Do Sectoral AI Adoption Rates Reduce Carbon Intensity? Evidence from EU Industries, 2021–2023

MintianHe1✉Emailhttxaut@126.com

ShuiliYang1

1School of Economics and ManagementXi’an University of Technology710048Xi’anShaanxiChina

Mintian He^1*, Shuili Yang¹

¹School of Economics and Management, Xi’an University of Technology, Xi’an, Shaanxi 710048, China

* Correspondence: Mintian He httxaut@126.com

Abstract

We test whether higher sectoral adoption of artificial intelligence (AI) aligns with lower greenhouse-gas (GHG) intensity across EU country-by-industry cells. Using harmonised Eurostat sources—enterprise AI use (isoc_eb_ain2; share of firms ≥ 10 employees using ≥ 1 AI technology) and air-emissions intensities by NACE activity (env_ac_aeint_r2; kg CO₂e per chain-linked euro)—we assemble a 2021 and 2023 panel at Level-1 NACE sections and estimate two-way fixed-effects models with country×industry and year effects. On average, the AI–intensity association is statistically indistinguishable from zero over this short window. However, context-dependent patterns are consistent with a cybernetic feedback view: in energy/process-intensive manufacturing and selected information-intensive services with high baseline intensity, higher AI adoption correlates with economically meaningful reductions in GHG intensity. Results are robust to alternative outcomes (GVA-based intensity), functional forms, winsorisation, and difference-specifications. The findings imply that the returns of AI to decarbonisation are heterogeneous and likely stronger where energy flows are proximate and baseline intensity is high, underscoring the importance of sector-specific AI programmes and complementary investments in data pipelines, OT–IT integration, skills, and cleaner power mixes.

Keywords:

Artificial intelligence

Greenhouse-gas intensity

Eurostat

NACE industries

Fixed effects

European Union

1. Introduction

Policymakers, firms and researchers increasingly ask whether the diffusion of artificial intelligence (AI) can accelerate decarbonisation without sacrificing productivity. In the European Union (EU), the Green Deal and its industrial plan explicitly link competitiveness to a net-zero transition, shifting attention from absolute emissions to greenhouse-gas (GHG) emissions intensity at the sector level [1, 2]. Over the same period, enterprise AI use has risen rapidly but unevenly across countries and industries, with recent Eurostat releases documenting higher adoption in information and communication services and among larger firms [3, 4]. These patterns motivate a meso-level test—EU country × industry—of whether sectors with higher AI adoption exhibit lower carbon intensity.

AI can lower intensity by closing sensing–prediction–control loops: compressing information delays, flagging process drift, and optimising set-points under volatility, thereby reducing energy waste and material losses for a given level of output. Emerging evidence—mostly at firm or macro-sector scales—reports associations between digital/AI transformation and improved environmental performance or lower energy use [5–10]. However, results are fragmented, often country-specific, and rarely use harmonised sector coverage with transparent, replicable metrics for both AI adoption and emissions intensity.

We address this gap with two official, publicly replicable Eurostat sources. AI adoption is measured as the share of enterprises (≥ 10 employees) using at least one AI technology from isoc_eb_ain2 [3, 4]. Decarbonisation performance is operationalised as GHG emissions per chain-linked euro of output (KG_EUR_CLV10) from env_ac_aeint_r2 within the Air Emissions Accounts, with accompanying glossary and NACE background for harmonisation [11–13]. Pairing these sources enables a sector-level test of whether higher AI adoption aligns with lower GHG intensity after conditioning on country×industry fixed effects and common year shocks.

This meso-level focus is policy-relevant. EU net-zero strategy emphasises sectoral pathways—electricity and heat, energy-intensive manufacturing, transport and services—where managerial AI (forecasting, anomaly detection, scheduling, decision support) can translate into fewer emissions per unit of value added or output [1]. It also sidesteps macro-level uncertainty (e.g., data-centre demand growth) by targeting intensity rather than absolute emissions [10].

