OIT - The Aggregation Paradox of AI

 ILO Brief 1

The Aggregation Paradox of AI: Why do micro-economic productivity gains from AI disappear at scale

 Research Brief April 2026

The Aggregation Paradox of AI: Why do micro-economic productivity gains from AI disappear at scale1

 Artificial Intelligence (AI) delivers large productivity gains at the task level (typically 10-70 per cent), with the strongest effects for less experienced workers and well‑defined, text‑intensive tasks.  At the firm level, evidence is more mixed and AI adoption remains uneven; productivity gains are concentrated in larger, digitally advanced enterprises, while many firms report little measurable impact beyond pilots.  At sectoral and macroeconomic levels, no clear AI‑driven productivity growth has yet appeared in official statistics, consistent with historical patterns of slow diffusion and delayed productivity gains (the “productivity J‑curve”) as well as persistent measurement gaps.  Translating micro‑level gains into aggregate productivity growth depends on broad diffusion, complementary investment in workplace re-organisation and skills, supportive macroeconomic conditions and effective competition policy.  Collective bargaining and social dialogue can shape AI adoption, working conditions and the distribution of productivity gains, including through provisions on transparency, training rights, work organization and data protection.  Historical experience with electrification and information and communications technology (ICT) shows that technological revolutions only raised aggregate productivity after substantial organizational change and institutional adaptation. AI is likely to follow a similar path, though with broader reach into cognitive and service‑sector tasks.  Without targeted policy interventions to improve on skills, digital infrastructure, social protection, competition and collective bargaining frameworks, AI risks widening productivity and income gaps across firms, workers and countries rather than closing them.

cognitive and service‑sector tasks.  Without targeted policy interventions to improve on skills, digital infrastructure, social protection, competition and collective bargaining frameworks, AI risks widening productivity and income gaps across firms, workers and countries rather than closing them.

1 The brief was prepared by Cheuk Yu Cheryl Chan (ENTERPRISE) and Khatia Shedania (RESEARCH) and reviewed by Caroline Fredrickson, Ekkehard Ernst (RESEARCH) and Dragan Radic (ENTERPRISE). Key points ILO Brief 2

The Aggregation Paradox of AI: Why do micro-economic productivity gains from AI disappear at scale  Introduction The rapid diffusion of artificial intelligence (AI) technologies, in particular the development of generative AI since late 2022, has reignited debates about the relationship between technological change and productivity growth. This comes against a backdrop of secular productivity stagnation across advanced economies: annual labour productivity growth in OECD countries averaged just 0.8 per cent during 2010–2019, roughly half the rate observed in the preceding decade

(ILO, 2023; OECD, 2024).

Prominent forecasts have attributed transformative economic potential to AI, yet they span more than an order of magnitude. Goldman Sachs (2023) projected that generative AI could raise global GDP by 7 per cent over a ten-year period. McKinsey Global Institute (2023) estimated annual productivity gains of 0.1–0.6 percentage points through 2040, while Acemoglu (2025) offered a more cautious estimate of roughly 0.5 per cent of total GDP over the next decade. This research brief surveys the empirical evidence on AI’s productivity impact, organised by level of aggregation: individual workers, firms, sectors, and the macroeconomy. It then explores the mechanisms through which micro-level productivity gains potentially translate into gains at higher levels and identifies the macroeconomic and institutional factors that would affect this aggregation process. This

