Precision Proactivity: Measuring Cognitive Load in Real-World AI-Assisted Work
Citation. Lepine, B., Kim, J., Mishkin, P., & Beane, M. (2026). Precision Proactivity: Measuring Cognitive Load in Real-World AI-Assisted Work. arXiv. https://arxiv.org/abs/2505.10742
Summary. Lepine et al. study how cognitive load shapes AI-assisted knowledge work by analyzing transcripts from 34 financial professionals completing a complex valuation task with GPT-4o. They propose a transcript-based measurement framework for estimating intrinsic and extraneous cognitive load using task decomposition, knowledge-graph structure, semantic coherence, and conversational dynamics. The main finding is not simply that AI help improves work: AI-generated content use was positively associated with output quality, but extraneous load showed a large negative association, model-initiated task switching was the strongest predictor of decline, and less experienced professionals both faced larger load penalties and gained more from AI content.
Key claims
- AI-generated content usage was positively associated with valuation-task quality across 1,178 participant-subtask observations, suggesting that professionals can benefit from model-produced content in complex knowledge work.
- Extraneous cognitive load had the largest negative association with quality, roughly three times the magnitude of intrinsic load in the authors’ core mixed-effects specifications.
- The authors identify a compensatory pathway in which AI-generated content partly offsets load-related deficits, but does not eliminate them.
- Extraneous load persisted within speakers and spilled asymmetrically into model responses, meaning the model tended to perpetuate user confusion more than correct it.
- Among proactive AI behaviors, model-initiated task switching was the strongest predictor of performance decline, followed by information sprawl and truncation.
- Expertise moderated the pattern: less experienced professionals faced larger penalties from load and gained more from AI-generated content, but were not the users who most increased uptake under load.
Evidence & limitations
- The evidence comes from 34 finance professionals performing a standardized, time-pressured financial valuation task with GPT-4o, analyzed at the participant-subtask level across 1,178 observations.
- The design is observational and correlational: the authors can identify strong associations among transcript-derived load indicators, AI-generated content use, and quality, but not definitive causal effects.
- Cognitive load is measured through computational proxies derived from conversational traces rather than direct self-report, physiological, or dual-task measures; the authors explicitly call for validation against established cognitive-load assessments.
- The measurement framework depends on domain-specific task decomposition and knowledge-graph representations, so porting it to other work domains requires rebuilding those components.
- The paper uses GPT-4o in one complex financial task, so generalization to other models, domains, expertise structures, and less time-constrained work remains open.