Anthropic Economic Index report: Uneven geographic and enterprise AI adoption

economic impactAI adoptionautomation and augmentationgeographyenterprise useClio

The third Anthropic Economic Index report tracks Claude.ai use from December 2024 to August 2025, adds the first country and US-state breakdowns through a new Anthropic AI Usage Index, and gives the first analysis of business use through Anthropic's API; it finds directive automation rising from 27% to 39% of conversations, per capita use strongly correlated with income, and 77% of API traffic showing automation patterns.

Ruth Appel, Peter McCrory, Alex Tamkin et al. · 2025-09-15 · original

The report's introduction notes that earlier technologies took years or decades to spread, and that early adoption has typically been concentrated in a few places and a few tasks before it broadens. The report asks whether AI is following that pattern, on a faster timeline, using Claude data. It is the third report of the Anthropic Economic Index, after the first (data from December 2024 and January 2025) and the second (February and March 2025). It extends the series in two directions: geographic breakdowns of Claude.ai use for more than 150 countries and every US state, and a first look at how businesses use Claude through Anthropic's first-party (1P) API. The report was published with a summary post, Anthropic Economic Index: Tracking AI's role in the US and global economy, an interactive website and an open dataset.

The report is an economics document rather than a machine learning paper. Its results are descriptive shares, ratios and regression slopes computed from samples of conversations classified by Claude, and its interpretations are offered with explicit hedging. The authors are Ruth Appel, Peter McCrory and Alex Tamkin (lead authors, contributing equally), with Miles McCain, Tyler Neylon and Michael Stern. Locations below refer to the report's chapters, figures and footnotes, which are numbered the same way in the PDF and on the web page.

Terms used in this note

  • O*NET. A US Department of Labor taxonomy of occupations and the tasks performed in them. Conversations are matched to O*NET tasks, which roll up into Standard Occupational Classification (SOC) major groups such as "Computer and Mathematical."
  • Bottom-up taxonomy. A hierarchy of request clusters that Claude builds from the conversations themselves, to capture uses that O*NET misses (Chapter 2, "Task usage patterns across countries").
  • Automation and augmentation. Automation covers directive use (the user hands over a task, which Claude completes with minimal back-and-forth) and feedback loops (the user automates a task and relays results back). Augmentation covers learning (asking for information or explanations), task iteration (working on a task together) and validation (asking for feedback) (Chapter 1, "Directive automation is accelerating").
  • Anthropic AI Usage Index (AUI). A geography's share of Claude.ai use divided by its share of the working-age population (ages 15 to 64). An AUI above 1 means more use than population alone predicts (Chapter 2).
  • Percentage point (pp). The arithmetic difference between two percentages, as opposed to the relative change.
  • Elasticity. The percentage change in one quantity associated with a 1% change in another, estimated as a slope on log-log axes.

Data

  • Claude.ai over time. Three samples: V1 (December 2024 and January 2025), V2 (February and March 2025) and V3 (August 2025) (Chapter 1, "Overview").
  • Geography. One million Claude.ai Free and Pro conversations from 4 to 11 August 2025, sampled at random from conversations not flagged for trust and safety review, located by IP address, with VPN, anycast and hosting traffic excluded (Chapter 2, footnote 2). The unit is a conversation, not a user. For privacy, cells with fewer than 15 conversations and 5 unique accounts are filtered out, and bottom-up request clusters need at least 500 conversations and 250 accounts (Chapter 2, footnote 1).
  • API. One million transcripts from August 2025, sampled at random from a pool of 1P API customers that accounts for roughly half of 1P API usage. Each record is a prompt and response, sometimes from the middle of a multi-turn session (Chapter 3, footnote 2).

All data are classified with privacy-preserving methods based on Clio.