productivity impact, organised by level of aggregation: individual workers, firms, sectors, and the macroeconomy. It then explores the mechanisms through which micro-level productivity gains potentially translate into gains at higher levels and identifies the macroeconomic and institutional factors that would affect this aggregation process. This brief also highlights lessons drawn from earlier generalpurpose technologies, including electrification and information and communications technology (ICT), and underscores the role of labour market institutions, including collective bargaining, in shaping AI diffusion and the distribution of productivity gains. Finally, it outlines key gaps in the existing literature on AI and productivity that warrant future research.  Empirical Evidence by Level of Aggregation Individual Worker Productivity The most robust and consistent evidence of AI-driven productivity gains comes from studies of individual workers using AI tools in controlled or quasi-experimental settings. A growing body of randomised controlled trials (RCTs) and natural experiments shows substantial effects on task-level output (ILO, forthcoming). Software development Peng et al. (2023) conducted an RCT with 95 professional developers and found that those assigned GitHub Copilot completed tasks 55.8 per cent faster than the control group. Similarly, a large-scale field experiment at Microsoft (Cui et al., 2024) involving thousands of developers found more modest but still substantial gains: AI-assisted developers completed about 26 per cent more pull requests per unit time, with the largest effects concentrated among less experienced programmers. Customer service Brynjolfsson, Li, and Raymond (2023) studied 5,179 customer support agents at a large software firm and found that access to a generative AI assistant increased worker productivity by 14 per cent on average, as measured by issues resolved per hour. Crucially, the distribution was highly uneven: novice and low-skilled workers experienced productivity gains of up to 35 per cent, while the most experienced agents showed negligible improvements. The points to the equalizing effect of AI by

measured by issues resolved per hour. Crucially, the distribution was highly uneven: novice and low-skilled workers experienced productivity gains of up to 35 per cent, while the most experienced agents showed negligible improvements. The points to the equalizing effect of AI by compressing the skill distribution among workers. Professional writing and knowledge work Noy and Zhang (2023) ran an experiment with 453 collegeeducated professionals (marketers, HR specialists, consultants) performing writing tasks. ChatGPT access reduced average completion time by 40 per cent and improved output quality as rated by blind evaluators. Another study by Dell’Acqua et al. (2023) on 758 Boston Consulting Group consultants found that those using GPT4 completed 12.2 per cent more tasks, 25.1 per cent faster, and with over 40 per cent higher quality. However, for tasks requiring integration of information that the AI tool could not directly access, AI users performed worse than the ILO Brief 3

The Aggregation Paradox of AI: Why do micro-economic productivity gains from AI disappear at scale control group, a phenomenon the authors labelled “falling inside the jagged frontier” (i.e., the boundary of what AI can and cannot yet do reliably). The extent to which such individual level gains persist in real workplaces depends in part on work organization, monitoring practices and negotiated expectations around workload and performance, which are increasingly the subject of emerging collective bargaining clauses on algorithmic management and the right to disconnect.

Scientific research Fang et al. (2024) documented that materials scientists using an AI-powered literature search and synthesis tool generated 44 per cent more novel material candidates in a fixed time window. However, while there is broad adoption of large language models for drafting, editing, and literature review in academic settings, the evidence of the productivity effects of AI in research remain scarce (Biswas, 2023). The individual-level evidence presented above consistently shows large, positive effects on task-specific throughput,

of large language models for drafting, editing, and literature review in academic settings, the evidence of the productivity effects of AI in research remain scarce (Biswas, 2023). The individual-level evidence presented above consistently shows large, positive effects on task-specific throughput, typically in the range of 10–70 per cent. Gains are largest for less experienced workers and for well-defined, textintensive tasks. Effects on complex, multi-step professional judgment are more ambiguous and occasionally negative. Firm-Level Productivity Firm-level evidence is growing but more heterogeneous than that at the individual-level. It typically draws on enterprise surveys, administrative data, and case studies, and may contend with selection effects, as firms that adopt AI are likely to be more productive or innovative to begin with. Survey data suggest that adoption remains in the single digits for the average firm, with strong gradients by sector and firm size. For instance, the US Census Bureau’s Business Trends and Outlook Survey (BTOS) indicates that approximately 5–6 per cent of US firms reported using AI in production as of late 2024, concentrated in information technology, professional services, and finance (McElheran et al., 2024). The European Commission’s Digital Economy and Society Index (DESI) similarly report single-digit adoption rates among EU firms, with a steep gradient by firm size. More recent evidence suggests, however, that uptake might have accelerated rapidly, with some surveys suggesting that almost 70 per cent of companies are using some form of AI, at least among advanced economies (Yotzov et al., 2026). Among adopters, survey evidence from McKinsey (2024) suggests that roughly 60 per cent report measurable cost reductions or revenue improvements, but the magnitude and attribution remain difficult to verify. Econometric studies Econometric analyses exploiting administrative and patent data provide more systematic evidence on the relationship between AI adoption and firm performance. Czarnitzki, Fernández, and Wastyn (2023) analysed European patent

