AI decision systems can either preserve or displace managerial discretion depending on design and governance; transparent, override-capable tools combined with aligned accountability tend to support human augmentation, while opaque or rigid systems produce functional substitution.

Main Finding

AI-enabled decision systems do not deterministically augment or substitute managerial judgment. Whether AI functions as augmentation (expands meaningful human discretion) or functional substitution (reduces humans to validators) is sociotechnically constructed: it depends on system design features (e.g., transparency, override friction) interacting with governance arrangements (accountability, incentives). The project distinguishes perceived agency (how managers feel about control) from enacted agency (how they actually behave) and shows these can diverge.

Key Points

• Research gap: Most work treats AI as a decision input; less attention to how AI reshapes managerial discretion and responsibility in practice.
• Core research question: How do AI-enabled decision systems reshape human agency, and under what conditions do they act as augmentation versus substitution?
• Augmentation–Substitution continuum: Augmentation preserves interpretive authority and genuine override power; substitution occurs when AI effectively determines outcomes while humans rubber-stamp outputs.
• Mechanisms studied:
  • Algorithmic transparency/explainability — affects interpretability and perceived control.
  • Override flexibility/friction — affects enacted willingness to deviate.
  • Accountability framing and incentive systems — shape legitimacy and frequency of deviation.
• Perceived vs enacted agency: Transparency may raise perceived agency without changing enacted behavior if governance or procedural friction discourages deviation.
• Expected practical design levers: explanation interfaces that support interpretation, low-friction override paths, clear accountability mapping, and performance metrics that reward contextual judgment.

Data & Methods

• Overall design: Sequential multi-phase mixed methods to move from causal mechanisms to organizational context to integrative theory.
  • Phase 1 — Experiment (mechanism testing):
    • Participants in manager-role simulations.
    • Manipulations: algorithmic transparency, recommendation strength, override flexibility, accountability framing.
    • Dependent measures: perceived control, responsibility attribution, confidence, likelihood of override.
    • Purpose: isolate causal effects and strengthen internal validity.
  • Phase 2 — Qualitative fieldwork (contextualized enacted agency):
    • Semi-structured interviews with managers, analysts, governance officers.
    • Organizational artifacts: policies, workflow diagrams, audit procedures, decision logs.
    • Purpose: observe when/why managers override and how norms and incentives shape behavior.
  • Phase 3 — Integrative model development:
    • Synthesize experimental and field findings into a conceptual Information Systems framework linking design + governance → perceived & enacted agency → augmentation/substitution outcomes.
• Validity approach: Triangulation across methods; explicit distinction between perceived and enacted agency to improve construct validity; experimental control for internal validity; fieldwork for external validity.

Implications for AI Economics

• Measurement & identification
  • Caution in interpreting adoption as automation: observable AI use can mask substitution produced by governance/friction rather than by model capability. Empirical measures of “automation” should capture override events, auditing burdens, and performance-metric incentives—not only algorithm deployment.
  • Need to distinguish perceived from enacted agency in microdata; decision logs and override records are valuable new data sources.
  • Natural experiments (e.g., regulatory changes such as the EU AI Act) and firm-level variation in override friction or accountability mapping are promising identification strategies for causal impact on labor and productivity.
• Labor demand and skill composition
  • Firms that design systems to augment (low-friction override, interpretable outputs) will preserve demand for managerial judgment and higher-skilled decision labor; substitution-prone designs may reduce discretionary skill requirements while retaining formal managerial roles.
  • Hidden substitution may cause slower, uneven labor displacement: job titles remain but task content and bargaining power change.
• Incentives, agency costs, and welfare
  • Incentive systems and accountability allocation are central to whether AI yields efficiency gains or simply reallocates responsibility without real discretion. Misaligned incentives can produce moral hazard (over-reliance on algorithmic outputs) or risk-averse conformity.
  • Firms face trade-offs: tight adherence to algorithms may improve measured consistency but can reduce adaptive, context-sensitive decisions that generate economic value.
• Productivity, adoption, and competition
  • The productivity gains from AI depend on sociotechnical design; two firms with identical models can realize different returns depending on override friction and governance.
  • Firms that preserve meaningful human-AI collaboration may generate trust-based competitive advantages (e.g., in customer-facing or high-stakes decisions), influencing market structure.
• Regulation and compliance costs
  • Regulations that mandate human oversight or transparency (e.g., EU AI Act) alter both design choices and governance costs. Compliance can encourage augmentation-oriented designs but may also create procedural frictions that unintentionally promote substitution unless crafted carefully.
  • Policy should consider not only model risk but organizational procedure design (audit burdens, approval chains) to avoid perverse incentives.
• Directions for empirical AI economics
  • Empirical studies should collect and exploit: override rates, documentation burden, audit-trigger thresholds, performance targets linked to algorithm adherence.
  • Evaluate heterogeneous effects: sectoral differences (HR, finance, operations), decision criticality, and regulatory regimes.
  • Examine long-run dynamics: how firms evolve governance to optimize trade-offs between speed, accountability, and discretionary expertise; and implications for occupational mobility and wage structure.

Short takeaway: Assessing AI’s economic impact requires moving beyond binary automation metrics to measure the sociotechnical environment—design, onboarding, incentives, and accountability—that determine whether AI augments or substitutes human economic agency.