Economic_Impacts_Paper

Which Economic Tasks are Performed with AI? Evidence from Millions of Claude Conversations Kunal Handa ∗ , Alex Tamkin ∗ , Miles McCain, Saffron Huang, Esin Durmus Sarah Heck, Jared Mueller, Jerry Hong, Stuart Ritchie, Tim Belonax, Kevin K. Troy Dario Amodei, Jared Kaplan, Jack Clark, Deep Ganguli Anthropic Abstract Despite widespread speculation about artificial intelligence’s impact on the future of work, we lack systematic empirical evidence about how these systems are actually being used for different tasks. Here, we present a novel framework for measuring AI usage patterns across the economy. We leverage a recent privacy-preserving system [Tamkin et al., 2024] to analyze over four million Claude.ai conversations through the lens of tasks and occupations in the U.S. Department of Labor’s O*NET Database. Our analysis reveals that AI usage primarily concentrates in software development and writing tasks, which together account for nearly half of all total usage. However, usage of AI extends more broadly across the economy, with∼36%of occupations using AI for at least a quarter of their associated tasks. We also analyzehowAI is being used for tasks, finding 57% of usage suggests augmentation of human capabilities (e.g., learning or iterating on an output) while 43% suggests automation (e.g., fulfilling a request with minimal human involvement). While our data and methods face important limitations and only paint a picture of AI usage on a single platform, they provide an automated, granular approach for tracking AI’s evolving role in the economy and identifying leading indicators of future impact as these technologies continue to advance. 1 Introduction Rapid advances in artificial intelligence suggest profound implications for the evolution of labor markets [Brynjolfsson et al., 2018b, Acemoglu, 2021, Trammell and Korinek, 2023, Hering, 2023, Comunale and Manera, 2024, Maslej et al., 2024]. Despite the importance of anticipating and preparing for these changes, we lack systematic empirical evidence about how AI systems are actually being integrated into the economy. Existing methodologies—whether developing predictive models [Webb, 2019, Eloundou et al., 2023, Kinder et al., 2024], conducting controlled studies of productivity effects [Peng et al., 2023, Noy and Zhang, 2023], or administering periodic surveys of users [Humlum and Vestergaard, 2024, Bick et al., 2024]—cannot track the dynamic relationship between advancing AI capabilities and their direct, real-world use across the economy. Here, we present a novel empirical framework for measuring AI usage across different tasks in the economy, drawing on privacy-preserving analysis of millions of real-world conversations on Claude.ai [Tamkin et al., 2024]. By mapping these conversations to occupational categories in the U.S. Department of Labor’s O*NET Database, we can identify not just current usage patterns, but ∗ Equal contribution. Authors above the line break are core contributors. Author contributions are listed in Appendix A. Direct correspondence to {kunal, atamkin, deep} @anthropic.com. Figure 1:Measuring AI use across the economy.We introduce a framework to measure the amount of AI usage for tasks across the economy . We map conversations from Claude.ai to occupational categories in the U.S. Department of Labor’s O*NET Database to surface current usage patterns. Our approach provides an automated, granular, and empirically grounded methodology for tracking AI’s evolving role in the economy.(Note: figure contains illustrative conversation examples only.) also early indicators of which parts of the economy may be most affected as these technologies continue to advance. 2 We use this framework to make five key contributions: 1.Provide the first large-scale empirical measurement of which tasks are seeing AI use across the economy (Figure 1, Figure 2, and Figure 3)Our analysis reveals highest use for tasks in software engineering roles (e.g., software engineers, data scientists, bioinformatics technicians), professions requiring substantial writing capabilities (e.g., technical writers, copywriters, archivists), and analytical roles (e.g., data scientists). Conversely, tasks in occupations involving physical manipulation of the environment (e.g., anesthesiologists, construction workers) currently show minimal use. 2.Quantify the depth of AI use within occupations (Figure 4)Only∼4%of occupations exhibit AI usage for at least 75% of their tasks, suggesting the potential for deep task-level use in some roles. More broadly,∼36%of occupations show usage in at least 25% of their tasks, indicating that AI has already begun to diffuse into task portfolios across a substantial portion of the workforce. 3. Measure which occupational skills are most represented in human-AI conversations (Figure 5). Cognitive skills like Reading Comprehension, Writing, and Critical Thinking show high presence, while physical skills (e.g., Installation, Equipment Maintenance) and managerial skills (e.g., Negotiation) show minimal presence—reflecting clear patterns of human complementarity with current AI capabilities. 2 We provide relevant data at https://huggingface.co/datasets/Anthropic/EconomicIndex/. 2 4.Analyze how wage and barrier to entry correlates with AI usage (Figure 6 and Table 2). We find that AI use peaks in the upper quartile of wages but drops off at both extremes of the wage spectrum. Most high-usage occupations clustered in the upper quartile correspond predominantly to software industry positions, while both very high-wage occupations (e.g., physicians) and low-wage positions (e.g., restaurant workers) demonstrate relatively low usage. This pattern likely reflects either limitations in current AI capabilities, the inherent physical manipulation requirements of these roles, or both. Similar patterns emerge for barriers to entry, with peak usage in occupations requiring considerable preparation (e.g., bachelor’s degree) rather than minimal or extensive training. 5.Assess whether people use Claude to automate or augment tasks (Figure 7)We find that 57% of interactions show augmentative patterns (e.g., back-and-forth iteration on a task) while 43% demonstrate automation-focused usage (e.g., performing the task directly). While this ratio varies across occupations, most occupations exhibited a mix of automation and augmentation across tasks, suggesting AI serves as both an efficiency tool and collaborative partner. Our methods provide an automated, granular, and empirically grounded approach for tracking AI usage patterns as both capabilities and societal usage evolve. This early visibility into emerging trends gives policymakers and civil society crucial lead time to respond to shifts in how AI transforms work. However, we acknowledge multiple key limitations (discussed in Section 4.1); for example, our usage data cannot reveal how Claude’s outputs are actually used in practice, and our reliance on O*NET’s static occupational descriptions means we cannot account for entirely new tasks or jobs that AI might create. 