AI-assisted work is saving time with an estimated annual labor-cost equivalent of $2.7 trillion, or 3.4% of the combined GDP of 86 countries, according to a new IMF working paper.

The figure does not mean artificial intelligence has added $2.7 trillion directly to global economic output. It values the hours that AI may save at the wages paid to workers performing those tasks.

Researchers Rachel Yuting Fan and Ha Minh Nguyen used five waves of the Anthropic Economic Index covering January 2025 through February 2026.

Each wave contained 1 million Claude conversations, which were classified by occupational task and combined with employment and wage data from the International Labour Organization.

Rich economies capture most of the measured gains

The study estimates that high-income economies account for 96% of the total labor-cost equivalent generated by AI use.

Measured relative to GDP, the estimated gain reached about 4.2% in high-income countries, compared with roughly 0.6% in middle-income economies and 0.1% in low-income countries.

The difference reflects more than higher wages in wealthy economies. AI use in poorer countries also remains concentrated among a comparatively small group of professional workers.

That limits the employment-scale effect that occurs when AI spreads into occupations employing larger sections of the population.

The authors found that virtually all measured AI value in some developing economies was generated within a narrow professional enclave.

By contrast, AI use in countries such as Australia, the United Kingdom and the United States was distributed across a broader range of occupations.

AI gains favor higher-paid occupations

The researchers created an AI concentration index to measure whether AI-assisted time savings were concentrated in higher-paid or lower-paid jobs.

The index was positive in nearly every country studied, indicating that the measured gains generally tilted toward occupations at the upper end of the wage distribution.

The United States recorded an index score of 0.49, suggesting that AI use had spread beyond its initial concentration in software and other highly paid professional work.

Tanzania recorded a score of 0.98, meaning that almost all measured AI gains were concentrated near the top of its occupational wage distribution.

Professional occupations employ fewer than 5% of workers in Tanzania, according to the paper. Agricultural, service and elementary occupations showed comparatively limited direct engagement with AI.

India and Kenya also showed heavily concentrated patterns, with index scores of 0.84 and 0.78, respectively.

AI use is gradually spreading to more jobs

The distribution became somewhat broader during the period studied.

Between August and November 2025, 32 of 108 countries, or 30%, recorded a decline in AI concentration.

Between November 2025 and February 2026, the number rose to 53 of 110 countries, or 48%.

AI use expanded from its early software-engineering base into education, sales and office work, the researchers said.

As that happened, the average wage associated with an AI conversation fell 5.5% over 13 months.

The decline did not reduce the paper’s aggregate valuation. Lower-paid occupations employ many more people, so their growing use of AI increased the total volume of potentially time-saving work.

Why it matters

Much existing research measures which jobs AI could theoretically affect. This paper instead uses observed conversations to examine where people are currently applying the technology.

That approach complements earlier StatsJournalist coverage of research using actual AI consumption data to measure how the technology is spreading through markets and the economy.

The distributional findings suggest that access to AI does not automatically produce broadly shared gains.

Countries with stronger AI regulatory readiness recorded larger estimated gains relative to GDP and a broader distribution across occupations.

Having English as an official language was also associated with a faster decline in concentration.

The authors suggest that English-language legal, educational and business documents may be better represented in the training data used by major language models. The relationship is an association and does not establish that language policy caused faster AI diffusion.

The $2.7 trillion figure is highly uncertain

The estimate depends on assumptions about the volume of AI conversations, Claude’s market share and the amount of time saved on each task.

Under more conservative assumptions, the annual labor-cost equivalent falls to $1.6 trillion. Under the paper’s most aggressive scenario, it rises to $6.1 trillion.

The occupational data come from Claude.ai web conversations, rather than a representative sample covering every AI provider.

The researchers scaled those conversations to approximate the wider AI market, but direct usage data from ChatGPT, Gemini, Copilot and other platforms were not analyzed in the same way.

The time required to perform tasks with and without AI was also estimated by Claude rather than measured through workplace experiments.

Each data wave covered only one week, and the analysis excludes enterprise API activity, which may be more heavily concentrated in technical and professional occupations.

The calculation also does not account for job losses, changes in wages or how workers and companies use the time that AI frees.

Infrastructure spending, electricity consumption, subscription fees and other costs of operating AI systems were excluded as well.

The $2.7 trillion estimate is therefore best understood as an early valuation of potentially saved labor time, not a direct measure of additional GDP, corporate profit or household income.

The IMF classifies the study as a working paper describing research in progress. Its findings represent the authors’ analysis rather than an official position of the IMF.