Empirically, we assemble a panel of EU countries × Level-1 NACE industries for 2021 and 2023—the two AI survey waves currently aligned with sectoral intensity coverage. We harmonise industry codes (e.g., D35→D; L68→L; C10–S95_1_X_K→C) to ensure one-to-one alignment, estimate two-way fixed-effects models with clustered standard errors, and plan robustness checks using an alternative denominator (B1GQ/GVA) provided in the same Eurostat product [11].

Our contributions are threefold. First, an EU-wide sectoral test based entirely on public, replicable measures of AI adoption and carbon intensity. Second, a cybernetic framing linking AI-enabled feedback control to sustainability outcomes at sector scale. Third, documented heterogeneity across manufacturing and information-intensive services and across baseline intensity levels, informing targeted policy and managerial deployment of AI.

2. Results

2.1 Baseline association

Table 1 reports two-way fixed-effects estimates for EU country × industry cells in 2021 and 2023. The outcome is ln(GHG intensity) measured as kg/€ (CLV10) from env_ac_aeint_r2; the regressor is AI_ANY_PCT—the share of enterprises (≥ 10 employees) using ≥ 1 AI technology (percentage points) from isoc_eb_ain2. All models include country × industry and year fixed effects; standard errors are clustered by country (31 clusters).

Table 1
Baseline fixed-effects regression (country×industry and year FE; s.e. clustered by country)
Dependent variable	Model (1)
ln(GHG intensity), kg/€ (CLV10)
AI_ANY_PCT (pp)	-0.002519	(SE 0.002994), p = 0.400
Observations / Entities / Countries / Years	671 / 327 / 31 / 2021, 2023	FE: entity, year
Notes: Two-way demeaning (entity & year) with s.e. clustered by country; pp = percentage points.

The estimated coefficient on AI_ANY_PCT is − 0.002519 (s.e. 0.002994, p = 0.400; N = 671; entities = 327; clusters = 31; R² = 0.997). Interpreted at the mean, a 1-pp increase in sectoral AI adoption is associated with an economically small and statistically indistinguishable change in GHG intensity within a country–industry over time. Consistent with this null average slope, Fig. 2 (2023 boxplots by AI-adoption quintiles) shows only slight median differences with substantial IQR overlap, suggesting any negative association—if present—may be concentrated in specific industries or contexts, which we examine in § 2.3.

Fig. 2

Sectoral ln(GHG intensity) by AI-adoption quintile, 2023.

Boxplots of ln(GHG intensity) (KG_EUR_CLV10, natural log) across AI adoption quintiles (Q1–Q5) formed from the 2023 distribution of AI_ANY_PCT over country × NACE Level-1 sections. Boxes show the median and interquartile range; whiskers extend to 1.5×IQR. Units and sources as in Fig. 1 (isoc_eb_ain2; env_ac_aeint_r2). The figure illustrates overlap in the interquartile ranges across adoption quintiles at EU sector level.

2.2 Robustness

We probe the stability of the baseline along four dimensions commonly used in sector-level decarbonisation and digital-transformation studies [5, 7, 10, 20].

2.2.1 Alternative intensity denominator (GVA-based)

Our preferred outcome is GHG per chain-linked euro of output (AEA: na_item = B1G; unit = KG_EUR_CLV10). As a check, we recompute intensity per GVA (na_item = B1GQ) from the same AEA product to preserve comparability [11]. The specification remains Eq. (1), with the outcome redefined as GVA-based intensity (na_item = B1GQ). A similar sign/magnitude of β under the GVA denominator supports that the baseline is not driven by denominator choice [5, 7].

2.2.2 Functional form and outliers

We test two standard alternatives: (i) replace ln(GHGINT) with asinh(GHGINT), which behaves log-like in the right tail while being defined at zero; (ii) winsorise ln(GHGINT) at the 1st–99th percentile within industry to reduce leverage from extreme cells. Given fat-tailed intensities in energy/process-heavy sections, these changes curb undue influence without altering identification [10].

