was counted. A simple rule is enough as a test: start only as many runs as you will review thoroughly in the next harvest block. After a week, look at how many results were left sitting.
Limits of the available studies
The brain fry study is based on self-report and a cross-section, comes from the USA and from large companies, and was published in a management magazine without peer review. The brain fry question is a single yes-no item with “ever”, not a validated questionnaire. The 33% more decision fatigue is a scale value. The mechanism behind it (ego depletion) could not be reliably shown in 2 large replication studies (Hagger et al. 2016, Vohs et al. 2021).
For the countermeasures, the picture is thinner. Fitz, Masicampo and Albulescu studied smartphones, lab tasks and breaks, not oversight of AI. I found no study on solo entrepreneurs with several agents. My own experience is a single data point. Anyone who runs similar workloads can test the rhythm as a hypothesis and measure it against their own harvest.
How I run AI runs in everyday work and what goes wrong is something I show in the AI Sync newsletter, which comes out every Sunday.
Everything on this page can be implemented without me. I add new tools and changes here, and every Sunday there is a tip about it in the newsletter.
To AI SyncI set up AI processes with you that deliver their results collected at fixed times for approval, and we define together who checks what. On site from Eisenstadt to Mattersburg and Wiener Neustadt, online across the whole German-speaking region.
Learn moreFrequently asked questions
What is AI brain fry?
AI Brain Fry is a term from an HBR study by researchers at the consultancy BCG (5 March 2026). It means mental fatigue from using or supervising AI tools beyond your own capacity. The term is not a medical diagnosis. The study is based on self-reports from 1,488 employees of large US companies.
How common is AI Brain Fry?
In the study, 14 % of AI users answered yes to the question of whether they had ever experienced it. This applies to AI users in large US companies, and the question was about “ever experienced”. The value varied widely by function: 26 % of employees in marketing functions, 6 % in legal functions. I know of no figure for solo entrepreneurs or SMEs in Austria.
Is AI Brain Fry the same as burnout?
No. The authors describe burnout as chronic work stress with emotional exhaustion, and brain fry as acute overload of attention and working memory. In the study, less routine work thanks to AI was associated with 15% lower burnout scores; the scores for mental fatigue were not lower. Media nevertheless often write about "AI burnout".
How many AI tools or agents at the same time is too many?
I did not find a fixed number in any study. In the study, perceived productivity rose clearly from 1 to 2 tools used at the same time, then more weakly, and dropped slightly from the fourth. What exactly was counted remains unclear. The developer Addy Osmani ties the number of agents to his review capacity and names a low single-digit number for most people. That is a rule of thumb from experience.
What helps against AI brain fry?
No tested measure exists for AI oversight. Other fields offer hints: bundled notifications, 3 times a day (Fitz et al. 2019, 237 people), a concrete plan for unfinished tasks (Masicampo and Baumeister 2011) and short breaks against fatigue (Albulescu et al. 2022). The study itself names handing off routine and setting limits for oversight. Whether this works for AI oversight is open.
Does AI make you tired or does it relieve you?
Both show up in the data. High rather than low oversight of AI went along with 12% more mental fatigue (relative difference in self-report). Less routine work went along with 15% lower burnout scores. In a field study of 5,172 support employees, those with an AI assistant resolved on average 15% more cases per hour (Brynjolfsson et al. 2025).
Changes to this post
- 10 November 2026First version.
