Latin America's challenge isn't adopting AI, but turning it into economic growth

The paradox: everyone's using it, few are gaining from it

Latin America has crossed the first threshold. AI use is no longer experimental. The IDB's May 2026 survey of non-AI-native startups found that 85% already use generative AI tools and 75% use predictive technologies, concentrated in marketing, product development, strategy, and R&D.¹ Broader industry surveys point the same way, with roughly 47% of companies implementing and another 37% actively experimenting.² By usage, the region is not behind — it is arguably ahead of its infrastructure.

The value side tells the opposite story. WEF research finds only 23% of Latin American organizations generate any economic value from AI, and just 6% report significant value — with roughly six in ten SMEs generating no measurable value at all.

On the startup side, only 12% consider off-the-shelf AI solutions a good fit for their needs.³ Adoption is high; value capture is low, and the whole story is the distance between them.

The gap is widest for small firms

That gap between use and value is not spread evenly. Value is concentrating among large, well-managed firms, while the long tail of SMEs adopts tools without changing outcomes. In Mexico, about 80–90% of firms of every size already have the digital basics: computers, internet and standard software. Yet only 15% of small firms use AI, against 44% of medium and large ones. In Argentina, 41.6% of SMEs use AI, but 56% of adopters are merely experimenting and 57% don't measure the impact. . In Brazil, AI adopters in São Paulo's are half as likely as G7 peers to train staff on AI (38% vs. about 75%) and a third as likely to have a dedicated AI team or senior AI role (17% vs. about 50%). Surveys in Chile and Costa Rica show the same pattern of broad but shallow use.⁴

This matters because the region's productivity problem lives in that long tail of small, often informal firms. AI that spreads only among a vanguard of large, globally connected companies will widen productivity gaps rather than close them. The most promising bridge is delivery through platforms: marketplace seller tools, fintech credit built on transaction histories, and public payment systems such as Brazil's Pix, which let AI-enabled services reach small firms at low cost.

Why the gap persists: it's how firms work, not which tools they have

The instinct is to blame infrastructure, where the gaps are real. LatAm's AI market is worth roughly US$12.7 billion and growing 28.1% a year. That sounds fast, but the global market is growing faster, at 31%, so the region's 5.8% share of global AI spending, already below its 6.6% share of world GDP, is shrinking.  Investment in generative AI is even lower. The region attracts $8.2 billion just 4.3% of the ＄190 billion invested worldwide.⁵

But infrastructure is not what binds today; behavior is. Firms say so themselves. Mexican firms that haven't adopted AI rank gaps in knowledge and skills ahead of cost. Argentine SMEs follow the same order, citing limited experience (46%), skills shortages (42%), and lack of AI knowledge (39%) ahead of cost (29%). Chile is the exception that proves the point. Cost is its most-cited hurdle (42%), with skills close behind (36%). Yet when adoption ran into trouble, Chilean firms most often responded by reorganizing processes or reassigning tasks (45%), well ahead of training (35%) or hiring specialists (10%).⁶ When firms cannot specify what to buy, cheaper tools do not help.

What does work is redesigning the job around the tool, not just adding the tool. McKinsey's case studies of Latin American firms show the pattern. A regional bank cut credit-decision time by 55% by rebuilding its process around 45 specialized agents, with analysts moving from doing evaluations to challenging them. A pipeline operator cut claims processing from six months to two weeks at 58% lower cost. A telecom turned a 600,000-call-a-month cost center into a revenue channel at three times typical outbound conversion. In every case, humans stayed accountable; the workflow changed.

Where value is being created and why there

If value depends on workflow redesign plus the right raw material, then it should concentrate where data and decisions already exist. It does — in two very different places.

Finance: Financial services are the region's most convincing AI success story, and structurally so. Large underbanked populations, fast-growing digital payments, and fintechs built on transactional data create exactly what AI needs: high-frequency decisions — fraud detection, credit underwriting, collections, personalized products — repeated millions of times. Nubank is perhaps the clearest example. The Brazilian digital bank surpassed 100 million customers across Brazil, Mexico and Colombia in 2024 and has since expanded beyond 114 million customers. Its AI infrastructure operates at a scale that increasingly resembles that of a global technology company: Nubank says its models can train on data from more than 100 million customers and their transaction histories,⁷ with training workloads processing trillions of tokens and Mercado Libre extends the same logic across marketplace, payments, advertising, credit, and logistics — one model-driven ecosystem optimizing the whole customer journey.

Physical industries: This is the less obvious point. Latin America is a leading agricultural and mineral producer, and its highest-value AI is analytical — systems that make physical industries more efficient. In agriculture, Kilimo (Argentina) optimized water use from climate, soil, and satellite data. The Linux Foundation quantifies the payoff at 20% water savings and 72 billion litres conserved.⁸ Buenos Aires's water utility, AySA, paired satellite radar with machine learning and reorganized its field crews around the model's output. It reached 93.4% accuracy in field checks, repaired more than 18,500 hidden leaks, and recovers about 28 million cubic meters of water a year.⁹

In mining, where the enabler and adopter roles meet, Chile's Codelco lifted annual profits by ＄80 million and extracted 8,000 additional metric tons of copper through optimized processing and predictive maintenance.¹⁰ Even here, the value sits in software, not machinery: McKinsey estimates software agents account for 80% of the 2030 automation prize, and still the majority within agriculture, manufacturing, and mining.