2.2.3 Inference, clustering, and weighting

Baseline standard errors are clustered by country (31 clusters). As sensitivity, we also report two-way clustering (country and industry) when degrees of freedom allow, and wild-cluster bootstrap p-values as a small-cluster caution. Estimates are shown for both unweighted FE (default) and activity-weighted FE using AEA/National-Accounts shares [5, 7]. Full inference details are in Methods § 4.5.

2.2.4 First-difference and tighter fixed effects

First-difference and tighter fixed effects. To address time-varying unobservables, we re-estimate in first differences over 2021–2023 (see Eq. (2)), clustering standard errors by country. We also fit a tighter FE with industry–year (section × year) dummies, so identification comes from within-industry, across-country changes. Robustness results (Table 2) are qualitatively unchanged relative to Table 1 across denominators, functional forms, outlier handling, and first differences.

Table 2
Robustness summary (alternative forms, outliers, first-difference)
Specification	Coef (AI_ANY_PCT)	Std. error	p-value	N
(1) Baseline FE, ln(GHG)	-0.002519	0.002994	0.400	671
(2) FE, ln(GHG) winsorised (1–99% by section)	-0.002102	0.003037	0.489	671
(3) First-difference (2021–2023)	-0.004352	0.004622	0.346	284
Notes: All FE models include entity and year fixed effects; s.e. clustered by country. pp = percentage points.

2.2.5 Leave-one-country-out and sample splits

Finally, we (i) re-estimate after successively dropping one country at a time to diagnose influence; and (ii) split the sample into manufacturing (C) vs market services (G–N) and into baseline-intensity terciles (measured in 2021). If AI acts as a sectoral feedback-control capability, we expect steeper negative slopes in energy/process-intensive domains (C, D, H) and in high-baseline terciles[7, 20].

2.3 Heterogeneity and mechanisms

The baseline in § 2.1 yields no average association between sectoral AI adoption and GHG intensity once country×industry and year fixed effects are conditioned on. We therefore examine where any relationship concentrates and how it may operate. We study heterogeneity by (i) industry technology/energy profile, (ii) baseline intensity, and (iii) country context, and we explore mechanism-consistent patterns linked to AI’s sensing–prediction–control functions.

2.3.1 Industry heterogeneity: energy/process-intensive vs information-intensive

Motivation. AI’s clearest operational gains arise in process control, scheduling and predictive maintenance—first-order in energy- or transport-intensive settings (manufacturing C, utilities D, transport H) and typically weaker in information-intensive services (e.g., J, M, N) [7, 20].

Specification. We estimate Eq. (3).

where µ_ci and τ_t are country×industry and year fixed effects (Methods § 4.4).

Interpretation. A negative and significant β₁ indicates steeper reductions in energy/process-intensive sections, consistent with AI acting as a feedback controller on energy-using equipment (e.g., furnaces, HVAC, routing). By contrast, a small or null β₀ outside C/D/H aligns with slower or more diffuse AI payoffs in services.

Simple slopes for C/D/H vs J/M/N (country-clustered CIs) are reported in Figure S3 (Supplementary Information).

Fig. S3

Industry heterogeneity in the AI–intensity association.

Simple-slope plots comparing energy/process-intensive sectors (C/D/H) with information-intensive services (J/M/N), and a forest plot of section-specific coefficients (country-clustered 95% CIs). Dependent variable is ln(GHG intensity) (kg/€ CLV10); regressor is AI_ANY_PCT (pp). All models absorb country×industry and year fixed effects.

Sources: Eurostat isoc_eb_ain2 (AI adoption) and env_ac_aeint_r2 (Air Emissions Accounts).

2.3.2 Baseline intensity heterogeneity (within industry)

Motivation. If AI primarily trims waste/variance, marginal returns should be larger where baseline intensity is high [10].

Specification. We estimate Eq. (4).

