Daily use,
shallow fluency.

The right question for 2026 is no longer whether AI is changing work in Brazil. It is where it has changed, and where it has not yet arrived. The answers, after eight months of research, are uncomfortable for anyone expecting linear transformation.

The report triangulates in-depth interviews (Brazilian HR managers and directors, C-levels and Stanford MBAs) with secondary sources from the Stanford Digital Economy Lab, McKinsey, Anthropic, PwC, BCG, Economist Impact, INSEAD, OECD and Drake Star. The framing is intentional. The MBAs show the workforce now entering the market; the leaders interviewed show the people recruiting, training and making decisions about that same workforce. What emerges is a country with adoption and anxiety in equal measure, and a management layer split between enthusiasm for pilots and the inertia of existing processes. Twelve numbers describe the landscape. Four findings frame the report.

The state in 12 numbers

The four findings

Every discussion about AI at work contains two parallel stories. The first is the expectation: a tsunami sweeping through offices, mass automation, immediate transformation. The second is the data, and it is more uncomfortable. Across every lens that matters (digital penetration, demand for skills, model capability, occupational exposure), there is a consistent distance between what is promised and what is measured. This chapter examines that distance from four angles.

The Brazilian paradox: fluent on mobile, slow in the enterprise

Brazil's starting point is not technological lag. Quite the opposite: Brazil has one of the populations most accustomed to digital interfaces in the world. According to DataReportal/GWI, 98.4% of Brazilian internet users access the internet by mobile phone, 93.9% used WhatsApp in the past month, and the country spends 53h30 per week consuming online media. INSEAD points in the same direction: more than 85% of Brazilians use the internet regularly, the country leads Latin America in GenAI adoption on LinkedIn with 110 thousand professionals listing GenAI skills, and São Paulo concentrates more than 540 thousand digital and AI professionals.

The paradox appears when this everyday fluency enters the enterprise. Demand for AI skills quadrupled in Brazil over three years, but organizational adoption is still narrow: in the OECD/BCG/INSEAD survey in São Paulo, only 7% of mid-sized and large companies in manufacturing and ICT were using AI; among users, 58% had just one or two applications. In other words: Brazilians already live inside digital interfaces, but Brazilian companies have not yet turned that familiarity into process, governance and productivity.

Demand quadrupled. Total share is still 1.1%.

PwC's AI Jobs Barometer for Brazil shows a country where the curve already exists, even if the base is small: job postings demanding AI skills went from 19 thousand in 2021 to 27 thousand in 2022, 56 thousand in 2023 and 73 thousand in 2024. As a share of the total job market, that means going from around 0.28% to 1.1% in three years.

Sector concentration is part of the story. In 2024, Information and Communication required AI in 4.2% of its postings (more than triple the national average). Professional, Scientific and Technical Activities stood at 1.5%. Finance and Insurance, around 3.5%. Manufacturing, Healthcare and Education remain below the average, despite the narrative suggesting otherwise. Where AI is discussed, it is not yet practiced.

Theoretical capability × Real-world use

This pattern is not uniquely Brazilian. The paper Labor Market Impacts of AI, by Massenkoff and McCrory (Anthropic, March 2026), measures the distance between what AI is theoretically able to do and what it actually does today. The metric combines three sources: the O*NET database of American occupational tasks, the theoretical β-scores from Eloundou et al. (the LLM's technical capability per task) and real-world Claude use in professional contexts. The result is a massive gap across all occupational categories.

In Computer & Math, 94% of tasks are theoretically exposed; only 33% appear in real-world use. In Business & Finance, 94% versus 28%. In Legal, 89% versus 15%. In Engineering, 85% versus 12%. About 30% of American workers have zero coverage in Anthropic's data: cooks, mechanics, bartenders, and other occupations whose raw material is not digital. At the other end, professionals in high-exposure occupations earn 47% more, are nearly four times more likely to hold a graduate degree (17.4% vs. 4.5%) and are 16 points more likely to be women. AI has arrived, but only in a narrow slice of work.

At the top of exposure, but demand is not falling

Anthropic's observed exposure ranking (a metric combining theoretical capability with real-world use) places programmers (74.5%), customer service representatives (70.1%) and data entry operators (67.1%) at the top. Then come medical records specialists, market research analysts, wholesale sales reps, financial analysts and software QA. The composition is informative. These are the professions whose raw material was already digitized. AI did not create that risk; it only activated it.