difficult to verify. Econometric studies Econometric analyses exploiting administrative and patent data provide more systematic evidence on the relationship between AI adoption and firm performance. Czarnitzki, Fernández, and Wastyn (2023) analysed European patent data and found that AI-patenting firms showed higher total factor productivity (TFP) growth than matched non-AI firms. By contrast, Alderucci et al. (2020) studied US firms with AIrelated patents and found positive associations with sales and employment but weaker links to productivity per se. Similarly, Babina et al. (2024) used job posting data to identify AI-investing firms and found that these firms subsequently experienced higher revenue growth and higher market valuations but did not find statistically significant effects on measured labour productivity (revenue per worker). Global firm-level evidence Global firm‑level evidence combining survey indicators with harmonised accounts provides a complementary perspective. A recent EIB working paper combines surveybased AI adoption indicators with harmonised enterpriselevel accounts for more than 12,000 enterprises across European countries and the United States. It finds that AI adoption raises labour productivity by approximately four per cent on average, without short-run job losses and with some evidence of higher wages in adopting firms. However, these gains tend to be concentrated in medium and large enterprises, whereas adoption rates are much lower among small firms (Aldasoro et al., 2026a)). A similar study in Germany using ZEW data also finds a positive and statistically significant association between AI use and firmlevel productivity, again with higher adoption and stronger effects among larger and digitally advanced firms (Czarnitzki et al., 2023). Recent international evidence based on harmonised business surveys provides an additional lens on diffusion and early outcomes. Yotzov et al. (2026) use firm-level survey data from almost 6,000 senior executives in more than 20 countries. They find that while a majority of larger firms report some form of AI

based on harmonised business surveys provides an additional lens on diffusion and early outcomes. Yotzov et al. (2026) use firm-level survey data from almost 6,000 senior executives in more than 20 countries. They find that while a majority of larger firms report some form of AI adoption, a substantial share (around four-fifths in their sample) report no measurable productivity gains, frequently citing integration challenges, lack of complementary skills, and organisational bottlenecks as key obstacles. Firms that do report positive impacts ILO Brief 4

The Aggregation Paradox of AI: Why do micro-economic productivity gains from AI disappear at scale associate AI primarily with efficiency improvements, cost reduction, and improved decision-making. Many firms anticipate greater effects over the next three years, with average expected productivity gains of around 1.4 per cent.

Cross-country comparisons reveal that AI adoption and associated productivity gains are systematically higher in digitally advanced economies with stronger complementary capabilities. This suggests that AI productivity outcomes are not solely technology-driven but are embedded in broader national ecosystems of innovation and institutional capacity. Collective bargaining and worker representation may also influence which firms are able to translate AI adoption into sustained productivity gains. Evidence from European workplaces suggests that when AI is adopted in firms with worker representation, the introduction of technologies with AI components is more often associated with better working conditions and provisions for training and consultation than in firms without such structures. These institutional arrangements can reduce implementation frictions and support the complementary investments needed for productivity improvements, particularly in larger enterprises where AI projects are more complex. Overall, firm-level evidence suggests that AI adoption is primarily associated with improved financial performance and, in some cases, higher productivity growth, but causal identification remains challenging. Many firms report difficulties in translating pilot-stage AI tools into enterprisewide productivity improvements. Gains are concentrated in larger firms and vary significantly across countries and institutional contexts. Sectoral Productivity