Where the value goes

The previous section looked at which sectors create value; this one ask who ends up keeping it. Follow the money, and the region's two roles look very different. Value from Latin America's AI economy lands in four places:

• Global model and cloud providers collect subscription and compute revenue that largely leaves the region. The tools also cost more here: the same request in Spanish or Portuguese typically needs 30–60% more tokens than in English, raising cost and latency.
• The physical supply chain is where the enabler story leaks. Only about a quarter of the copper mined in the region is refined there, and less than 10% of rare earths are processed locally. Data-center announcements are vast, from ByteDance's planned $39 billion in Brazil to a $25 billion, 500-megawatt Stargate letter of intent in Patagonia. But data centers create few permanent jobs, are not the same as compute that local firms can actually use and can shift costs onto other grid users when hyperscalers receive discounted power.
• Data-owning local platforms such as Nubank and Mercado Libre retain the most, because they own the customer, the data, and the decision loop.
• Sector-specific builders, companies that build AI tailored to local industries such as agriculture, mining, credit and utilities, fill the gap left by generic global tools, which only 12% of startups say fit their needs.

The implication: enabler value is real but largely exported, unless hosting terms on power, water, local suppliers, and compute access are negotiated to keep a share at home. Adoption value is domestic by nature. That is why the adoption story matters more for the region itself.

Constraints that keep value shallow

Beyond how firms work, three further limits keep value shallow, even where adoption is enthusiastic. Talent is wide but thin. Demand for AI fluency, the practical ability to use AI rather than build it, grew roughly 11-fold between 2023 and 2025, about twice the US and European pace, mostly in non-STEM roles. But that growth starts from a low base: explicit AI skills still appear in fewer than 3% of job postings, and more than four in ten digitally connected workers cannot write a prompt an AI system can act on.¹¹ Specialized talent, meanwhile, is leaking to more mature markets. Wide fluency lets the region use AI; thin depth keeps value capture shallow.

Computing power is the next bottleneck. Today the main constraint is how firms work; once they fix that, computing capacity becomes the limit. Only a handful of countries have a strong data-center base, and a minority meet international sustainability standards. For a region whose edge is clean electricity, that is an obvious place to invest.

Sovereignty means adaptation, not ownership. Owning every layer of AI, from chips to models, is out of reach and would waste scarce capital. The realistic path is open-weight models fine-tuned on local data, diversified suppliers, and governments acting as anchor clients for local builders. LATAM-GPT, trained in Spanish and Portuguese on regional data, and Maritaca AI's Sabiá models, trained on Brazilian legal texts, show the direction.

A regional race, not a single story

Investors who treat Latin America as a single market will misread where value lies. The region splits into three types of economy: agriculture-led, such as Guatemala and Honduras; manufacturing and supply-chain-integrated, such as Mexico and the Dominican Republic; and service- and knowledge-oriented, such as Chile and Uruguay. Each faces a different AI opportunity.

McKinsey's country dashboards show tight automation potential but divergent trajectories: Mexico offers the largest absolute prize — an estimated ＄204 billion by 2030, a fifth of it in manufacturing — yet its AI-skills demand grows more slowly than peers. Brazil follows at ＄135 billion with faster-rising skills demand. Chile shows the fastest labor-market adaptation — demand for AI-fluency skills is rising more steeply there than anywhere else in the region. There is no single Latin American AI model; countries are building advantages in different layers. 

The size of the prize, and the risks

Latin America's defining economic problem is weak productivity, which grew only about 0.4% a year from 2000 to 2024. AI is the most credible lever it has to change that. Estimates of the gain vary widely because each one measures something different. The broadest, from the WEF, covers the whole economy: effective adoption could lift productivity by 1.9–2.3% a year and add $1.1–1.7 trillion in value annually, roughly 60% of it from analytical AI.¹² McKinsey looks only at automation and puts the gain at $450 billion by 2030, falling to $230 billion if adoption is slower. The most cautious estimate, from OECD simulations for Mexico and Colombia, finds income per person would be only 0.3–2.0% higher after ten years than without AI.

The range shows that the outcome depends on execution. And even the cautious OECD estimate matters: across the whole 2015–25 decade, GDP per person in the region grew just 5.5%.

The risks are concrete: high capital costs and currency volatility, continued talent flight, dependence on foreign model and cloud pricing, and local pushback over data-center water and power use. Further out, cheaper robotics will test labor-intensive jobs: 26.8% of the region's employment is exposed to routine physical automation, versus 20.1% in the United States.

Three indicators will show whether the gap is closing:

1. The share of firms reporting significant value (6% today).
2. The share of adopters moving from experimentation to scale.
3. How much data-center capacity is genuinely accessible to local firms.

What this all adds up to

A century ago, Latin America met the Second Industrial Revolution unable to "learn how to learn," and lost ground for generations. Access to the grid did not make a factory productive; redesigned work did. AI poses the same test. The region has crossed the first threshold: adoption is no longer experimental. The second is harder, and only a small minority are clearing it, because value comes from redesigning operations, not buying tools. The winners will be the firms that rebuild their workflows around AI and turn adoption into measurable gains in revenue, cost, and reach. The same logic applies to countries, and to the capital that backs them.

Latin America was never going to be the fourth corner of the AI triangle, and it does not need to be. It already anchors the physical foundation of the global build-out. Its task is to become exceptionally good at using AI across the industries that anchor its economy, and to negotiate a fairer share of the value it helps create. Whether it converts heavy usage into deep value will decide if it remains a supplier and consumer of the AI revolution or becomes one of its most important beneficiaries.