AI systems will have a significant impact on the way work is organized in the coming years. In response, we are launching the Anthropic Economic Index , an initiative to study the impact of AI on labor markets and the economy over time.
The index’s initial report provides unique data and analysis based on millions of anonymized conversations with Claude.ai , painting the clearest picture yet of how AI is being deployed to perform real-world tasks in the modern economy.
We have also made the dataset used for this analysis publicly available so that researchers can build on and extend our findings. Different perspectives are needed to inform policy responses to the upcoming labour market transformation and its impact on employment and productivity. We therefore invite economists, policy experts and other researchers to express their views on the Index.
Key findings from the first study published in the Economic Index report include:
- Today, AI applications are concentrated primarily in software development and technical writing tasks. More than a third of occupations (about 36%) use AI in at least one-quarter of their tasks, while about 4% of occupations use AI in three-quarters of their tasks.
- The use of AI is predominantly complementary (57%), where AI collaborates with humans and enhances their capabilities, compared to automation (43%), where AI directly performs a task.
- AI use is more common in jobs associated with middle and high salaries, such as programmers and data scientists, and less common in both the lowest-paid and highest-paid jobs. This likely reflects both the limitations of current AI capabilities and practical barriers to using the technology.
Below are additional details of our initial findings.
Where and how AI is used in the economy, based on real-world usage data from Claude.ai . The numbers represent the percentage of Claude conversations related to specific tasks, professions, and categories.
Analysis of the implementation of AI in various areas of work
Our new paper continues a tradition of research into the impact of technology on the labour market – from Hargreaves’ spinning machine during the Industrial Revolution to today’s robots in the auto industry. In this study, we focus on the ongoing impact of AI. We do not survey users about their use of AI or attempt to predict the future – we have direct data on how AI is being used today.
Analysis of tasks by profession
Our research began with an important insight in the economics literature : Sometimes it makes sense to analyze the tasks of jobs rather than the jobs themselves. Jobs often involve common tasks and skills: for example, visual pattern recognition is performed by designers, photographers, security personnel, and radiologists.
Some tasks are easier to automate or augment with new technology than others. Therefore, AI is expected to be deployed selectively across different tasks in different occupations, and task analysis, along with occupational analysis as a whole, will provide a more complete picture of how AI is integrated into the economy.
Using Clio to Map AI Usage to Tasks
This study was made possible by Claude’s analytics capabilities and Clio , an automated analytics tool that enables us to process Claude conversations while preserving user privacy. We used Clio to analyze approximately one million Claude conversations (specifically, conversations within the free and professional plans of Claude.ai ) and categorized them into job-related task categories.
We selected the tasks according to classifications developed by the U.S. Department of Labor, which maintains a database of about 20,000 specific job tasks known as the Occupational Information Network, or O*NET . Clio matched each conversation to the O*NET task that best represented the role of AI in that conversation (see the diagram below). We then used the same grouping scheme to group the tasks into relevant occupations and to form a few general categories, such as “education and library science,” “business and finance,” and so on.
The process by which Clio transforms Claude conversations (strictly confidential; top left block) into categories of professional tasks (middle top) and professions/professional categories defined by the ONET schema (top right block). The resulting data is then used for various analytics (bottom row; detailed below; Clio aggregates a large number of conversations and groups them into higher categories for analysis. It is important to note that to preserve user privacy, this process is done automatically, without researchers having access to the original conversations. You can read more about Clio here .).
Results
AI usage by occupational type. The highest prevalence of AI in our dataset is in the “computer and math” category of tasks and occupations, which heavily employ software developers. 37.2% of queries sent to Claude fall into this category, which includes tasks such as modifying software, debugging code, and troubleshooting network issues.
The second largest category is “art, design, sports, entertainment and media” (10.3% of queries), which mostly reflects the use of Claude for various types of writing and editing. As expected, occupations with a high proportion of physical labor, such as in the “agriculture, fishing and forestry” category (0.1% of queries), are minimally represented.
We also compared the performance of our dataset with the level of representation of each occupation in the overall labor market. The comparison of the performance is presented in the chart below.
For each occupation type, the percentage of conversations with Claude is shown in orange, and the percentage of workers in the US economy (according to the US Department of Labor’s ONET) is shown in grey.
