## wpiea2025067-print-pdf

## Source details

**Canonical URL:** [wpiea2025067-print-pdf](https://www.imf.org/-/media/files/publications/wp/2025/english/wpiea2025067-print-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2025/english/wpiea2025067-print-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2025/english/wpiea2025067-print-pdf.pdf.json)

---

### 1. Introduction
- Objective: estimate the size of effects of AI on total factor productivity (TFP) across 31 European countries over the medium term and examine impeding effects of regulation in Europe.
- Context:
  - Europe has experienced lackluster productivity growth and a large productivity gap vis-à-vis the US (IMF, 2024).
  - Widespread view that Europe may be falling behind the US and China in AI development and adoption, partly due to a more stringent regulatory environment.
- Approach:
  - Uses Acemoglu (2024) framework to estimate medium-term productivity gains from AI; interprets "medium term" as 5 years.
  - Quantifies uncertainty by combining a comprehensive set of available estimates of AI exposure of individual tasks to deliver 44 scenarios.
  - Calibrates country- and sector-specific AI adoption rates using Svanberg et al. (2024) estimates and regression evidence of drivers of AI adoption in Europe.
  - Examines regulation effects: national occupation licensing and training requirements, data privacy laws, and the EU AI Act; assumes regulation halves AI capabilities for affected tasks.

### 2. Stylized facts
- Generative AI (genAI) diffusion has been historically fast: genAI (measured by ChatGPT users) reached 100 million users in months versus years or decades for past general-purpose technologies.
- Rapid user uptake does not imply broad use across firm tasks or broad firm-level AI adoption.

### 3. Methodology
- Model: Acemoglu (2024) (based on Acemoglu and Restrepo 2018, 2019, 2022).
  - Production requires a fixed set of tasks; tasks performed with capital or labor.
  - Two channels for AI productivity gains: automation (substitution) and task complementarity (augmentation).
  - Model suited to medium-term, non-transformational productivity effects; not designed to capture long-run transformational changes.
- Three key calibration parameters:
  1. AI exposure of occupations (share of tasks AI is capable of performing).
     - Example benchmark: Eloundou et al. (2024) measure yields a wage-bill-weighted share of exposed tasks in the U.S. of 19.9 percent.
  2. AI adoption rate (share of AI-exposed tasks where benefits exceed costs).
     - Acemoglu (2024) uses Svanberg et al. (2024) to assume benefits exceed costs for 23 percent of AI-exposed tasks.
  3. Labor cost savings when AI is applied to a task.
     - Acemoglu (2024) calibrates automated-task labor cost reduction at 27 percent; given labor costs account for 53 percent of output, this implies 15 percent total cost savings.
- Calculation: productivity gain estimated as product of the three parameters (example: 19.9%, 23%, and 15% yield Acemoglu’s 0.71% cumulative medium-term TFP increase for the US over a 10-year horizon in Acemoglu).

### 4. Key quantitative findings and scenarios
- Scope: analysis for 31 European countries, 44 exposure scenarios, country-specific adoption calibrated to economic characteristics.
- Main drivers of AI adoption across European sectors and countries:
  - Wage levels are the main driver of AI adoption (stronger than capital costs, industry concentration, digitalization, or human capital).
- Preferred (most plausible) scenario:
  - Europe-wide cumulative TFP effect over the medium term (5 years): 1.1 percent.
  - This preferred Europe estimate exceeds Acemoglu’s US estimate by almost 60 percent (Acemoglu’s US estimate: 0.71 percent cumulative over 10 years).
  - Heterogeneity across countries:
    - Higher-income countries tend to have much larger estimated TFP gains than lower-income economies.
    - Example: Luxembourg — preferred scenario cumulative gains could be 2 percent (about twice the European average and more than 4 times larger than those in Romania).
    - Upside risk: Luxembourg’s productivity gains could be more than twice as high if AI turns out to be more capable than in the preferred scenario.
- Uncertainty quantification:
  - 44 scenarios constructed from available AI exposure estimates to capture a range of plausible AI capabilities and exposures.
- Comparison to other macro estimates:
  - McKinsey (2023) and Goldman Sachs (Hatzius et al., 2023) estimate cumulative GDP gains above 35 percent for advanced economies and 7 percent globally over a 10-year period, respectively.
  - Commission de l’Intelligence Artificielle (2024) infers potential growth impacts up to 1.3 annually by drawing parallels to past general-purpose technologies.
  - IMF (2024) and Cazzaniga et al. (2024) estimate annual growth impacts of up to 0.8 percentage points.
  - Acemoglu (2024) estimates much smaller gains (<0.7 percent cumulatively over 10 years); alternative assumptions can increase those estimates (Aghion and Bunel, 2024).

### 5. Results — variation in medium-term productivity gains
- Method and scenarios:
  - Applied Acemoglu (2024) baseline and constructed alternative scenarios varying AI capabilities and adoption rates; combined to form 44 scenarios.
- Average and cross-country variation:
  - Average cumulative medium-term productivity gain: close to 0.8 percent in Europe versus around 0.7 in Acemoglu (2024) for the US.
  - Substantial cross-country variation under the Acemoglu (2024) baseline: around 0.5 percent in Romania to close to 1 percent in Luxembourg.
  - Higher-income countries tend to have larger gains due to prevalence of white-collar services with higher AI exposure.
- Sensitivity to AI exposure measures:
  - Conservative exposure assumptions mute gains and reduce variation.
  - Optimistic exposure assumptions imply large upside risks; e.g., Luxembourg could be above 3 percent cumulatively.
- Sensitivity to AI adoption rates:
  - Acemoglu (2024) baseline assumes 23 percent cost-effective adoption; calibrated Europe-wide adoption scenario: 18 percent in Europe, 5 percentage points lower than US baseline.
  - Allowing country-sector variation in adoption (wage-driven) increases cross-country range: high-wage countries revised upward; lower-income countries revised downward.
- Scenario distribution examples:
  - Norway: median scenario indicates the largest gains; one scenario shows productivity gain around 5.1% cumulatively, conservative scenarios near zero.
  - Luxembourg: inter-quartile range spans from 0.8% to 2.95%.
  - Romania: lowest median gain and lowest uncertainty; interquartile range approximately 0.3% to 0.7%.

