Could AI reduce inequality within firms?

For decades, automation and globalization appeared to exacerbate inequality within developing countries, particularly affecting low-skilled manufacturing workers whose jobs were relocated to the Far East. The winners were the spreadsheet set — clerical, managerial, and professional workers whose tasks were too complex or too intangible to offshore.

There is some evidence that the AI revolution may redress some of these effects. A new study by Gustavo de Souza (“Artificial Intelligence in the Office and the Factory,” CEPR, 2025) suggests that artificial intelligence could impact the office more significantly than the factory floor. Using uniquely detailed data from Brazil, de Souza finds that AI adoption has increased overall employment and even narrowed inequality among workers. But this has been achieved by replacing routine office work and boosting jobs for low-skilled production workers.

Brazil offers an unusual window into this shift. Since 1987, every commercial software product in the country has been registered with its National Institute of Industrial Property (NIIP), thanks to a copyright law that encourages firms to log their code. That creates a kind of digital census: tens of thousands of software filings covering every industry, every decade.

De Souza identifies all programs that utilize artificial intelligence — from predictive maintenance systems to administrative decision-making tools — and then links them to employment data across various occupations. Unlike patent filings or job postings, the NIIP data track real products actually deployed in factories and offices.

Firstly, this allowed us to acknowledge two very important pieces of evidence. First, the “AI boom” wasn’t confined to Silicon Valley; it had quietly gone global: around 2013 — the dawn of modern machine learning — AI development in Brazil exploded, growing sevenfold by 2022. Second, AI use was not limited to the realm of applications useful to office workers. When firms register their software with the NIIP, they must indicate its intended application. Managerial uses (information management, administration, etc.) account for only 27 percent of AI registrations. Production-related uses — manufacturing, maintenance, agriculture — represent a comparable 40 percent. In short, AI has hit the shop floor as much as the office floor. This suggests that AI isn’t confined to corporate headquarters; it is running conveyor belts, scheduling repairs, and guiding robotic arms. It’s not only learning from data — it’s learning from machines.

To see how these patterns translate into jobs, de Souza calculates each occupation’s exposure to AI: the textual similarity between what a worker does and what AI software is designed to do. As shown in Figure 1, employment in highly exposed occupations has risen faster than in less-exposed ones since 2013.

Figure 1. Employment and AI exposure

Note: 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. De Souza (2025)

The most “AI-exposed” jobs are not disappearing; they are expanding. This is happening very differently in offices and factories. In offices, AI automates tasks once performed by clerical and managerial staff: performance monitoring, maintenance planning, inventory control, putting at risk the “digital middle management.” In factories, by contrast, AI acts as a complement, making machines more productive and easier to operate. A good example is imachine, a predictive maintenance system that analyzes real-time sensor data to detect failures before they occur. It schedules repairs, assists operators, and can cut downtime by 50 percent (Agoro 2025; Benhanifia et al. 2025). The result: more machines running, more output — and more workers needed to run them.

Figure 2. Effect of AI by occupation (De Souza, 2025)

As shown in Figure 2, employment in production jobs increases, while employment in administration decreases. The factory expands as the office contracts.

De Souza finds that a one-standard-deviation increase in AI exposure raises employment by about 2 percent immediately and 7 percent after three years. The beneficiaries are not coders or data scientists, but younger, less educated, and lower-ability workers — the people who suffered most from the last wave of globalization. This also shows its effect on the wage distribution: pay at the top drops by about 2.8 percent, while wages at the bottom remain stable. AI, in other words, is compressing the wage gap rather than widening it, by making middle management less wealthy and reducing inequality between office workers and shop floor employees. Employment growth is most pronounced among machine operators, as AI reduces the skill required for complex tasks. Algorithms replace expertise, allowing less experienced workers to handle sophisticated machinery.

To appreciate how remarkable this reversal is, it helps to recall what happened during the last great transformation of work — the era of globalization.

From the 1980s onward, rich countries outsourced low-skill, labor-intensive production to cheaper regions. Factory jobs vanished, industrial towns hollowed out, and the political consequences still echo through populist movements today. Meanwhile, demand surged for clerical, managerial, and professional labor — the people who could coordinate supply chains, design products, and manage global operations. The world split into makers and managers, and the rewards flowed to the latter.

Now, as de Souza’s work makes clear, AI may be inverting that pattern. The revolution in algorithms is doing to the office what globalization did to the factory: automating the middle. It is the accountants, schedulers, and middle managers — not the machinists — whose roles are being redefined.

The reason is structural. Globalization expanded the scale of production, demanding more coordination and paperwork. AI automates coordination itself. The very layers of management that globalization multiplied are now being made redundant by software that can plan, forecast, and supervise.

In factories, by contrast, AI has become an enabler. Predictive maintenance and computer vision make machines simpler to use; natural-language interfaces allow workers to give commands without special training. Tasks once done only by skilled technicians can now be handled by line operators.

For once, technology seems to be democratizing skill instead of destroying it.

In Brazil, the balance tilted toward inclusion as AI arrived in a context where manufacturing still mattered and low-skill labor was plentiful. In economies where manufacturing has already been hollowed out, the effects may differ.

The study implies that policy can shape the outcome. If AI tools are deployed not just to replace workers but to amplify human capabilities, the result can be more jobs, not fewer. Training programs, open data standards, and investment in digital infrastructure can all widen access to these productivity gains.

The real threat is not that AI will destroy work, but that its benefits will be captured by a few firms or geographies. The lesson from globalization is clear: distribution matters as much as innovation.

For firms, this implies that AI introduction is not the end of labor but its recomposition. Productivity gains will still require human presence — just in different places, doing different things.

Artificial intelligence, like globalization before it, will redraw the social map of work. But it need not do so cruelly. As Figure 6 in de Souza’s paper shows, the gains from AI can accrue to those at the bottom of the wage ladder, not just the top.

That’s a reason for cautious optimism. The future of work may not be one of mass displacement, but of rebalanced opportunity — a world where the machinist, not the manager, gains from the latest wave of innovation.

If the last 40 years were about shipping jobs abroad, the next twenty may be about bringing them back home — with algorithms as the new assembly line.