and, in some cases, higher productivity growth, but causal identification remains challenging. Many firms report difficulties in translating pilot-stage AI tools into enterprisewide productivity improvements. Gains are concentrated in larger firms and vary significantly across countries and institutional contexts. Sectoral Productivity Sectoral productivity analyses attempt to trace AI’s impact through industry-level data. To date, the evidence is limited and largely indirect. Acemoglu et al. (2022) constructed an AI exposure index across US industries and examined subsequent productivity dynamics. They found that industries with higher AI exposure did not exhibit systematically faster productivity growth, though the analysis period (2010– 2018) preceded the generative AI wave. Filippucci et al. (2024a), using OECD data, found a positive correlation between digital technology adoption intensity and TFP growth at the industry level across European economies, but the contribution of AI specifically was difficult to disentangle from broader digitalisation. In specific sectors, more tangible evidence exists. In healthcare, studies have shown that AI diagnostic tools can match or exceed specialist accuracy in radiology and pathology (Topol, 2019), though system-wide productivity effects depend on integration into clinical workflows, regulatory frameworks, and reimbursement structures. In financial services, AI-driven automation of compliance, fraud detection, and trading has expanded, but aggregate productivity growth in the sector has not been captured in official statistics (Bessen, 2022). Macroeconomic Productivity At the macroeconomic level, there is no clear empirical evidence that AI has contributed to productivity growth. Labour productivity growth in the United States recovered somewhat during 2023–2025 relative to the immediate post-pandemic period, but attribution to AI remains speculative and cannot be separated from cyclical factors and pandemic-related compositional effects (Syverson, 2024). Similarly, other advanced economies, including most European countries and Japan have not exhibited any discernible change in productivity trends that might be

post-pandemic period, but attribution to AI remains speculative and cannot be separated from cyclical factors and pandemic-related compositional effects (Syverson, 2024). Similarly, other advanced economies, including most European countries and Japan have not exhibited any discernible change in productivity trends that might be attributed to AI adoption. The Solow Paradox, referring to Robert Solow’s 1987 observation that “you can see the computer age everywhere but in the productivity statistics”, remains an apt characterisation of the current situation. Brynjolfsson, Rock, and Syverson (2021) argued that this pattern is precisely what one would expect during the early diffusion phase of a general-purpose technology: the “productivity J-curve” hypothesis posits that measured productivity may initially stagnate or even decline as firms invest in complementary intangible capital before eventually seizing productivity gains. A further measurement challenge arises from the nature of AI’s output. Quality improvements in services, new product varieties, and consumer surplus from free digital tools are poorly captured by conventional GDP and productivity statistics (Aghion et al., 2019). If AI primarily raises the quality of services rather than their measured quantity, official statistics will understate true productivity gains. ILO Brief 5

The Aggregation Paradox of AI: Why do micro-economic productivity gains from AI disappear at scale

 Table 1: Summary of Empirical Evidence on AI and Productivity

Level Evidence Strength Key Sources Main Findings Individual worker Strong — RCTs, natural experiments (50+ studies) Peng et al. (2023); Brynjolfsson et al. (2023); Noy & Zhang (2023); Dell’Acqua et al. (2023) 10–70% task-level throughput gains; largest for novices and routine text tasks; ambiguous for complex judgment Firm Moderate — surveys, matched employeremployee data, patent/job posting

Dell’Acqua et al. (2023) 10–70% task-level throughput gains; largest for novices and routine text tasks; ambiguous for complex judgment Firm Moderate — surveys, matched employeremployee data, patent/job posting analysis Babina et al. (2024); Czarnitzki et al. (2023); EIB (2026); Yotzov et al. (2026) Positive association with revenue growth and TFP; causal identification difficult; many firms stuck in pilot stage Sector Weak — industry-level panel data, AI exposure indices Acemoglu et al. (2022); Filippucci et al. (2024a) No robust evidence of differential productivity acceleration in AIexposed sectors (pre-GenAI data) Macroeconomy Very weak — national accounts, growth accounting Syverson (2024); Brynjolfsson, Rock & Syverson (2021) No detectable AI contribution to aggregate productivity; consistent with J-curve hypothesis and measurement issues  From Micro to Macro: The Aggregation Problem A central question emerges from the evidence presented above: why have substantial individual-level productivity gains not been translated into measurable macroeconomic productivity acceleration? Understanding the mechanisms of aggregation, also accounting for barriers preventing productivity gains at scale, is essential for anticipating whether and when AI’s productivity potential will materialise at scale.

