Deni Sugaipov, Researcher at the Gaidar Institute’s Mathematical Modeling of Economic Processes Department, analyzed a study by Anthropic, the developer of the Claude neural network, based on calculations using data from the Anthropic Economic Index and the Research Center for Economic and Social Research (RMES) at the National Research University Higher School of Economics (HSE).
Most estimates of AI’s impact on the labor market are based on job descriptions. However, Anthropic uses a different source—millions of real-world user conversations with Claude. Since 2025, the company has been running the Anthropic Economic Index project based on this data and regularly publishes reports on which professions and tasks people use AI for.
For each profession, Anthropic compares two metrics: the proportion of tasks that AI could theoretically help solve, and the proportion of tasks for which it is actually used. For example, in programming and mathematics, AI could be applied to solve 94% of tasks, but in practice, it is used for only one-third of them.
Artificial intelligence is used most extensively by programmers, call center operators, and financial analysts. However, in the U.S., the technology has not yet led to a rise in unemployment. Nevertheless, there is an effect: young professionals aged 22–25 are 14% less likely to seek employment in professions where AI is used most actively.
Comparison with estimates for Russia
In a study by the Laboratory for Mathematical Modeling of Economic Processes at the Gaidar Institute covering Russia, these professions were also identified as among those most vulnerable to AI adoption; however, based on RMEZ data for 2021–2024, we found no sustained impact of this vulnerability on real wage growth.
How much time does AI save?
Anthropic also asked Claude to estimate, for each dialogue, how long the task would have taken a person with and without AI, and whether it was successfully solved. On average, it took about three hours without AI versus 15 minutes with AI. However, only two-thirds of the tasks are successfully solved, and the more complex and time-consuming the task, the less likely it is to be solved. With this in mind, the company calculated that, if current AI models were widely adopted, they could contribute about 1 percentage point per year to labor productivity growth in the U.S. over the next ten years.
Conclusion
Conclusion AI is used far less than is technically possible. It appears that, for now, it is not so much taking away jobs as it is making it harder for newcomers to enter the profession.