Anthropic Economic Index: Analyzing AI's Impact on Labor Markets

Anthropic has launched the Anthropic Economic Index, an initiative and open-source dataset designed to track the real-world effects of AI on labor markets and the economy. By analyzing millions of anonymized conversations on Claude.ai, the initiative provides empirical data on which occupational tasks are being augmented or automated by AI, moving beyond surveys to observe actual usage patterns.

AI Usage Patterns Across Occupations

AI adoption is currently concentrated in technical and creative fields, with a significant disparity between AI usage and the actual size of the workforce in those sectors.

  • Primary Adoption Hubs: The "computer and mathematical" category accounts for 37.2% of queries, focusing on software modification, code debugging, and network troubleshooting. The "arts, design, sports, entertainment, and media" category follows at 10.3%, primarily for writing and editing tasks.
  • Low Adoption Areas: Occupations involving high physical labor, such as "farming, fishing, and forestry," show the lowest representation at 0.1% of queries.
  • Workforce Disparity: There is a notable gap between workforce representation and AI use. For example, computer and mathematical jobs represent only 3.4% of U.S. workers but account for 37.2% of Claude conversations.

Depth of Integration and Wage Correlation

AI is currently diffused across many tasks rather than replacing entire jobs, with usage patterns correlating strongly with specific salary brackets.

  • Task-Level Diffusion: Approximately 36% of occupations use AI for at least 25% of their associated tasks. However, only 4% of occupations use AI for 75% or more of their tasks, suggesting that very few jobs are being entirely automated.
  • The "Mid-to-High Wage" Peak: AI usage is most prevalent in mid-to-high median salary ranges (e.g., computer programmers and copywriters). Conversely, both the lowest-paid roles and the highest-paid roles (such as obstetricians) show low rates of AI use, which Anthropic attributes to both current capability limits and practical barriers.

Augmentation vs. Automation

Real-world usage data indicates that AI is more frequently used to enhance human capability than to replace it entirely.

  • Augmentation (57%): A majority of tasks involve AI collaborating with the user. This includes "Task Iteration" (31.3%) for brainstorming and refinement, "Learning" (23.3%) for knowledge acquisition, and "Validation" (2.8%) for verifying work.
  • Automation (43%): A smaller portion of tasks involve AI directly performing the work. This is split between "Directive" tasks (27.8%), which are complete delegations with minimal interaction, and "Feedback Loops" (14.8%), where completion is guided by environmental feedback.

Methodology: The Clio Analysis Tool

To map conversations to economic data, Anthropic developed "Clio," an automated analysis tool that preserves user privacy while categorizing interactions.

  1. Task Mapping: Clio analyzes conversations and matches them to the U.S. Department of Labor's O*NET database, which contains approximately 20,000 specific work-related tasks.
  2. Categorization: Tasks are grouped into occupations, which are then further aggregated into broader occupational categories (e.g., business and financial).
  3. Data Source: The analysis was performed on approximately one million conversations from Claude.ai Free and Pro plans.

Study Limitations

Anthropic notes several caveats regarding the representativeness of the data:

  • Intent Ambiguity: It is not always clear if a user is performing a professional task or a personal hobby.
  • Usage Ambiguity: A task that appears as "automation" (e.g., asking for a full memo) may actually be "augmentation" if the user heavily edits the output.
  • User Base Bias: The data only includes Claude.ai Free and Pro users, excluding API, Team, and Enterprise users. Additionally, as Claude is marketed as a strong coding model, coding tasks may be overrepresented.
  • Modality Limits: Because Claude cannot generate images, creative uses involving visual media are not captured.

Sources