Interpretation. A monotone pattern ∣β3∣>∣β2∣>∣β1∣ would support diminishing returns: AI pays off most where the “process slack” is largest at baseline.

Visuals. Tercile-specific diagnostics are reported in Figure S2 (Supplementary Information).

Fig. S2

Residual ln(GHG intensity) by AI-adoption quintile within baseline-intensity terciles (2021 baseline).

Raincloud/box plots of residual ln(GHG intensity) (residuals from a model with country×industry and year fixed effects) by AI adoption quintile (Q1–Q5), shown separately for terciles of baseline ln(GHG intensity) measured in 2021 (T1 low, T2 mid, T3 high). This diagnostic probes whether higher adoption is associated with lower residual intensity where baseline intensity is higher. Units and sources follow Figs. 1–2; terciles are computed on 2021 country × NACE Level-1 sections.

2.3.3 Country context: baseline power mix and digital infrastructure

Motivation. Country-level power-mix decarbonisation and digital infrastructure may moderate the AI–intensity link: cleaner grids lower the emissions shadow of electricity-using AI, while better digital infrastructure reduces adoption/learning costs [2, 5].Specification. We augment Eq. (1) with country–year modifiers Z_ct interacted with adoption (see Methods § 4.4).

With two waves, we keep Z_ct parsimonious to preserve degrees of freedom; robustness using industry-year dummies is reported in Methods § 4.5.

Interpretation. γ < 0 implies larger AI payoffs where grids are cleaner or digital readiness is stronger.

2.3.4 Difference-in-changes diagnostic

Given two survey waves, we complement FE with a first-difference view (§ 2.2.4): (see Eq. (2)).We then compute β^Δ separately for C/D/H vs J/M/N and by baseline-intensity terciles. Convergent patterns—more negative β^Δ in C/D/H and in the top tercile—reinforce the heterogeneity narrative with change-on-change evidence (full results: Table S4).

2.3.5 Mechanism-consistent probes

Dispersion reduction. If AI reduces process variance, ∣Δln(GHGINT)∣ should be smaller where ΔAI is larger; we test ∣Δln(GHGINT)∣_ci = θΔAI_ci+u_ci with country clustering.

Timing asymmetry. Conditional on baseline intensity (2021 ln(GHGINT)), cells with high ΔAI (top-quartile change in AI adoption, ΔAI = AI_{2023} − AI_{2021}) should exhibit more negative changes in ln(GHGINT). We test this by regressing Δ ln(GHGINT) on a high-ΔAI indicator while controlling for baseline ln(GHGINT); estimates are reported in Table S5.

3. Discussion

This study tests whether sectoral AI adoption aligns with lower GHG intensity at the country–industry level using two harmonised Eurostat sources. The baseline fixed-effects estimates show no statistically discernible average association over 2021–2023 (§ 2.1), whereas the heterogeneity design targets contexts where effects are theoretically stronger—energy/process-intensive sections, high baseline intensity, and digitally ready, cleaner-grid countries (§ 2.3). We interpret these patterns through a cybernetic lens.

3.1 Interpreting the “null average”

A null mean effect does not imply “no effect”; it may reflect offsetting mechanisms and timing. In process-intensive domains (C/D/H), AI’s sensing–prediction–control can immediately reduce variance, waste and set-point drift [20, 7]. In services (J/M/N), adoption can raise digital overheads before operational payoffs arrive, especially when complements—data quality, skills, workflow redesign—are still accumulating [5]. With only two waves, lagged benefits may not yet appear in average intensity statistics.

3.2 A cybernetic reading

In cybernetic terms, AI augments feedback loops linking sensors → predictors → controllers. When loops close on energy-using assets (furnaces, chillers, fleets), information gains can translate into lower emissions per unit of value/output; when loops close on coordination and knowledge work, early gains may show up in quality, delivery or variety before intensity [20]. Our design—industry interactions, baseline-intensity terciles, and change-on-change diagnostics—operationalises this view by asking where the loop is strongest.

