What Sorts Countries Under AI (Part 2)

This is Part 2 of “What Sorts Countries Under AI”. To see how the archetype develops, please see What Sorts Countries Under AI (Part 1).

To recap, using employment and task data across 147 countries, Part 1 shows that translating employment shares into token-demand shares systematically inflates language- and judgment-heavy professions (legal, healthcare, computing) while compressing physical, capital-gated work (farming, cleaning, transportation). Clustering each country's occupational shift pattern reveals three archetypes: Professional Amplification, Service Reorganization, and Agricultural Compression, determined by a country's existing occupational mix rather than its income level, meaning countries with very different wealth levels can land in the same structural category.

This chart lets you see that translation directly: pick a scope to compare each country's share of employment against its share of theoretical token demand, colored by which archetype it falls into.

Before this archetype is treated as a structural finding, it should survive an attempt to explain it away. Three competing, simpler explanations are available: that the archetypes are just industry composition under a different name; that they are just income; that they are just current AI adoption, described backward. Each test below is an attempt to eliminate one of those explanations. If the archetypes remain informative after all three, they are more than a descriptive clustering: they capture a structural feature of labour markets that industry, income, and adoption each fail to fully account for on their own.

Test One: Is It Just Industry Composition?

The archetypes come from an occupational reconstruction what tasks people actually perform. The first test asks whether a country’s industry composition (i.e., which sector employs them) already contains the same information. If it does, the occupational layer is redundant.

Conceptually, occupations classify jobs by the tasks and duties workers perform, whereas industry statistics classify the economic activity of the employer⁠. The first test asks whether the coarser classification already contains the same information.

A model predicting archetype from nothing but a country’s World Bank agriculture and industry employment shares, tested on data it hadn’t seen, classifies 76.9% of 130 countries correctly. Against a baseline of always guessing the most common archetype, 34.6%, that’s real signal. A McFadden pseudo-R² of 0.49 says the same thing: for a two-variable model, industry composition is doing substantial work.

One variable does almost all the work: agriculture share of employment is significant at p<0.001 against both other archetypes; industry share isn’t significant at all (p=0.917, p=0.146). In addition, agricultural compression is identified with 86% precision and 84% recall, close to a clean rule. Service reorganization gets 72% precision and just 64% recall, the weakest of the three by a wide margin, and where nearly all the model’s errors live.

The ternary plot makes this visible. Agricultural compression (blue) pulls cleanly toward the agriculture corner, largely its own territory. Professional amplification (green) and service reorganization (yellow) sit tangled together across the same services-heavy region, sharing the upper part of the triangle almost point for point. Industry composition can tell you whether a country’s professional base is thin or substantial in aggregate, not what kind of work fills a large services sector once it exists. The same sector includes hospitals and law firms alongside hospitality, retail, logistics, and domestic work, even though service employment spans occupations with fundamentally different task content and skill requirements⁠. Japan and the UAE sit in that same overlapping cluster, indistinguishable by industry share alone. The archetypes’ real contribution is legible exactly where this chart shows yellow and green refusing to separate.

Test Two: Is It Just Income?

The more obvious objection: richer countries have more lawyers and doctors, so perhaps the archetypes are simply GDP per capita with extra steps. Structural transformation usually shifts employment away from agriculture and toward services as countries get richer⁠. The question is whether income fully explains that transformation, or whether occupational structure contains information of its own.

