Brazil leads AI adoption among professionals, but lacks people to put it into production

Three numbers, one shared gap

Between October 6 and 7, three surveys landed in the press almost at the same time and, together, form a clear picture: Brazil is not behind in the AI-at-work race, it is ahead.

Luria recorded a 183% increase in the base of people using its AI platform between January and August 2026, compared to the same period in 2025, with query volume rising 240%. Microsoft's Work Trend Index 2026 placed 27% of Brazilian AI users in the "frontier professionals" category, against an average of 16% across the ten countries surveyed, including the United States, Japan, and India. Fundação Getúlio Vargas (FGV, a Brazilian economic research institution), in a study of 3,000 companies cited by Jornal Nacional (Brazil's main television newscast), measured a jump from 28.7% to 45% in the share of Brazilian companies using AI between 2025 and 2026.

In summary: Brazil does not have a problem of appetite for AI. It has, according to the same three studies, a problem of capacity to sustain that appetite with engineering.

What the studies call an obstacle

The FGV study, presented by Rodolpho Tobler, an economist at FGV Ibre, asks companies directly whether they face any obstacle in implementing AI. Half answered that they face none, a better result than in 2025. But, out of every 100 companies surveyed, 25 pointed to a lack of qualified professionals as the hurdle.

Microsoft arrives at a number that helps explain this gap: organizational factors, such as culture, manager support, and internal policy, account for 67% of the influence on AI's perceived impact, against 32% attributed to individual behavior. In practice, this means that even where the professional already knows how to operate advanced tools, the company often lacks a process, data, or system ready to absorb that use.

The Work Trend Index itself quantifies the size of the bottleneck: only 19% of global users are at the "true frontier," where individual capability and organizational readiness advance together. Another 10% have high knowledge but run into unprepared companies.

Frontier is not a prompt, it's engineering

Microsoft's definition of "frontier professional" deserves the attention of anyone who codes, because it does not describe someone who simply prompts a chatbot well. This category includes people who use AI agents in multi-step tasks, redesign processes around the technology, and create usage flows that other people on the team can repeat.

This is, by definition, engineering work: orchestrating chained calls, versioning prompts and agent configurations, deciding where human review fits in, measuring token cost per task, and building observability to know when the agent errs. Among the Microsoft 365 Copilot users analyzed in the study, 49% of interactions already involve cognitive activity, information analysis, problem solving, evaluation of alternatives, not repetitive automation. This kind of use breaks down without infrastructure behind it.

Here lies the difference between the nice-looking adoption number and adoption that sustains production. McKinsey, in a figure cited by Luria itself in its survey, shows that nearly nine in ten organizations worldwide say they use AI in some function, but only 44% managed to scale it across the entire company, and just 6% reach high performance. In Brazil, Deloitte found that 23% of companies managed to carry forward at least 40% of the pilots they started.

Who has already hit this wall

Jornal Nacional (Brazil's main television newscast) brought two cases that illustrate the problem outside the spreadsheet. At Wise, a company that uses AI to monitor pre-salt wells (Brazil's deep offshore oil fields) at a depth of 7,000 meters, operations director João Vitor Bernardino described the difficulty of hiring people who already arrive ready:

In our experience, it's very difficult to find a professional who meets all the requirements we need by searching the market. Training them in-house is a solution we found, and it has addressed this challenge very well.

João Vitor Bernardino, operations director at Wise

At Mediquimíca, a pharmaceutical company in Juiz de Fora (MG), president Alexandre França described a mapping of more than 3,000 internal processes to identify where AI can reduce errors, implemented gradually, process by process, not as a one-time replacement. The two cases have something in common: neither company went to the market to buy a ready-made professional. Both trained their people in-house.

The counterpoint: isn't this just the pitch of someone selling a solution?

It's worth putting the most interested source in this framing under suspicion. Luria's CEO, Allan Miranda, argues that the real problem is not which model a company chooses, because that changes every six months, but what stays recorded within the company when the model changes or when the person who made the best analyses leaves:

The question that separates companies today is not which AI model they chose, because that changes every six months. It's what stays within the company when the model changes or when the person who does the best analyses moves to another area or even another company.

Allan Miranda, CEO of Luria

It's a correct reading, but it's also the natural pitch of someone who sells precisely a platform to record that institutional memory. The honest counter-argument is that much of the "AI talent shortage" in the Brazilian market is inflated by those who profit from selling training, consulting, or governance software for this bottleneck. FGV's numbers, however, don't come from an AI vendor: they are 25 out of every 100 companies pointing to a lack of qualification as a real obstacle, in a survey that also recorded half of companies with no obstacle at all. The problem exists, it's just not the size that AI course marketing sometimes suggests.

What changes for those who build

For engineering squads, the practical takeaway from the three surveys is not "learn to use AI," Brazil is already doing that faster than the United States, Japan, and India, according to Microsoft itself. It's learning to productize AI: turning one person's individual use of a good prompt into a system that survives that person leaving, with tests, logging, cost control, and a fallback for when the agent errs.

Those who work with data and backend currently have more demand than supply exactly at this junction: it's not the data scientist who trains a model from scratch, nor the advanced chatbot user, it's the engineer who knows how to get a multi-step agent running reliably in production, with observability, and without breaking when the model vendor changes the API.

The cases of Wise and Mediquimíca suggest that, in the absence of this ready-made professional in the market, training in-house has become a strategy, not a plan B. For those deciding where to invest study time or hiring budget in this end of 2026, this is the gap least worth ignoring: it's arriving faster than the labor market can fill on its own.