Predictions about generative artificial intelligence (AI) often focus on extremes, from mass job displacement to sweeping productivity gains. New usage-level data from Anthropic offers a more measured picture of what is happening inside real workflows today.
In its latest Economic Index report, Anthropic analyzes anonymized Claude.ai and first-party API interactions from November 2025 to track how AI is used at the task level, how autonomous those interactions are, and how performance changes as tasks become more complex.
Instead of relying on surveys or self-reported adoption, the report examines millions of actual interactions and classifies them across more than 3,000 distinct work tasks. It introduces five “economic primitives” — task complexity, user and AI skill, autonomy, success rates, and purpose of use.
The data shows that while AI is a useful tool, it’s hardly the job replacer some feared, as businesses remain cautious in its implementation, while open to the possibility of saving employees time in their workday. This dovetails with PYMNTS Intelligence data that also found roughly 7 in 10 workers who use AI for their jobs say their workplace encourages its use, while fewer than 1 in 10 say their employer actively discourages it.
Small Share of Tasks Accounts for Large Usage
The report finds that the top 10 tasks account for 24% of all Claude.ai conversations, even though more than 3,000 unique work tasks were identified in the dataset. Concentration is even higher in enterprise environments: among first-party API usage, the top 10 tasks represent 32% of interactions, up from 28% in the prior reporting period.
Software-related work dominates this clustering. Modifying software to correct errors alone accounts for 6% of Claude usage, making it the single most common task. More broadly, computer and mathematical tasks together make up roughly one-third of all interactions, underscoring how strongly AI use is tied to technical, well-scoped activities.
Advertisement: Scroll to Continue
Beyond software, usage is broadening but remains uneven. Educational tasks such as coursework assistance and instructional content creation now represent about 15% of Claude interactions, up from earlier in the year. Writing, editing and other arts and media tasks have also grown as a share of usage, signaling gradual diversification even as overall use remains concentrated in a narrow task set.
Collaboration Accounts for Most AI Interactions
A central finding of the report is that most AI use still involves close human oversight. In the November 2025 sample, 52% of Claude conversations were classified as augmentation, meaning users worked with the model, guiding prompts, reviewing outputs, and making final decisions, the company said. By comparison, 45% of interactions were classified as automated, reflecting more directive, hands-off requests.
This balance marks a shift back toward collaboration after earlier reporting periods showed a higher share of automated interactions. While automation remains more common in enterprise API traffic, where models are embedded into scripted workflows, even their usage is concentrated in a limited set of tasks rather than broad end-to-end processes.
Uneven Adoption
Geographic adoption mirrors existing economic and workforce patterns. Overall usage is led by a small group of countries, including the United States, India, Japan, the United Kingdom and South Korea.
Within the United States, usage is also uneven. The top five states account for about 50% of all Claude interactions, despite representing only 38% of the working-age population. At the same time, lower-usage states are growing faster. If current trends persist, Anthropic estimates that usage per capita could converge across U.S. states within two to five years, a pace far faster than the diffusion of many earlier general-purpose technologies.
The findings align with broader workforce sentiment. As PYMNTS reported on research from OpenAI, 75% of workers said AI had improved the speed or quality of their output.