What it finds: Claude.ai use over time (Chapter 1)

Education and science grow, business shrinks

Computer and mathematical tasks remain the largest group at 36% of use. In the introduction's figures, educational tasks rose from 9.3% to 12.4% and scientific tasks from 6.3% to 7.2%; the chapter text rounds these to 9% to 12% and 6% to 7%. Business and financial operations fell from 6% to 3% and management from 5% to 3% (Introduction; Chapter 1; Figure 1.1). These are shares of use, so a falling share can coexist with a rising number of conversations.

Product features shape tasks

Some task shares moved with new features between V2 and V3 (Chapter 1, "New capabilities are shaping usage patterns"): searching electronic sources rose from 0.03% to 0.49% after web search was released in March, and internet-based research from 0.003% to 0.27% after the Research mode in April (footnote 1). Developing instructional materials rose from 0.2% to 1.5%, which the report links to adoption among educators, and creating multimedia documents from 0.16% to 0.55%.

Within coding, creating new code rose from 4.1% to 8.6% (+4.5pp) while debugging and error correction fell from 16.1% to 13.3% (−2.8pp in the chapter; the introduction gives −2.9pp), "a net 7.4pp shift toward creation over fixing code." The authors suggest users may be achieving more in a single exchange (Chapter 1; footnote 3 lists the component tasks).

Directive use rises past augmentation

Directive conversations rose from 27% in V1 to 39% in V3, mainly at the expense of task iteration and learning, making this "the first report where automation usage exceeds augmentation usage" (Chapter 1; Figure 1.2). Footnote 4 flags a measurement change: V3 was classified with Claude Sonnet 4 and V2 with Claude Sonnet 3.7. Re-classifying V3 with Sonnet 3.7 still showed a significant rise, but to 45% automation rather than 49%, and the shift appears across a wide range of occupational categories rather than coming from a change in task mix.

The report offers two explanations it cannot yet separate: better models that need fewer refinements, or users learning to delegate ("learning-by-doing"). It notes that they have different labor-market implications: the first suggests more displacement risk for workers whose tasks are automated, the second greater demand for workers who adapt (Chapter 1). Changes in the user base are a third possibility ("Looking ahead").

What it finds: geography (Chapter 2)

Countries

The United States has the largest share of Claude.ai use (21.6%), followed by India (7.2%) and Brazil (3.7%) (Figure 2.1). Adjusted for working-age population, small technologically advanced countries lead: Israel has an AUI of 7, followed by Singapore (4.57), Australia (4.10), New Zealand (4.05) and South Korea (3.73) (Figure 2.2, countries with at least 200 observations). The United States scores 3.62, Canada 2.91 and the United Kingdom 2.67, while France (1.94), Japan (1.86) and Germany (1.84) are lower. Many lower- and middle-income countries are below 1, including Bolivia (0.48), Indonesia (0.36), India (0.27) and Nigeria (0.2).

Across countries, a 1% higher GDP per working-age capita is associated with a 0.7% higher AUI (Figure 2.4). The report lists possible reasons: digital infrastructure, economic structure (more knowledge work in richer economies), regulation, awareness of and access to Claude, and trust in AI.

US states

California has 25.3% of US use, followed by New York (9.3%), Texas (6.7%) and Virginia (4.0%). Adjusted for population, the District of Columbia leads with an AUI of 3.82, then Utah (3.78), California (2.13), New York (1.58) and Virginia (1.57) (Figure 2.5). Footnote 7 reports that a notable fraction of Utah's use appeared to be possibly associated with coordinated abuse, reflected in a higher directive score; the authors ran robustness checks and believe it does not drive the results.

Across states, income explains less than half the variation in adoption, yet use rises faster with income: each 1% of state GDP per capita is associated with a 1.8% higher AUI (Chapter 2). The authors suggest that other regional differences, including how well Claude's capabilities fit the occupational composition of the local workforce, play a larger role (Chapter 2, "Task usage patterns across the United States"). California shows disproportionate IT requests, digital marketing and translation; Florida business advice and fitness; DC document editing, information and job applications, with help on job applications 1.84 times as common in DC as in the US overall (Figures 2.9 and 2.10).