Looking at the data, income does most of the work you’d expect it to. Log GDP per capita alone explains 77.8% of the variation in token demand per worker across 141 countries, a correlation of 0.88. However, income alone does not predict transformation. Adding archetype to the income model raises explained variance to 85.6%, a gain too large to be chance (F=37.0, p=1.4×10⁻¹³). The reason is systematic: professional amplification countries have deeper professional work than their income alone would suggest; agricultural compression countries have shallower professional work than their income alone would suggest (F=8.89, p=2.3×10⁻⁴ across archetypes). Income sets the volume. Archetype sets the shape. That residual shape is economically plausible because occupational structures can be "path dependent", reflecting institutional and specialization patterns built long before current income levels.⁠

The countries furthest from what income alone predicts make the point concretely. Somalia’s occupational structure looks far richer in professional terms than a GDP per capita near $660 would suggest. Russia and Belarus both sit well above their income-predicted intensity, consistent with Soviet-era legacy of a state-built professional workforce: the Semashko health system left post-Soviet countries with large, centrally organized medical systems⁠, while Soviet higher education was oriented toward producing specialists for industrial and state planning⁠. At the other end, a cluster of Sub-Saharan African economies, Guinea-Bissau, Burundi, Mozambique, Rwanda, Ethiopia, Zimbabwe, sit well below what their income would predict, occupational structures thinner in professional work than their GDP per capita alone implies. They are simply the largest gaps left over once income has done all the explaining it can.

Test Three: Is It Just Current Adoption?

The simplest version of this test asks: do countries with deeper professional work, more healthcare, legal, and computing-heavy economies, actually use more AI today?

Yes. Adoption rises in three clean steps: 10.5% average in agricultural-compression countries, 18.0% in service-reorganization countries, 31.7% in professional-amplification countries. Structure alone explains about half the difference in adoption rates across 110 countries (R²=0.51), and each step up is statistically solid, not a coincidence of a few outlier countries.

But there’s an obvious complication: rich countries tend to have both deep professional economies and high AI adoption. So does structure actually explain anything, or are we just seeing “rich countries do more of everything”? Historically, income has been the dominant predictor of new-technology adoption speed across two centuries of technologies studied.

Take income out of the picture first, see how well a country’s wealth alone predicts its AI adoption. Then ask: once income has already done its job, does occupational structure explain anything more?

It doesn’t. Once income is accounted for, structure adds essentially zero extra predictive power. The two are too tangled together to separate here: countries that built deep professional economies are, largely, the same countries that got rich doing it, the same trickle-down pattern documented for technologies going back centuries.

The result draws a real line under what this analysis has shown so far. Structure explains something income can’t when it comes to what kind of work an economy produces, that was Test Two’s finding, and it holds. It has not yet been shown to explain how fast a country adopts AI, once income is already known. Those are different claims. Right now, the data supports the first and not the second.

What still stands, and matters for what comes next, is which countries don’t follow either pattern well. Russia and Belarus have exactly the kind of deep professional structure that should predict high adoption, and don’t have it, both sit under a U.S. export-control regime that presumes denial for advanced AI chips and model weights. The UAE has adopted far more than either its income or its structure would suggest on their own, built on a $30 billion, government-backed national AI infrastructure program rather than organic demand from its occupational base. Either demand with no outlet yet or adoption with insufficient foundation yet is worth watching more closely than the country’s adoption number alone would suggest.

What Three Tests Establish

The current AI-country debate has a blind spot: rankings compare countries by adoption, model usage, chip capacity, and investment. These rankings ignores a lot of national contexts, and most important among them is what kind of economy it is arriving in.

That omission helps explain a real puzzle in the adoption data. The United States, home to the frontier labs, ranks outside the world's top twenty in population-level AI use, while the UAE and Singapore lead by a wide margin, a pattern the International AI Safety Report confirms is now a stable feature of the data. Research at the St. Louis Fed has started asking why adoption diverges even across similarly rich countries, but the question is posed at the level of workers and firms. The question the analysis focuses, instead, on the level of what an economy is structurally built to do.

Test One shows the result is not industry composition in disguise. Agriculture explains most of the agricultural-compression archetype; industry classification loses power exactly where economies turn services-heavy, because hospitals, law firms, hotels, and warehouses share a sector code while sharing nothing about the work itself. Occupation, instead of the sectoral economic activity, is the relevant unit.

