Brain Fry
Fluency: Orchestration
In March 2026, the Harvard Business Review published BCG research on a pattern researchers were seeing across organizations: workers who had integrated AI deeply into their workflows reported a specific kind of exhaustion. Not from the work itself — from managing the AI. Prompting, verifying, iterating, switching contexts. The article called it brain fry.
The underlying cognitive science is older and more rigorous: cognitive load theory, attention residue, and the foreman problem — with honest notes on where the evidence is preliminary.
4+
The number behind this guide
AI tools running simultaneously. Managing them becomes the job.
BCG/HBR (2026): workers managing 4+ AI tools concurrently spend more cognitive resources coordinating than thinking. The productivity promise inverts.
Cognitive load. The fixed budget.
Sweller, Ayres & Kalyuga (2011) identified three types of cognitive load. AI management creates a new category of extraneous load that competes with the intrinsic load of the task itself.
Intrinsic load
The inherent complexity of the task. A complex analysis has high intrinsic load regardless of what tools you use. This is fixed by the task.
AI can reduce intrinsic load for well-defined subtasks. This is the genuine productivity gain AI offers.
Extraneous load
Load generated by how the task is presented or managed — interface friction, format overhead, unnecessary steps. This is where AI orchestration lives.
AI management adds new extraneous load: prompting, verification, iteration, context-switching, output review. It doesn't automatically disappear just because AI handles part of the task.
Germane load
Load that produces learning — the cognitive work of building schemas and understanding. This is productive load.
AI may reduce germane load by shortcutting the productive struggle that builds skill. This connects to the Deference Reflex.
items in human working memory — Cowan (2001), the 'magical mystery four'
BCG/HBR named the pattern: brain fry from AI orchestration overhead
Peer-reviewed studies on AI-specific cognitive overload — still largely absent
Bainbridge: automation overhead predates AI — cognitive cost of managing automated systems
Users become foremen.
When AI handles the execution and humans supervise, the cognitive role shifts from worker to foreman. Foreman work has its own overhead — and it doesn't always add up to less total cognitive labor.
In a construction analogy: a foreman who manages ten unreliable workers has more cognitive overhead than a skilled worker who does the job themselves. They must track each worker's output quality, verify it against specifications, identify errors, redirect work, and integrate inconsistent pieces. If the workers were highly reliable, the overhead might be worth it. If they're unreliable, the overhead may exceed the benefit.
AI systems are unreliable in specific, predictable ways. They hallucinate, produce plausibly-wrong outputs, fail on multi-step reasoning, and require verification that users cannot skip if they care about quality. The foreman work — prompting, reviewing, catching errors, revising, integrating — is real cognitive labor that doesn't show up in the "AI did this automatically" narrative.
The BCG/HBR research documented workers who reported spending more time managing AI than the underlying task would have taken. Honest accounting of cognitive overhead is the point — not whether AI is good or bad.
Attention residue.
Sophie Leroy (2009) documented that incomplete tasks generate ongoing cognitive interference — attention residue that impairs next-task performance. Parallel AI threads create this continuously.
What attention residue is
When you switch from task A to task B without completing A, part of your working memory continues processing task A. This 'attention residue' is not consciously experienced but measurably impairs performance on task B. Leroy and Schmidt (2016) extended this: the effect compounds with multiple incomplete tasks.
What this means for AI workflows
An AI that is generating a long output in a background tab while you work on something else is an incomplete task creating attention residue. Multiple open AI threads — a common pattern in multi-agent or multi-tool workflows — create compounding residue that continuously taxes working memory.
The context-switching tax
Each context switch between AI threads, between AI and human work, between prompting and reviewing — adds switching overhead. Cowan's working memory limit (approximately four items) is not expanded by AI tools. Managing more contexts means each context gets less cognitive bandwidth.
Productivity gains coexist with fatigue.
Net productivity gains from AI are documented. Brain fry is also documented. These are not contradictory — they describe different effects on different workers in different contexts.
