Which Economic Tasks are Performed with AI? Evidence from Millions of Claude Conversations

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 Al's evolving role in the economy. (Note: figure contains illustrative conversation examples only.)

• Framework Overview: The figure serves as an overview of the research framework. The key aspect is the mapping of conversations from Claude.ai, an AI assistant, to occupational categories as defined in the U.S. Department of Labor's O*NET (Occupational Information Network) database. The O*NET database is a comprehensive resource that describes various occupations and the tasks, skills, knowledge, abilities, and other characteristics associated with them.
• Mapping Process: The figure illustrates the process of connecting user conversations with the AI to specific tasks within those occupational categories. This involves analyzing conversations to identify the relevant tasks being performed and then categorizing these tasks according to the O*NET framework. The illustrative conversation examples provided in the figure (though not numerically specified in the caption) provide concrete instances of how these connections are made.
• Additional Framework Aspects: The framework also includes the use of wage data and augmentative vs. automative task categorizations. Augmentative tasks refer to how AI can be used to augment human capabilities, whereas automative tasks are those where AI can be used to automate tasks. Finally, the figure also includes a section on skills breakdown, that are relevant to human-AI conversations.

• Alignment with Objective: The figure's caption aligns with the study's objective of providing an empirical measurement of AI usage across different economic tasks. By grounding the analysis in the O*NET database, the study leverages a standardized and well-established framework for categorizing occupations and tasks. The mention of an 'automated, granular, and empirically grounded methodology' suggests a rigorous approach to data collection and analysis, enhancing the credibility of the findings.
• Temporal Context: The caption's reference to 'current usage patterns' acknowledges the dynamic nature of AI adoption and the need for ongoing monitoring. While the caption itself does not delve into specific methodological details, it sets the stage for a more in-depth discussion of the data analysis techniques employed in the study, which should be elaborated upon in the methods section.
• Limitations: It would strengthen the scientific validity to briefly mention the limitations inherent in using conversation data from a single AI platform (Claude.ai) and the potential biases that may arise from this specific data source.

• Clarity and Summary: The caption effectively summarizes the core purpose of Figure 1, highlighting its role in measuring AI usage across the economy and linking it to the methodology used (Claude.ai conversations mapped to O*NET). The parenthetical note is helpful in managing reader expectations, clarifying that the figure contains illustrative examples rather than comprehensive data.
• Technical Language: The phrase 'automated, granular, and empirically grounded methodology' is informative for a scientific audience, conveying the rigor and detail of the approach. However, for a broader audience, a slightly more accessible phrasing might improve understanding without sacrificing precision.

Figure 2: Hierarchical breakdown of top six occupational categories by the amount of AI usage in their associated tasks. Each occupational category contains the individual O*NET occupations and tasks with the highest levels of appearance in Claude.ai interactions.

• Overall Structure: The figure presents a series of bar charts showing the relative amount of AI usage within the top six occupational categories. Each category is broken down into individual occupations listed in the O*NET database, showing the percentage of total conversations associated with each. For example, 'Computer and Mathematical' occupations show 37.2% of all conversations, with specific titles like 'Computer Programmers' (6.1%) and 'Software Developers, Systems Software' (5.3%) listed below.
• Granularity of Data: Within each occupation, specific tasks are also listed, along with their percentage contribution to that occupation's AI usage. For instance, within 'Computer and Mathematical' occupations, 'Develop and maintain software' accounts for 16.8% of AI usage, while 'Program and debug computer systems and machinery' accounts for 6.9%. This provides a granular view of AI application within each field.
• O*NET Explanation: The 'O*NET occupations' are classifications from the U.S. Department of Labor's Occupational Information Network (O*NET) database. This database categorizes jobs based on required skills, knowledge, and activities. The percentage values represent the proportion of Claude.ai conversations that are associated with tasks falling under each specified occupation or task category, providing a measure of AI usage within those areas.

