Artificial intelligence is now at the centre of economic research and policy debates, with a wide range of scenarios for its implications for productivity and labour (Acemoglu 2025, Aghion and Bunel 2024, Jones 2026). Despite headlines about large AI infrastructure investments and large language model (LLM) improvements, economy-wide effects will depend on how widely typical firms adopt AI and deploy it in their businesses.

Real-time evidence from AI users is especially useful in the current environment. In a recent paper (Baslandze et al. 2026), we survey senior corporate decision-makers about what AI is already doing inside their firms and what they expect it to do next. Our evidence complements recent survey-based work on AI adoption across countries and firms (Aldasoro et al. 2026, Bick et al. 2026, Bonney et al. 2026 Ropele and Tagliabracci 2026, Yotzov et al. 2026), while focusing on the productivity and labour market mechanisms inside typical US companies. We also draw comparisons between the current AI wave and the earlier IT boom.

Data on corporate executives

We survey 734 senior financial executives, mostly CFOs, through The CFO Survey and supplemental panels with FEI, Nasdaq, and Duke University. The data have broad coverage across states, firm sizes, and all major nonfarm nongovernmental industries.

A key advantage of the survey is that we can ask corporate decision-makers directly about realised and expected AI-attributed changes in employment, revenue, productivity, and capital, rather than inferring AI effects from standard outcomes that also reflect many other shocks. Our focus is on the typical companies that might be using AI, not the small set of hyperscalers and AI producers that dominate the headlines.

AI adoption is broad but shallow

Figure 1 summarises the first fact about AI adoption. AI investment is already broad across sectors and is expected to expand rapidly in 2026. In the survey, AI investment includes expenditures or financial investments in AI, ranging from subscriptions (e.g. to ChatGPT) to services, training, software, hardware, and internal development.

This broad definition is part of the story. Adoption should not be confused with deep, firm-specific investment. Per-employee spending is low, and most spending (65% in smaller firms and 54% in large firms with more than 500 employees) takes the form of subscriptions, services, and training rather than hardware purchases or internal software development (Panel A). Many firms, therefore, adopt AI in a way that looks less like the classic IT investment boom and more like renting intangible capital from upstream providers.

Among firms that had not invested in AI in 2025, the main barriers are immature technology, untrained workers, privacy concerns, and uncertainty about AI’s capabilities (Panel B).

Figure 1 AI adoption

Notes: Panel A reports the share of firms investing in AI in 2025 and expected to invest in 2026. Panel B reports reasons for not investing among firms that had not invested in AI in 2025. Large firms are those with at least 500 employees. 
Source: Baslandze et al. (2026).

Return of the Solow productivity paradox?

Companies already report positive labour productivity effects from AI, and they expect larger gains going forward. We construct an implied measure of AI-attributed labour productivity growth from firms’ reported AI-attributed changes in revenues and employment. This measure was already positive in 2025 (about 0.6%) and is expected to reach about 1.8% in 2026. Gains are largest in high-skill services and finance, where they exceed 2% (shaded bars in Figure 2).

At the same time, Figure 2 shows a wedge between implied and reported productivity metrics. Firms consistently report larger productivity gains from AI than those implied by contemporaneous changes in revenue and employment. This wedge likely reflects delayed output realisation and quality improvements that are not yet captured in measured revenues.

This pattern is reminiscent of the Solow productivity paradox, whereby transformative technologies are widely perceived as important well before their effects are fully reflected in productivity statistics. The productivity J-curve literature makes a similar point: new general-purpose technologies often require learning, organisational change, and complementary investments before the gains appear real (Brynjolfsson et al. 2021).

Figure 2 Reported and implied productivity effects of AI

Notes: The figure compares reported and implied AI-attributed labour productivity growth across sectors for 2025 and expected 2026. Implied productivity is based on firms’ reported AI-attributed changes in revenues and employment. 
Source: Baslandze et al. (2026).

The near-term labour story is reallocation, not mass layoffs

The employment results are more nuanced than the public debate often suggests (Figure 3). In the near term, AI has not led to – and is not expected to lead to – meaningful reductions in aggregate employment. Firm-size- and sector-weighted employment is expected to decline by less than 0.4% due to AI in 2026. Effects of this magnitude may be hard to detect in aggregate statistics and are concentrated among large companies in certain sectors.

The heterogeneity is important. Large firms expect employment reductions, especially in finance and high-skill services, while smaller firms often anticipate modest AI-driven employment growth. More meaningful adjustments are expected in workforce composition – away from routine clerical work and toward skilled technical roles. By 2028, the share of employees performing routine clerical work is expected to decline by more than two percentage points, mostly among large firms, with partially offsetting increases in skilled technical roles, mostly among smaller firms.

Figure 3 AI and employment

Notes: Panel A reports expected AI-driven employment changes by sector and firm size. Panel B reports expected changes in workforce composition by task and firm size. 
Source: Baslandze et al. (2026).

