What Sorts Countries Under AI (Part 1)

Two countries can reach identical AI adoption rates and become different economies afterward. The adoption figure, the number that appears in every market report and policy brief, tells you where AI has arrived. It says almost nothing about what it lands on.

What it lands on is hidden in the structure of work: in how many workers draft, diagnose, coordinate, teach, classify, advise, calculate, inspect, clean, drive, farm, or build. Before AI becomes productivity, revenue, or investment, it first encounters this occupational structure. The same technology therefore does not translate into the same economic pressure everywhere.

This paper maps that pre-adoption terrain across 147 countries, using occupational employment and task-structure data to estimate where, and through what mechanism, each economy’s work is reorganised when it becomes AI-compute demand. All estimates are structural (i.e., theoretical token-addressable demand). This figure does not equal actual adoption, API revenue, or realised productivity. The US-proxy assumption used to bridge ILO employment categories and O*NET task structures is the main source of methodological uncertainty. Comparative patterns across countries are considerably more reliable than absolute country totals.

What total demand hides?

The countries with the most total theoretical token demand are, in order: India, United States, Indonesia, Brazil, Russia, Japan, Pakistan, Nigeria, Mexico, Bangladesh.

Readers familiar with those countries will notice immediately that this is also, more or less, a list of the world's largest workforces. India ranks 1st in total demand and 116th in demand per worker. Japan ranks 6th in total and 45th per worker. Nigeria ranks 8th in total and 114th per worker. The statistic is doing two entirely different jobs at once: counting workers and measuring task intensity. Conflating them produces a map tells you almost nothing useful about where AI's structural effects will be strongest.

Separate the two, and a different picture emerges. Demand per worker varies 3.1 times across the 147 countries in the dataset, from approximately 810,000 tokens per worker at the top to 260,000 at the bottom, and this variation is entirely explained by what workers do. That is the signal worth following.

What Happens When Employment Becomes Demand

Every occupation group shifts when employment shares are translated into token-demand shares. The shifts are consistent across countries and the magnitudes are large enough to reorganize a country's economic map entirely.

Legal Occupations hold 0.5% of global employment but 3.3% of global token demand, a 6.4x overrepresentation. Healthcare Practitioners hold 2.4% of employment but 7.0% of demand. Both occupation groups are dense with language, documentation, judgment, advice, classification, and procedural reasoning. But overrepresentation should not be confused with frictionless adoption. Survey evidence from Bick, Blandin, and Deming finds that legal and business occupations show among the highest theoretical AI exposure, yet the realized adoption lags behind, suppressed by confidentiality constraints, liability exposure. The Anthropic Economic Index finds clinical occupations remain underrepresented in actual AI usage relative to their structural weight, held back by barriers around data fragmentation, regulation, governance, workforce capacity, and trust, as documented by the OECD. Legal and healthcare are high-(theoretical) demand, high- (adoption) friction sectors: structurally central to AI-compute demand, but not necessarily the first places where that demand is fully realised.

The compression is equally systematic. Building and Grounds Cleaning employs nearly one in five workers globally. In the translation to token demand, it generates 0.62x its proportional share. Farming translates at 0.52x. Transportation at 0.65x. The employment map and the demand map no longer resemble one another. What drives the compression is that the pathway into physical work runs through capital stock, sensors, and hardware replacement cycles rather than through software adoption. A transport worker's exposure to AI is gated by whether the vehicle they operate has been built or retrofitted to carry it. Only 5.6% of fleet operators have deployed AI broadly as it must travel through machinery that turns over on decade-long cycles. The same logic applies in agriculture and cleaning: the work is physical, context-dependent, and reaches AI only when the equipment around it does. Token-addressable tasks are a small fraction of what these workers actually do.

The translation therefore happens twice: first at the level of occupations, then at the level of countries. What this means for any individual country depends entirely on its occupational mix. A country where 12% of workers are in healthcare, legal, and computing occupations undergoes a very different translation than one where 34% are in farming. The question is whether these differences produce a continuous spectrum of outcomes, or whether countries cluster into a smaller number of recognizable patterns.

Three Transformations

Each country inherits the same occupational multipliers, but applies them to a different employment structure. The result is a country-specific shift vector: some occupations gain weight in the demand map, others lose it.

Clustering the full occupational shift vectors of 130 countries reveals three stable transformation archetypes. The archetypes describe the dominant mechanism through which an economy’s existing work is reorganized when translated into theoretical AI-compute demand. Every economy experiences some degree of professional amplification, service reorganization, and agricultural compression. The archetype simply identifies which mechanism dominates.