### 6. Preferred scenario (detailed)
- Preferred assumptions:
  - Occupational AI exposure: baseline task-based estimates from Eloundou et al. (2024).
  - AI adoption rates: country- and sector-specific rates that vary according to wages, shifted upward to keep Europe-wide average at 18% (‘adjusted country-sector 1’).
- Preferred-scenario outcomes:
  - Europe average cumulative medium-term productivity gain: around 1.1 percent.
  - The 1.1 percent is almost 60 percent larger than the Acemoglu (2024) baseline estimate for Europe.
- Drivers of cross-country differences (decomposition):
  - Three drivers: AI adoption rate (varies across countries/sectors), industry composition of value added, occupational composition within industries.
  - Heterogeneity in adoption rates (driven by relative cost of labor) explains most cross-country variation.
  - Industry composition is second-largest driver.
  - Occupational composition within industries accounts for minimal variation.
  - Non-linearities: the sum of drivers does not equal preferred-scenario gains; residual often negative for countries with largest estimated gains.

### 7. The role of regulation — modeling and quantified illustrative impacts
- Regulations modeled:
  - Regulated occupations (national occupation-level licensing and training requirements) using European Commission Regulated Professions Database.
  - EU AI Act: two effects modeled (systems above a computational threshold discouraged for ‘hard’ tasks; additional requirements discourage AI use in tasks classified as ‘high-risk’).
  - Data privacy laws: modeled as discouraging AI use in three data-intensive sectors: Information and Communication (NACE J), Financial and Insurance Activities (NACE K), and Human Health and Social Work Activities (NACE Q).
- Modeling assumption for illustrative effects:
  - AI exposure is reduced by 50 percent in tasks, occupations, and sectors affected by regulation.
- Quantified impacts (relative to the preferred scenario):
  - Data privacy laws reduce productivity gains of AI by 10 percent.
  - The EU AI Act reduces productivity gains of AI by around 15 percent.
  - National occupation-level regulation reduces productivity gains of AI by around 15 percent.
  - The combined effect of all three reduces productivity gains from AI by well over 30 percent.
- Note: these regulatory scenarios are hypothetical illustrations; not all regulations are fully implemented or applicable in all countries.

### 8. Conclusions and policy implications
- Main conclusions:
  - Estimated cumulative productivity gains from AI for European economies over the medium term are around 1.1 percent in the preferred scenario.
  - This 1.1 percent estimate is almost 60 percent above the estimates for the US from Acemoglu (2024).
- Distributional implications:
  - Significant variation across countries; higher-income countries are predicted to benefit more due to larger high-exposure service sectors and higher wages that incentivize adoption.
  - AI may slow income convergence within Europe.
  - Higher-income countries face larger uncertainty around potential gains.
  - Micro evidence suggests AI boosts productivity of low-skill workers more than higher-skilled workers (Brynjolfsson et al., 2023).
- Caveats and uncertainty drivers:
  - Model captures incremental productivity changes, not longer-term transformational effects.
  - Results sensitive to assumptions about adoption rates, costs of AI systems, labor cost savings per exposed task, and timing.
  - If costs fall faster or AI-as-a-service expands, adoption rates and gains could be much larger.
- Policy implications:
  - Policies that affect costs of AI adoption (facilitating AI-as-a-service, reducing deployment costs) could materially affect adoption and productivity gains.
  - Regulatory choices that limit AI exposure in tasks, occupations, or sectors can reduce productivity gains (analysis estimates effects under exposure being 50 percent lower due to regulation).
  - Attention to cross-country heterogeneity and potential widening of income differences within Europe is warranted when designing redistribution, labor market, and training policies.
- Research and data priorities:
  - Need for country- and sector-specific calibration of AI adoption drivers (wage levels, industry composition) to understand cross-country heterogeneity.

