What if artificial intelligence isn’t coming for your job, but perhaps offering a career change instead? In Brazil, that’s exactly what new data suggest. Drawing on a national registry of nearly every commercial AI program developed since the 1980s, in a recent paper (de Souza 2025) I find that artificial intelligence is used not just in the office, but on the factory floor as well. There, it optimises processes, runs quality control, and guides equipment operation. Among office workers, AI reduces employment and wages, particularly for middle-wage earners. But in production, it increases employment of low-skilled and young workers operating machinery. These results suggest that AI displaces routine office tasks while making machines more productive and easier to operate, leading to a net increase in employment.
To see how AI can increase factory jobs while automating office work, consider a concrete example. imachine, a predictive maintenance system developed by a Brazilian tech firm, analyses real-time data from sensors embedded in factory equipment to detect failures before they occur, schedule repairs, and assist with machine operation. These tools can cut unplanned downtime by as much as 50% and make complex machinery easier to operate (Agoro 2025, Benhanifia et al. 2025). That keeps machines running longer and increase demand for workers to operate them. However, with the same breath that imachine raises productivity on the factory floor, it automates tasks in the office. Performance analysis, maintenance planning, and inventory control are now done by imachine rather than humans. This case captures the broader pattern in the data: AI increases employment in low-skilled production roles by making machines easier to operate and more productive, while reducing demand for routine office work through automation.
The administrative software registry
The reason why I can tell vivid stories about AI development in Brazil is that most commercial AI applications are part of an administrative software registry. Since 1987, Brazilian firms have registered nearly every commercial software they create with the National Institute of Industrial Property (NIIP), thanks to a law granting copyright protection to registered code. Registering commercial software is a standard industry practice: 96% of software development firms have at least one program registered. Among firms with more than 20 workers, 99.9% of them have registered at least one software.
I collected data on all AI-related software ever registered with the NIIP, containing de- tailed information on ownership, programmers, intended use of the software, and technical features. Compared to patents or job ads, which are common metrics of AI adoption in the literature, the NIIP registry has the advantage of containing detailed technical and application information on actual software products, including software developed by specialised IT companies and adopted elsewhere.
The AI boom
The NIIP dataset offers two key insights into AI development. First, there was a marked boom in AI development around 2013, following global breakthroughs in machine learning. Figure 1 shows this surge by plotting the number of unique firms with at least one AI software. Importantly, about 30% of these firms are technology providers that build AI tools for multiple external clients. In some cases, the same software may be deployed across dozens or even thousands of firms. As a result, the number of developers understates the true scale of AI adoption. Therefore, I consider Figure 1 as illustrative of the trend in AI adoption but not of its levels.
Starting around 2013, the development of AI software accelerated rapidly. The number of firms owning AI technologies increased sevenfold, reaching 1,434 by 2022. This period, often referred to as the AI boom, is marked by a surge in innovation and investment in artificial intelligence (Toosi et al. 2021, Chauvet 2018, Sevilla et al. 2022, Bughin 2017).
Figure 1 Number of firms owning AI software
AI is used in management and production
The second fact coming out of the NIIP dataset is that AI is used in both management and production. When firms register their software, they indicate its application domain, i.e. a list of areas where the software is intended to be used. Figure 2 presents the ten most common broad application domains for AI software. Managerial uses, such as information management and administration, are the most frequent but account for only 27% of registrations. AI is also widely applied in fields like healthcare, manufacturing, and agriculture, which together represent 16.7% of the total.
Figure 2b groups all application domains into three categories: management, production, and academic. AI registrations are nearly evenly distributed between management and production applications, each accounting for about 40% of filings. This distributional balance challenges the prevailing view that AI primarily targets high-skill, white-collar occupations (Frey and Osborne 2017, Webb 2020, Felten et al. 2021). With a comparable share of AI tools applied in production, blue-collar workers may be just as exposed to AI as administrative workers.