Task mix and collaboration by country

As AUI rises, use shifts away from computer and mathematical tasks toward education, office and administrative work, arts, and the sciences, although the pattern is noisy (Figure 2.7). Coding is over half of all use in India against roughly a third globally, yet software development is the most common use in every country. In the bottom-up taxonomy, US users disproportionately ask about household management, job searching and medical guidance; Brazilian users about translation and legal services; Vietnamese users about software development and education; Indian users almost exclusively about software development (Figure 2.8).

Controlling for each country's task mix through a partial regression (footnote 9), higher-AUI countries use Claude more collaboratively and lower-AUI countries more for automation (Figure 2.11). The authors call this "somewhat counter-intuitive" and speculate about cultural and economic factors, or early adopters in each country being more willing to delegate, saying more research is needed.

What it finds: business use through the API (Chapter 3)

Context from public data

According to the US Census Bureau's Business Trends and Outlook Survey, the share of US firms using AI to produce goods or services rose from 3.7% in fall 2023 to 9.7% in early August 2025; in the Information sector about one firm in four reported AI use, roughly ten times the rate in Accommodation and Food Services (Figure 3.1).

What businesses use Claude for

Coding dominates. Among the top 15 bottom-up clusters, about half of API traffic, most relate to software; debugging web applications and resolving technical issues are each about 6%, about 5% concerns developing and evaluating AI systems, marketing materials are 4.7% and processing business and recruitment data 1.9% (Figure 3.2). In O*NET terms, 44% of API traffic maps to computer and mathematical tasks, more than 8 percentage points above Claude.ai, and office and administrative tasks are about 10%. Education and library tasks fall from 12.3% on Claude.ai to 3.6% in the API, and arts and entertainment from 8.2% to 5.2% (Figure 3.3).

Use is highly concentrated on both surfaces: the bottom 80% of tasks account for 12.7% of Claude.ai use and 10.5% of API use, The text gives Gini coefficients of 0.84 and 0.86; the legend of Figure 3.4 gives 0.822 for Claude.ai and 0.842 for the API (Chapter 3, "Occupational segmentation vs. task specialization"; Figure 3.4). The authors suggest an "O-Ring" mechanism, in which any single barrier (insufficient reasoning, inaccessible data, missing regulatory approval) can block adoption of a task.

Automation dominates

77% of API transcripts show automation patterns, mostly directive, against 12% augmentation, while Claude.ai is close to an even split. Across tasks, 97% are automation-dominant in the API against 47% on Claude.ai (Chapter 3, "Automation vs. augmentation among API transcripts"; Figure 3.5).

Context limits complex tasks

For each task, average input and output lengths were divided by the average across tasks to give input and output token indices. Tasks at the 90th percentile of output length produce more than 4 times the output of tasks at the 10th (Chapter 3, "The more Claude does, the more Claude needs to know"; examples in Table 3.1). Across tasks, each 1% increase in input length is associated with a 0.38% increase in output length (Figure 3.7). The authors read this as strongly diminishing returns to context and suggest that complex deployments might be constrained more by access to information, including tacit knowledge held by staff, than by model capability. Footnote 10 names another possible contributor: the degradation some models show at long context lengths.

Cost matters little

Tasks typical of computer and mathematical occupations cost more than 50% more than sales tasks yet dominate use, and across occupational categories higher-cost tasks have higher use, with an estimated elasticity of 3 (Figure 3.8). Controlling for occupational category and collaboration mode, each 1% increase in cost is associated with a 0.29% decrease in use, so a 10% cost reduction would raise use by only about 3% (Figure 3.9; footnote 13). The authors call this "a preliminary exploration" and suggest that capability and the value of automating a task matter more than price.