### 9. Appendix highlights — regression, adoption, micro evidence, and regulatory specifics
- Key regression findings (determinants of AI adoption):
  - Wage (2016) coefficients include 0.75*** (SE 0.22) and 0.52*** (SE 0.10); wage (2021): 0.57** (SE 0.24).
  - Paper uses coefficient 0.52 from Column (2) for calibrations.
  - Labor cost: 0.58*** (SE 0.17).
  - Cost of capital: mixed results including -0.96** (SE 0.45) and -1.34*** (SE 0.44) across specifications.
  - Concentration, computer use, energy cost, and human capital show limited or non-robust relationships once wages are controlled for.
  - # observations vary by column: 265, 265, 252, 265, 226, 226, 202, 153, 265, 232, 102, 109.
- Economy-wide adoption estimates and scenarios:
  - 2023 wage-based estimate (Column (2)) average estimated adoption rate: 13.8 percent (across countries and sectors, weighted by value added).
  - Medium-term scenarios to reach average 18 percent:
    - Rescaled: (2023 estimate) × 18/13.8.
    - Shifted: (2023 estimate − 13.8) + 18.
  - Using a 50 percent US–Europe wage gap and slope 0.52 implies adoption around 5-percentage points lower in Europe than the US; Acemoglu’s 23 percent for the US implies 18 percent for Europe.
  - Country examples (shifted method): Bulgaria 2023 adoption 4.0 percent → shifted 8.2 percent; Switzerland 2023 adoption 25.1 percent → shifted 29.3 percent.
- High-adoption AI-as-a-service scenario:
  - Svanberg et al. (2024) estimate US medium-term adoption could be ~80 percent if firms rent AI services.
  - Europe under AI-as-a-service: 80 percent US implies 75 percent for Europe as a whole.
  - Scenario construction: 2023 wage-based average = 13.8 percent; future scenario uses shifting formula (2023 estimate − 13.8) + 75 to ensure average adoption = 75 percent.
- Descriptive statistics (selected):
  - AI adoption: no. countries 30; no. sectors 11; no. country-sectors 298; mean 9.9; st dev. 10.2; min 0.0; max 84.7.
  - wage (2016): mean 17.0; st dev. 13.1; min 1.5; max 70.6.
  - wage (2021): mean 19.6; st dev. 13.0; min 1.7; max 72.2.
  - labor cost: mean 21.6; st dev. 16.8; min 1.8; max 89.5.
  - cost of capital: mean 2.6; st dev. 1.0; min 1.2; max 6.0.
  - concentration: mean 30.4; st dev. 21.4; min 0.0; max 90.6.
  - computer use: mean 58.5; st dev. 24.2; min 17.3; max 100.0.
  - energy cost: mean 0.1; st dev. 0.0; min 0.1; max 0.2.
  - human capital: mean 39.3; st dev. 21.2; min 4.2; max 89.7.
- Microeconomic evidence on labor cost savings from AI (selected randomized trials and studies):
  - Brynjolfsson et al. (2023): customer support issues resolved per hour increased by 14%, including a 34% improvement for novice and low-skilled workers.
  - Dell’Acqua et al. (2023): consultants below average improved 43%; above average 17%.
  - Doshi et al. (2023): writers improved novelty 6.7% and usefulness 6.4%.
  - Kanazawa et al. (2022): AI navigation narrowed productivity gap between high- and low-skilled taxi drivers by 14%.
  - Korinek (2023): 40% of tested research tasks were highly useful with LLMs.
  - Noy and Zhang (2023): simple writing tasks 40% faster and 18% quality improvement.
  - Peng et al. (2023): programmers completed tasks 55.8% faster with GitHub Copilot.
  - Schoenegger et al. (2024): forecasters gained 24% to 28% improvement in prediction accuracy.
  - Kalliamvakou (2022): 88% of developers reported increased perceived productivity with GitHub Copilot.
- Regulation and matching:
  - European Commission Regulated Professions Database lists 563 distinct generic professions regulated in at least one country; authors matched 366 to ISCO-08 4-digit titles.
  - Matched regulated professions (percent of unique 4-digit ISCO code), selected values preserved:
    - CZ: 32.1; SK: 26.1; HR: 24.5; AT: 24.0; LU: 22.4; CH: 22.4; PL: 21.5; DE: 21.5; HU: 21.0; SI: 20.6; LI: 19.9; PT: 19.4; GB: 18.5; IT: 17.1; FR: 17.1; BE: 16.6; DK: 15.7; GR: 15.5; ES: 14.8; NL: 14.8; CY: 13.4; RO: 13.4; IS: 17.6; FI: 12.9; IE: 12.7; NO: 12.5; SE: 11.3; MT: 11.1; EE: 11.1; LT: 7.6; LV: 7.6; BG: 6.7.
- EU AI Act specifics (Appendix 6):
  - The Act specifies requirements based on level of risk and computational power (capacity cap).
  - Implementation timing: authors simplify by assuming the AI Act is in effect or fully anticipated by firms.
  - Computational power threshold: defined at 10 25 FLOPS.
  - Footnote timing detail: Chapter III Section 1 Article 6 Point 1 will enter into force in August 2027 and Section 4 will enter into force in August 2025.
- Data privacy laws and sector adjustments:
  - Three industries adjusted for data privacy intensity: Information and Communication (NACE J), Financial and Insurance Activities (NACE K), Human Health and Social Work Activities (NACE Q).
  - Empirical adjustment: occupation-level AI exposure multiplied by 0.5 in these sectors to reflect GDPR and similar constraints.

*Source: wpiea2025067-print-pdf — https://www.imf.org/-/media/files/publications/wp/2025/english/wpiea2025067-print-pdf.pdf*

### 1. Introduction ........................................................................................................

### wpiea2025067-print-pdf - 1. Introduction .................................................................................................

### 1. Introduction
- Objective: estimate the size of effects of AI on total factor productivity (TFP) across 31 European countries over the medium term and examine impeding effects of regulation in Europe.
- Context:
  - Europe has experienced lackluster productivity growth and a large productivity gap vis-à-vis the US (IMF, 2024).
  - Widespread view that Europe may be falling behind the US and China in AI development and adoption, partly due to a more stringent regulatory environment.
- Approach:
  - Uses Acemoglu (2024) framework to estimate medium-term productivity gains from AI; interprets "medium term" as 5 years given model characteristics and authors’ assumptions.
  - Quantifies uncertainty by combining a comprehensive set of available estimates of AI exposure of individual tasks to deliver 44 scenarios.
  - Calibrates country- and sector-specific AI adoption rates using Svanberg et al. (2024) estimates and regression evidence of drivers of AI adoption in Europe.
  - Examines regulation effects: national occupation licensing and training requirements, data privacy laws, and the EU AI Act; assumes regulation halves AI capabilities for affected tasks.