3 Nevertheless, this framework offers a foundation for understanding AI’s evolving impact on the economy. While our methods are imperfect, they provide a systematic way to track usage patterns and identify leading indicators of economic effects across different sectors. As AI capabilities and adoption continue to advance, we believe this kind of empirical measurement will be crucial for understanding and preparing for the technology’s broader economic implications. 2 Background and Related Work Our work builds on many lines of research attempting to model, measure, and forecast AI’s impact on the economy. Economic foundations and the task-based frameworkA wide body of work in economics has proposed theoretical models to understand the impact of automation on the labor market. Most notably, Autor et al. [2003], Autor [2013] argue for modeling labor markets through the lens of discretetasks which can be performed by either human workers or machines—for example,debugging codeor cutting hair. Building on this framework, Autor [2015] shows that while technologies automate some tasks, they often augment human capabilities in others due to complementarity between humans and machines, leading to higher demand for labor. In addition Acemoglu and Restrepo [2018] use this framework to explore a model where automation technologies can create entirely new tasks in addition to displacing old tasks. Forecasting the impact of AI on labor marketsAnother branch of work leverages the task- based framework to predict the future prevalence of automation across the economy, often based on descriptions of tasks and occupations from the O*NET database of occupational information provided by the U.S. Department of Labor [National Center for O*NET Development, 2025b]. For example, Frey and Osborne [2017] fit a gaussian process classifier to a dataset of 70 labeled occupations to predict which occupations are subject to computerization. Brynjolfsson et al. [2018a] hire human annotators to rate 2,069 detailed work areas in the O*NET database, focusing specifically on their potential to be performed by machine learning. Webb [2019] analyzes the overlap between patent documents and job task descriptions to predict the "exposure" of tasks to AI, finding highest exposure in high-education, high-wage occupations—a pattern partially reflected in our empirical usage data, though we find peak usage in mid-to-high wage occupations rather than at the highest wage levels. 3 Though our methodology using Clio [Tamkin et al., 2024] will allow us to detect such emerging patterns of work as they arise. 3 Felten et al. [2023] focus specifically on large language models, estimating exposure by using a dataset that links human abilities to different occupations. Eloundou et al. [2023] also consider exposure of tasks to language models, using language models themselves to obtain more granular estimates of exposure at the level of individual tasks—an approach we follow in our work. When considering the impacts of language model-powered software, they conclude that around half of all tasks in the economy could one day be automated by language models. While our empirical usage data shows lower current adoption (∼36%of occupations using AI for at least a quarter of their tasks), the patterns of usage across tasks largely align with their predictions, particularly in showing high usage for software development and content creation tasks. Real-world studies of AI usageTo complement these forecasts based on human or machine judgment, another body of work attempts to gather concrete data to understand how AI is currently being adopted across the labor market. For example, studies show rapid AI adoption across differ- ent sectors and countries: research from late 2023 found that half of workers in exposed Danish occupations had used ChatGPT, estimating it could halve working times in about a third of their tasks [Humlum and Vestergaard, 2024], while a subsequent study in August 2024 found that 39% of working-age US adults had used generative AI, with about a quarter using it weekly [Bick et al., 2024]. Moreover, further research has attempted to measure the breadth and depth of this usage, with studies finding positive effects of generative AI tools on productivity for a wide range of individual domains, including software engineering [Peng et al., 2023, Cui et al., 2024], writing [Noy and Zhang, 2023], customer service [Brynjolfsson et al., 2023], consulting [Dell’Acqua et al., 2023, for tasks that were suitable for AI], translation [Merali, 2024], legal analysis [Choi and Schwarcz, 2023], and data science [Wiles et al., 2024]. We bridge these separate approaches to perform the first large-scale analysis of how advanced AI systems are actually being used across tasks and occupations. We build on the task-based framework, but rather than forecasting potential impacts ("exposure" of occupations to AI), we measure real-world usage patternsusing Clio [Tamkin et al., 2024], a recent system that enables privacy-perserving analysis of millions of human-AI conversations on a major model provider. This allows us to complement controlled studies of AI productivity effects in specific domains with a comprehensive view of where and how AI is being integrated into work across the economy. Our methodology enables tracking these patterns dynamically as both AI capabilities and societal adoption evolve—revealing both present day usage trends as well as leading indicators of future diffusion. 3 Methods and analysis To understand how AI systems are being used for different economic tasks, we leverage Clio [Tamkin et al., 2024], an analysis tool that uses Claude [Anthropic, 2024] to provide aggregated insights from millions of human-model conversations. We use Clio to classify conversations across occupational tasks, skills, and interaction patterns, revealing breakdowns across these different categories. All analyses draw from conversation data collected during December 2024 and January 2025. See Appendices B, E and F for more details and prompts, including validating the composition of our dataset and how we perform classification in cases with large numbers of categories (e.g. O*NET tasks). 3.1 Task-level analysis of AI usage Using Clio on a dataset of one million Claude.ai Free and Pro conversations, 4 we analyzed each interaction to map it to its most relevant task category in the O*NET database. Because there are nearly