Sophisticated AI Use and AI Brain Fry Look Identical From the Outside

In March 2026, two studies described the most advanced AI users inside large enterprises. The first was a KPMG and University of Texas at Austin analysis of more than 1.4 million prompts and responses from about 2,500 employees over eight months, published in Harvard Business Review by Nick Hallman, Zach Kowaleski, Anu Puvvada, and Jaime J. Schmidt as “What the Best AI Users Do Differently”. It celebrated the top 5% of users and named the behaviors that set them apart. The second was a Boston Consulting Group survey of 1,488 full-time U.S. workers, published two weeks earlier in HBR by Julie Bedard, Matthew Kropp, Megan Hsu, Olivia T. Karaman, Jason Hawes, and Gabriella Rosen Kellerman as “When Using AI Leads to ‘Brain Fry’”. It introduced a concept it called “AI brain fry” and warned about the cognitive cost of intensive AI use.

If you read both papers in the same week, as I did, you notice the awkward thing immediately. They are largely describing the same behaviors. Long initial prompts. Frequent model switching. Intense back-and-forth with the model. Delegation of complex multi-step tasks. Use of multiple tools at once. The KPMG paper presents this as the recipe for high-leverage AI work. The BCG paper presents it as the path to mental fog, decision errors, and a 39% relative jump in intent to leave among workers it describes as top AI users.

Neither study acknowledges the tension, because each is answering a different question on a different sample. Underneath both sits a question neither paper resolves cleanly. If the behavioral signature of sophisticated AI use matches the behavioral signature of AI burnout, what actually separates the two groups?

Table of Contents

• What KPMG measured: the four signatures of sophisticated use
• What BCG measured: AI brain fry and the cost of oversight
• Same behaviors, opposite outcomes
• Where the difference actually lives
• What this looks like in a one-person operation
• The scaffolding hypothesis
• How to avoid AI brain fry without an enterprise program

What KPMG measured: the four signatures of sophisticated use

KPMG had reached an enviable adoption baseline. Nearly 90% of employees used AI regularly. The company could measure usage frequency, prompt counts, and tokens, but those metrics left open what leaders actually wanted to know: whether AI was producing better work, and from whom.

Working with a research team at UT Austin, the company had OpenAI’s o1 reasoning model evaluate every conversation in the eight-month archive and flag more than 50 behavioral variables, which the team then refined. The point of the exercise was to find observable signals that predicted high-quality use, signals you could spot without running the same expensive evaluation pipeline every week.

Four behaviors separated the top tier from everyone else. Sophisticated users were ambitious in how they framed problems, writing longer initial prompts and sticking with the same problem across many turns. They treated the model as a reasoning partner, using techniques like role definition, worked examples, iterative refinement, and explicit self-verification in place of one-shot questions. They delegated complex multi-step tasks with explicit constraints and success criteria. And they used AI across a broad range of work, switching intentionally between models depending on the task.

Two of the secondary findings are worth holding on to. First, only about 5% of users met the sophisticated bar despite the 90% adoption rate. Second, the best users were often above manager level, which surprised the authors given the common assumption that junior employees, being more comfortable with the tools, would lead AI fluency. The authors draw a distinction between comfort and sophistication. My reading is that experience with framing problems and evaluating outputs is the harder skill to develop, and the seniority finding is what that looks like in data.

The recommendation is what you would expect from a major-firm study. Translate the behaviors into role-specific playbooks. Move training from “how to use the tool” to scenario-based work that practices delegation, verification, and refinement. Treat AI as a thinking instrument across the full job, beyond a few discrete productivity tasks.

What BCG measured: AI brain fry and the cost of oversight

BCG ran a broader survey, 1,488 full-time U.S. workers at large companies across roles and industries, and asked about cognitive outcomes alongside AI use patterns. About 14% of participants using AI for work reported what the authors call AI brain fry. The paper’s working definition is mental fatigue from use or oversight of AI tools that exceeds a person’s cognitive capacity. The rate varied by function, from 6% in legal roles to 26% in marketing.

