AI model-maker Anthropic has analyzed millions of conversations with its chatbot Claude. The findings show that programmers are by far the most frequent AI users, while employees at both the lowest and highest ends of the salary scale have largely stayed away.
This conclusion stems from Anthropic’s newly released Economic Index, the dataset for which is now available under an open-source license. It marks the first time an AI company has explicitly disclosed how users interact with its chatbot. Previously, we only had benchmarks for tasks performed by Claude, ChatGPT, and Gemini without knowing whether all those listed capabilities were truly being utilized.
The Economic Index dataset consists of more than four million Claude conversations from December 2024 and January 2025. A system called Clio stored and anonymized all this information. Interactions via the Claude API were excluded from the study, meaning many business-related use cases were left out.
(Programming) tool
The Index compares AI usage across various occupational fields with each field’s share of the overall U.S. workforce. Put differently, it reveals whether certain professions consult the Claude chatbot disproportionately often. The categorization into occupations is based on the content of the outputs themselves rather than linking users to their actual jobs. Because Anthropic examined the data anonymously, it does not know each user’s line of work. The company relies on the O*NET database to define occupational fields, as O*NET maps occupations based on their constituent tasks and is primarily used to match job seekers to roles that align with their existing skills.
A striking 37.2 percent of all Claude interactions involve computer and mathematical tasks—an outlier, since only 3.4 percent of the U.S. workforce is employed in those specialized areas (such as programmers, software engineers, and mathematicians). Unsurprisingly, more manual professions remain distant from GenAI: truck drivers, fishermen, farmers, and installers seldom discuss their work with Claude.
Another important finding from Anthropic is that 57 percent of all Claude users clearly use the chatbot for help—i.e., augmentation—rather than fully automating tasks. Even the remaining 43 percent calls for caution: in many of those instances, Claude might independently produce dozens of lines of code or a long text, but users often go on to revise or expand it. Anthropic categorizes such iterative modification of AI output as augmentation, but it cannot measure how extensive that process is.
Augmentation versus automation
Anthropic breaks augmentation and automation down further. Iteration of tasks (31.3 percent) and educational purposes (23.3 percent) are by far the most common forms of augmentation used. Users who allow Claude to work more autonomously typically provide what Anthropic calls a “directive” (27.8 percent)—for example, to write an email or perform a geometric calculation, according to the authors of the paper. Feedback loops (14.8 percent) also occur frequently, such as when programmers tweak a piece of code across multiple successive inputs and outputs.
Anthropic describes its Economic Index as the clearest reflection yet of actual AI usage. There is some merit to that claim. Organizations like Cognizant and Capgemini have mainly focused on the future impact of AI on employment, highlighting its potential and predicting a “hockey stick moment” when AI becomes so advanced that it replaces entire occupations. They stress that companies must be ready for that possibility.
Looking at the current reality, we see a cautious, measured introduction of AI. On a broad scale, we already have some idea of how various professions use ChatGPT and its AI siblings—Gemini, Claude, and DeepSeek. Combined with Anthropic’s Economic Index, we get a clearer picture of how these tools are shaping working life.
That picture points to AI as an assistant, not a job destroyer—echoing what Cognizant CEO Ravi Kumar S told Techzine last year. In fact, he argued that AI might actually create jobs by offloading tasks people are less inclined to do themselves.
But a little nuance
While Anthropic’s openness is commendable, a few caveats remain. All indications suggest that most AI users are found elsewhere, not on Claude. In popular perception, ChatGPT is to AI chatbots what “googling” is to search engines. Claude never became the most widely adopted AI app; that distinction goes to ChatGPT as well as China’s DeepSeek—and recently Mistral in certain regions.
This matters because Claude occupies a niche. Online discussions show Anthropic’s chatbot is especially popular among programmers, who praise its coding guidance as that of a “true expert.” This is despite benchmarks often favoring other AI tools for coding tasks. Whatever the reason—Claude’s style, approach, or temperament—it resonates particularly with software developers.
This skews the Anthropic Economic Index in a clear direction. On top of that, data from Claude’s API—which might reveal a lot of corporate use—is excluded here. The precise extent of that bias is unknown. Ideally, we would see similar datasets from GitHub Copilot, ChatGPT, Microsoft’s Copilot suite, Google’s AI tools, and so on. That is unlikely to happen, so for now, we can conclude only that Claude mainly serves as an AI programming assistant—and certainly not as an all-purpose AI agent poised to replace everyone’s jobs.
Also read: Google’s Gemini 2.0 holds a trump card against its rivals