AI Already Saves $2.7 Trillion of Work a Year. Most of the World's Workers Aren't in the Room

$2.7 trillion a year. That is the value two IMF economists just placed on the working time AI currently saves worldwide — about 3.4% of the combined GDP of the 86 countries they could measure. [Fact] And in Tanzania, nearly all of that value is generated inside occupations employing fewer than 5% of the workforce. [Fact]

Both numbers come from the same July 2026 working paper: "Aggregate Gains from AI and Their Distribution: Global Evidence from Usage Data" by Rachel Yuting Fan and Ha Minh Nguyen (IMF WP 2026/147 — staff research, the standard disclaimer applies: it does not represent the IMF's official view). The headline figure will travel far. The map of who is left out deserves to travel with it.

Watching usage instead of guessing exposure

Most estimates of AI's effect on jobs score occupations against what AI could do — the exposure-index approach we examined when BLS added AI exposure categories to its projections. Fan and Nguyen flip the question: which jobs are actually using AI, in which countries, and how is that changing?

Their raw material is five releases of the Anthropic Economic Index, spanning January 2025 to February 2026, each sampling one million Claude conversations that Anthropic's classifier maps to O*NET occupational tasks. [Fact] From the third wave onward the conversations are geocoded: in the February 2026 release, 176 countries received a geographic code, 117 met the minimum-conversation threshold, and 86 countries with usable ILO wage data — together covering 79% of world GDP — form the core sample. [Fact] One notable absence: China, where Anthropic does not operate. [Fact]

The aggregate measure, labor cost equivalent (LCE), is disarmingly simple: hours saved per conversation, times the number of conversations, times the local wage for that occupation. For "modify existing software," the estimated completion time falls from 3.54 hours human-only to 0.30 hours with AI — 3.24 hours saved per conversation. [Fact] Scale up under the paper's median assumption (200 million Claude conversations per week, a 25% market share, so roughly 800 million AI conversations economy-wide) and you get $2.7 trillion a year — with an honest sensitivity range of $1.6 to $6.1 trillion depending on those market assumptions. [Fact]

Run the division yourself and the number becomes more tangible: 800 million weekly conversations is about 41.6 billion a year, which works out to roughly $65 of labor time saved per AI conversation on average — a calculation the paper implies but never states. [Estimate]

Two curves moving in opposite directions

Here is the finding we think matters most for individual workers. The average wage of the occupations using AI is falling: a US-wage-weighted index of AI usage dropped from $82,353 in January 2025 to $77,845 in February 2026, a 5.5% decline, as Computer & Math occupations lost 4.9 percentage points of global conversation share while Education gained 4.1 points, Sales 2.7, and Office/Admin 1.6. [Fact] Yet total LCE more than doubled over the last six months of data, from $1.2 trillion to $2.7 trillion annualized. [Fact]

The resolution of that paradox is employment scale. Teachers, salespeople, and administrative staff earn less per hour than software developers — but there are vastly more of them, so lower-value conversations multiplied across a much larger workforce push the total up. Growth in conversation volume accounted for about 72% of the LCE increase in one quarter and 84% in the next. [Fact]

The software core is dissolving outward. That is the paper's most hopeful sentence, even though it never quite writes it.

A Gini coefficient for AI's gains

The paper's second invention is an AI concentration index (ACI), adapted from health economics: it is positive when AI's time savings tilt toward higher-paid occupations relative to their employment share, zero when neutral. In nearly every country, in every wave, it is positive. [Fact]

The spread is what stings. The United States sits at 0.49; India at 0.84; Kenya at 0.78; Tanzania at 0.98. Across sub-Saharan Africa and South Asia, ACIs cluster between 0.90 and 1.00 — virtually all measured gains flow to the top of the wage ladder — while Western Europe and North America range from 0.35 to 0.55. [Fact] Between countries, the tilt is just as steep: high-income economies capture 96% of total LCE, worth about 4.2% of their GDP, versus roughly 0.6% for middle-income and 0.1% for low-income countries. [Fact]

Even the "broad" American distribution is less broad than it sounds. In the paper's own worked example, Professionals — 23.2% of US employment — capture 83.6% of AI's time savings, and walking up the wage ladder through Technicians, which covers 65.1% of US employment, accumulates only 10.7% of the gains. [Fact] Do the ratio: the lower two-thirds of the US wage ladder is capturing about one-sixth of its proportional share. [Estimate]