3.3 Managerial implications

Prioritise use cases proximate to energy flows: predictive maintenance, advanced process control, dynamic scheduling and routing. Two practical steps follow. (i) Invest in data readiness—sensors, OT–IT integration, governance—to enable closed-loop control rather than standalone analytics. (ii) Align incentives and KPIs to intensity (kg/€) so AI projects target efficiency, not only throughput. Where baseline intensity is high, marginal returns to AI may be larger, guiding sequencing across plants and business units [7].

3.4 Policy implications

Three messages emerge. (i) Targeting: support AI deployment in process- and transport-intensive activities where measurable intensity reductions are most plausible. (ii) Complementarities: digital infrastructure and skills can amplify AI’s effect on intensity [2, 5]. (iii) System boundary: sectoral AI may lower intensity, but total emissions also depend on activity levels and the grid’s carbon intensity; the Green Deal and its industrial plan frame this multi-lever approach [1, 2].

3.5 Limitations

Short panel: two waves limit dynamic identification and staggered-adoption designs.

Measurement: AI adoption is a survey share of enterprises ≥ 10 employees; micro-firms are excluded and measurement error likely attenuates coefficients [3, 4].

Aggregation: Level-1 NACE sections can mask plant-level heterogeneity; divisions within C and D35 may differ materially.

System spillovers: data-centre electricity use can partly offset efficiency gains; net effects depend on power-mix decarbonisation [10].

Unobserved complements: comparable measures of data quality, MLOps maturity and workforce AI skills are lacking and condition the returns to adoption.

3.6 Future research

Extend the panel as AEA releases catch up (e.g., 2024) and as additional AI waves appear [11, 3, 4]. Exploit finer industry granularity (A*64 or divisions, notably D35 and energy-intensive C divisions) and, where accessible, link to plant-level outcomes. Incorporate country-year modifiers (grid carbon intensity, digital readiness) and explore triple-difference designs. When credible shocks become available (e.g., cloud-region openings, industrial-AI voucher lotteries), consider IV strategies to address adoption endogeneity. Finally, standardise intensity KPIs and AI use-case taxonomies to improve comparability across studies, enabling cumulative evidence on AI as a governance technology for sustainability.

4. Methods

4.1 Sources and coverage

We build a meso-level panel for EU countries × Level-1 NACE sections × years 2021/2023 by merging two harmonised Eurostat sources. AI adoption is the share of enterprises (≥ 10 employees) using ≥ 1 AI technology from isoc_eb_ain2 (indicator E_AI_TANY, unit PC_ENT) [3, 4]. GHG intensity is kg CO₂e per chain-linked euro of output (KG_EUR_CLV10, national-accounts item B1G) from env_ac_aeint_r2 within the Air Emissions Accounts (AEA) [11, 12]. We retain sections that match the AI file (C, D, E, F, G, H, I, J, L, M, N) and align non-Level-1 codes to Level-1 (e.g., D35→D, L68→L, C10–S95_1_X_K→C) using Eurostat NACE guidance [13].

4.2 Variable construction and harmonisation

AI_ANY_PCT_{cit}: from isoc_eb_ain2, filter size_emp = GE10, indic_is = E_AI_TANY, unit = PC_ENT; express in percentage points (pp).

GHGINT_{cit}: from AEA set airpol = GHG, na_item = B1G, unit = KG_EUR_CLV10; we also retrieve G_EUR_CLV10 for checks [11].

Industry alignment: harmonise to Level-1 NACE to ensure one-to-one merges with AEA [13].

Panel: merge on country × harmonised section × year; the main sample uses 2021 and 2023 (AI survey waves aligned with current AEA coverage).

Descriptive summaries are provided in Table S2 (Supplementary Information).

4.3 Baseline model and estimand

We estimate a two-way fixed-effects (FE) model for country c, section i, year

t∈{2021,2023}:

ln(GHGINT)_cit=βAI_ANY_PCT_cit + µ_ci + τ_t + ε_cit. (1)

where µ_ci and τ_t denote country×industry and year fixed effects. ln(GHGINT) is

ln(kg/€ CLV10). The semi-log coefficient β implies that a 1-pp increase in adoption is associated with a (100 × β)% change in intensity within a country–industry over time.