• Empirical Support: The figure provides empirical support for the claim that AI usage is concentrated in specific occupational categories, particularly software engineering. The hierarchical structure allows for a detailed examination of AI adoption at different levels of granularity, from broad categories to specific tasks.
• Mapping Accuracy: The validity of the figure depends on the accuracy of the mapping between Claude.ai conversations and O*NET occupations and tasks. This mapping process should be clearly described in the methodology section, including any validation steps taken to ensure the reliability of the assignments. It would also be valuable to discuss potential sources of error or bias in the mapping process.
• Category Selection: The choice of the 'top six' occupational categories should be justified based on a clear and objective criterion (e.g., overall AI usage). Presenting data for a larger number of categories or including a category for 'other' occupations would provide a more comprehensive picture of AI adoption across the economy.

• Clarity of Purpose: The caption clearly states the purpose of Figure 2: to present a hierarchical breakdown of AI usage across the top six occupational categories. The phrase 'highest levels of appearance in Claude.ai interactions' is specific and accurately reflects the data source and metric used.
• Accessibility: For a general audience, the term 'hierarchical breakdown' might benefit from a brief explanation. However, for a scientific audience familiar with data visualization, the term is likely sufficient. The caption could be enhanced by explicitly stating the criteria used to determine the 'top six' categories (e.g., overall AI usage).

Figure 3: Comparison of occupational representation in Claude.ai usage data and the U.S. economy. Results show most usage in tasks associated with software development, technical writing, and analytical, with notably lower usage in tasks associated with occupations requiring physical manipulation or extensive specialized training. U.S. representation is computed by the fraction of workers in each high-level category according to the U.S. Bureau of Labor Statistics [U.S. Bureau of Labor Statistics, 2024].

• Visual Representation: The figure likely consists of a set of horizontal bar graphs. Each bar represents an occupational category. There are two bars for each occupation, one represents the percentage of Claude.ai conversations that are associated with that occupation, and the other represents the percentage of the US workforce that are in the same occupation. This allows for a direct comparison of AI usage and real-world employment.
• Key Trends: The caption highlights a key trend that the occupations with the most usage in Claude.ai are tasks associated with software development (37.2%), technical writing (10.3%), and analytical tasks. The occupations with the lowest usage in Claude.ai are tasks associated with physical manipulation or extensive specialized training.
• BLS Data: The U.S. Bureau of Labor Statistics (BLS) data provides a baseline for understanding the overall composition of the U.S. workforce. The BLS collects data on employment, unemployment, earnings, and other labor market characteristics. By comparing the AI usage data with the BLS data, the researchers can identify occupations that are disproportionately represented in AI interactions.

• Methodological Soundness: Comparing AI usage data with the overall occupational distribution in the U.S. economy is a scientifically sound approach for identifying potential biases and understanding the broader implications of AI adoption. This comparison helps to contextualize the AI usage patterns and assess whether AI is being used in a representative manner across different sectors.
• Data Comparability: The validity of the comparison depends on the accuracy and comparability of the occupational categories used in the Claude.ai data and the BLS data. It is essential that the researchers have carefully mapped the occupational categories to ensure consistency and avoid introducing bias. The methodology section should provide details on this mapping process.
• Limitations of BLS Data: It is important to acknowledge potential limitations in the BLS data, such as the level of granularity in the occupational categories and the potential for measurement error. The researchers should also consider whether the BLS data accurately reflects the current state of the U.S. economy, given that labor market conditions can change rapidly.

• Clarity of Purpose: The caption clearly indicates the purpose of Figure 3: to compare the distribution of occupations represented in the Claude.ai usage data with the overall occupational distribution in the U.S. economy. This comparison provides valuable context for interpreting the AI usage data, highlighting potential biases or over/under-representation of certain sectors.
• Summary of Findings: The caption effectively summarizes the key findings, noting the high representation of software development, technical writing, and analytical tasks, and the low representation of occupations involving physical manipulation or specialized training. This provides a concise overview of the main trends revealed by the figure.
• Source Transparency: The reference to the U.S. Bureau of Labor Statistics (BLS) as the source for U.S. representation data is crucial for transparency and allows readers to assess the reliability and validity of the comparison. Specifying that the U.S. representation is computed as 'the fraction of workers in each high-level category' provides a clear definition of the metric used.

