Generative AI (GenAI) has become a focal point in debates on productivity and the future of work. Yet systematic firm-level evidence remains scarce, reflecting the technology’s novelty and the speed of its diffusion. Task-based studies document sizeable performance gains in specific activities (Noy and Zhang 2023, Brynjolfsson et al. 2025, Maršál and Perkowski 2025). However, translating micro evidence into estimates of aggregate productivity gains remains challenging (Brynjolfsson et al. 2021, Bergeaud 2024, Acemoglu 2025, OECD 2025).

In this column, we draw on the Bundesbank Online Panel – Firms (BOP-F), a representative survey of the German corporate sector. Using the Q2 2025 wave, covering more than 7,000 firms in manufacturing and services, we document firms’ past, current, and planned GenAI adoption on both the extensive and intensive margins, associated costs, and firms’ perceived effects on key economic outcomes, including productivity, employment by skill group, and wages (Falck and Nagengast 2026).

Two-speed diffusion: Fast take-up, gradual deepening

On the extensive margin, the survey results point to a rapid expansion of GenAI adoption from 2024 to 2026 in the German corporate sector. The share of firms reporting that they use – or expect to use – generative AI increased from 26% in 2024 to 44% in 2025, reaching 56% in 2026 (Figure 1, left panel).

Figure 1 GenAI adoption in German firms

Notes: Firms reported GenAI intensity in intervals (0%; 1–5%; 6–10%; 11–20%; 21–40%; 41–60%; and >60% of working hours). To compute the average intensity, we assign each firm the midpoint of its selected interval (0%, 3%, 8%, 15.5%, 30.5%, 50.5%, and 60% for the open-ended top bin). The lower and upper bounds shown in the figure are computed analogously by assigning the lower and upper endpoints of each interval. For the open-ended top category (>60%), we use 60% as an upper cap. All statistics are computed using firm weights to obtain representative estimates.

A key advantage of our survey is that it goes beyond a binary yes/no adoption measure and directly quantifies usage intensity, defined as the share of total working hours during which employees use generative AI. The right panel of Figure 1 shows that among firms reporting GenAI use throughout 2024–2026 (‘early adopters’), the average share of working time involving GenAI rises from about 7.5% in 2024 to 10.2% in 2025 and is expected to increase to 12.6% in 2026. Over the same period, the average across all GenAI-using firms increases more gradually, from about 7.5% in 2024 to around 8.9% by 2026, because entrants typically start with comparatively low intensity (around 6–7% in their first year of adoption) and scale up only over time. Put differently, much of the growth in effective exposure to GenAI comes from deepening use within incumbent adopters, while compositional change, with new adopters starting small, mechanically dampens the rise in average intensity as diffusion continues.

Less prompt for the buck: Sizeable GenAI spending, diminishing marginal gains

A natural question is how the rapid diffusion documented above affects AI-related expenditures, an aspect on which there is still very little information. To capture this, we asked firms to estimate their total GenAI-related expenditure as a share of annual sales, covering both one-off and recurring outlays, such as external service providers, hardware and software, internal and external personnel costs, training, and licensing and maintenance fees (Figure 2, left panel). Among firms actually using or planning to use the technology, average spending increases from roughly 1.0% of sales in 2024 to 1.2% in 2025 and 1.5% in 2026. For the German economy as a whole, a back-of-the-envelope calculation suggests that AI-related expenditures rise from 0.3% of aggregate sales in 2024 to 0.5% in 2025 and reach 0.8% in 2026. This is economically non-negligible, likely putting GenAI spending in the same ballpark as some ‘classic’ digital investment categories in macro data.

Figure 2 GenAI spending by German firms

Notes: Right panel: The very high-expenditure tail is difficult to characterise because the top category is open-ended (i.e., right-censored at ≥10% of sales). Hence, we disregard this category in the figure.

Spending and use intensity are closely related, but over the observed range, the cost-intensity schedule is concave (Figure 2, right panel). Firms in the lowest cost brackets (including zero-cost users) already report non-trivial GenAI use, consistent with trial adoption based on free tools or low-tier subscriptions. As expenditure rises through the mid-range, intensity increases but at a declining marginal rate, suggesting diminishing returns once easily scalable applications have been implemented and organisational complements become binding (Bresnahan et al. 2002, Brynjolfsson et al. 2021). Early adopters and information/communication firms lie systematically above the aggregate relationship, consistent with stronger complementary capabilities and a task mix that is more amenable to GenAI.