### 2. Stylized Facts
- Generative AI (genAI) diffusion has been historically fast: genAI (measured by ChatGPT users) reached 100 million users in months versus years or decades for past general-purpose technologies.
- Rapid user uptake does not imply broad use across firm tasks or broad firm-level AI adoption.

### 3. Methodology
- Model: Acemoglu (2024) (based on Acemoglu and Restrepo 2018, 2019, 2022).
  - Production requires a fixed set of tasks; tasks performed with capital or labor.
  - Two channels for AI productivity gains: automation (substitution) and task complementarity (augmentation).
  - Model suited to medium-term, non-transformational productivity effects; not designed to capture long-run transformational changes (new industries, accelerated scientific discovery).
- Three key parameters for calibrating productivity gains:
  1. AI exposure of occupations (share of tasks AI is capable of performing, including automation and complementarity).
     - Example benchmark: Eloundou et al. (2024) based measure yields a wage-bill-weighted share of exposed tasks in the U.S. of 19.9 percent.
  2. AI adoption rate (share of AI-exposed tasks where benefits exceed costs).
     - Acemoglu (2024) uses Svanberg et al. (2024) to assume benefits exceed costs for 23 percent of AI-exposed tasks.
  3. Labor cost savings when AI is applied to a task.
     - Acemoglu (2024) calibrates automated-task labor cost reduction at 27 percent; given labor costs account for 53 percent of output, this implies 15 percent total (labor and capital) cost savings.
- Calculation: productivity gain estimated as product of the three parameters (example: 19.9%, 23%, and 15% yield Acemoglu’s 0.71% cumulative medium-term TFP increase for the US, over a 10-year horizon in Acemoglu).

### 4. Key quantitative findings and scenarios
- Scope: analysis for 31 European countries, 44 exposure scenarios, country-specific adoption calibrated to economic characteristics.
- Main drivers of AI adoption across European sectors and countries:
  - Wage levels are the main driver of AI adoption (stronger than capital costs, industry concentration, digitalization, or human capital).
- Preferred (most plausible) scenario results:
  - Europe-wide cumulative TFP effect over the medium term (5 years): 1.1 percent.
  - This preferred Europe estimate exceeds Acemoglu’s US estimate by almost 60 percent (Acemoglu’s US estimate: 0.71 percent cumulative over 10 years).
  - Heterogeneity across countries:
    - Higher-income countries tend to have much larger estimated TFP gains than lower-income economies.
    - Example: Luxembourg — preferred scenario cumulative gains could be 2 percent (about twice the European average and more than 4 times larger than those in Romania).
    - Upside risk: Luxembourg’s productivity gains could be more than twice as high if AI turns out to be more capable than in the preferred scenario.
- Uncertainty quantification:
  - 44 scenarios constructed from available AI exposure estimates to capture a range of plausible AI capabilities and exposures.
- Comparison to other macro estimates (reported in the introduction):
  - McKinsey (2023) and Goldman Sachs (Hatzius et al., 2023) estimate cumulative GDP gains above 35 percent for advanced economies and 7 percent globally over a 10-year period, respectively.
  - Commission de l’Intelligence Artificielle (2024) infers potential growth impacts up to 1.3 annually by drawing parallels to past general-purpose technologies.
  - IMF (2024) and Cazzaniga et al. (2024) estimate annual growth impacts of up to 0.8 percentage points based on labor reallocation and capital share changes.
  - Acemoglu (2024) estimates much smaller gains (<0.7 percent cumulatively over 10 years); alternative assumptions can increase those estimates (Aghion and Bunel, 2024).

### 5. The role of regulation (summary of analysis and assumptions)
- Regulations considered: national occupation licensing and training requirements, data privacy laws, and the EU AI Act.
- Assumption for regulatory impact: regulation halves AI capabilities for tasks it affects.
- Findings on regulation:
  - Combined adverse effects of national occupation-level regulation, the EU AI Act, and data privacy laws on productivity could be significant.
  - Occupational restrictions and the EU AI Act have the largest negative effects on productivity gains.
  - Data privacy laws affect industries with high AI exposure (e.g., IT and financial services) but their productivity-inhibiting effect is somewhat smaller than occupational regulation and the EU AI Act.

### 6. Policy implications (concise)
- Medium-term outlook:
  - AI is unlikely to be a silver bullet to significantly boost sluggish productivity growth in Europe or to fully close the productivity gap with the United States over the medium term.
  - AI may slow income convergence within Europe because gains tend to be larger in more advanced economies.
- Trade-offs:
  - Policymakers face trade-offs between regulations that protect privacy and safety and maximizing productivity gains from AI; this analysis is not a comprehensive cost-benefit evaluation of those regulations but informs the balance policymakers must consider.
- Research and data priorities implied:
  - Need for country- and sector-specific calibration of AI adoption drivers (wage levels, industry composition) to understand cross-country heterogeneity in productivity gains.

