Wages & Compensation

Bottom Line

Across natural experiments (studies that exploit policy or external shocks), AI and automation raise wages in some settings and compress or stall them in others. Wage gaps widen. Gains cluster where work complements AI; routine and low-complementarity roles face wage pressure. Firm size, formality, regulation, and local labor supply shape outcomes, so averages hide sharp divergence.

What This Means in Practice

• Set differentiated pay bands. Pay premiums for advanced digital/AI and complementary human skills. Expect wage pressure in routine roles. Fund equity adjustments and mobility paths to retain staff.
• If you roll out AI, budget for targeted upskilling and gain-sharing. Track pay for non-complementary roles to catch compression early and reduce turnover.
• For platform and other nonstandard work, set wage floors and clarify worker classification. Use targeted collective action to counter wage suppression in posted-price markets (the platform sets prices).
• In workforce policy, prioritize formal-sector apprenticeships and portable benefits tied to AI-complementary skills. Do not rely on generic retraining alone; it rarely shifts workers into less automated jobs but can aid wage recovery.
• Manage place-based risk. Monitor local AI exposure and talent crowding. Where exposure is high and wages stall, pair employer incentives with local skill-building.

What the Research Finds

1) Wage levels respond heterogeneously to AI and automation adoption

• Italy: firms pushed to automate by imports saw average wages about 4% higher five years later, with bigger gains in small firms and for managers and white-collar workers Bisio et al. (2026).
• EU: workers in more easily digitized sectors earned about €0.52 more per hour (≈4.6%) in before-after comparisons with similar sectors Bieliaieva et al. (2026).
• Within establishments, pay for non-AI tasks and non-senior roles can fall as AI use rises Aoki et al..
• Beijing neighborhoods with high exposure to generative AI saw wage stagnation after ChatGPT’s release despite inflows of high-skilled workers, in neighborhood-level before-after comparisons He et al. (2026).
• Within Chinese firms, AI development is associated with a higher skill premium and a lower labor share of income, especially in non-state and more digitalized firms Liang et al. (2026).
• A cross-country study using external variation links AI adoption to higher top-quintile wages and modestly lower middle-quintile wages A. T. D..
• In Chinese A-share firms, executive pay rises with AI adoption, partly via productivity and financing channels Shen (2026).
• Matched before-after firm analyses link generative AI adoption to shifts in pay structure within firms O'Connor.

2) Skill premia and polarization: who gains, who loses

• Jobs requiring AI and new digital skills pay more Jaumotte et al..
• Korea: specialized digital skills carry ≈14% wage premiums vs ≈6% for general digital literacy, larger in big conglomerates Zhang et al. (2026).
• Colombia: a formal-sector “augmentation premium” appears for AI-complementary cognitive skills; informal workers do not capture these rents Espinal Maya (2026).
• Reviews report 15–22% premiums for AI-augmented workers, with downward pressure on routine tasks and rising inequality Dehouche; Mncube; Draper.
• Task-level evidence points to polarization: high-skill roles gain while mid-skill routine roles see wage compression Palake (2026).
• Models predict faster technology creation increases wage premia and manager–worker pay gaps, especially in dense labor markets Hassan et al.; Farach.

3) Inequality along formality, gender, caste, and place

• Formality: AI-complementary premiums appear in formal jobs but not informal work in Colombia Espinal Maya (2026). In China, digitalization raised incomes mainly via wages, with larger gains in urban and highly digital regions Xiong & Li.
• Gender: in Indonesia, routine-biased change briefly narrowed the gender wage gap as women moved to non-routine interpersonal roles, then widened it as returns in female-concentrated jobs fell Jamil et al. (2026). Simulations for Sweden suggest generative AI could widen the gender wage gap given pre-AI occupational sorting Gardberg et al..
• Caste: in India, graduates from Scheduled Castes/Tribes are less represented in AI-exposed occupations within districts; AI-exposed roles pay up to 20% more, implying wider caste wage gaps absent intervention Mishra (2026).
• Place: high-exposure Beijing districts saw wage stagnation post-ChatGPT despite talent inflows, driven by task de-skilling and crowding He et al. (2026).

4) Platforms, institutions, and compensation design

• Platform work pays less than comparable employment. Median adjusted gig pay is ≈22% lower than traditional jobs, though top-decile earners can see premiums; employee reclassification reduces supply but raises hourly pay ≈31% for those remaining Han. In posted-price markets, wages can be strongly suppressed; small, targeted worker coalitions imposing price floors can restore linear spending Stoica et al. (2026).
• Mobility constraints can suppress skilled-worker pay. Large U.S. green-card backlogs are associated with ≈12.2% lower H-1B wages Balamurugan. Public retraining under the U.S. Workforce Innovation and Opportunity Act rarely moves participants into less automated occupations; observed gains look like wage recovery, with apprenticeships performing best Jacobs & Canedy (2026).
• In Latin America and the Caribbean, non-wage labor costs average 51% of formal wages and formalization costs average 88% above informal wages, limiting how AI-driven wage gains translate into total compensation Alaimo et al..

5) Complementary human skills and compensation signals

• Employers pay more for complementary human skills, analytical thinking, teamwork, resilience, supervision, especially in AI-intensive workplaces Stephany et al. (2026); Zhang et al. (2026). Reviews and employer data show premiums for workers who use AI to augment their work Dehouche; Yu & Yu.
• AI-inferred personality and trait signals predict compensation and sorting among MBAs at scale Guenzel et al..
• In customer support and software, generative AI is associated with higher productivity; supervisory and verification skills likely earn higher returns than routine execution Horn Sarun (2026); Horn Sarun (2026).

New since the cutoff, recent natural experiments reinforce earlier patterns: wage gains at adopters concentrated in specific groups and firm types Bisio et al. (2026); sectoral wage increases with digitalization Bieliaieva et al. (2026); widening skill premiums within firms Liang et al. (2026); and wage stagnation where generative AI crowding is highest He et al. (2026). Alongside establishment-level declines for non-complementary roles Aoki, the balance points to unequal wage impacts shaped by complementarity, formality, and local crowding.

What We Still Don't Know

• Causal long-run wage paths beyond two to five years after adoption across economies; most evidence is short to medium term in a few regions and sectors.
• How compute-cost dynamics map to human wage bounds where AI and humans are substitutes; current evidence is theoretical without wage-setting calibration Zhu.
• The net wage impact of AI on informal workers and SMEs in low- and middle-income countries; early evidence shows limited augmentation rents in informality, but coverage is sparse Espinal Maya (2026).
• Distributional effects of platform governance experiments; theory and natural experiments suggest large pay impacts, but rigorous multi-platform, multi-jurisdiction evaluations are rare Stoica et al. (2026); Han.
• Interactions between immigration policy, AI diffusion, and wage setting for high-skill roles; current analyses document backlogs and associated wage suppression but lack large-scale natural experiments Balamurugan.