But the tilt is easing. Countries with a falling ACI rose from 32 of 108 (30%) to 53 of 110 (48%) within six months, and 26 countries reversed from rising to falling concentration — 12 of them middle-income. [Fact] The sobering footnote: 20 of 81 countries showed no change at all, because every one of their classified AI conversations mapped to a single occupational group. [Fact]

What predicts the spread? English, mostly

Cross-country regressions say AI regulatory readiness — the regulatory sub-index of the IMF's AI Preparedness Index — is the dominant predictor of how much AI value a country captures relative to GDP, and richer, better-prepared countries also spread the gains more broadly. [Fact] But neither income nor readiness predicts whether concentration is falling. What does is whether English is an official language. [Fact]

The authors' interpretation: countries that write their legal codes, curricula, and business documentation in English have their institutional knowledge well represented in LLM training data, so AI gives locally accurate answers across many occupations rather than serving only elites who can work in English. [Claim] The coefficient survives excluding the United States, so this is not just an early-adopter artifact. [Fact] For every country whose institutional life runs in a lower-resource language — Korea included — that mechanism is worth sitting with: the constraint on broad AI gains may be less about your infrastructure than about whose documents the models were trained on.

The measurement problem this paper cannot escape

Now the uncomfortable part. Last week we covered an NBER paper that surveyed 20,000 American workers and found that chatlog-based occupation measures correlate only 0.11 to 0.34 with survey-measured adoption — and barely correlate with each other. This IMF estimate is built entirely on one vendor's chatlogs. Worse, the hours-saved variable is estimated by Claude itself, not measured experimentally — the model is, in a real sense, grading its own time savings. [Claim] The authors flag this openly, and their bounding exercise (swinging mean hours saved from 1.80 to 3.81) keeps the total within $2.5–2.9 trillion because high-volume tasks are the precisely estimated ones. [Fact] Still: each wave covers a single week of consumer web conversations, enterprise API traffic is excluded, and the ACI is likely a lower bound on true concentration since API usage skews even harder toward high-wage computing work. [Fact]

That excluded enterprise channel hides one more striking ratio. A footnote estimates Anthropic's API alone at roughly $1.14 trillion a year in labor value, from about 31 billion conversations inferred at $0.34 of revenue each. Set those side by side: buyers pay about $0.34 per conversation for something the paper values at roughly $37 of saved labor time — a wedge of about 100× between AI's price and its estimated labor value. [Estimate] If even a modest fraction of that wedge is real, the adoption pressure on every employer is enormous.

Full disclosure cuts both ways: the occupation pages on this site render exposure metrics descended from the same Anthropic Economic Index lineage this paper uses — so every critique above also applies, in part, to our own dashboard. And the LCE is a partial-equilibrium accounting of time saved at current wages; it says nothing about displacement, wage compression, or whether those gains reach workers or their employers — a question the authors explicitly leave open.

What this means for your paycheck

Three practical readings. First, if you work in education, sales, or office administration, the diffusion is arriving in your occupational group now — those are the three groups gaining conversation share fastest, and the $65-per-conversation arithmetic means the value being created in your field scales with how many of your tasks AI can touch. Second, "gains flowing to your occupation" is not the same as gains flowing to you: whether that $2.7 trillion shows up in paychecks or in margins depends on bargaining and on whether AI complements or substitutes your specific tasks. Third, if you work in a country where AI usage still maps to a single professional enclave, the binding constraint documented here is institutional — training-data representation and regulatory readiness — not personal effort, which is both a relief and a warning about where to direct it.

The data ends in February 2026. AI is trickling down the wage ladder, faster where the documents speak English. It has not yet reached the bottom — in twenty countries, it has not even left the first room.

Sources

• Fan, R. Y. & Nguyen, H. M. (2026). "Aggregate Gains from AI and Their Distribution: Global Evidence from Usage Data." IMF Working Paper 2026/147, July 10, 2026. https://www.imf.org/en/publications/wp/issues/2026/07/10/aggregate-gains-from-ai-and-their-distribution-global-evidence-from-usage-data-577586 (full text read via IMF eLibrary; underlying Anthropic Economic Index releases are public on Hugging Face, wages from ILOSTAT)

This article was produced with AI-assisted analysis. All figures were verified against the working paper's full text; our own derived calculations are tagged as estimates.