Figure 2 AI software registrations by intended use
AI exposure correlates with employment growth
To measure the exposure of each occupation to AI, I calculate the text similarity between the tasks workers perform and the descriptions of AI software. This measure captures the overlap between AI applications and worker’s tasks, either due to task replacement or complementarity, and evolves over time according to the number of AI software developed in Brazil.
Figure 3 plots the employment share of occupations in the top and bottom deciles of the 2022 AI exposure distribution, normalising both series to 1 in 2003. From 2003 to 2012, the two employment trajectories moved in parallel. Starting in the first year of the AI boom (2013), when the development of AI software in Brazil exploded, employment in high-exposure occupations began to rise relative to low-exposure ones. By 2022, employment share in the most exposed occupations increased by 20%.
Figure 3 Occupations more exposed to AI grew faster after AI boom
Notes: Figure plots the log employment of occupations in the top and bottom deciles of the 2022 AI-exposure distribution, normalizing both series to 1 in 2003.
Instrument: AI ease of development
To identify the causal effect of AI on employment, I construct an instrument that exploits differences in how easy it is to build AI software for different occupations over time. The key idea is that AI software becomes cheaper to develop when certain programming languages gain popularity. But this cost decline doesn’t affect all types of AI equally because each programming language is better suited for certain applications. For example, COBOL, a language commonly used in banking systems due to legacy code, has declined in popularity, making it harder to find programmers and resources, while R, used for statistical analysis, has grown rapidly and benefits from abundant support. As a result, it is now relatively more costly to build AI tools for banking automation than for statistical tasks. Therefore, bank tellers are relatively less exposed to AI when compared to archivists because it is relatively more costly to build banking automation software, which requires COBOL, a dying programming language. The instrument builds on this intuition to construct an AI easy-of-development shifter for each occupation over-time.
Figure 4 Illustration of the instrument
Notes: This figure illustrates how the instrument is constructed.
AI increases employment of low-skilled workers
I find that AI leads to a net increase in employment, primarily by expanding job opportunities for low-skilled workers. Figure 5 show that occupations more exposed to AI experience higher employment growth, with a one standard deviation increase in AI exposure raising employment by about 2% in the current period, and by as much as 7% after three years. Importantly, this growth is not evenly distributed: AI disproportionately boosts employment among younger, less educated, less experienced, and lower ability workers.
Figure 5 Effect of AI on employment
Figure 6 shows the effect of AI in different deciles of the wage distribution. AI reduces inequality by lowering wages at the top of the wage distribution. A one standard deviation increase in AI exposure has no significant effect at wages in the lowest decile, while wages in the 9th decile fall by about 2.8%.
Figure 6 AI decreases wages at the top of the wage distribution
The finding that AI rises employment among low-skilled workers and decreases wages at the top suggests that it acts as a skill-replacing technology. By substituting for expertise, it allows lower-skilled workers to perform tasks that once required significant experience. This shift reduces the barriers to entry in high-exposed occupations, increased the hiring of lower- skilled workers, and erodes the wage premium for high-skilled individuals. This conclusion is supported by multiple micro-level experiments but has been shown to hold in scale only now (Kanazawa et al. 2022, Brynjolfsson et al. 2025, Gruber et al. 2020, Choi et al. 2023, Dell’Acqua et al. 2023, Noy and Zhang 2023, Peng et al. 2023).
AI shrinks the office and expands the factory
AI has sharply contrasting effects on the office and the factory. As shown in Figure 7, it significantly increases employment in production-related occupations, such as manufacturing, maintenance, and agriculture, while reducing employment in administrative jobs. The expansion in factory employment is driven by a shift toward low-skilled workers: AI enables younger, less educated, and less experienced individuals to take on tasks that previously required more training. In contrast, the decline in administrative employment is not accompanied by any change in worker composition. These findings indicate that AI acts as a substitute for labour in routine office tasks but as a complement to low-skilled labour in production settings.