*Source: wpiea2025067-print-pdf — https://www.imf.org/-/media/files/publications/wp/2025/english/wpiea2025067-print-pdf.pdf*

### 4. Results

### 4. Results

### 4.1 Variation in Medium-term Productivity Gains
- Method and comparisons
  - Applied the same methodology as in Acemoglu (2024) to Europe (the ‘Acemoglu (2024) baseline’) to compare productivity gains between the US and European countries using the same measure of occupation-level AI exposure, AI adoption rate, and labor cost savings.
  - Constructed alternative scenarios varying AI capabilities (AI task-level exposure measures) and AI adoption rates; combined all assumptions to form 44 different scenarios to quantify uncertainty.
- Key findings on magnitudes and cross-country variation
  - Average cumulative medium-term productivity gain: close to 0.8 percent in Europe versus around 0.7 in Acemoglu (2024) for the US.
  - Substantial cross-country variation under the Acemoglu (2024) baseline: around 0.5 percent in Romania to close to 1 percent in Luxembourg.
  - Higher-income countries tend to have larger gains driven by higher prevalence of white-collar services (e.g., financial services) that are more exposed to AI.
- Sensitivity to alternative AI exposure measures
  - More conservative exposure assumptions mute productivity gains and reduce cross-country variation.
  - More optimistic exposure assumptions imply large upside risks; e.g., in Luxembourg productivity gains could be above 3 percent cumulatively over the medium term—more than 2 percentage points higher than under the Acemoglu (2024) baseline.
- Sensitivity to AI adoption rate
  - Acemoglu (2024) baseline assumes AI is cost-effective to adopt for 23% of firms in every country and sector (based on Svanberg et al., 2024).
  - Calibrated Europe-wide adoption scenario: 18 percent in Europe, 5 percentage points lower than under the Acemoglu (2024) baseline for the US.
  - Two alternative scenarios allow adoption rates to vary across countries and sectors according to wages while maintaining the average Europe-wide adoption rate of 18 percent (one shifts wage-based estimates upward; the other rescales them upward).
  - Allowing country-sector variation in adoption rates increases the cross-country range of estimated productivity gains: high-wage countries (e.g., Luxembourg, Norway, Switzerland) tend to see upward revisions; lower-income countries tend to see lower gains.
- Full scenario distribution (44 scenarios)
  - Productivity gains vary significantly across countries and across scenarios.
  - Norway: in the median scenario expected to gain the most; one scenario shows productivity gain around 5.1% cumulatively over the medium term, while conservative scenarios place effects near zero.
  - Luxembourg: largest inter-quartile range across scenarios, spanning from 0.8% to 2.95%.
  - Romania: lowest median productivity gain and lowest uncertainty; interquartile range approximately 0.3% to 0.7%.

### 4.2 Preferred Scenario
- Choice of preferred assumptions
  - Occupational AI exposure measure: baseline task-based estimates from Eloundou et al. (2024).
  - AI adoption rates: country- and sector-specific rates that vary according to wages, with the sector-country adoption estimates shifted upward to keep the average adoption rate in Europe at 18% (referred to as ‘adjusted country-sector 1’).
- Preferred-scenario results
  - The diamonds in Figure 5 show the preferred scenario outcomes, which vary significantly across countries due to economic structure and wage-driven adoption rates.
  - Europe average cumulative medium-term productivity gain under the preferred scenario: around 1.1 percent.
  - This 1.1 percent is almost 60 percent larger than the Acemoglu (2024) baseline estimate for Europe.
- Drivers of cross-country differences (decomposition under the preferred scenario)
  - Three drivers considered: AI adoption rate (varies across countries and sectors), industry composition of value added, and occupational composition within each industry.
  - Heterogeneity in adoption rates (driven by relative cost of labor) explains most cross-country variation.
  - Industry composition is the second-largest driver: countries with larger shares of high-exposure industries have higher expected impacts.
  - Occupational composition within industries accounts for only minimal variation.
  - The framework is non-linear; the sum of the three drivers does not equal the preferred-scenario gains. The residual (difference between sum of drivers and the preferred-scenario gains) is often negative, especially for countries with the largest estimated productivity gains (i.e., the sum of the three drivers overpredicts the preferred-scenario gains in many cases).

### 4.3 The Role of Regulation
- Three types of regulation considered and how they are modeled
  - Regulated occupations (national occupation-level regulation): licensing and training requirements from the European Commission Regulated Professions Database. Assumed to discourage AI use in those occupations.
  - EU AI Act: introduced in 2024. Two modeled effects:
    - Systems above a computational threshold: assumed to be discouraged for tasks Acemoglu classifies as ‘hard’ (e.g., diagnosing and treating a patient).
    - Additional regulatory requirements for tasks classified as ‘high-risk’: assumed to discourage AI use in occupations performing such tasks.
  - Data privacy laws: modeled as discouraging AI use in three data-intensive sectors—Information and Communication (NACE Rev. 2 code J), Financial and Insurance Activities (NACE Rev. 2 code K), and Human Health and Social Work Activities (NACE Rev. 2 code Q).
- Modeling assumption for illustrative effects
  - For illustrative purposes, AI exposure is reduced by 50 percent in tasks, occupations, and sectors affected by regulation (i.e., AI exposure is half of what it would be otherwise).
- Quantified impacts (relative to the preferred scenario)
  - Data privacy laws reduce the productivity gains of AI by 10 percent.
  - The EU AI Act reduces the productivity gains of AI by around 15 percent.
  - National occupation-level regulation reduces the productivity gains of AI by around 15 percent.
  - The combined effect of all three regulatory areas reduces the productivity gains from AI by well over 30 percent.
- Note on scope
  - These regulatory scenarios are hypothetical illustrations; not all regulations are fully implemented or applicable in all countries in the sample.

*Source: wpiea2025067-print-pdf - 4. Results.*

### 5. Conclusions and Policy Implications

### 5. Conclusions and Policy Implications

### Main findings on productivity gains
- Estimated cumulative productivity gains from AI for European economies over the medium term are around 1.1 percent in the preferred scenario.
- This 1.1 percent estimate is almost 60 percent above the estimates for the US from Acemoglu (2024).
- The main reasons for the cross-region difference are Europe’s sectoral composition and different assumptions about AI capabilities.