Three findings from that paper matter for the comparison with KPMG.

The first is that intensive AI oversight, the requirement to monitor and steer AI tools directly, carried measurable cognitive cost. Workers reporting high oversight expended about 14% more mental effort, reported 12% more mental fatigue, and reported 19% more information overload than workers reporting low oversight. A second predictor was whether AI had increased their workload, and the authors read the two together as a wider sphere of accountability.

The second is that perceived productivity rose as workers went from one to two and then three concurrent AI tools, with smaller gains at the third, and dipped beyond three. The authors tie this to the familiar cost of multitasking.

The third is the link to outcomes that matter to a CFO. Workers who endorsed AI brain fry reported 33% more decision fatigue, scored 39% higher on the frequency of major errors at work, and showed active intent to leave at 34% compared with 25% among those who did not, a 39% relative increase. The authors frame this as a retention risk concentrated among their heaviest AI users.

The recommendations from the BCG paper sit on a different plane from the KPMG ones. Redesign jobs and tools holistically instead of stacking AI oversight on top of human oversight. Set explicit expectations about whether AI is meant to absorb workload or expand it. Shift metrics from activity to impact. Build worker skills in problem framing and prioritization. Treat human attention as a finite, allocated resource.

Same behaviors, opposite outcomes

If you place the two papers side by side, the behavioral overlap is striking. Long, structured initial prompts in KPMG language are intensive oversight in BCG language. Multi-model switching in KPMG is multi-tool juggling in BCG. Delegating complex multi-step tasks in KPMG maps directly onto the multi-agent oversight load that BCG is warning about.

The KPMG paper never uses the words “fatigue,” “burnout,” or “errors.” The BCG paper avoids the word “sophisticated” and describes its high-intensity users as “today’s superstars,” people the company has to keep. That describes the same kind of person KPMG calls a sophisticated user.

So one of two things is true. Either the studies are talking about different populations doing the same things at different intensities, which is partly true and still leaves the gap unexplained. Or they are measuring two cross-sections of one population and missing the variable that decides which side you land on.

I think it is the second.

Where the difference actually lives

The KPMG paper hints at the missing variable without naming it. The recommendations section talks about role-specific playbooks, peer-led champion networks, scenario-based training, and expectations tailored by group. Those are organizational artifacts, outside any individual user’s behavior, and they absorb cognitive work the user would otherwise have to do alone.

The BCG paper hints at the same thing from the negative side. At the manager level, people whose managers make time for their AI questions report 15% lower mental fatigue than workers whose managers do not, and workers expected to figure AI out on their own report 5% higher fatigue. At the team level, organized integration of AI into team workflows lowers mental strain, while team pressure to use AI raises it. At the organizational level, workers who feel their company values work-life balance report 28% lower mental fatigue, and workers who expect to be asked for more output because of AI report 12% higher fatigue.

Read against each other, the two studies describe one population doing the same things in two different structural environments. With scaffolding, the behaviors compound into sophistication. Without scaffolding, the same behaviors compound into cognitive overload, decision errors, and turnover.

The variable is the structural context the behaviors run inside. The BCG authors come close to saying this themselves when they conclude that what matters is how workers, teams, leaders, and organizations shape AI use, more than how much AI any one person uses.

What this looks like in a one-person operation

I work as a solo operator across multiple projects, with AI as my primary technical partner. Most days I run something close to the behavioral profile both papers describe. Long initial prompts, often a full design specification. Multi-model switching, generally per task class. Intense iterative back-and-forth on complex deliveries. Delegation of multi-step work with explicit constraints. Use of AI across writing, analysis, infrastructure, and product specs.

By the BCG description I should be a brain fry candidate. Some weeks I am. In the weeks I am fine, the volume of AI work is about the same. What differs is what is sitting around the AI work.

The structural artifact I rely on most is a persistent knowledge base I have written about previously. It is a small pile of documents that establish who I am, who the AI is, what each project is for, what the safety rules are, and what decisions belong to me. It exists outside my head. The AI reads it at the start of every session. So do I, when I need to.