Software engineering is the most provocative example. Programmers sit at the top of theoretical AI exposure (74.5%), and yet global demand for engineers keeps rising. TrueUp's dashboard, cited by Lenny Rachitsky in March 2026, shows 67,665 open engineering positions at global tech companies, up 78% from the 2024 trough. The curve fell with the launch of ChatGPT, hit its low in early 2024 and has been rising since. In 2026, even after the launch of Opus 4.5 and Claude Code, it is the highest number of postings in three years. Theoretical exposure is not the same thing as role extinction.

There is a wide distance between using AI and operating AI. The interviews in this study show that most Brazilian professionals, including C-level executives, are still firmly in the first group. The reasons are structural.

The most revealing evidence is not in the quantitative research, but in the language. In interviews with HR directors and MBAs, the phrase that comes up most often to describe current use is some variation of "superficial", "basic" or "email drafting". An HR director at a company in the pharmaceutical sector put it bluntly. The use is ChatGPT for common tasks, not for strategic performance improvement. The difference is not semantic. It is about ROI.

The quantitative data confirms the interview pattern. In a survey of 1,466 professionals conducted by Brazilian career outlet The News: Better Work, only 3% of respondents said they were satisfied with the way they use AI at work today. The dissatisfaction is not with the tool. It is with the level of proficiency they have reached on their own.

The AI adoption cycle cannot keep up with the technology

A recurring pattern in the interviews explains the distance between use and fluency. Highly qualified professionals (some with backgrounds in Big Tech, strategy consulting, or executive leadership) describe the same dilemma. They realize they could automate a workflow, but the setup time for the automation seems to exceed the gain. The phrase captured is nearly identical across three separate interviews: "I'll just do it manually".

On top of that is a gray zone between using AI as an individual and using AI as an employee. The same professional who uses ChatGPT naturally outside work often does not know what they can put into a tool inside the company, which data is sensitive, who approves an automation, or which platform is allowed. Extra caution around sensitive data is rational, but it has a cost: procurement and enterprise sales cycles move in months, while the tools change in weeks. The user, however fluent individually, ends up without resources, direction or institutional cover to use AI as a productivity lever.

The mismatch is measurable. The METR Time Horizon measures the duration of software tasks that agents can complete with 50% reliability. In four years, the curve went from seconds (GPT-3.5, March 2022) to hours of equivalent human work. Technical capability doubles every few months; the enterprise cycle (procurement, governance, process change) is still measured in quarters and years. That difference in clock speed is what produces shallow fluency: the technology arrives before the operating model.

In the Brazilian public sector, the same pattern appears as self-teaching at scale. The Public Sector AI Adoption Index 2026 shows that 63% of Brazilian public servants started using AI at work in the past year, almost always on their own initiative; 52% describe their learning as entirely self-taught. That is a cultural strength, but it is also an institutional risk: when the learning curve grows from the bottom up, governance, security and ROI measurement arrive late.

The cognitive cost of thinking like an automator (choosing a tool, connecting APIs, debugging prompts, validating outputs) is real. And it is exactly that cost that separates those who use AI from those who operate AI. What is missing is not technical training. It is a mindset shift around risk and testing. The phrase comes directly from the HR director at a company in the pharmaceutical sector, and it echoes across almost every interview.

What external data confirms

The BCG study with 480 consultants and 44 data scientists is the best experimental confirmation of this distance. When non-technical consultants received GenAI to perform data cleaning in Python, they jumped from 37% of the data scientist benchmark to 86% (a 49 percentage point gain). The result depended on one condition: some level of code experience. "Code experience is a key success factor for workers using GenAI," BCG concludes.

Another implication appears in the speed at which the skills themselves change. PwC global data shows that the skills required in AI-exposed jobs are changing 66% faster than in others, and that rate is 2.5× higher than the previous year's. In Brazil, the effect appears in a specific way. Professionals in the quartile most exposed to AI register a +1.6 net skill change, compared with +0.7 for the least exposed. Put simply: those close to AI are relearning constantly. Those far from it stay frozen.

There is a direct analogy with digital literacy in the 1990s. Back then, knowing how to use email and Word was the minimum bar to sit at the table, not a differentiator. Knowing how to use AI is becoming that condition. The difference is speed. Digital literacy had a decade to spread. AI fluency has two or three years.

To move beyond the macro trend, the report interviewed HR directors in three segments of the Brazilian market. The portraits that follow show real initiatives, with scale, partnerships and metrics. Names were removed; the substance was preserved.

Click on each portrait to expand the detail. Together, the three form a spectrum that runs from the scale operator in transformation to the regulatory fortress.

