AI that augments human judgment is spreading faster than substitute technologies and is associated with net job gains: regions exposed to automation-type AI lose employment while regions exposed to complementarity-type AI gain it, yielding a small positive net employment effect once the task-complementarity channel is accounted for.

Main Finding

Adding a complementarity zone (AI as prediction that augments human judgment) to the Acemoglu–Restrepo task-based framework reverses the usual “robots → net job losses” conclusion for AI. Using Eurostat adoption data across EU NUTS2 regions, the paper finds that (i) AI that complements humans is being adopted much faster than automation AI (2.6×), (ii) regions exposed to automation-AI lose employment (ExpoAuto: β = −5.19, t ≈ −2.8), (iii) regions exposed to complementary-AI gain employment (ExpoComp: β = +2.85, t ≈ +3.6), and (iv) when calibrated into the general-equilibrium model the net employment effect of observed AI adoption is slightly positive (+0.042 percentage points), versus a substantially negative effect for robots (−0.37 pp).

Key Points

• Theory
  • Extends Acemoglu–Restrepo task model by inserting a Zone II (complementarity): tasks where AI prediction lowers prediction costs and raises the value/productivity of subsequent human judgment.
  • Decomposes AI effects on employment into four channels: displacement (−), complementarity/augmentation (+), productivity-driven demand expansion (+), and reinstatement/new-task creation (+).
  • Predicts opposite signs for automation vs. complementarity exposure: ExpoAuto < 0, ExpoComp > 0; net sign depends on relative magnitudes.
• Empirical patterns
  • Complementarity AI (text mining, NLG, ML, image recognition) adoption grew ≈2.6× faster than automation AI (autonomous robots, workflow automation) across EU countries 2021–2024.
  • Main NUTS2 regression (N = 59, ΔEmployment 2019–2023): ExpoAuto β = −5.19 (t ≈ −2.8); ExpoComp β = +2.85 (t ≈ +3.6). Long-run (2015–2024) coefficients are larger in magnitude.
  • Structural decomposition estimates show all four theoretical channels with expected signs; complementarity and new-task effects are estimated precisely, productivity effect less precisely.
• Robustness and replication
  • ExpoComp result robust to controls (routine-task share, sectoral HHI, trade exposure) and subsamples (excluding capitals, western/eastern Europe), remains positive in every subsample.
  • Pre-trend test: ExpoComp has no predictive power in 2015–2017 (t ≈ 1.5) but becomes significant in 2019–2023, consistent with the pandemic-era AI acceleration; ExpoAuto shows pre-existing negative trend (manufacturing decline).
  • First-stage diagnostics: instrument for ExpoComp is strong (F ≈ 11); ExpoAuto instrument is weak (F ≈ 0.9), partly reflecting confounding deindustrialization.
  • US state-level replication (31 states) yields qualitatively similar pattern: ExpoAuto negative, ExpoComp positive.
• Limitations
  • Relatively small regional sample (N = 59) and limited post-adoption time span constrain power and causal identification.
  • Automation results confounded by pre-existing manufacturing decline; first-stage weak for automation instrument.
  • Zone classification may require revision as AI capabilities evolve; firm-level microdata and longer time series would improve precision.

Data & Methods

• Data sources
  • Eurostat AI adoption surveys: isoc_eb_ai (AI by technology type, 7 categories, 27 countries, 2021–2025), isoc_eb_ain2 (AI by NACE sector), isoc_r_eb_ain2 (AI by NUTS2 region), and lfst_r_lfe2en2 (employment by sector and NUTS2 region, 2008–2024).
  • Regression sample restricted to 59 NUTS2 regions with complete data.
• Technology → model-zone mapping
  • Zone I (Automation): Workflow automation, Autonomous robots (Δp30 ≈ +0.75 pp).
  • Zone II (Complementarity): Text mining, Natural language generation (NLG), Machine learning (ML), Image recognition (Δp30 ≈ +1.92 pp).
  • Zone III (New tasks): ICT security, (marketing used as a proxy) (Δp30 ≈ +1.54 pp).
• Exposure construction
  • Bartik-style exposure: baseline (2015) regional employment shares by sector ℓri interacted with changes in the 30th percentile (Δp30) of AI adoption across countries (uses p30 as the technology frontier, following Acemoglu & Restrepo).
  • Two distinct exposure indices: ExpoAuto (sum over automation technologies) and ExpoComp (sum over complementarity technologies).
• Empirical strategy
  • Cross-regional regressions of ΔEmployment% on ExpoAuto and ExpoComp (separately and jointly), with robustness checks controlling for routine-task intensity, sectoral concentration (HHI), trade exposure, and excluding capital regions.
  • Pre-trend (placebo) checks using 2015–2017 employment changes.
  • IV first-stage diagnostics reported; ExpoComp instrument strong (F>10).
• Calibration
  • Calibrated general-equilibrium model uses standard parameters (ε, η, π, θ, φ) from literature; inserts estimated Δp30 for Auto and Comp to compute quantitative employment and wage effects. Resulting net employment effect +0.042 pp (vs. robots −0.37 pp).

Implications for AI Economics

• Conceptual
  • The prediction–judgment distinction matters: many modern AI applications are prediction tools that augment (not replace) human judgment; task-based models must allow for a complementarity zone to capture that.
  • General-equilibrium feedbacks (wages, demand, new tasks) can flip the net sign: observed adoption dominated by augmenting AI implies a net positive employment effect despite displacement in some tasks.
• Measurement & identification
  • Disaggregating AI by technology type (automation vs. complementarity) is essential to identify heterogeneous labor-market impacts; aggregated AI measures can mask offsetting channels.
  • Bartik-style regional exposure using technology-type adoption at a frontier percentile (p30) is a useful empirical device but subject to sectoral pre-trend confounds (notably in manufacturing).
• Policy
  • Policies that (i) incentivize adoption of AI that augments human judgment, (ii) support worker skills that complement AI (judgment, oversight, collaboration), and (iii) facilitate the creation and diffusion of new human-centric tasks may preserve or increase employment.
  • Because automation effects are concentrated in declining manufacturing regions, policy should pair AI deployment with targeted transition support (retraining, local development) to avoid localized hardship.
• Research agenda
  • Need for firm- and worker-level microdata, richer time series, and evolving classification of AI technologies to improve causal identification of displacement vs. augmentation.
  • Further work should quantify the magnitudes and persistence of the productivity channel (conversion of cost savings into jobs) and map which occupations/tasks are most likely to shift from prediction to judgment demand.

Summary takeaway: when AI is treated empirically and theoretically as a prediction technology that often augments human judgment, the labor-market impact can be materially different — and in current European adoption patterns, mildly positive — than the negative employment effect inferred from robot-driven automation alone.