### Cross-country variation and distributional implications
- There is significant variation in the productivity gains from AI across European countries.
- Higher-income countries are predicted to benefit more because:
  - They have larger sectors like white-collar services that benefit most from AI.
  - Their higher wages create stronger incentives for firms to adopt AI.
- As a result, AI could slow down the rate of income convergence within Europe.
- Although higher-income countries look set to enjoy higher productivity gains from AI, they also face larger uncertainty around these potential gains.
- At the microeconomic level, evidence suggests AI boosts the productivity of low-skill workers far more than that of higher-skilled workers (Brynjolfsson et al., 2023), highlighting potential heterogeneity between macro and micro results.

### Key calibration and data points used in the analysis
- Average of the sample AI-exposed value added in the preferred scenario is 6.8%.
- For the United States, Acemoglu (2024) calibrates the AI adoption rate at 23 percent.
- Svanberg et al. (2024) estimate the rate of AI adoption in the US could be around 80%, about three-and-a-half times as large as the 23% assumed in Acemoglu (2024) baseline.
- Eurostat data indicate around 8 percent of all European firms used at least one type of AI in 2023.
- Regression evidence: every one euro per hour increase in wages is associated with a 0.75 percentage point increase in the share of firms using AI in each country and sector, controlling for unobserved factors.
- Cost estimates for the median firm’s AI system reported by Svanberg et al. (2024) are between 2 and 3 million US dollars.
- Svanberg et al. (2024) baseline cost assumptions include a discount rate of 5 percent a year, an annual reduction in computing costs of 22 percent a year, and a useful life of 5 years for an AI system.

### Caveats and uncertainty drivers
- The model (Acemoglu, 2024) captures incremental productivity changes rather than larger transformational effects of AI (for example, faster scientific progress).
- Results are sensitive to assumptions about:
  - AI adoption rates (which depend on costs and availability of business-to-business AI service providers).
  - Costs of AI systems (chips, energy, system development and maintenance).
  - Labor cost savings per exposed task (alternative microeconomic estimates exist; see Appendix 4).
  - Timing: productivity gains may take longer to materialize if exposure estimates assume improvements relative to existing AI technologies.
- If costs fall faster or AI service provision expands, adoption rates and productivity gains could be much larger (see the Svanberg et al. (2024) counterfactual).

### Longer-term and structural considerations
- Over the longer term, productivity gains are likely to be larger than the medium-term estimates because AI could:
  - Lead to structural transformation of the economy.
  - Create new industries and value chains.
  - Accelerate R&D activity through generating novel research ideas and investigating them (Si et al., 2024).
  - Generate local spillovers from the development of AI itself.
- Considering these transmission channels could lead to more permanent growth effects rather than one-off level effects from automating existing tasks (Frey et al., 2024).

### Policy implications (implicit from findings)
- Policies that affect the costs of AI adoption (for example, facilitating business-to-business AI service provision or reducing deployment costs) could materially affect the scale of adoption and productivity gains.
- Regulatory choices that limit AI exposure in tasks, occupations, or sectors can reduce productivity gains (the paper also estimates effects under exposure being 50 percent lower due to regulation).
- Attention to cross-country heterogeneity and potential widening of income differences within Europe is warranted when designing redistribution, labor market, and training policies.

*Source: 5. Conclusions and Policy Implications, wpiea2025067-print-pdf*

### Appendix Table A2: Determinants of AI adoption over the long run.

### Appendix Table A2: Determinants of AI adoption over the long run

### Key regression findings
- Primary dependent variable: the proportion of firms within each country–sector using at least one type of AI in 2023.
- Wage effects (main result):
  - wage (2016): coefficients reported across specifications, including 0.75*** (SE 0.22) and 0.52*** (SE 0.10).
  - wage (2021): 0.57** (SE 0.24).
  - The paper uses the coefficient 0.52 from Column (2) for calibrations.
- Other estimated coefficients (selected):
  - labor cost: 0.58*** (SE 0.17).
  - cost of capital: 0.35; -0.96** (SE 0.45); -1.34*** (SE 0.44) across specifications.
  - concentration: -0.03 (SE 0.02); -0.02 (SE 0.04); -0.02 (SE 0.05).
  - computer use: -0.05 (SE 0.06); 0.05 (SE 0.04); -0.03 (SE 0.05).
  - energy cost: -6.12 (SE 17.34); -25.61 (SE 19.70).
  - human capital: 0.09 (SE 0.11); 0.003 (SE 0.05); 0.01 (SE 0.07).
- Fixed effects and sample sizes:
  - Country FE: included in many specifications; Sector FE: included in most specifications.
  - # countries: 30, 30, 27, 30, 25, 25, 24, 24, 30, 28, 16, 17 (by column).
  - # sectors: 11, 11, 11, 11, 11, 11, 11, 7, 11, 10, 7, 7 (by column).
  - # observations: 265, 265, 252, 265, 226, 226, 202, 153, 265, 232, 102, 109 (by column).

### Interpretation and robustness
- Wages:
  - Strong positive relationship between AI adoption and wages is the most robust finding.
  - The wage–AI relationship underpins the paper’s cross-country calibrations for Europe relative to the US.
- Cost of capital:
  - Some evidence of negative effects (e.g., negative and statistically significant coefficient in Column (6)), but result is sensitive to specification (not robust to exclusion of sector-specific intercepts).
  - Data on cost of capital vary by country only, preventing inclusion of country fixed effects in those specifications.
- Market concentration:
  - No statistically significant relationship between AI adoption and concentration (Column (7)).
- Computer use / digitalization:
  - No significant relationship between AI adoption and the proportion of firms using computers (Column (8)).
  - Positive correlations with some EU digitalization indices may reflect mechanical inclusion of AI measures in those indices.
- Energy cost:
  - No statistically significant relationship between country-specific electricity costs and AI adoption (Column (9)).
  - Notes that large language model development is energy-intensive, but many adoption decisions involve renting services rather than building models.
- Human capital:
  - Initial positive association between tertiary-educated workers and AI adoption disappears once wages are controlled for (Column (10)), suggesting human capital may proxy for wages.