In Test Two, income sets the volume of structural demand. The dataset finds real, patterned divergence between countries at the same income level, best explained by occupational structures inherited from history rather than current GDP. That residual is consistent with a broader concern now surfacing at places like the Milken Institute: adoption has gone global, but its productivity payoff hasn’t followed evenly, without attributing it to occupational structure.

Test Three delivers the harder verdict. Once income is known, occupational structure adds nothing to predicting who adopts AI fastest today. What does explain it is closer to policy than to economics: the UAE out-adopt because of intense government investments that may be years ahead of any organic demand its occupational structure would generate. That is a government buying its way into a different race than the one this paper measures.

Which is the point. Adoption measures where AI is being used right now. Structural demand measures how much of an economy is actually positioned to be reorganized by it. These are two different races, and most of the current discourse is scoring only one of them. A country can lead the adoption race by spending its way there, as the UAE has. A country can also sit on deep unrealized structural demand and show up nowhere in this quarter’s usage rankings.

From Structural Demand to Investment Strategy

If the previous sections are right, the investment question changes. Instead of asking where demand is highest today, the sharper question is where AI can still reshape an economy tomorrow.

A different screen - the realization gap - follows naturally from the analysis. Compare a country’s structural AI demand with its current adoption. Where structural demand substantially exceeds realized adoption, the economy may contain underappreciated transformation potential waiting for the right conditions to emerge. Where adoption has already run ahead of structural demand, expectations may already be pricing in more economic transformation than the underlying labour market can easily support. The opportunity is the realization gap between structural demand and realized use.

Then, the archetypes suggest what to look for. For professional-amplification economies, their bottlenecks lie in regulation, liability, interoperability, compliance, and enterprise integration, making healthcare AI, legal technology, governance software, and workflow platforms the more relevant themes. Service-reorganization economies depend more on organizational transformation: SME software, business-process automation, payments, workforce platforms, and enterprise digitization. Agricultural-compression economies follow a different pathway altogether. Here, AI reaches work through capital equipment, shifting attention toward machinery, sensors, robotics, connectivity, satellite services, and equipment finance. The archetypes also suggest different catalysts: what closes the realization gap. In professional economies, realization may accelerate after changes in regulation, reimbursement, or data-sharing rules. In service economies, the trigger is more likely to be organizational adoption and workforce capability. In agricultural economies, the decisive event is often the replacement of physical capital rather than the deployment of another model.

Cut across all of this is a question of horizon. The framework also separates two investment horizons: a short-duration book rides the adoption race: policy-driven, sovereign-infrastructure-linked, fast, and reversible by the next political decision; a long-duration book rides the structural-demand race: slower, occupation-driven, and mispriced for longer precisely because almost no one is measuring it directly yet.

To illustrate the framework, consider one economy from each archetype. Japan is a professional-amplification economy, where the bottleneck is its integration into regulated professional workflows. The closest investment expression is therefore companies such as M3, which sits directly inside physician workflows and digitizes physician documentation and clinical workflows. India is a service-reorganization economy. Here the opportunity lies in reorganizing enterprise work itself, making business-process specialists such as Genpact (claims and finance operations), WNS (insurance and travel workflows), and Firstsource (healthcare revenue-cycle management) a cleaner expression of the theme than diversified IT firms whose AI exposure is spread across many businesses. Nigeria illustrates the opposite case. The most relevant precision-agriculture, equipment-finance, and rural-digitization companies remain private, leaving public investors with indirect proxies such as C & I Leasing and Airtel Africa. This reinforces the conclusion that in agricultural-compression economies, the bottleneck is capital stock rather than software, so the cleanest investment expressions may not yet be listed.

The framework ultimately suggests a simple rule: do not invest where AI is most visible. Invest where the bottleneck preventing structural demand from becoming realized demand is most likely to disappear. Therefore, begin with the realization gap, identify the bottleneck that prevents structural demand from becoming realized demand, and invest as close to that bottleneck as public markets allow.

Cover photo by Brett Andrei Martin on Unsplash