What Work is AI Actually Doing? Uncovering the Drivers of Generative AI Adoption
Citation. Agarwal, P., Agarwal, H., & Rana, A. (2025). What work is AI actually doing? Uncovering the drivers of generative AI adoption. arXiv. https://doi.org/10.48550/arXiv.2510.23669
Summary. Agarwal, Agarwal, and Rana analyze four million Claude AI interactions mapped to O*NET tasks and score those tasks on routine, cognitive, social intelligence, creativity, domain knowledge, complexity, and decision-making dimensions. They find that observed GenAI use is concentrated in a small share of tasks and is highest for non-routine, cognitively demanding, creative, and complex work. The paper extends observed-use work by proposing task archetypes that distinguish dynamic problem solving, procedural and analytical work, and standardized operational tasks.
Key claims
- AI usage was sharply concentrated across O*NET tasks: the median task accounted for 0.006% of AI conversations, 75% of tasks were below 0.017%, and the top 5% of tasks accounted for 59% of all interactions.
- Tasks attracting AI use had a demanding knowledge-work profile, with high median cognitive demand (8.2), complexity (8.0), and decision making (7.8), and lower median routineness (4.2).
- The granular task parameters most positively associated with AI usage were idea generation (Spearman’s rho = 0.173), information processing (rho = 0.157), and originality (rho = 0.151); the strongest negative associations were outcome predictability (rho = -0.135) and repetition frequency (rho = -0.131).
- The authors identify three task archetypes: Procedural & Analytical Work, Dynamic Problem Solving, and Standardized Operational Tasks. Dynamic Problem Solving had the highest mean AI usage share (3.31%), followed by Procedural & Analytical Work (2.45%), while Standardized Operational Tasks had the lowest usage.
- Social intelligence was weaker and more statistically decoupled from observed AI usage than cognitive, complexity, decision-making, creativity, and routine dimensions, suggesting that interpersonal requirements were not a primary adoption driver in this dataset.
Evidence & limitations
- Evidence comes from the Anthropic Economic Index dataset of four million Claude AI interactions mapped to O*NET tasks, with LLM-based scoring of 35 task parameters across seven task-characteristic dimensions.
- The study is an arXiv preprint and provides descriptive, correlational evidence about observed Claude AI use; it does not establish causal effects on productivity, wages, employment, skill development, or durable organizational redesign.
- The task-characteristic scores are model-generated rather than human-rated, and the data come from one AI platform, so platform selection and measurement error may shape the observed archetypes and correlations.
- Conversation-level use should not be read as formal organizational adoption: a task may attract AI interactions without being redesigned, governed, or tied to accountable work outcomes.