Figure 7 Effect of AI on employment for different occupations
AI increases employment among machine operators
Moreover, AI increases employment among machine operators and decreases inequality across occupations. Figure 8 shows heterogeneity in the effect of AI across different occupations. AI increases employment most strongly in occupations involving machine operation, where it leads to an influx of younger, less educated, and less experienced workers.
Figure 8 AI has larger employment effects in machine-operating jobs
In my paper, I also show that AI reduces wages more in occupations that initially had higher average wages and education levels. These results suggest that AI lowers barriers to entry and allows less qualified workers to take on roles once reserved for specialists.
Conclusion: AI increases employment and decreases inequality
These results are consistent with AI affecting the labour market in two distinct ways. In the factory, AI increases employment of low-skilled workers by making machines more productive and easier to operate. In the office, however, it automates tasks previously done by workers. Because the effect on production workers dominates, AI increases employment and decreases inequality – a far more positive outcome than the Terminator-like conjecture that many make nowadays.
Authors’ note: This column represents my opinions and not those of the Federal Reserve Bank of Chicago or the Federal Reserve System.
References
Agoro, H (2025), “Reducing Downtime in Production Lines Through Proactive Maintenance Strategies”.
Benhanifia, A, Z B Cheikh, P M Oliveira, A Valente, and J Lima (2025), “Systematic review of predictive maintenance practices in the manufacturing sector,” Intelligent Systems with Applications 26, 200501.
Brynjolfsson, E, D Li, and L Raymond (2025), “Generative AI at Work”, The Quarterly Journal of Economics 140: 889–942.
Bughin, J (2017), “The new spring of artificial intelligence: A few early economies”, VoxEU.org, 21 August.
Chauvet, J-M (2018), “The 30-Year Cycle In The AI Debate”.
Choi, J H, D Schwarcz, and K E Yeh (2023), “AI Assistance in Legal Analysis: An Empirical Study”, Legal Studies Research Paper 23-22, University of Minnesota Law School.
Dell’Acqua, F, E McFowland III, E Mollick et al. (2023), “Nav- igating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality”, Working Paper 24-013, Harvard Business School.
De Souza, G (2025), “Artificial Intelligence in the Office and the Factory: Evidence from Administrative Software Registry Data”, Federal Reserve bank of Chicago Working Paper 2025-11.
Felten, E, M Raj, and R Seamans (2021), “Occupational, industry, and geographic exposure to artificial intelligence: A novel dataset and its potential uses”, Strategic Management Journal 42: 2195–2217.
Frey, C B and M A Osborne (2017), “The future of employment: How susceptible are jobs to computerisation?”, Technological Forecasting and Social Change 114: 254–280.
Gruber, J, B R Handel, S H Kina, and J T Kolstad (2020), “Managing Intelligence: Skilled Experts and AI in Markets for Complex Products”, NBER Working Papers 27038.
Kanazawa, K, D Kawaguchi, H Shigeoka, and Y Watanabe (2022): “AI, Skill, and Productivity: The Case of Taxi Drivers”, CIRJE F-Series No, CIRJE-F-1202, CIRJE, University of Tokyo.
Noy, S and W Zhang (2023), “Experimental evidence on the productivity effects of generative artificial intelligence”, Science 381: 187–192.
Peng, S, E Kalliamvakou, P Cihon, and M Demirer (2023), “The Impact of AI on Developer Productivity: Evidence from GitHub Copilot”.
Sevilla, J, L Heim, A Ho, T Besiroglu, M Hobbhahn, and P Villalobos (2022): “Compute Trends Across Three Eras of Machine Learning”, in Proceedings of the 2022 International Joint Conference on Neural Networks, pp. 1–8.
Toosi, A, A G Bottino, B Saboury, E Siegel, and A Rahmim (2021), “A Brief History of AI: How to Prevent Another Winter (A Critical Review)”, PET Clinics 16: 449–469.
Webb, M (2020), “The Impact of Artificial Intelligence on the Labor Market”.