Channels of Aggregation (a) Diffusion and adoption breadth The most direct pathway from individual to aggregate gains is through broad adoption across the economy. Given that individual-level experiments are conducted with workers who have access to AI tools and are motivated to use them, the external validity of these results depends on the share of the workforce with equivalent access, skills, and incentives. Current adoption rates of 7 per cent of US

that individual-level experiments are conducted with workers who have access to AI tools and are motivated to use them, the external validity of these results depends on the share of the workforce with equivalent access, skills, and incentives. Current adoption rates of 7 per cent of US firms (and lower in most European economies) mean that even substantial task-level gains would be diluted at the ILO Brief 6

The Aggregation Paradox of AI: Why do micro-economic productivity gains from AI disappear at scale aggregate level.2 Historical analogy is instructive: electricity took roughly 30 years from initial commercial deployment in 1882 to measurable aggregate productivity impact

(David, 1990). Similarly, the personal computer diffused for over a decade before contributing visibly to US productivity growth in the late 1990s (David, 1990; Devine, 1983; Brynjolfsson, Rock and Syverson, 2021; Filippucci et al., 2024a). (b) Within-firm scaling Even within adopting firms, pilot-to-production scaling is a major bottleneck. Individual workers may demonstrate impressive gains in controlled settings, but organisational deployment requires integration into existing workflows, IT infrastructure, data governance, change management, and quality assurance. Agrawal, Gans, and Goldfarb (2022) emphasise that AI adoption requires “system-level redesign” of production processes, not merely substitution of AI for human effort at isolated task nodes. This redesign entails fixed costs and organisational friction that especially smaller firms find prohibitive. (c) Reallocation effects Aggregate productivity growth reflects not only within-firm improvements but also the reallocation of labour and capital toward more productive firms and sectors. If AI disproportionately benefits already-productive frontier firms, and if competitive pressure or market exit pushes less productive firms to shrink, reallocation can amplify the aggregate effect. Conversely, if AI adoption is concentrated in sectors with low aggregate weight or if the benefits accrue mainly as rents to monopolistic firms without

disproportionately benefits already-productive frontier firms, and if competitive pressure or market exit pushes less productive firms to shrink, reallocation can amplify the aggregate effect. Conversely, if AI adoption is concentrated in sectors with low aggregate weight or if the benefits accrue mainly as rents to monopolistic firms without inducing market-level restructuring, the reallocation channel may remain weak (Bresnahan and Trajtenberg, 1995; De Loecker, Eeckhout and Unger, 2020; Bessen, 2022). (d) Task restructuring and new task creation Acemoglu and Restrepo (2019) formalise a framework in which automation displaces workers from existing tasks, but aggregate productivity growth also depends on the creation of new tasks in which humans have a comparative advantage. If AI automates routine cognitive tasks without generating sufficient new productive roles, the net aggregate effect may be smaller than task-level gains

2 Note that there is a substantial difference between AI adoption rate among workers and adoption rates of companies: suggest. The extent of new task creation is, by its nature, difficult to forecast and measure in real time. (e) General equilibrium and price effects Productivity gains in specific activities may be partially offset by general equilibrium effects. If AI reduces the cost of producing certain outputs (e.g. written content, code, customer service), prices may fall, shifting expenditure towards less AI-affected activities with lower productivity growth, a variant of Baumol’s cost disease. Whether this “structural change burden” dominates depends on demand elasticities and the breadth of AI’s applicability across the economy (Baumol, 2012; Filippucci et al., 2024b). Recent multisector general equilibrium simulations find that AI could raise aggregate TFP growth by around 0.25–0.60 percentage points per year, conditional on broad diffusion and complementary investment. (Filippucci et al., 2024b). These aggregation channels are mediated by institutions, including wage‑setting systems and collective bargaining

