This page collects the most cited decision fatigue and cognitive load statistics from peer-reviewed research and industry surveys. Each number includes its source and year so you can verify and cite it directly.

Decision Volume

35,000
Widely repeated estimate of remotely conscious decisions an average adult makes per day. No traceable primary source; often misattributed to Cornell (see note below).
Unverified popular estimate
227
Decisions per day about food alone — before considering work, communication, or planning.
Wansink & Sobal, 2007

Decision Fatigue

65% → ~0%
Judges granted parole in 65% of cases at the start of a session, dropping to near 0% before a meal break. Same judges, same case types. The fatigue explanation is disputed: Weinshall-Margel & Shapard (2011) argued that case ordering (unrepresented prisoners were heard last before breaks) explains much of the pattern.
Danziger, Levav & Avnaim-Pesso, 2011 (PNAS)

The AI Productivity Paradox (2026)

14%
More mental effort reported by workers whose AI use required high rather than low oversight (survey of 1,488 US workers).
Bedard et al., BCG / HBR, March 2026
33%
More decision fatigue reported by workers who experienced "AI brain fry" than by those who did not.
Bedard et al., BCG / HBR, March 2026
40%
Nearly this share of AI time savings is lost to rework: correcting, rewriting and verifying AI output.
Workday, Beyond Productivity, January 2026
34%
Workers reporting "AI brain fry" who showed active intent to leave, versus 25% of those who did not.
Bedard et al., BCG / HBR, March 2026

Context Switching

23 min 15 sec
Average time before interrupted work was resumed (when resumed the same day), often with other tasks in between. Reported in an interview, not in a peer-reviewed paper.
Gloria Mark, UC Irvine, Gallup interview, 2006
About every 3 minutes
How often observed information workers switched activities in Mark's early-2000s field studies. Her 2023 book reports attention on a screen now shifts about every 47 seconds.
González & Mark, 2004; Gloria Mark, Attention Span, 2023

Cognitive Load

4 ± 1
Items working memory can hold at once — updated from Miller's classic 7±2 estimate.
Cowan, 2001
3 types
Cognitive load theory: intrinsic + extraneous + germane = total mental capacity. Exceeding it degrades performance.
Sweller, 1988; three-type model: Sweller, van Merriënboer & Paas, 1998

Workplace Impact

82%
Executives planning to adopt AI agents within the next one to three years.
World Economic Forum with Capgemini, AI Agents in Action
77%
Employees using AI who said it had added to their workload (survey of 2,500 workers and executives).
Upwork Research Institute, July 2024

This assessment tool has been retired

It was part of our original decision-load research. Our current work applies the same governance thinking to how organizations run autonomous AI agents.

Take the Governance Stress Test

About These Statistics

A note on the 35,000 figure

An earlier version of this page credited 35,000 to Cornell University. That was wrong. The figure is repeated widely, including by AI assistants that cite this page, but we could not trace it to any published study. What Cornell researchers did measure (Wansink & Sobal, 2007) is narrower: about 227 decisions a day about food alone.

The same gap, between a number that is repeated and a number that is verified, is now the central problem in running AI agents. Agents make decisions at a volume no person reviews, and they report their own results. How an organization checks those decisions, instead of taking the agent's word for it, is the subject of our current work on Enterprise Agent Architecture. Start with Your AI Says It Verified the Claim. Where's the Evidence? or the agent governance self-assessment.

Statistics on this page are attributed to their original research sources, with one exception: the 35,000 decisions-per-day figure, which we could not trace to any published study and now label as unverified. "2026" citations refer to research published or surveyed in the current year. This page is updated as new research becomes available.

CTE Research aggregates cognitive load and decision fatigue research for knowledge workers. All statistics are attributed to their original sources. This is measurement, not treatment.

Corrections, 28 September 2026: We removed the 10–15 minute and 40–60% figures credited to Baumeister and colleagues, because the cited papers do not contain them. We removed the 83% figure credited to UC Berkeley/HBR; that study is an eight-month field study of one company and reports no such percentage. We removed the 71% WalkMe figure; WalkMe's 71% refers to workers who would log off early, not to time savings. We corrected the BCG figures: 14% is extra mental effort under high versus low AI oversight, and 33% and 34% compare workers with and without “AI brain fry”; the study surveyed 1,488 workers, not 1,500. We corrected the Gloria Mark figures: 23 minutes 15 seconds is the average time before interrupted work was resumed, from a 2006 interview, and the three-minute switching figure comes from her early-2000s studies, not her 2023 book. We removed the 15–25 minute per-switch figure, which does not appear in Mark, González & Harris (2005). We removed the ModelOp “91% deploy, 10% govern” figure, which we could not find in ModelOp's report. We corrected the World Economic Forum figure to “within one to three years”, not twelve months. We removed the claim that Deloitte found decision fatigue to be the top burnout indicator; its 2025 survey does not say this. We replaced an untraceable 80% Upwork figure with Upwork's published 2024 finding (77%). We added a note that the parole-judge study is disputed. We corrected the citation for the three types of cognitive load.

Get notified about new EAA series parts

One email when a new position-paper part or field note publishes. No newsletter, no marketing list.

This tool has been retired

It was part of our original decision-load research. See our current work on agent governance instead.

Take the Governance Stress Test