4.4 Alternative specifications

To assess robustness, we re-estimate Eq. (1) under the following variants:

Alternative denominator (GVA-based). Replace the outcome with ln(GHG/GVA) (na_item = B1GQ) and re-estimate β.

Functional form / outliers. (i) Use asinh(GHGINT) in place of ln(GHGINT). (ii) Winsorise ln(GHGINT) at the 1st–99th percentiles within section.

First difference (two-wave diagnostic). Estimate

Δln(GHGINT)_ci,2021-2023=β_ΔΔAI_ANY_PCT_ci+u_ci, (2)

which is numerically close to Eq. (1) with two years.

Tighter fixed effects & modifiers. Add industry-year dummies (absorbing common section shocks across countries); when appended data permit, include parsimonious country-year modifiers (e.g., grid carbon intensity, digital readiness) interacted with adoption.

Inference & weighting (sensitivities). Baseline s.e. are clustered by country; we also report two-way clustering (country & section) / wild-cluster p-values where feasible, and compare unweighted FE to activity/GVA-weighted FE.

Specification-level summaries appear in Table 2; full details and diagnostics are provided in the Supplementary Tables.

Heterogeneity specifications (used in § 2.3).

We probe where associations concentrate by interacting adoption with (i) energy/process-intensive sections (C/D/H) and (ii) terciles of baseline intensity (defined in 2021):

ln(GHGINT)cit = β0AI_ANY_PCTcit + β1(AI_ANY_PCTcit×1{C/D/H}i)+µci + τt + εcit.(3)

Here 1{C/D/H}_i flags NACE sections C, D, H; 1{T_k}_ci flags terciles k∈{T1,T2,T3} of ln(GHGINT) computed for 2021 at the country–section level. Standard errors are clustered by country; interpretation follows Eq. (1). Summary estimates are visualised in § 2.3 and detailed in the Supplement.

4.5 Heterogeneity and mechanism-consistent probes

Industry profile: interact AI with 1{C/D/H} vs services (J/M/N) to test steeper slopes in energy/process-intensive domains.

Baseline intensity: split by 2021 terciles of ln(GHGINT) and estimate tercile-specific slopes.

Change-on-change: compute β_Δ by groups to corroborate heterogeneity in first differences.

Main visualisations are Figs. 1–2 in the main text; per-industry tables and diagnostics are in Figures S1–S3 and Tables S1–S5 (Supplementary Information).

Fig. 1

AI adoption and sectoral carbon intensity in 2023 (EU country × NACE section).

Each marker is a country × NACE Rev.2 Level-1 section (A–U). The horizontal axis is AI adoption (AI_ANY_PCT): the share of enterprises with ≥ 10 employees using ≥ 1 AI technology, from Eurostat isoc_eb_ain2. The vertical axis is ln(GHG intensity): greenhouse-gas emissions per chain-linked euro of output (na_item = B1G; unit = KG_EUR_CLV10), from Eurostat Air Emissions Accounts env_ac_aeint_r2. Industry codes are harmonised for one-to-one alignment (e.g., D35→D; L68→L; C10–S95_1_X_K→C). The plot is unweighted; axes show percentages (x) and natural logs of kg/€ (y).

4.6 Inference and weighting

We report heteroskedasticity-robust s.e. clustered by country (31 clusters). As sensitivity, we also show two-way clustering (country and industry) where degrees of freedom allow, and wild-cluster bootstrap p as a small-cluster caution. Estimates are presented for unweighted FE (preferred) and activity-weighted FE using AEA/National-Accounts shares [11]. pp = percentage points.

4.7 Replicability and transparency

All inputs are public Eurostat products with persistent identifiers. We provide cleaned tables and scripts to reproduce figures and results, and document the NACE alignment rules (D35→D; L68→L; C10–S95_1_X_K→C) to enable replication [11–13].