GitHub Copilot randomized controlled trials and Brynjolfsson, Li & Raymond (2023) both document productivity gains from AI assistance — especially for less experienced workers who benefit most from AI completing tasks at the frontier of their capabilities. These gains are real, replicated, and should not be dismissed.
Brain fry, as documented in the BCG/HBR research, describes a specific subgroup: workers who have integrated AI deeply and continuously, who manage multiple AI tools, and who are responsible for verification of AI outputs across many tasks. For this group, cognitive overhead may grow faster than productivity gains.
The design constraint is real: productivity measurements that count AI-generated outputs without counting the cognitive cost of managing and verifying them are measuring the wrong thing.
Evidence quality — read this first
Brain Fry is the most preliminary domain in this framework.
The BCG/Harvard Business Review paper (March 2026) is consulting research, not peer-reviewed. It describes patterns observed in BCG client organizations, which are not a representative sample of knowledge workers.
The cognitive load theory it invokes (Sweller et al. 2011) is well-established and peer-reviewed. The application to AI-specific orchestration is theoretically sound but not yet empirically validated in controlled studies.
The attention residue literature (Leroy 2009, 2016) is peer-reviewed and solid. The extension to AI contexts is inferred, not directly studied.
The Kosmyna et al. MIT EEG study (2025) is a preprint with methodological critiques. It should not be cited as evidence without noting its preliminary status.
In short: the underlying mechanisms are real. The specific AI applications are theoretically plausible but empirically thin. We name these limits rather than paper over them.
What's your AI cognitive load?
12 questions about how you use AI tools — parallel threads, verification habits, context switches, and how you feel after heavy AI use. Produces a cognitive load profile with specific recommendations.
Take the workflow audit →Action for every level of influence.
For yourself
- Close AI threads before switching to a new task. Leaving a half-finished AI conversation open generates attention residue — it competes for working memory even when you're not looking at it.
- Batch AI interactions rather than running them continuously. Interrupting independent work to check AI outputs adds context-switching costs that compound across the day.
- Designate AI-free thinking time for complex work. Some problems — complex reasoning, synthesis, creative work — are hurt by early AI involvement that narrows the solution space before exploration.
For knowledge workers
- Audit which AI interactions actually need AI. Many tasks that get routed to AI would be faster and less mentally taxing to do independently. The overhead of prompting, verifying, and revising is real.
- The foreman problem: if you're spending more time reviewing AI outputs than you would spend doing the task, your efficiency calculation is wrong. AI should reduce total cognitive work, not shift it.
- Identify your cognitive peak hours and protect them from AI overhead. High-load AI management (multi-agent, verification-heavy) should not compete with your best thinking time.
For organizations
- Measure AI-related cognitive overhead explicitly. Worker satisfaction surveys that ask about AI tool friction, task-switching burden, and 'always-on' AI pressure capture what productivity metrics miss.
- Resist the pressure to add more AI tools without removing cognitive load elsewhere. 'More AI tools' ≠ 'less cognitive work' — orchestration overhead scales with tool count.
- Review your AI verification requirements. If you require humans to verify all AI outputs without providing time to do so, you've added extraneous load without adding capacity.
For researchers and developers
- The BCG/HBR 2026 findings are consulting research, not peer-reviewed. Longitudinal, peer-reviewed study of AI cognitive load in knowledge work is urgently needed.
- Cognitive load theory provides a rigorous framework for AI tool design. Extraneous load (from AI verification, prompt iteration, context management) should be a design metric, not an afterthought.
- Design AI tools that reduce total cognitive work, not just shift it. Intrinsic load (the task itself), extraneous load (the interface and verification), and germane load (learning) must all be measured.
For Educators
Teaching AI orchestration and cognitive load?
Facilitation guide for the workflow audit, discussion of the evidence quality distinction, and notes on teaching preliminary findings responsibly.
Research & further reading.
Want CPAI to deliver AI workflow training to your organization?
We help organizations design AI workflows that reduce cognitive overhead, not just shift it.