The composition of GenAI spending is striking. For most users, one-off implementation costs – such as external consulting and hardware – account for less than 25% of total expenditure, with only a modest increase expected over time (Figure 3, left panel). One notable exception is larger firms, which tend to report a higher fixed-cost share, consistent with a more setup-heavy implementation model (Figure 3, right panel). Overall, this pattern indicates that GenAI is integrated primarily through an ‘operating expense’ model, such as recurring subscriptions and permanent IT staff. This shift also has implications for measurement and cyclicality: because many AI-related outlays are recorded as intermediate consumption rather than capital (Highfill et al. 2025), investment-based measures may understate technology deepening and mechanically overstate the productivity residual, while a service-flow model can make usage more responsive to cash-flow conditions than sunk IT investment (DeStefano et al. 2025).

Figure 3 Share of fixed costs for GenAI

Great expectations: Productivity up, stronger high-skill demand, rising wages

Two findings stand out in firms’ assessments of GenAI’s economic effects (Figure 4). First, among adopters (and near-term adopters), productivity expectations are decisively positive. The share of (current or prospective) GenAI users expecting labour productivity to increase by at least 2% rises from 46% (2024) to 51% (2025) and 54% (2026); around one-quarter even expect gains of 5% or more, while only about 4–5% foresee productivity losses.
These magnitudes are not economy-wide forecasts, but they echo the macro literature’s generally optimistic view of AI’s growth potential (Bergeaud 2024, Acemoglu 2025, OECD 2025).

Figure 4 Expected effects of GenAI use in German firms

Second, firms’ labour-market expectations are modestly positive overall with net gains in high-skill jobs and a gradual upward drift in wages. For high-skill employment, about two-thirds expect changes within ±1%, but the tails are asymmetric: in 2026, 28% anticipate growth of at least 2%, versus 8% expecting a decline of at least 2%, consistent with GenAI being perceived as complementary to high-skill work. Low-skill employment expectations are close to balanced on average and concentrated on ‘no change’.

Wages are mostly expected to be stable, yet tilt upward over time: the share of firms expecting wage growth of ≥2% rises from 19% in 2024 to 26% in 2026. In Acemoglu’s (2025) task-based framework, this pattern suggests that adopters expect displacement to be more than offset by task complementarities and the expansion of human tasks – an interpretation worth flagging, given exposure-based measures that emphasise sizeable potential impacts in white-collar tasks (Eloundou et al. 2023).

Authors’ note: This column represents the authors’ personal opinions and does not necessarily reflect the views of the Deutsche Bundesbank or the Eurosystem.

References

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

Bencivelli, L, L De Masi, E Falck, A Fernández Cerezo, S Formai, I H Bricio, E Mattevi, and A Nagengast (2026), “Embracing AI in Europe: New evidence from harmonized central bank business surveys”, VoxEU.org, 6 January.

Bergeaud, A (2024), “The past, present and future of European productivity”, paper prepared for ECB Forum on Central Banking.

Bresnahan, T F, E Brynjolfsson. and L M Hitt (2002), “Information technology, workplace organization, and the demand for skilled labor: Firm-level evidence”, Quarterly Journal of Economics 117(1): 339–76.

Brynjolfsson, E, D Li, and L Raymond (2025), “Generative AI at work”, The Quarterly Journal of Economics 140(2): 889–942.

Brynjolfsson, E, D Rock, and C Syverson (2021), “The productivity J-curve: How intangibles complement general purpose technologies”, American Economic Journal: Macroeconomics 13(1): 333–72.

DeStefano, T, R Kneller, and J Timmis (2025), “Cloud computing and firm growth”, Review of Economics and Statistics 107(6): 1638–51.

Eloundou, T, S Manning, P Mishkin, and D Rock (2023), “GPTs are GPTs: An early look at the labor market impact potential of large language models”, arXiv:2303.10130.

Falck, E, and A Nagengast (2026), “Scaling generative AI: Diminishing marginal returns and firm-level outcomes”, SSRN.

Highfill, T, D Wasshausen, and G Prunchak (2025), “Concepts and challenges of measuring production of artificial intelligence in the US economy”, BEA Working Paper Series WP2025-1.

Maršál, A, and P Perkowski (2025), “Task-based returns to generative AI: Evidence from a central bank”, VoxEU.org, 31 July.

Noy, S, and W Zhang (2023), “The productivity effects of generative artificial intelligence”, VoxEU.org, 7 June.

OECD (2025), “Macroeconomic productivity gains from artificial intelligence in G7 economies”, OECD Artificial Intelligence Papers No. 41.