A new intensive study from Anthropic researchers (the developers of Claude AI) provides large-scale empirical analysis of how AI is actually being used across different economic tasks. The research, analyzing over four million Claude conversations through a privacy-preserving system, offers concrete data on which occupations are adopting AI, how deeply it’s being integrated, and whether it’s being used to automate or augment human capabilities.
While there has been widespread speculation about AI’s impact on work, until now systematic evidence about how these systems are actually being utilized has been lacking. This study helps bridge that gap by mapping real-world usage patterns to occupational categories in the U.S. Department of Labor’s O*NET Database.
The findings reveal that AI usage primarily concentrates in software development and technical writing tasks. Computer and Mathematical occupations show the highest associated AI usage rate, comprising 37.2% of all queries, followed by Arts, Design, Entertainment, Sports, and Media occupations at 10.3%. Education-related occupations were also highly represented. As expected, occupations requiring physical manipulation of the environment, such as Transportation and Healthcare Support occupations, show minimal usage.
The key metrics on AI adoption and usage patterns include:
- 36% of occupations show AI usage in at least a quarter of their associated tasks
- Only 4% of occupations exhibit AI usage across 75% or more of their tasks
- 57% of AI interactions demonstrate augmentative patterns (e.g., learning or iterating on outputs)
- 43% show automation-focused usage (e.g., fulfilling requests with minimal human involvement)
- Peak usage occurs in mid-to-high wage occupations, with notably lower usage at both wage extremes
- Usage is highest in Job Zone 4 (considerable preparation needed, typically bachelor’s degree) rather than roles requiring minimal or extensive training
The researchers found that cognitive skills like Critical Thinking, Reading Comprehension, Writing, and Programming showed high presence in AI interactions, while physical skills like Equipment Maintenance and Installation showed minimal presence. This aligns with current AI capabilities and suggests natural boundaries in how AI can be applied across different types of work.
Interestingly, the study revealed clear specialization in how different AI models are used. Claude 3.5 Sonnet demonstrated more usage for coding and software development tasks, while Claude Opus showed higher usage for creative and educational work such as producing content and developing educational materials. This specialization suggests how different AI models might serve distinct professional needs.
The study acknowledges important limitations. The data represents only Claude.ai Free and Pro conversations, cannot fully capture how outputs are used in practice, and is limited by O*NET’s static occupational descriptions which cannot account for entirely new tasks or jobs that AI might create. Additionally, the research team notes that their analysis cannot determine whether users are professionals in the fields corresponding to their queries.
Looking ahead, this research suggests that AI integration into work is happening selectively rather than comprehensively. Instead of wholesale automation of entire occupations, we’re seeing AI adoption concentrated in specific tasks where current capabilities align well with job requirements. This nuanced picture offers an empirically grounded view of how AI is transforming work – not through dramatic displacement, but through targeted integration that varies significantly across different types of tasks and occupations.