Archetype 1: Professional Amplification (57 countries)

Japan's healthcare workforce is 5.3% of employment. In the translation to token demand, it becomes 16.3%, three times its employment weight. For instance, Japan has committed ¥22 billion to diagnostic AI while Japanese legal-tech platforms are moving into contract review, research, and drafting workflows, precisely the documentation-heavy professional work that the demand map elevates.

Similar to Japan, across the 57 countries in this archetype, Healthcare Practitioners gain a mean of +6.1 percentage points above their employment share and Legal Occupations gain +4.0. Computer and Mathematical occupations rise. Average tokens per worker reach 668,000, the highest of the three archetypes.

The mechanism is straightforward. These economies already employ large shares of workers in healthcare, legal, education, management, and computing. When work is translated into AI-compute demand, those occupations claim a disproportionately large share of the economy’s structural demand.

Members include Western and Northern Europe, the US, UK, Australia, Japan, Russia, Israel, Singapore, Chile, and Egypt, a group defined by a shared occupational inheritance. Specifically, Chile and Egypt show why this is not simply a rich-country category: Chile enters through a broad professional and higher-skill occupational base, while Egypt enters through the legacy of credentialed public-sector employment in education, health, and administration.

Archetype 2: Service Reorganization (54 countries)

These economies have a different structural problem. Building and Grounds Cleaning falls 8.0 percentage points below its employment share in the translation, far larger compression than in Archetype 1. Management, Healthcare, and Educational Instruction all gain, but from a professional base averaging just 4% of employment rather than 12%.

The ceiling is lower because the foundation is thinner. Coordination and knowledge work rise relative to routine labour. Informal employment is estimated at 53% of the workforce in Latin America and 77% across Asia: manual, non-linguistic work concentrated in precisely the occupations with the lowest translation multipliers. The structural pressure accumulates in a smaller professional base and runs into the same headwinds converting it into realized use. Average tokens per worker stands at 506,000.

Members include India, Indonesia, Pakistan, Bangladesh, Vietnam, Mexico, Turkey, Thailand, the UAE, Colombia, Peru, and the Philippines. But within this archetype the compression mechanism is not uniform. Wealth, however, does not change which archetype a country belongs to. India and the UAE sit $45,000 apart in income per capita and in the same structural category. In both, Building and Grounds Cleaning is the single largest transformation driver, compressing at -7.2pp in India, -7.7pp in the UAE. The difference is what that sector weighs: in India it employs 32.4% of the workforce where the sector’s collapse in the translation dominates everything else. In the UAE it employs 11.3%, a large compressor but one that a substantial Management sector partially offsets, compression that is structurally entrenched by design: migrant workers constitute 90% of the UAE’s private sector workforce, concentrated in construction, hospitality, and domestic work, precisely the occupations with the lowest token-demand multipliers. That offset is not enough, and not in the right occupational categories, to cross into amplification. Therefore, occupational structure determines where AI-compute demand accumulates first.

Archetype 3: Agricultural Compression (36 countries)

In agricultural economies, AI shrinks the largest employer: Farming, Fishing and Forestry compresses by a mean of -13.8 percentage points across archetype 3, against -2.5 points in archetype 2 and -0.9 in archetype 1. Sales, Educational Instruction, and Management gain, but from a professional base averaging 2.6% of employment. Average tokens per worker is 423,000, about 16% lower than that of Archetype 2 and 37% lower than that of Archetype 1.

The mechanism is mathematical domination. Farming employs 34% of the average archetype 3 workforce. At a translation multiplier of 0.52x, a sector that large reorganizes the entire occupational profile regardless of what the professional sector does. At the global level, existing AI applications in agriculture are concentrated in precision farming, predictive analytics, and crop and soil monitoring, where AI augments machinery and production systems, primarily living in the advanced economies, rather than the worker’s day-to-day tasks, especially in the developing world.

Members include Nigeria, Ethiopia, Uganda, Zambia, Zimbabwe, Mozambique, the DRC, and most of Sub-Saharan Africa's lower-income tier. Nigeria illustrates the mechanism cleanly: farming holds 21% of employment against a professional sector of under 7%. In the translation, that professional sector still more than doubles its weight, rising to 16% of token demand, the same uplift seen in Archetype 1 economies. What differs is scale: the base beneath it is too thin to outweigh farming's collapse. Despite AI-adjacent tools and initiatives, for instance, a Nigerian mobile advisory app with visible rice yields increase, the tool sits inside the narrow advisory slice of the farm work instead of affecting the physical tasks that dominate it. Where farming and other physical occupations account for a large share of employment, the uplifts from the professional occupations are not large enough to become the economy’s dominant transformation mechanism.

If occupational structure largely determines where AI-compute demand emerges, why do countries with similar structures adopt AI at dramatically different rates? Spoiler alert: the difference lies in the conditions required to realize its structural potential.

To be continued.

Cover photo by Arlington Research on Unsplash