Import-driven adoption of automation boosts wages at adopters—about a 4% average gain five years after an automation spike—largely concentrated in small firms and among older, managerial and white-collar staff, while small firms see a notable rise in within-firm wage inequality (≈7.5%).

Main Finding

Investment in automation-related capital goods raises wages at adopting firms, but the effects are concentrated in small firms. Using Italian linked employer–employee and import data (2011–2019), the authors find an automation adoption premium of ~10% in raw comparisons and ~3% after controlling for worker sorting. In a causal, event-based difference-in-differences design, adopters’ average wages rise ≈4% by five years after an automation spike; this effect is driven by micro and small firms (≈5% wage increase), while medium and large firms show no significant wage change. Adoption also raises within-firm wage dispersion (e.g., ≈7.5% higher variance in 10–19 employee firms). Wage gains concentrate among older workers (45+), managers and white-collar employees; blue-collar and many other groups see stagnant wages (no group shows negative effects).

Key Points

• Raw automation wage premium ≈10%; after controlling for worker characteristics, the premium falls to ≈3%.
• Causal adoption effect (staggered DID, Callaway & Sant’Anna): average wages up ≈4% five years post-adoption.
• Firm-size heterogeneity: effects concentrated in small firms:
  • Micro/small (1–19 employees): ≈5% wage gain by year 5.
  • Small (20–49): similar but smaller increases.
  • Medium/large: effects statistically insignificant.
• Within-firm inequality rises after adoption:
  • Wage variance +≈7.5% in 10–19 employee firms; +≈3.8% in 20–49 firms.
  • Top-decile/median ratios show comparable upper-tail widening.
• Distributional winners: workers aged 45+, managers/middle managers, white-collar staff. No evidence of negative wage effects for any worker group.
• Robustness: results hold under alternative adoption definitions, spike thresholds, sample restrictions (to address intermediaries), and checks for anticipation effects.

Data & Methods

• Data (2011–2019):
  • Commercio Estero (COE): transaction-level import data (HS6 codes) — used to identify automation-related capital goods (robots, CNC machines, embedded automation).
  • ASIA-Employment Register: linked employer–employee data with worker demographics, occupation, contract type.
  • Labour Cost of Employees Register (LCER): wages and tenure.
  • Structural Business Statistics Register (Frame-SBS): firm financials (revenues, value added, wage bill).
• Sample: 139,941 importing firms (≈40% of Italian employment; ≈70% of manufacturing employment).
• Identification of adoption:
  • Define automation “spikes” as import events ≥3× a firm’s average automation imports (lumpy investment logic).
  • Treatment modeled as staggered, binary, absorbing (once adopted).
• Estimation strategy:
  • Mincer-style regressions to estimate cross-sectional wage premia (with and without worker controls).
  • Event-study / difference-in-differences using Callaway & Sant’Anna (2021) to estimate dynamic causal effects and accommodate heterogeneous timing and effects; condition on firm observables and test for anticipation.
  • Decomposition of effects across firm-size bins and worker subgroups (occupation, age, tenure, contract, gender).
• Robustness checks: alternative spike thresholds, alternative technology categories, restrictions to address trade intermediation, and tests for anticipation.

Implications for AI Economics

• Firm-size heterogeneity matters for technological wage impacts: policy and empirical work on AI/automation should not pool firms indiscriminately. Small firms can both raise average pay and amplify within-firm inequality after adopting automation.
• Distributional consequences are intra-firm as much as inter-firm: automation/AI can increase within-firm wage dispersion, shifting rents toward older, managerial, and white-collar workers. Analyses of AI’s labor-market effects should measure within-firm wage structure—not only mean wages or aggregate employment.
• Complementarities and skill-biased returns: the concentration of gains among managers and white-collar staff implies that automation/AI complements certain skill bundles and organizational roles. Policies aimed at skill-upgrading and reallocation should be targeted by worker type and firm size.
• Small-firm policy priorities: since SMEs account for the bulk of firms and employment in many economies, the finding that small adopters exhibit both wage gains and rising inequality suggests targeted interventions—e.g., incentives for inclusive adoption practices, support for internal upskilling, guidance on pay-setting—to ensure broader sharing of productivity gains.
• Caution for generalization to modern AI: the study identifies adoption via imported capital goods (robots, CNC, embedded automation) up to 2019. The results are informative for physical and industrial automation and likely relevant for some forms of AI that act as complementary capital, but the labor-market dynamics of cloud-based AI services and software-driven automation may differ. Future AI-era studies should replicate this firm-size heterogeneous approach and track software-driven adoption channels and spillovers.
• Research priorities: extend the approach to post-2019 AI diffusion, examine employment and reallocation/spillover effects across local labor markets, and investigate mechanisms inside firms (compensation policies, productivity changes, task reallocation) that mediate wage and inequality outcomes.