European software firms that integrate AI into their enterprise systems tend to underperform financially on average, but leadership traits decisively shape outcomes; well‑aligned upper management can mitigate or reverse the adverse payoff.

Main Finding

AI integration in European software firms is associated with lower short-run financial performance. The negative effect is stronger for publicly listed firms, but can be moderated—positively—by age diversity in top management and, unexpectedly, negatively by manager gender diversity and the presence of PhD-holding managers.

Key Points

• Core result: Disclosure of AI-based systems is significantly negatively associated with log net profit (baseline RE estimate β ≈ -0.218; translates to ~19.6% decrease in profitability).
• Public firms: Interaction term AI×Public is negative and significant (β ≈ -0.575), indicating amplified adverse effects for listed companies.
• Upper-echelons moderation:
  • Manager age diversity positively moderates the AI → profit relationship (AI×age-diversity β ≈ +0.122; supports the idea that generational mix helps capture AI benefits).
  • Manager gender diversity unexpectedly negatively moderates AI’s effect (AI×gender-diversity β ≈ -1.485).
  • Number of PhD managers also negatively moderates AI’s effect (AI×PhD-count β ≈ -0.444).
• Robustness: Results hold under alternative specifications (FE and pooled checks), heteroskedasticity-robust SEs, and an IV 2SLS test (using managers-with-PhD and number-of-trademarks as instruments for AI disclosure).

Data & Methods

• Sample: Panel data from Bureau van Dijk Orbis (2012–2022); initial pool 7,508 European firms; focused on software sector: 736 companies and ~8,107 firm-year observations; final analytic sample after preprocessing 10,418 observations.
• AI measure: Binary indicator (1/0) based on voluntary textual disclosures (custom AI keyword dictionary across strategy, operations, policies, industry descriptions); 23.3% of firms flagged as AI users. SMOTE used to rebalance classes.
• Outcome: Firm financial performance measured as log(net profit). Missing profit values imputed using MICE; winsorized (1st–99th percentiles); log-transformed.
• Controls: Firm age, firm size (log employees+1), profit volatility (sd), intangible assets (log), country and year fixed effects.
• Estimation: Random effects panel regressions (preferred over FE to allow time-invariant covariates and partial pooling), clustered robust standard errors. Interaction models to test moderation by Public status and upper-echelons measures (manager age SD, proportion female managers, count of PhD managers).
• Diagnostics: GVIF for multicollinearity, Breusch-Pagan test for heteroskedasticity (used robust SEs), IV 2SLS to address endogeneity concerns.

Implications for AI Economics

• Short-run value creation is not guaranteed: Empirical evidence that AI integration can reduce profitability in software firms—cautioning against assuming immediate economic returns from AI adoption.
• Importance of integration depth and strategy alignment: Findings support the claim that lightweight or pilot AI implementations—especially when misaligned with existing systems and strategy—can produce negative financial externalities.
• Market signaling and disclosure: Public disclosure of AI use may magnify investor/stakeholder reactions or compliance costs, worsening short-term performance for listed firms.
• Role of managerial composition: Upper-echelons matter—age-diverse leadership can help realize AI gains (suggesting complementary skills and balanced risk horizons), while gender diversity and higher managerial academic credentials produced counterintuitive negative moderation in this study, indicating complex behavioral/governance dynamics that require further investigation.
• For policymakers and investors: Monitor not only AI adoption counts but integration quality, governance, and managerial makeup. Policies and due diligence should differentiate experimental/marketing disclosures from substantive, integrated AI deployment.
• Research implications: Necessitates richer measurement of AI (depth, use-cases, capex vs. opex), dynamic/longer-horizon studies to capture potential delayed returns, and deeper causal identification of when and why upper-echelons characteristics alter AI’s economic payoff.

Limitations to bear in mind: AI use is inferred from voluntary disclosures (measurement noise); SMOTE rebalancing and MICE imputation introduce modeling choices; IV strategy relies on instruments (PhD managers, trademarks) that may not be perfect; sample restricted to European software firms, so external validity beyond this sector/region is limited.