Figure 7: Distribution of automative behaviors (43%) where users delegate tasks to AI, and augmentative behaviors (57%) where users actively collaborate with AI. Patterns are categorized into five modes of engagement; automative modes include Directive and Feedback Loop, while augmentative modes are comprised of Task Iteration, Learning, and Validation.

• Visual Representation: Figure 7 likely presents a pie chart or bar graph showing the distribution of automative and augmentative behaviors. The two main categories are 'automative behaviors' and 'augmentative behaviors'. Automative behaviors are those where the user delegates tasks to the AI, and they are made up of 'Directive' and 'Feedback Loop' modes. Augmentative behaviors are those where the user collaborates with the AI, and they are made up of 'Task Iteration', 'Learning', and 'Validation' modes.
• Key Finding: The key finding is that augmentative behaviors (57%) are slightly more prevalent than automative behaviors (43%). This suggests that users are more likely to actively collaborate with AI than to simply delegate tasks to it. This may reflect the current limitations of AI capabilities, or a preference for human control and oversight in certain tasks.
• Automative Behaviors: 'Automative behaviors' are those where the user delegates tasks to the AI, and they are made up of 'Directive' and 'Feedback Loop' modes. 'Directive' mode is where the human gives the AI a single instruction to complete, without much interaction. 'Feedback Loop' mode is where the human and AI engage in iterative dialogue, with the human mainly providing feedback.
• Augmentative Behaviors: 'Augmentative behaviors' are those where the user collaborates with the AI, and they are made up of 'Task Iteration', 'Learning', and 'Validation' modes. 'Task Iteration' mode is where the human and AI iteratively refine a task, with the human refining AI outputs. 'Learning' mode is where the human seeks understanding and explanation from the AI, and 'Validation' mode is where the human uses AI to check or validate work.

• Significance of Findings: The figure provides empirical evidence on the relative prevalence of automation and augmentation in human-AI interactions. This distinction is important for understanding the potential impact of AI on the labor market, as automation may lead to job displacement, while augmentation may enhance human productivity and creativity.
• Classification Method: The validity of the figure depends on the accuracy of the method used to classify conversations into the different modes of engagement. This classification method should be clearly described in the methodology section, including any validation steps taken to ensure the reliability of the classifications. It should also be shown that the categorization is robust and accounts for edge cases.
• Potential Variations: It is important to acknowledge that the distribution of automative and augmentative behaviors may vary depending on the specific tasks being performed and the characteristics of the users. Further analysis is needed to explore these potential variations and identify the factors that influence the choice of collaboration mode.

• Clarity of Distribution: The caption clearly presents the distribution of automative and augmentative behaviors, providing the percentages for each category (43% and 57%, respectively). This offers a concise overview of the overall balance between these two modes of AI usage.
• Mode Categorization: The caption effectively lists the five modes of engagement and assigns them to either the automative or augmentative category. This provides a clear and structured understanding of the different types of human-AI interactions being analyzed. The use of specific names for each mode (Directive, Feedback Loop, Task Iteration, Learning, Validation) enhances the clarity and memorability of the taxonomy.
• Accessibility: For a broader audience, brief examples of each mode of engagement would enhance the caption's accessibility. However, for a scientific audience familiar with the concepts of automation and augmentation, the current level of detail is likely sufficient.

Table 2: Analysis of AI usage across occupational barriers to entry, from Job Zone 1 (minimal preparation required) to Job Zone 5 (extensive preparation required). Shows relative usage rates compared to baseline occupational distribution in the labor market. We see peak usage in Job Zone 4 (requiring considerable preparation like a bachelor's degree), with lower usage in zones requiring minimal or extensive preparation.

