AI’s gains are large and rising, but unevenly shared

How large is the labour cost saved by AI, and how is it distributed across occupations? Using five waves of the Anthropic Economic Index from January 2025 to February 2026, this column constructs two novel measures from observed AI usage across countries over the world. The labour cost equivalent values the time currently saved by AI at $2.7 trillion annually (3.4% of GDP), concentrated in a small professional enclave in developing economies, with concentration declining in many countries.

Who benefits from using AI, and how large are the labour-saving gains? The question now runs through policy debates among central banks and policymakers. 

Survey-based studies find that most firms report no measurable impact on employment or productivity so far (Yotzov et al. 2026). Exposure indices, which measure which jobs AI could affect, find that advanced economies face greater exposure but are better positioned to benefit thanks to stronger digital infrastructure and human capital (Cazzaniga et al. 2024, Cerutti et al. 2025).

Micro-level experiments find that AI raises productivity, with the largest gains among less experienced and lower-skilled agents (Brynjolfsson et al. 2025). Access to ChatGPT reduced time on writing tasks by 40% and compressed the quality distribution, benefiting lower-ability workers most (Noy and Zhang 2023). Our recent paper (Fan and Nguyen 2026) offers a different approach. Rather than asking which jobs AI could affect, we observe which jobs are using AI, in which countries, and how the pattern is changing.

Drawing on five waves of the Anthropic Economic Index, a record of millions of Claude AI conversations from January 2025 to February 2026, we ask two questions: how large is the labour cost saved by AI, and how is it distributed across countries and occupations? The answer to the first is large. AI is already saving time worth about$ 2.7 trillion a year, roughly 3.4% of GDP across 86 countries, up from $1.2 trillion six months earlier. But those gains are unevenly shared, concentrated in rich countries and, within countries, in the highest-paid jobs.

That figure is a labour cost equivalent (LCE), which values AI’s time savings at each country’s own wages. It is an indicative measure of the productivity gains implied by current AI usage.

As AI spreads from software development into education, office work, sales, and other occupations, the typical AI conversation increasingly serves lower-paid occupations. Thus while the usage-weighted average wage fell 5.5% over 13 months, thanks to the higher employment in those lower-wage occupations, the aggregate LCE more than doubled over the same period.

Across countries, the gains are not evenly shared. A companion paper (Fan 2026) finds that per-capita AI usage varies more than 200-fold across countries, tracking income, economic structure, scientific output, and AI-policy readiness.

The aggregate gains are themselves steeply tilted by income. Relative to GDP, AI’s measured gains are about seven times larger in high-income countries (4.2% of GDP) than in middle-income economies (roughly 0.6%), and about ten times larger again than in low-income countries (0.1%). High-income countries account for 96% of the total LCE while employing a much smaller share of the world’s workers.

Figure 1 Aggregate AI gains: Scale, trend, and incidence by income

Note: Panel (a) shows the global usage-weighted wage index (left axis, R1–R5) and the annualized LCE for the 86 sample countries (right axis, R3–R5). Panel (b) shows LCE as a share of GDP by income group in February 2026, with the number of sample countries in each group.
Sources: Anthropic Economic Index; US Bureau of Labor Statistics; ILOSTAT; IMF World Economic Outlook; authors’ calculations.

AI’s realised value is concentrated both across countries (towards richer ones) and within countries (towards higher-paid occupations).

Within countries, the gains tilt toward the higher-paid occupations, and this tilt is greatest in countries where incomes are lowest. Our AI concentration index (ACI) measures whether AI usage is concentrated in higher-paid occupations relative to their employment share. A higher ACI, say close to one, suggests that almost all of AI’s gains flow to the highest-paid occupations; on the other hand, an ACI close to 0 would suggest that gains are neutral between higher-paid and lower-paid occupations.

In nearly every country in every wave, we find AI’s gains tilt toward higher-paid occupations. In the high-income US, where clerical, service and sales, and craft and trade occupations together represent a visible portion of AI’s time savings, the ACI is 0.44. In low-income Tanzania it is nearly 1, as almost all time savings flow to the top of the wage distribution, where professional occupations account for fewer than 5% of employment. In Uganda and Cambodia, virtually all AI usage-based value is generated in a small professional enclave, while in Australia and the UK, gains are spread more broadly.

Both patterns can be predicted by income (Figure 2). Richer countries capture more AI value relative to GDP, and the level of gains rises with income (panel a). They also spread those gains more widely, with the ACI falling along a steep and robust income slope (panel b). In sub-Saharan Africa and South Asia, where GDP per capita is below $5,000, ACIs cluster near 0.9 to 1, and virtually all AI usage-based value is generated at the top. In Western Europe and North America, where GDP per capita is above $50,000, they mostly range from about 0.4 to 0.6, and drop to 0.17 in Norway, reflecting AI adoption across a broader occupational range.

Figure 2 AI gains and their distribution

Panel (a)

Panel (b)

Note: R5 (February 2026), 86 sample countries. Panel (a) plots LCE as a share of GDP; panel (b) plots the ACI (0 = gains neutral between higher- and lower-paid occupations; 1 = all at the top of the wage ladder). Both panels are shown against GDP per capita (PPP, current international dollars; IMF World Economic Outlook, April 2026, 2025 values), on a log scale, the same income measure as the regressions.
Sources: as above.

Initially, AI’s measured gains were concentrated in already high-paid occupations. As AI diffuses through more occupations, those gains are broadening, and the ACI is declining in a growing number of countries. Between August and November 2025, 23% of the countries saw their ACI fall; between November 2025 and February 2026, 42% of them did. And 28 countries whose ACI was rising through November 2025 reversed course by February 2026, 12 of them middle-income economies.

Figure 3 The tilt is reversing: AI gains are spreading more broadly over time

Note: Change in the ACI across two consecutive three-month windows, 86 sample countries.
Sources: as above.

What determines who captures AI’s value, and whether it is broadening? AI regulatory readiness is the dominant predictor of how much value a country captures relative to its GDP, alongside a larger service sector. But what predicts whether the ACI is falling is language and economic structure: countries where English is an official language, and whose institutional knowledge is well represented in the training data of today’s large language models, saw their ACI fall faster, while countries with larger service sectors saw it rise. That leaves much of the developing world at a disadvantage.

Expanding the breadth and quality of non-English training data could raise the distributional returns to AI in Francophone West Africa, the Arabic-speaking Middle East, and Southeast Asia. And because regulatory readiness lifts both the value a country captures and how evenly it is shared, investment in digital infrastructure, skills, and adaptable regulation does double duty.

These figures value the time AI saves at current wages, and they cover one provider’s consumer usage, so we read them as indicative. But the direction holds across every check: AI’s gains are large, rising, and unevenly shared. AI is trickling down, but it has not yet reached the bottom.

Source : VOXeu