There is a recurring narrative about AI and work: "it is going to eliminate jobs." The latest data show that is not what is happening, at least not yet. What is happening is subtler, and probably more serious. AI is absorbing the tasks that historically trained junior professionals.

A first piece of evidence comes from Anthropic Nowcasting in March 2026. Cross-referencing data from the American Current Population Survey with the exposure metric built from real-world Claude use, the authors reach a dual conclusion. Unemployment in exposed occupations has not risen systematically since the launch of ChatGPT. The aggregate effect is statistically indistinguishable from zero. But there is a significant hidden effect. Hiring of young workers aged 22 to 25 in exposed occupations fell by around 14% after 2022.

The "canaries in the coal mine" metaphor comes from another reference: the paper Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence, by Erik Brynjolfsson, Bharat Chandar and Ruyu Chen (November 2025), supported by the Stanford Digital Economy Lab. Using ADP payroll administrative data, the authors show a relative drop of 16% in employment of workers aged 22 to 25 in AI-exposed occupations, after controlling for firm-level shocks, while employment of more experienced professionals remained stable. For software developers aged 22 to 25, the drop was nearly 20% from the late-2022 peak. The signal is not mass layoffs. It is that the first rung is no longer bringing people in.

The first equivalent signal in Brazilian primary data came in 2025. In an independent analysis by Elisa Pereira published in October, the occupations most exposed to AI in Brazil registered an average decline of 5.4% in hiring in the first three quarters of 2025, relative to what was projected by the 2023 to 2024 historical average adjusted for the growth of low-exposure occupations. The effect is heterogeneous. Collections fell 21.6% (about 10 thousand fewer hires), administrative support fell 5.4% (~29 thousand), and accounting assistants dropped 3.5%. On the other hand, HR analysts grew 8.5% and active telemarketing 8.3%. It is the first national signal that the entry door has begun to close here too, in specific occupations, before appearing in the aggregate.

Economist Impact, in the report The AI Glass Floor, describes the dynamic in one sentence: "AI may not eliminate jobs, but it can block the door of entry into them". The intuition is direct. The tasks a junior analyst used to do to learn the trade (building spreadsheets, doing background research, writing first drafts, processing simple tickets) are precisely the most automatable. Without those tasks, there is no learning ramp. Without a ramp, there is no hiring. Without hiring, there is no next generation of senior professionals.

Why Brazil is at greater risk

Three structural factors make the ladder without a first rung sharper in Brazil than in the US. First, demographics: Brazil is already in transition (median age of 35 according to the 2022 IBGE Census), but still concentrates around 35 to 40 million people between 18 and 29 years old. Proportionally, that age group weighs more here than in the US: around 17 to 20% of the Brazilian population versus ~15% of the American population. Every percentage point of decline in the job-finding rate applied to this cohort equals hundreds of thousands of young people fewer entering the formal market.

Second, informality. The floor was already fragile. Professionals who historically entered the market through formal internships and migrated to CLT employment now compete with platform freelancers and self-employed professionals who use AI to scale. The result is margin compression and reduced need for juniors within formal structures.

Third, university. As heard in the interview with the HR director at a company in the healthcare sector: "universities are not teaching AI in the curriculum, and companies are not training what makes a good analyst in the AI era". The Brazilian education system is three to five years behind in practical training. Graduates in applied social sciences are entering the market without operational AI fluency, and losing out to peers who learned in a bootcamp or through self-teaching.

The combination that hits hard

The data connect in an uncomfortable way. In Brazil, demand for AI skills quadrupled; but the most exposed occupations are growing less in number of postings (264% vs. 291% for the least exposed) and require 135% more skill adaptation. Abroad, the wage premium for those who master AI more than doubled in a year (from 25% to 56%). And Anthropic shows that workers in the most exposed quartile earn, on average, 47% more than those with zero exposure, are 4× more likely to hold a graduate degree, and are less unionized.

Put simply: those who master AI are being well paid and retained. Those who do not yet have a door through which to learn will not get in.

The pipeline that eats itself

The effect is not a sequence of isolated mistakes. It is a single composite dynamic, looping back on itself. Seven steps reinforce one another.

1. Senior professionals need fewer and fewer juniors for operational and routine tasks. AI absorbs that layer of work.
2. Companies stop hiring juniors. Entry-level postings in exposed occupations fall by around 15%.
3. Juniors don't accumulate practical experience. They don't become seniors.
4. In five to ten years, current seniors retire, change companies, or become consultants.
5. There is no new generation ready to replace them, and the technology keeps advancing.
6. Demand for seniors explodes. The AI wage premium jumps from 25% to 56% in a year.
7. Back to step 1. But worse.