• Table Structure: Table 2 likely presents a breakdown of AI usage across different 'Job Zones'. Job Zones are categories defined by the U.S. Department of Labor's O*NET database, and they represent the amount of preparation needed for a human to perform the duties of a given occupation. The table likely consists of several columns, with each row corresponding to a different Job Zone.
• Relative Usage Rates: The table shows 'relative usage rates compared to baseline occupational distribution'. This means that the researchers are comparing the percentage of AI conversations associated with each Job Zone to the percentage of the U.S. workforce that is in the same Job Zone. This comparison reveals whether AI usage is over- or under-represented in different Job Zones, relative to their overall presence in the labor market.
• Key Finding: The key finding is that AI usage 'peaks in Job Zone 4'. This means that the relative usage rate is highest for occupations requiring considerable preparation, such as a bachelor's degree. The 'lower usage in zones requiring minimal or extensive preparation' suggests that AI tools may not be well-suited for either very simple or very complex tasks.

• Analytical Approach: Analyzing AI usage in relation to occupational barriers to entry is a valuable approach for understanding the factors that influence AI adoption. The use of Job Zones as a measure of barriers to entry provides a standardized and well-defined framework for this analysis.
• Mapping Accuracy: The scientific validity of the table depends on the accuracy of the mapping between AI conversations and Job Zones. This mapping process should be clearly described in the methodology section, including any steps taken to ensure the reliability of the assignments. Potential sources of error or bias in the mapping process should be acknowledged.
• Limitations of Job Zones: It is important to consider potential limitations in the Job Zone classification system, such as the granularity of the categories and the potential for subjectivity in assigning occupations to different zones. The researchers should also acknowledge that other factors, such as the availability of AI tools for different occupations and the regulatory environment, may also influence AI adoption.

• Clarity and Conciseness: The caption provides a clear overview of Table 2's purpose: to analyze AI usage in relation to occupational barriers to entry. It clearly defines the range of Job Zones being considered (1 to 5) and provides a concise summary of the key finding: peak usage in Job Zone 4.
• Accessibility: The parenthetical descriptions of each Job Zone (e.g., 'minimal preparation required', 'extensive preparation required') enhance the caption's accessibility for a broader audience. The explicit mention of 'a bachelor's degree' as an example of the preparation required for Job Zone 4 further clarifies the meaning of the different levels of preparation.

Figure 4: Depth of AI usage across occupations. Cumulative distribution showing what fraction of occupations (y-axis) have at least a given fraction of their tasks with AI usage (x-axis). Task usage is defined as occurrence across five or more unique user accounts and fifteen or more conversations. Key points on the curve highlight that while many occupations see some AI usage (~36% have at least 25% of tasks), few occupations exhibit widespread usage of AI across their tasks (only ~4% have 75% or more tasks), suggesting AI integration remains selective rather than comprehensive within most occupations.

• Cumulative Distribution: Figure 4 presents a cumulative distribution function (CDF). A CDF plots the probability that a real-valued random variable X takes on a value less than or equal to x. In this case, the y-axis shows the cumulative fraction of occupations, and the x-axis shows the fraction of tasks within each occupation that utilize AI. So, for any point on the curve, the y-value tells us the proportion of occupations for which at least that proportion of their tasks are being performed using AI.
• Axis Interpretation: The x-axis represents the 'fraction of tasks with AI usage'. A value of 0.25 on the x-axis means that 25% of the tasks associated with a particular occupation are being performed with the assistance of AI. The y-axis indicates the 'fraction of occupations'. A value of 0.36 on the y-axis at the x-axis value of 0.25 means that 36% of occupations have at least 25% of their tasks performed with AI.
• Task Usage Definition: The caption defines 'task usage' as 'occurrence across five or more unique user accounts and fifteen or more conversations'. This is a threshold applied to filter out tasks that are only performed sporadically or by a small number of users. It implies that the data used to generate the figure only considers tasks with a substantial level of adoption across multiple users of Claude.ai.