could raise aggregate TFP growth by around 0.25–0.60 percentage points per year, conditional on broad diffusion and complementary investment. (Filippucci et al., 2024b). These aggregation channels are mediated by institutions, including wage‑setting systems and collective bargaining arrangements. Where collective bargaining supports wage coordination and worker participation, productivity gains can feed into wage growth without undermining firms’ incentives to invest, while also facilitating agreement on work reorganization and training that are required to realize AI‑related efficiency gains. Conversely, fragmented bargaining structures and weak worker voice may limit information flows about AI deployment, increase resistance to organizational change, and concentrate gains among a narrow set of firms and workers. Conditions for Successful Aggregation Drawing on the preceding analysis and the broader literature on technology diffusion and productivity, the translation of micro-level AI productivity gains into macroeconomic impact depends on a set of interrelated conditions.

https://www.brookings.edu/articles/mind-the-gap-ai-adoption-ineurope-and-the-us/ ILO Brief 7

The Aggregation Paradox of AI: Why do micro-economic productivity gains from AI disappear at scale

 Table 2: Conditions for Translating Micro-Level Gains to Macroeconomic Impact Condition Mechanism Current Assessment Adoption breadth Share of firms and workers with effective access to AI tools Currently low (7 per cent of firms in the US); steep gradients by size, sector, and country Complementary human capital Workforce skills to use, adapt, and manage AI effectively Demand for AI-complementary skills outpacing supply; training systems lagging Complementary intangible investment Organisational redesign, data infrastructure, new business processes Concentrated in large firms; high fixed costs create barriers for micro, small and medium-sized enterprises (MSMEs) Digital and physical infrastructure

Complementary intangible investment Organisational redesign, data infrastructure, new business processes Concentrated in large firms; high fixed costs create barriers for micro, small and medium-sized enterprises (MSMEs) Digital and physical infrastructure Broadband connectivity, cloud computing access, computing capacity Uneven across and within countries; energy constraints emerging Competitive dynamics Market structure incentivising productive use and diffusion of AI Risk of concentration effects if AI reinforces winner-take-all dynamics Labour market adjustment Smooth reallocation of workers from displaced tasks to new productive roles Depends on mobility, training, social protection; risk of prolonged friction Regulatory environment Rules governing data use, AI deployment, liability, and standards Rapidly evolving; uncertainty may delay investment; under-regulation risks misallocation Measurement adequacy Statistical capacity to capture AI-related quality and variety gains Conventional productivity statistics likely undercount true gains in services

Of the conditions listed above, three currently represent the most binding constraints. First, adoption breadth remains low, mechanically limiting aggregate impact regardless of task-level gains. Second, complementary intangible investment, including the organisational redesign and data infrastructure required to embed AI into core production, is costly and concentrated in large firms. Third, measurement inadequacy means that official statistics are likely to understate AI’s true productivity contribution, potentially for years. These three constraints help explain why the Goldman Sachs scenario (7 per cent GDP uplift) remains a distant upper bound under current conditions, while Acemoglu’s cautious 0.5 per cent figure may be closer to the near-term baseline if diffusion and ILO Brief 8

The Aggregation Paradox of AI: Why do micro-economic productivity gains from AI disappear at scale complementary investments do not accelerate substantially.

Cross-Country Heterogeneity and Collective Bargaining Emerging evidence highlights significant variation across

The Aggregation Paradox of AI: Why do micro-economic productivity gains from AI disappear at scale complementary investments do not accelerate substantially.

Cross-Country Heterogeneity and Collective Bargaining Emerging evidence highlights significant variation across countries, sectors, and institutional settings in how AI exposure and adoption translate into employment and productivity outcomes. Cross-country comparisons in the recent firm-level surveys show that firms in Northern and Western Europe tend to exhibit higher AI adoption and larger productivity payoffs than firms in Southern and Eastern Europe, reflecting differences in digital readiness and complementary capabilities (Aldasoro et al., 2026a). Collective bargaining institutions are one element of this broader ecosystem. Earlier studies by Hübler and Jirjahn, (2001) and Svarstad and Kostøl, (2022) find posit