### Economy-wide adoption estimates and scenarios (Appendix Figure A1 and discussion)
- 2023 wage-based estimate:
  - Based on specification in Column (2); average estimated adoption rate of 13.8 percent (across countries and sectors, weighted by value added).
- Medium-term scenarios to reach average 18 percent:
  - Rescaled scenario (orange triangles): use formula (2023 estimate) × 18/13.8.
  - Shifted scenario (gray crosses): use formula (2023 estimate − 13.8) + 18.
- US–Europe wage differential and implication:
  - Yearly wages are estimated to be around 50 percent higher in the US (footnote calculation).
  - Using the 50 percent wage gap and the slope coefficient 0.52 from Column (2), the paper finds AI adoption should be around 5-percentage points lower in Europe than in the US.
  - Acemoglu’s 23 percent calibration for the US implies an 18 percent adoption rate for Europe as a whole.
- Country examples (using the shifted method to reach average 18 percent):
  - Bulgaria: 2023 adoption 4.0 percent; shifted medium-term adoption = 8.2 percent (=4.0 + 4.2).
  - Switzerland: 2023 adoption 25.1 percent; shifted medium-term adoption = 29.3 percent (=25.1 + 4.2).

### High-adoption AI-as-a-service scenario (Appendix Figure A2)
- Svanberg et al. (2024) estimate:
  - If firms can rent AI services (AI-as-a-service), US medium-term adoption could be ~80 percent (vs 23 percent if each firm must build its own systems).
- Europe under AI-as-a-service:
  - Given lower wages, an 80 percent US adoption implies a 75 percent adoption rate for Europe as a whole, with cross-country variation by wages.
- Scenario construction in Figure A2:
  - 2023 wage-based average = 13.8 percent.
  - Future scenario uses shifting formula (2023 estimate − 13.8) + 75 to ensure average adoption = 75 percent.
  - Rescaling method would produce adoption rates > 100 percent and is not shown.

### Descriptive statistics (Appendix Table A3, selected rows)
- AI adoption:
  - no. countries 30; no. sectors 11; no. country-sectors 298; mean 9.9; st dev. 10.2; min 0.0; max 84.7; source: Eurostat Digital Economy and Society.
- wage (2016):
  - no. countries 37; no. sectors 17; no. country-sectors 580; mean 17.0; st dev. 13.1; min 1.5; max 70.6; source: Eurostat.
- wage (2021):
  - no. countries 29; no. sectors 17; no. country-sectors 455; mean 19.6; st dev. 13.0; min 1.7; max 72.2; source: Eurostat.
- labor cost:
  - no. countries 37; no. sectors 17; no. country-sectors 581; mean 21.6; st dev. 16.8; min 1.8; max 89.5; source: Eurostat.
- cost of capital:
  - no. countries 27; no. sectors 18; no. country-sectors 482; mean 2.6; st dev. 1.0; min 1.2; max 6.0; source: European Central Bank.
- concentration:
  - no. countries 26; no. sectors 16; no. country-sectors 361; mean 30.4; st dev. 21.4; min 0.0; max 90.6; source: Eurostat Structural Business Statistics.
- computer use:
  - no. countries 26; no. sectors 8; no. country-sectors 201; mean 58.5; st dev. 24.2; min 17.3; max 100.0; source: Eurostat Digital Economy and Society.
- energy cost:
  - no. countries 34; no. sectors 18; no. country-sectors 605; mean 0.1; st dev. 0.0; min 0.1; max 0.2; source: Eurostat Energy Statistics.
- human capital:
  - no. countries 34; no. sectors 18; no. country-sectors 529; mean 39.3; st dev. 21.2; min 4.2; max 89.7; source: Eurostat.

### Microeconomic evidence on labor cost savings from AI (Appendix Table 4, selected results)
- Brynjolfsson et al. (2023) — Customer Care Agent / customer support:
  - Randomized trial with tailored GPT conversational assistant; on average issues resolved per hour increased by 14%, including a 34% improvement for novice and low-skilled workers.
- Dell’Acqua et al. (2023) — Individual contributor-level consultants:
  - Randomized trials with access to GPT-4; 43% improvement for consultants below average; 17% enhancement for those above average.
- Doshi et al. (2023) — Writers:
  - Randomized trial with Gen AI tool; improvements in novelty and usefulness of 6.7% and 6.4%, respectively.
- Kanazawa et al. (2022) — Low-skilled Taxi Drivers:
  - Quasi-experimental study of AI navigation system; narrows productivity gap between high- and low-skilled drivers by 14%.
- Korinek (2023) — Academic Economists:
  - Evaluation of LLMs for research tasks; 40% of research tasks experimented were found to be highly useful.
- Noy and Zhang (2023) — White-collar occupations / simple writing tasks:
  - Randomized trial using ChatGPT 3.5; 40% faster completion and 18% improvement in quality scores.
- Peng et al. (2023) — Software Developers / programming:
  - Randomized trial using GitHub Copilot; programmers completed tasks 55.8% faster.
- Schoenegger et al. (2024) — Data Forecasters:
  - Randomized trial with frontier LLMs; forecasters gained a 24% to 28% improvement in prediction accuracy.
- Kalliamvakou (2022) — Software Engineers:
  - Survey and participation in trial; 88% of developers reported increased perceived productivity when using GitHub Copilot.