The prompts and classifiers used

The report has no principles or constitution; the text that shapes its results is the classification machinery. Conversations and transcripts are assigned to O*NET tasks, bottom-up clusters and collaboration modes by Claude through Clio, and footnote 4 shows that the choice of classifier model alone moved the measured V3 automation share from 45% to 49%. Table 3.1's summaries of short- and long-output tasks were produced by Claude Sonnet 4 with what the caption calls "minimal guidance": "The columns should be ‘Example tasks’, ‘Index Value’, ‘Summary’ where you provide a summary". The report says Claude associated output length with task complexity, which the authors then adopt as "a sensible, easily measured proxy" while noting it does not capture all dimensions of complexity (Chapter 3).

Limitations and open questions

The report states or implies these limits:

  • Classifier change between reports. The rise in automation is partly confounded with the switch from Sonnet 3.7 to Sonnet 4 as classifier (Chapter 1, footnote 4).
  • Sampling. Small countries and states have few observations; figures are restricted to countries with at least 200 and states with at least 100, and tier assignment for low-usage countries is uncertain. The authors propose stratified sampling in future work (Chapter 2, footnote 4).
  • Abuse in the data. Part of Utah's usage was possibly associated with coordinated abuse; the authors' robustness checks led them to believe it does not drive the results (Chapter 2, footnote 7).
  • Conversations, not users. A heavy user contributes many conversations, though the authors note that their past work found sampling at random and sampling stratified by user gave similar results (Chapter 2, footnote 2).
  • API coverage. The API sample covers customers accounting for about half of 1P API usage, and geographic data are available only for Claude.ai (Introduction, "Open source data").
  • Proxies and correlations. Output length stands in for complexity, and the cost and income results are correlations; the cost regression is described as preliminary.
  • Interpretation. The report repeatedly offers alternative explanations (model improvement, learning-by-doing, changing user base) without distinguishing them.

The report itself lists open questions for users of the data: local labor-market consequences, what determines adoption across and within countries, the role of cost in enterprise deployment, and why firms automate some tasks and not others (Introduction).

Research questions

  1. How much of the rise in directive use is a measurement effect? The classifier change moved the automation share by four points. Re-classifying V1 and V2 with the V3 classifier, and validating mode labels against human annotation, would separate measurement from behavior.
  2. Does rising directive use reflect better models or user learning? The report names both. Following cohorts of users over time, and comparing new and experienced users on the same model, would distinguish them.
  3. Why do lower-adoption countries automate more, after controlling for task mix? The authors speculate. Comparing early and late adopters within the same country, and survey data on users' occupations, would test the early-adopter explanation.
  4. Is the 0.38 input-output elasticity a property of tasks or of pricing? API users have an incentive to minimize tokens. Comparing the same relationship on Claude.ai, where users pay a subscription, would show how much reflects cost-driven prompting.
  5. Does the context bottleneck limit adoption? The authors infer it from lengths. Measuring adoption in firms before and after centralizing documentation or connecting data sources would test whether access to context drives use of complex tasks.
  6. How stable are the geographic patterns? Repeating the AUI analysis in later reports, with stratified samples for small countries, would show whether adoption gaps narrow or widen, the question the report raises about economic divergence.

How it connects

References

  • Appel, R., McCrory, P., Tamkin, A., Stern, M., McCain, M., and Neylon, T. (2025, September 15). Anthropic Economic Index report: Uneven geographic and enterprise AI adoption. Anthropic. https://www.anthropic.com/research/anthropic-economic-index-september-2025-report. PDF: https://assets.anthropic.com/m/218c82b858610fac/original/Economic-Index.pdf
  • Anthropic. (2025, September 15). Anthropic Economic Index: Tracking AI's role in the US and global economy [summary post]. https://www.anthropic.com/research/economic-index-geography
  • External sources (Gallup, the Census Bureau's Business Trends and Outlook Survey, and the economics literature in the footnotes) are cited as the report describes them and were not read for this note.