Depth of AI use within professions. Our analysis shows that only a few professions use AI in all tasks: only about 4% of professions use AI in at least 75% of their tasks. However, moderate use of AI is much more common: about 36% of professions use AI in at least 25% of their tasks.
As predicted, there is no evidence of full automation of jobs in this dataset: AI is being deployed in a distributed manner across many tasks, with a stronger impact on some task groups than others.
AI Use and Salary Levels. The O*NET database provides median salaries for each of the listed occupations. We added this information to our analysis, allowing us to match occupations’ median salaries to the level of AI use in their tasks.
Interestingly, both low- and very high-paying jobs are characterized by low levels of AI use (these jobs tend to involve a lot of manual labor, such as hair washers or midwives). It is the middle- and upper-middle-income jobs, such as programmers and copywriters, that are among the largest users of AI in our data.
Median annual salary (x-axis) and percentage of conversations with Claude by respective profession (y-axis). Illustrative examples are provided for some professions.
Automation vs. augmentation: We also looked at how tasks are performed, dividing them into “automation” (where the AI directly performs the task, such as formatting a document) and “add-on” (where the AI collaborates with the user to perform the task).
Overall, there is a slight advantage of augmentation – 57% of tasks are performed with the participation of AI as an assistant, while 43% of tasks are performed based on automation. In other words, in slightly more than half of the cases, AI does not replace a person, but works together with them – participating, for example, in verification (for example, rechecking the user’s work), training (helping the user to acquire new knowledge and skills), and the iterative process (helping in brainstorming or re-executing generative tasks).
Percentage of Claude conversations involving augmentation versus automation, and a breakdown of task subcategories within each group. The subcategories are defined in our document as follows: Directive Execution: Completely handing over a task to the AI with minimal interaction; Feedback: Performing a task with guidance from feedback; Task Iteration: Collaborative refinement; Learning: Acquiring knowledge and understanding; Validation: Verifying and improving the work performed.
Limitations of the study
Our study provides unique insights into how AI is changing labor markets. However, like any study, it has its limitations, including:
- We cannot say for sure that using Claude to complete a task is necessarily work-related: a request for writing or editing advice could have been made for work or for personal creativity.
- We also don’t know how users use Claude’s responses. For example, are snippets of code copied? Are responses verified or accepted without question? Some of the data we perceive as automation may in fact be an add-on: for example, a user might instruct Claude to write a full memo (which would be classified as automation) and then edit it themselves (which would be an add-on).
- The analysis used data from Claude.ai Free and Pro plans, not API, team, or enterprise users. Although Claude.ai datasets do contain irrelevant work conversations, we used a language algorithm to filter only the conversations related to professional tasks, which helps minimize this issue.
- The sheer number of different tasks means that Clio’s system may have misclassified some conversations (details on the analysis validation are provided in Appendix B of the full document).
- Claude does not generate images (except indirectly through code), so some creative uses may not be reflected in the data.
- Given that Claude is positioned as a cutting-edge model for programming, one would expect coding-related tasks to be overrepresented in the data. Therefore, we do not claim that our dataset represents a comprehensive picture of AI use in general.
Conclusions and directions for future research
The use of AI is rapidly expanding and the models are becoming more sophisticated. The picture of the labor market can change dramatically in a very short time. For this reason, we plan to periodically repeat many of our analyses to track the social and economic changes that will inevitably occur. The results and corresponding data sets will be published regularly as part of the Atropic Economic Index.
Long-term studies can provide new insights into the impact of AI on the labor market. For example, we can track changes in the depth of AI use across occupations. If we find that AI is used in only a few tasks, and only a few occupations use it for the vast majority of tasks, the future may be one in which most current occupations evolve rather than disappear. We can also monitor the ratio of automation to augmentation, which can serve as a signal to identify areas where automation is becoming dominant.
Our research provides data on how AI is being used, but does not provide specific policy recommendations. Answers to questions about how to prepare for the impact of AI on the labor market cannot be provided by academic research alone – they must be informed by the data, values, and experiences of a variety of experts. We are interested in using our new methodology to explore these questions further.
Open data and invitation to discussion
The main contribution of this study and the Atropic Economic Index is the development of a new methodology that provides detailed data on the impact of AI. We are sharing the dataset used for the analyses described above immediately and plan to publish other similar datasets in the future as they become available.