### Regulation and adoption (Appendix 5)
- Regulated professions can slow AI adoption because:
  - They require formal qualifications, human oversight, licensing, and may entail extensive regulatory processes before AI use.
  - Integration of AI in regulated fields can be slower and more cumbersome, limiting productivity gains in those occupations.
- Data source and matching:
  - The European Commission Regulated Professions Database (covering EU27, Iceland, Norway, Switzerland, and the UK) lists 563 distinct generic professions regulated in at least one country.
  - The authors matched 366 of the 563 regulated professions to ISCO-08 4-digit occupational titles; 197 could not be matched due to overly general or overly specific names.
  - The authors note potential undercoverage in the database and take the database at face value despite possible gaps.

*Source: Appendix Table A2 and accompanying text, Appendix Figures A1–A2, Appendix Tables A3 and 4, and Appendix 5 from the provided IMF working paper content.*

### Appendix Table 4: Matched Regulated Professions to

### Appendix Table 4: Matched Regulated Professions to

### ISCO-08 mapping and methodology
- ISCO-08 is a four-level hierarchical classification system with 436 unit groups at the lowest level, represented by a 4-digit code.
- These unit groups encompass over 3,000 preferred occupational titles, indicating that each unit group covers multiple professions.
- Each occupational title may have up to 89 alternate titles, all of which are used to match regulated professions.
- Sources: European Commission Regulated Professions Database; ILOSTAT.

### Matched regulated professions (Percent of unique 4-digit ISCO code)
- CZ: 32.1
- SK: 26.1
- HR: 24.5
- AT: 24.0
- LU: 22.4
- CH: 22.4
- PL: 21.5
- DE: 21.5
- HU: 21.0
- SI: 20.6
- LI: 19.9
- PT: 19.4
- GB: 18.5
- IT: 17.1
- FR: 17.1
- BE: 16.6
- DK: 15.7
- GR: 15.5
- ES: 14.8
- NL: 14.8
- CY: 13.4
- RO: 13.4
- IS: 17.6
- FI: 12.9
- IE: 12.7
- NO: 12.5
- SE: 11.3
- MT: 11.1
- EE: 11.1
- LT: 7.6
- LV: 7.6
- BG: 6.7

(Note: country–value pairs are presented as in the source table; order preserved where possible.)

### EU AI Act (Appendix 6) — scope, timing, and risk thresholds
- The EU AI Act was proposed by the European Commission in April 2021; it aims to create a regulatory framework that supports innovation while establishing a consistent legal structure for AI within the EU, ensuring that AI technologies are developed and used safely, ethically, and with respect for fundamental rights.
- The Act specifies requirements based on level of risk (implemented through defining high-risk systems) and level of computational power (implemented through a risk threshold of computational power, which the source refers to as ‘capacity cap’).
- Implementation timing: While the legislation will take effect in different stages (for instance the section on high-risk systems will enter into force mostly in August 2026), for simplification the authors assume that the EU AI act is in effect already or fully anticipated by firms so that they already abide by the regulation.
- Risk levels and practical effects:
  - The AI Act prohibits and regulates unacceptable or high-risk systems while imposing lighter to minimal obligations for limited or minimal-risk applications.
  - High-risk systems refer to systems posing significant risk to safety in critical infrastructure or safety components of products; assessment of access to training, jobs, essential private and public services, and immigration control; and enforcement or administration of law, justice, and democratic processes.
  - High-risk systems face stricter requirements and obligations, which could substantially limit or delay the use of AI technologies in certain sectors and occupations that use or work with such high-risk systems.
- Computational power threshold:
  - The AI Act defines a risk threshold of general-purpose AI model based on the computational power used to train the models (at 10 25 floating point operations per second or FLOPS).
  - European Commission (2024) highlights that this is 10 times lower than a similar threshold in the United States that was included in the now revoked Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence and that a lower threshold can greatly influence the complexity, efficiency, and effectiveness of AI models trained and used.
  - European Commission (2024) notes that some state-of-the-art models already exceed the current threshold of the Act.
- Footnote timing detail:
  - EU AI Act Chapter III Section 1 Article 6 Point 1 will enter in to force in August 2027 and Section 4 will enter into force in August 2025.

### Data privacy laws and sector adjustments (Appendix 7)
- Three industries selected for potential impact of data privacy laws due to data intensity:
  - Information and Communication (NACE Rev. 2 code J)
  - Financial and Insurance Activities (NACE Rev. 2 code K)
  - Human Health and Social Work Activities (NACE Rev. 2 code Q)
- Rationale and literature:
  - These sectors are identified based on studies of the GDPR’s impact: Arcuri (2020) for the financial sector; Prasad and Perez (2020) and Chen et al (2022) for digital and technology companies; Yuan and Li (2019) for the health sector.
  - Examples from the literature:
    - Yuan and Li (2019) find that hospitals providing digital health services had to make more costly investments in data protection practices once the GDPR entered into force, which ultimately affected their financial performance.
    - Prasad and Perez (2020) suggest that AI use could be limited by the GDPR, posing a challenge particularly for digital services companies.
    - Arcuri (2020) notes that while greater investments in data protection infrastructures are necessary, these may ultimately boost productivity.
- Empirical adjustment applied in analysis:
  - In these sectors, the occupation-level AI exposure is multiplied by 0.5, implying that the share of tasks exposed to AI in a given occupation is half that of the same occupation performed in an industry not affected by data privacy laws.

*IMF WORKING PAPERS Artificial Intelligence and Productivity in Europe — Appendix Table 4*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2025/english/wpiea2025067-print-pdf.pdf_