Federal agencies with more AI-exposed jobs are shifting staff from routine to expert roles while compressing wages, suggesting reallocation under public-sector constraints rather than mass layoffs. The pattern reflects institutional features—employment protections, standardised pay, and political oversight—that shape how technology-driven change plays out in government.

Main Finding

Agencies with higher concentrations of AI-exposed occupations (2019–2024) show systematic within-agency workforce reallocation: declines in routine/administrative employment shares, expansion of expert/professional roles, and wage-compression effects. These patterns are consistent with a political-economy model in which institutional constraints (civil-service protections, standardised pay, political oversight) mediate technological pressures, producing reallocation rather than straightforward displacement.

Key Points

• Exposure vs. deployment: The paper studies AI occupational exposure (AIOE) — task-level susceptibility to AI — not measured AI use. Because occupational exposure scores are fixed, observed agency-level exposure changes come from compositional shifts across occupations, not from changing exposure measures or measured adoption.
• Empirical patterns:
  • Substitution in routine tasks: Within-agency increases in AI exposure are associated with declines in routine/administrative employment shares. Baseline estimate: ΔA → ΔRoutine coefficient ≈ −0.301 (s.e. 0.030); with richer controls ≈ −0.198 (s.e. 0.049). A one-standard-deviation quarterly increase in exposure (0.0718) corresponds to ~1.42 percentage-point decline in routine share; a 0.10 increase corresponds to ~1.98 pp decline.
  • Complementarity for expert roles: Compositional shifts toward more AI-exposed occupations are associated with expansions in expert/professional employment shares (the paper reports symmetric positive associations).
  • Wage compression: Shifts in occupational composition under standardised public pay schedules generate compression in within-agency wage distributions (authors document wage-compression effects tied to compositional changes).
  • Heterogeneity across agencies: Different agencies exhibit diverse trajectories (e.g., Air Force, Navy, Veterans Affairs show steady increases in AI-exposed employment; Agriculture and Interior show little trend; discrete reorganisations cause step-changes in some agencies).
• Theory: The authors build a political-economy model (drawing on Garicano, Niskanen, Moe, etc.) with strategic interactions between budget-maximising agency directors and electoral-sensitive overseers, incorporating institutional frictions and political cost functions. The model yields four qualitative propositions matching the empirical patterns: routine substitution, expert complementarity, wage compression from compositional change, and political-equilibrium constraints on responses.
• Limitations acknowledged by authors:
  • Non-causal: Because occupational exposure is fixed and composition is endogenous to agency mission and constraints, the analysis documents associations from compositional reweighting but does not identify causal effects of AI deployment.
  • Exposure measure caveats: AIOE is constructed primarily from private-sector task descriptions (O*NET); it may overstate automation potential in accountability-intensive public-sector roles.
  • Decoupling of exposure and actual AI use: Institutional procurement, oversight, and legal constraints can prevent potential exposure from translating into deployed automation.

Data & Methods

• Exposure index: AI Occupational Exposure (AIOE) from Felten et al. (2021) — occupation-level task-to-AI-capability mapping, fixed over time in authors’ analysis.
• Employment data: Quarterly federal employment records from OPM’s FedScope Employment Cube, covering 2019Q1–2024 (23 quarters) and all federal agencies.
• Constructed measures:
  • Agency-level AI exposure = employment-share-weighted average of fixed occupation AIOE scores.
  • Routine vs. expert shares: Derived from personnel management classifications (occupational series, grades, appointment types).
  • Wage-compression proxies: Within-agency wage-structure measures (log salary changes, distributional proxies).
• Identification strategy:
  • Rely on time variation in agency-level exposure produced only by occupational reweighting (since occupation scores are fixed).
  • Changes-on-changes regressions: relate within-agency changes in routine/expert shares and wages to within-agency changes in AI exposure; include quarter fixed effects and cluster standard errors at the agency level.
  • Controls and robustness: Specifications add controls for changes in headcount, log salary, workforce composition (age, education, employment arrangements). Authors report robustness checks using alternative exposure measures and discuss potential measurement issues in appendices.
• Decomposition logic: ΔA_it = Σ_a_o Δw_io,t (since a_o fixed), so observed ΔA reflects composition (COMP) term only; TECH and CROSS terms are zero by construction.

Implications for AI Economics

• Institutional context matters: The labor-market effects of AI depend critically on institutional features. Standard private-sector models (free wage/employment adjustment) may mis-predict outcomes in public bureaucracies where pay grids, hiring protections, and political oversight constrain adjustments.
• Reallocation over displacement in constrained settings: In the federal government, adjustment appears to occur mainly via occupational reallocation (toward higher-exposure expert roles and away from routine positions) rather than mass layoffs—implying different distributional and transition dynamics than in competitive markets.
• Wage dynamics differ: Standardised public pay structures can generate wage compression when workforce composition shifts toward occupations with different market valuations; this has implications for retention and recruitment of technical talent and for incentive design in public organisations.
• Policy and governance:
  • Workforce planning: Agencies need targeted training and recruitment strategies to manage transitions (grow expert capacity, retrain routine workers for higher-complexity roles).
  • Procurement and oversight: Because exposure does not equal deployment, governance choices (procurement rules, procurement speed, accountability safeguards) will shape realized automation and its effects.
  • Democratic accountability: Political constraints that slow or condition AI deployment can preserve procedural safeguards but may also slow productivity gains; balancing efficiency and legitimacy is central.
• Research directions for AI economics:
  • Link exposure to measured deployment: Combining exposure indices with procurement/adoption data would allow causal assessment of AI’s labor impacts in the public sector.
  • Cross-sector comparisons: Comparing constrained public organisations vs. private firms can reveal how institutional frictions alter occupation-level adjustment and wage responses.
  • Micro-level task analysis: Task-level time-use and process-change data within agencies could show how AI augments vs substitutes at the task margin.
  • Distributional and dynamic modelling: Extend political-economy models to endogenise hiring, promotion, and retention under pay schedules to forecast long-run capacity and inequality effects within public organisations.

Limitations to keep in mind when using these results: the study measures exposure (not deployed AI), is observational about compositional change (not causal about technology effects), and uses exposure scores derived primarily from private-sector task descriptions that may misstate public-sector susceptibility.