That knowledge base does three things that matter for the brain fry question. It externalizes context so I am free of holding it in working memory while I oversee a session. It makes the boundary between AI-owned decisions and human-owned decisions explicit, so I am spared deciding whether to intervene in the moment. And it gives me a place to land structured corrections after incidents, so the system gets less surprising with use. In April 2026 an AI session merged a 30-commit feature into the main branch of one of my sites when it belonged on a paused work branch. Because the intended branch was written down, the mismatch was obvious and the fix was a single revert. I did not have to audit every commit to work out what had gone wrong.

The second artifact is a habit I impose on the AI. I require it to justify decisions and narrate as it goes. That increases the surface area of each step and reduces the number of post hoc audits I need to run. Many of the AI brain fry accounts in the BCG paper read like post hoc audits to me: people double-checking AI work after the fact, when the cognitive cost of catching up has already been paid.

The third artifact is verification baked into specific high-risk skills. The literature review skill I built has an explicit verification loop, added after a single-pass extraction produced a confidently wrong figure on its first live use. The verification step looks like friction. What it replaces is a more expensive piece of friction, which is me independently rechecking everything later or, worse, skipping the recheck and being wrong.

Nothing here is unique to my setup. It is a small, deliberate scaffolding around AI work, the same kind both HBR papers gesture at without quite naming, applied to a one-person operation instead of a large professional-services firm.

The scaffolding hypothesis

The reading I am proposing is that the behavioral signatures KPMG identified (ambitious framing, treating AI as a reasoning partner, delegating complex tasks with constraints, broad cross-task use) are necessary and insufficient. They define the upper bound of what is possible with AI. Whether you reach it depends on something else.

What separates sophistication from brain fry is whether the behaviors run inside scaffolding that absorbs cognitive load. In a corporate environment that scaffolding is role-specific playbooks, peer champion networks, organized team integration, manager support, and communicated workload expectations. In a solo operation it is a knowledge base, a decision-rights document, a narrate-and-justify habit, and verification loops where they earn their keep.

The hypothesis is testable. Neither paper tested it, because each was answering a different question. KPMG did not measure brain fry. BCG did not measure prompt complexity or delegation patterns. A study that measured both would, on this reading, find that behavioral intensity predicts brain fry only weakly once you control for structural support.

I do not know yet whether that prediction would hold. What I do know is that the recommendations of the two papers, read together, already point that direction. KPMG asks leaders to make the desired behaviors “visible, discussable, and expected.” BCG’s first lesson is to redesign “jobs, work, and tools holistically” around shared human and AI responsibility. Those are the same prescription with different starting points.

How to avoid AI brain fry without an enterprise program

Most readers of either paper are in no position to commission role-specific playbooks or to redesign jobs holistically. The actionable form of the scaffolding hypothesis is smaller than that, and it works at any scale.

Three things follow.

First, write the structural context before you run the agent. The single biggest fatigue reducer in my experience is having a design specification the AI can read at the start of a complex delegation. The cost of writing the spec is paid once. The cost of holding the structural context in your head while the AI works is paid every minute the work is running.

Second, make verification a habit. Verification feels like friction when you put it in. It feels like sanity when the AI is wrong and you catch it the same day instead of two weeks later. The BCG finding that AI brain fry goes with more frequent major errors says the same thing from the opposite side: missing verification compounds.

Third, treat your concurrent AI tool count as a span-of-control problem. The BCG productivity-by-tool-count curve is the most concrete operational guidance in either paper. Perceived productivity gains shrink at the third tool and dip past three. That number is a useful default for anyone who finds themselves stacking agents without an obvious reason.

The headline finding from each paper, taken alone, is incomplete. KPMG looks at sophisticated use without measuring its cost. BCG looks at the cost of intensive use without measuring its sophistication. Together they say something more useful than either says separately. The behavioral signature of high-leverage AI work and the behavioral signature of cognitive overload are nearly identical. The variable that decides which one you produce is the structural support around the behaviors. That variable is something you can build, and most people who look like top users without it eventually find themselves looking like burnout statistics instead.