• Appropriateness of CDF: The use of a cumulative distribution function (CDF) is appropriate for visualizing the distribution of AI usage across occupations. The CDF effectively illustrates the proportion of occupations that have at least a certain fraction of their tasks performed with AI. The CDF is a standard statistical tool for visualizing distributions, and its use is well-justified in this context.
• Justification of Thresholds: The definition of 'task usage' as 'occurrence across five or more unique user accounts and fifteen or more conversations' is a reasonable approach for filtering out noise and focusing on tasks with substantial adoption. However, the specific thresholds (five users and fifteen conversations) should be justified based on methodological considerations and sensitivity analyses. The impact of varying these thresholds on the overall results should be explored.
• Complementary Analyses: The CDF provides a valuable overview of the depth of AI integration across occupations, but it does not reveal information about which specific tasks are being performed with AI or the impact of AI on task performance. Complementary analyses that delve into the specific tasks and their associated outcomes would provide a more complete picture of AI adoption.

• Clarity and Conciseness: The caption provides a clear and concise summary of the figure's content, including the type of visualization (cumulative distribution), the variables represented on each axis (fraction of occupations vs. fraction of tasks with AI usage), and the definition of 'task usage'. The inclusion of key data points (36% and 4%) further enhances its informativeness.
• Key Takeaway: The caption effectively highlights the main takeaway from the figure: that AI integration remains 'selective rather than comprehensive'. This provides a valuable interpretation of the data and guides the reader's understanding of the figure's significance.
• Accessibility: For a broader audience, the term 'cumulative distribution' might benefit from a brief, intuitive explanation. However, for a scientific audience familiar with statistical visualizations, the term is likely sufficient. The caption effectively uses approximations (~36% and ~4%) to convey the key data points without overwhelming the reader with precise numbers.

Figure 5: Distribution of occupational skills exhibited by Claude in conversations. Skills like critical thinking, writing, and programming have high presence in AI conversations, while manual skills like equipment maintenance and installation are uncommon.

• Visual Representation: The figure likely presents a horizontal bar chart displaying the distribution of various occupational skills. Each bar represents a specific skill (e.g., critical thinking, writing, programming), and the length of the bar indicates the percentage of Claude.ai conversations in which that skill is exhibited.
• O*NET Skills: The skills are derived from the O*NET database, which identifies 35 occupational skills that are essential for workers to perform tasks across different jobs. These skills encompass a wide range of abilities, including cognitive, interpersonal, and physical skills.
• Skill Distribution: The caption notes that skills like critical thinking, writing, and programming have a high presence in AI conversations. This suggests that AI is being used to support tasks that require these cognitive abilities. The figure probably shows that skills like equipment maintenance and installation are uncommon, which suggests that AI is not frequently used for tasks that require physical manipulation.

• Empirical Support: The figure provides empirical support for the claim that AI interactions are more strongly associated with cognitive skills than with manual skills. This finding aligns with the broader trends observed in AI adoption, where AI is often used to augment or automate tasks that require cognitive abilities.
• Skill Identification Method: The validity of the figure depends on the accuracy of the method used to identify the occupational skills exhibited in Claude.ai conversations. This method should be clearly described in the methodology section, including any validation steps taken to ensure the reliability of the skill assignments. Further analysis should also consider whether Claude's responses are reflecting actual skill performance, or simply reflecting default conversational behaviors.
• Limitations of O*NET Skills: It is important to acknowledge potential limitations in the O*NET database, such as the comprehensiveness of the skill list and the potential for overlap between different skills. The researchers should also consider whether the O*NET skills accurately reflect the current demands of the labor market.

• Clarity and Summary: The caption clearly summarizes the figure's content, highlighting the distribution of occupational skills exhibited in Claude.ai conversations. The use of specific examples (critical thinking, writing, programming, equipment maintenance, and installation) enhances the clarity and informativeness of the caption.
• Key Takeaway: The caption effectively conveys the main takeaway from the figure: that cognitive skills are more prevalent in AI interactions than manual skills. This provides a valuable insight into the nature of AI adoption and its potential impact on different types of work.

Figure 6: Occupational usage of Claude.ai by annual wage. The analysis reveals notable outliers among mid-to-high wage professions, particularly Computer Programmers and Software Developers. Both the lowest and highest wage percentiles show substantially lower usage rates. Overall, usage peaks in occupations within the upper wage quartile, as measured against U.S. median wages [US Census Bureau, 2022].