To better understand potential labour reallocation, we use executives’ open-ended responses about the roles and responsibilities they expect AI to replace or enhance. Figure 4 depicts word cloud summaries. Office and administrative support roles show the most negative exposure, consistent with the automation of routine clerical activities such as data entry, transaction processing, customer service, and basic accounting. Enhancement is often associated with marketing, accounting, finance, and management. Some functions, including customer service, appear in both categories, suggesting that AI may complement some tasks while replacing others within the same occupation. Our data provide insight into nuanced task reallocations within jobs and the direction of labour demand that AI exposure measures alone cannot determine.

Figure 4 Tasks and roles enhanced or replaced by AI

Notes: The word clouds summarise open-ended responses on tasks and roles enhanced or replaced by AI; word size reflects frequency within each category. 
Source: Baslandze et al. (2026).

More on mechanisms and why the AI wave looks different from the 1990s IT wave

Historically, labour productivity gains from new technologies are substantially from capital deepening: more capital per worker raises output per worker, which was central to the US productivity resurgence associated with the 1990s IT wave (Jorgenson et al. 2008). In contrast, in our data capital deepening accounts for only about 15% of near-term AI-attributed labour productivity growth among surveyed firms. This relatively small role is not surprising: for the typical firm, AI spending is often subscription-based and expensed rather than capitalized. Although the economy-wide capital-intensive buildout is large (Rubinton and Patro 2026), much of it happens upstream among technology firms building data centres and cloud infrastructure, while user firms access AI through cloud services, software subscriptions, and other rented digital infrastructure.

The mechanisms are also distinct. We find that labour productivity gains are most strongly associated with innovation- and demand-oriented channels: developing or improving products and services, and reaching or serving customers more effectively. These demand-driven gains raise revenues without leading to broad near-term employment declines. Cost reduction and labour-saving applications occur, especially among larger firms, but they are not the dominant short-run channel in our data.

Policy lessons

The macroeconomic footprint of AI includes the large capital buildout by technology firms, but it also includes ordinary adoption decisions by user firms. The first policy lesson is the importance of measurement: tracking AI adoption in typical firms and keeping a pulse on changes in productivity, work, and labour demand across the broader economy. To keep up with rapid technological change, frequent surveys with in-depth questions need to be designed and analysed. While surveys usually focus on existing firms, understanding the entry of new firms may also be an important margin.

A second lesson is that workforce policy should focus on task and job reallocation, not only on headcounts. The near-term aggregate employment effects appear modest, but the composition of work is changing. Training, diffusion of practical AI skills, and support for workers in routine clerical roles may therefore be beneficial.

Finally, the realisation of full productivity gains may take time. Early AI use may improve workflows, efficiency, product quality, and customer reach before these gains substantially increase revenue per worker. As with earlier general-purpose technologies, the first years are likely to combine exploration and adaptation, slow diffusion, and delayed productivity measurement.

References

Acemoglu, D (2025), “The simple macroeconomics of AI”, Economic Policy 40(121): 13-58.

Aghion, P and S Bunel (2024), “AI and Growth: Where Do We Stand?”, policy note, June.

Aldasoro, I, L Gambacorta, R Pál, D Revoltella, C Weiss and M Wolski (2026), “How AI is affecting productivity and jobs in Europe”, VoxEU.org, 17 February.

Baslandze, S, Z Edwards, J Graham, T McClure, M Sparks, B Meyer, S Waddell and D Weitz (2026), “Artificial Intelligence, Productivity, and the Workforce: Evidence from Corporate Executives”, Journal of Finance: Insights and Perspectives, forthcoming (see also CEPR Discussion Paper 21313).

Bick, A, A Blandin, D Deming, N Fuchs-Schündeln and J Jessen (2026), “Differences in AI adoption in Europe and the US: Explanations and implications for productivity growth”, VoxEU.org, 9 April.

Bonney, K, C L Breaux, E Dinlersoz, L S Foster, J C Haltiwanger and A A Pande (2026), “The Microstructure of AI Diffusion: Evidence from Firms, Business Functions, and Worker Tasks”, NBER Working Paper No. 35141.

Brynjolfsson, E, D Rock and C Syverson (2021), “The Productivity J-Curve: How Intangibles Complement General Purpose Technologies”, American Economic Journal: Macroeconomics 13(1): 333-372.

Jones, C I (2026), “A.I. and Our Economic Future”, NBER Working Paper No. 34779.

Jorgenson, D W, M S Ho and K J Stiroh (2008), “A Retrospective Look at the U.S. Productivity Growth Resurgence”, Journal of Economic Perspectives 22(1): 3-24.

Ropele, T and A Tagliabracci (2026), “The economic impact of artificial intelligence: Evidence from Italian firms”, Bank of Italy Occasional Paper No. 1005.

Rubinton, H and B A Patro (2026), “Tracking AI’s contribution to GDP growth”, Federal Reserve Bank of St. Louis On the Economy, 12 January.

Yotzov, I, J M Barrero, N Bloom, P Bunn, S J Davis, K M Foster, A Jalca, B H Meyer, P Mizen, M A Navarrete, P Smietanka, G Thwaites and B Z Wang (2026), “Firms predict an AI productivity boom is coming”, VoxEU.org, 12 March.