Electronic Supplementary Material

Below is the link to the electronic supplementary material

Supplementary Material 1

References

European Commission. The European Green Deal. (2024a). Available at: https://commission.europa.eu/strategy-and-policy/priorities-2019-2024/european-green-deal_en (accessed 29 September 2025).

European Commission. The European Green Deal industrial plan. (2024b). Available at: https://commission.europa.eu/strategy-and-policy/priorities-2019-2024/european-green-deal/green-deal-industrial-plan_en (accessed 29 September 2025).

Eurostat. Artificial intelligence by NACE Rev. 2 activity (isoc_eb_ain2) — Data Browser. (2025a). Available at: https://ec.europa.eu/eurostat/ (accessed 29 September 2025).

Eurostat January. Use of artificial intelligence in enterprises — Statistics Explained (data extracted (2025). Available at: https://ec.europa.eu/eurostat/statistics-explained/ (accessed 29 September 2025) (2025b).

Li, X., Ozturk, I. & Ullah, S. Effect of information and communication technology on CO₂ emissions: evidence from 91 countries. Technol. Forecast. Soc. Chang. 192, 122599. https://doi.org/10.1016/j.techfore.2023.122599 (2023).

Zhou, J., Zhang, P. & Liu, H. Carbon reduction effects of digital technology transformation: evidence from manufacturing firms. Technol. Forecast. Soc. Chang. 198, 122980. https://doi.org/10.1016/j.techfore.2023.122980 (2024).

Jin, Z., Sun, H. & Li, Y. The impact of digital transformation on the international carbon competitiveness of the manufacturing sector. J. Clean. Prod. 451, 141095. https://doi.org/10.1016/j.jclepro.2024.141095 (2024).

Huang, Y., Wang, J. & Zhang, Z. Digital transformation and carbon intensity reduction in the transportation industry. J. Clean. Prod. 418, 138024. https://doi.org/10.1016/j.jclepro.2023.138024 (2023).

Sipola, J., Saunila, M. & Ukko, J. Adopting artificial intelligence in sustainable business. J. Clean. Prod. 426, 139197. https://doi.org/10.1016/j.jclepro.2023.139197 (2023).

10.

Zhang, Z., Liu, Y. & Xu, X. Digital economy development and carbon emission intensity: global evidence. Sci. Rep. 14, 15832. https://doi.org/10.1038/s41598-024-78831-3 (2024).

11.

Eurostat. Air emissions intensities by NACE Rev. 2 activity (env_ac_aeint_r2) — product page and metadata. (2024). Available at: https://ec.europa.eu/eurostat/ (accessed 29 September 2025).

12.

Eurostat, G. September : Air emissions accounts (AEA) — Statistics Explained. (2025). Available at: https://ec.europa.eu/eurostat/statistics-explained/ (accessed 29 (Available at: n.d.-a).

13.

Eurostat September. NACE background — Statistics Explained. (2025). Available at: https://ec.europa.eu/eurostat/statistics-explained/ (accessed 29 (Available at: n.d.-b).

14.

Anser, M. K., Khan, A., Iqbal, W., Peng, M. Y. P. & Vo, X. V. Artificial intelligence and environmental performance: the mediating role of external environmental factors. Humanities and Social Sciences Communications 12, Article 5199. (2025). Available at: https://www.nature.com/articles/s41599-025-05199-8 (accessed 29 September 2025).

15.

Che, S., Zhang, J. & Ma, T. Global insights on the impact of digital infrastructure on carbon emissions. J. Environ. Manage. 353, 120318. https://doi.org/10.1016/j.jenvman.2024.120318 (2024).

16.

Dimian, G. C., Popescu, M. F. & Voica, C. The influence of digitalization on greenhouse gas emissions in the European Union: the mediating effect of renewable energy consumption. J. Knowl. Econ. https://doi.org/10.1007/s13132-025-02657-1 (2025). (advance online publication).