Cycle modeled from data from Anthropic, PwC and Economist Impact

Consulting reports like to present AI as a one-way vector. The interviews show something else. The terrain is full of tensions, and each one is both an obstacle and an opportunity. Mapping these paradoxes is more useful than listing best practices.

The opportunities below show where AI can become operational capability. The challenges show the locks preventing that capability from scaling.

The conclusion of this report is not that Brazilian companies need to "use more AI." That is already happening, even if unevenly. The conclusion is more uncomfortable: individual use has advanced faster than the institutional capacity to guide, measure and absorb that use.

The problem, therefore, is not choosing between enthusiasm and caution. It is designing the bridge between the two. In large companies, that bridge runs through three concrete decisions.

First: name the operational owner. AI cannot keep circulating between HR, IT, Compliance and business units as everyone's topic and no one's responsibility. The company needs someone with a mandate to define guardrails, prioritize workflows, measure results and resolve the conflict between risk and speed.

Second: pick a few real workflows. Adoption is not proven by active licenses, training completed or number of prompts. It is proven when an important process becomes faster, cheaper, more reliable or more auditable. Three well-chosen workflows say more about maturity than fifty scattered pilots.

Third: start from the business objective. AI adoption is never an end goal. It is a means to reduce cycle time, increase conversion, decrease error, accelerate onboarding, improve quality or free up team capacity. If a workflow cannot say which business metric it wants to move, it is still a tech experiment, not an operational strategy.

This report is more useful when its construction is clear: which sources were cross-referenced, what type of inference was made and what is out of scope. This section documents the research in six layers, from framing to acknowledged limitations.

Framing and period

Research conducted over eight months, closing in May 2026. The focus is the intersection between the Brazilian labor market and the global AI movement, with a dual lens: the workforce entering the market (Stanford MBAs, recent graduates, career transitions) and the leaders who recruit, train and make decisions about that workforce (HR directors, C-levels, board members). The framing is intentional: it avoids both the bias of research only with decision-makers and the bias of research only with talent.

Primary sources

More than 50 hours distributed across in-depth interviews with HR managers and directors, C-levels and Stanford MBAs; unstructured conversations with talent and ecosystem operators; observations and discussions in MBA program classes; and analysis of internal materials gathered throughout the research. Three Brazilian company portraits were built from formal interviews with the HR leadership of each company, with subsequent confirmation of numbers and initiatives against the original transcript.

Secondary sources

Review of reports, academic papers, public datasets, specialized newsletters and external materials on AI, work, enterprise adoption, HR Tech, productivity, education and the Brazilian market. The main references cited in the report:

Triangulation

Most of the quantitative data on occupational exposure, hiring contraction and technical capability comes from American research (Anthropic, Stanford, METR, BCG/BU/OpenAI). Brazil does not yet have equivalent series at the same scale of academic production. The report's approach was to cross those findings with Brazilian data whenever available (PwC AI Jobs Barometer Brazil, OECD/BCG/INSEAD São Paulo, INSEAD Knowledge Brazil, Pereira's CAGED analysis) and with qualitative evidence from interviews to verify whether the pattern observed abroad was reproduced locally. When Brazilian data did not exist, the international source is presented as directional evidence, not as local measurement.

Data treatment and attribution

All quantification was referenced to the original source. When a chart in the report visually reproduces a figure from a paper whose authors did not publish the underlying numerical table (the case of Figure 2 from Massenkoff & McCrory, chapter 02), the values were anchored on the explicit points cited in the body of the paper, with the remainder estimated by visual inspection. The chart source notes that limitation. The charts about youth employment in exposed occupations (chapter 05) are stylized illustrations of the relative magnitude reported by the papers, not original datapoints; this is stated in the chart source.

Acknowledged limitations

Three limitations the reader should keep in mind. First: the three company portraits are intentional and narrative, not a statistically representative sample of the Brazilian market; they serve to illuminate patterns, not to quantify them. Second: no proprietary sample survey was conducted, which limits the report's ability to produce new Brazilian estimates; the work depends on third-party surveys (PwC, OECD/BCG/INSEAD, The News: Better Work) for quantitative data on Brazil. Third: the speed of the AI market makes any conclusion sensitive to the time window. The data in this report reflects the state of the art through May 2026; periodic revisions will be necessary.