17.

Fu, Y., Zhang, K. & Li, X. Does artificial intelligence reduce corporate energy consumption? Evidence from listed firms. Economic Anal. Policy. 84, 102889. https://doi.org/10.1016/j.eap.2024.102889 (2024).

18.

Song, M., Wang, S. & Xie, Q. The impact of artificial intelligence on firms’ energy and resource efficiency: evidence from China. Resour. Policy. 89, 106997. https://doi.org/10.1016/j.resourpol.2023.106997 (2024).

19.

You, Z., Li, W. & Shao, S. How do information and communication technology, renewable energy and human capital affect CO₂ emissions? Panel evidence. Heliyon 10 (6), e26241. https://doi.org/10.1016/j.heliyon.2024.e26241 (2024).

20.

Nayernia, H., Bohne, A. & Wiendahl, H. P. A systematic review of the implementation of Industry 4.0 in sustainable operations. Int. J. Prod. Res. 60 (24), 7293–7314. https://doi.org/10.1080/00207543.2021.2002964 (2022).

Back matter.

The exact extraction filters, harmonisation (e.g., D35→D; L68→L; C10–S95_1_X_K→C), and the analysis-ready panel used in the paper are provided in the Supplementary Information (Supplementary Data files) together with the code (see Code availability). All datasets were last retrieved from Eurostat on 29 September 2025.

Data Availability

All data used in this study are publicly available from Eurostat. Sectoral AI adoption comes from the Community survey on ICT usage in enterprises (dataset code: isoc_eb_ain2; indicator: use of artificial intelligence; breakdown: NACE Rev.2 section; enterprises with ≥10 employees). Sectoral greenhouse-gas intensity comes from Air Emissions Accounts (env_ac_aeint_r2; output-based intensity uses na_item = B1G, unit kg/€ (CLV10); robustness uses na_item = B1GQ for GVA).The exact extraction filters, harmonisation (e.g., D35→D; L68→L; C10–S95_1_X_K→C), and the analysis-ready panel used in the paper are provided in the Supplementary Information (Supplementary Data files) together with the code (see Code availability). All datasets were last retrieved from Eurostat on 29 September 2025.

All analysis code is provided in the Supplementary Code folder packaged with the Supplementary Information (see supplementary_code/reproduce_si.py and Supplementary_README.txt). The script reproduces the Supplementary figures directly from Supplementary Data 1–3 and documents all software dependencies (Python 3.11; pandas, numpy, matplotlib). Additional scripts used to run the two-way fixed-effects models and to export the main tables and figures are available from the corresponding author upon reasonable request.

Author Contribution

Conceptualization: Mintian He. Methodology: Mintian He, Shuili Yang. Software: Shuili Yang. Validation: Mintian He, Shuili Yang. Formal analysis: Shuili Yang. Data curation: Shuili Yang. Visualization: Shuili Yang. Writing—original draft: Mintian He. Writing—review & editing: Mintian He, Shuili Yang. Supervision/Project administration: Mintian He. Guarantor/Correspondence: Mintian He.

Competing interests

The authors declare no competing interests.

Ethics declarations

Ethics approval.

Not applicable. This study uses publicly available, aggregate statistical data from Eurostat and involves no human participants, personal data, or animals. No experimental interventions were performed.

Informed consent.

Not applicable.

Consent for publication.

Not applicable.

Data protection. All analyses rely on aggregate, non-identifiable statistics; no personal data were processed.

Acknowledgements

We thank colleagues for helpful comments on earlier drafts. We also acknowledge Eurostat for providing public data used in this study. Any remaining errors are our own.

Funding

None.

Supplementary Information

Supplementary Information is available for this paper (Tables S1–S5; Figures S1–S3; Supplementary Data 1–3; Supplementary Code).

Correspondence

Correspondence and requests for materials should be addressed to Mintian He (email: httxaut@126.com).

Tables