Algorithmic managers can both help and hinder productivity: software that preserves method and decision-making autonomy, offers developmental monitoring, and provides actionable transparency supports sustainable performance, whereas systems that narrow method discretion or use punitive surveillance erode long-term outcomes.

Main Finding

Algorithmic management produces an "autonomy paradox": digital managerial systems can simultaneously increase some forms of worker freedom (e.g., scheduling flexibility) while constraining others (method and decision-making autonomy). Whether algorithmic practices improve or undermine performance depends on which autonomy facet is affected, the psychological mechanisms triggered (e.g., perceived fairness, stress, empowerment), and contextual moderators (employment relation, dependence on algorithmic income, framing of monitoring). Algorithmic systems are most likely to support sustainable performance when workers retain meaningful method and decision autonomy, monitoring is developmental rather than punitive, and transparency is actionable (explanations, contestability, human review).

Key Points

• Conceptual clarity:
  • Autonomy is disaggregated into three facets: scheduling autonomy, method autonomy, and decision-making autonomy.
  • Performance outcomes are separated into psychological states (engagement, stress, perceived fairness), behavioral performance (efficiency, quality, adaptive performance), and sustainability outcomes (exhaustion, turnover intention, long-run viability).
• Five groups of algorithmic practices identified:
• Allocation and nudging (task offers, surge/bonuses, prompts)
• Method standardization (routing, scripts, enforced procedures)
• Surveillance and feedback (real-time monitoring, dashboards)
• Rewards and sanctions (ratings, deactivation, automated penalties)
• Transparency and voice (explanations, appeal mechanisms, human oversight)
• Mechanisms and tensions:
  • Scheduling flexibility can be genuine autonomy or coerced via nudges and incentives; the latter leads to overwork and fatigue.
  • Method standardization can raise short-run efficiency but reduce adaptive performance and intrinsic motivation.
  • Monitoring evaluated as developmental fosters learning and performance; monitoring framed as punitive increases stress, defensive behavior, and turnover risk.
  • Transparency matters only when it is actionable (explainability, contestability, human review); mere information can be demotivating or meaningless.
• Evidence base:
  • Structured integrative review (2015–Jan 2026): 358 records → 276 screened → 94 full texts → 41 retained sources (empirical, conceptual, policy documents).
  • Heterogeneous literature: platform qualitative studies, field studies in logistics/warehouses, experiments on electronic monitoring, governance documents.
• Limitations:
  • Single-author coding (no interrater reliability reported).
  • Heterogeneous methods and contexts; synthesis is thematic and proposition-driven rather than meta-analytic.

Data & Methods

• Review design:
  • Structured integrative review across Scopus, Web of Science, PsycINFO, ABI/INFORM, ACM Digital Library; backward/forward citation tracing.
  • Core search terms combined algorithmic management, autonomy-related constructs, and performance outcomes.
• Screening and coding:
  • PRISMA-style screening reported (numbers above).
  • Thematic coding matrix captured: algorithmic practice, autonomy facet, psychological mediator, outcome category, moderators, evidence type, and level of analysis.
  • Coding proceeded in three rounds to establish five practice groups, map to autonomy facets, and identify mediators/outcomes.
• Types of evidence in corpus:
  • Qualitative platform worker studies (rich experience data).
  • Field studies in warehouses/logistics measuring productivity, monitoring, quitting.
  • Experimental and meta-analytic evidence on electronic monitoring and surveillance.
  • Policy/legal/regulatory documents for governance prescriptions.
• Analytical approach:
  • Mechanism-focused, middle-range framework yielding six testable propositions (not all empirically settled).
  • Emphasis on cross-context comparison to explain divergent empirical findings.

Implications for AI Economics

• Modeling labor supply and incentives:
  • Distinguish scheduling autonomy from method/decision autonomy in labor supply models; nudges and dynamic pricing convert nominal choice into de facto constraints and alter effective wage/effort responses.
  • Incorporate non-wage incentives (nudges, quest bonuses) as behavioral constraints that can increase hours but raise long-run disutility (fatigue, burnout).
• Productivity and welfare accounting:
  • Short-run gains from method standardization and monitoring may obscure long-run costs (reduced adaptive performance, higher turnover, health-related productivity loss). Economic evaluations should account for sustainability externalities.
  • When estimating returns to AI-enabled management, include heterogeneous effects across performance dimensions (speed vs. quality vs. adaptability) and time horizons.
• Platform and market design:
  • Algorithmic opacity and unilateral sanctions (e.g., deactivation) affect bargaining power, reservation wages, and worker outside options; models of platform competition should incorporate these labor market frictions.
  • Transparency that is merely informational may not improve welfare unless paired with contestability and human oversight—policy/regulatory interventions should target actionable transparency.
• Regulation and governance:
  • Policy levers (mandating human review, contestability, explainability standards, limits on automated sanctions) can alter platform incentives and labor market outcomes; economic analysis should evaluate incidence and compliance costs.
  • Regulation may shift platforms toward more developmental monitoring, changing optimal investment in algorithmic systems and altering firm productivity-cost trade-offs.
• Empirical strategy recommendations for AI economists:
  • Identify natural experiments or platform policy changes (e.g., introduction/removal of real-time monitoring, changes in deactivation rules, transparency interventions) to estimate causal effects on hours, output quality, turnover, and health indicators.
  • Use difference-in-differences, RCTs (where feasible), IV strategies, and panel/longitudinal data to separate short-run productivity from long-run sustainability effects.
  • Structural models should allow autonomy facets to enter utility and production functions separately; calibrate using combined administrative, platform API, time-use, and survey data (psychological states).
  • Measure heterogeneity by employment type (contractor vs employee), income dependence on platform, task complexity (routine vs adaptive), and worker skill.
• Policy-relevant research questions:
  • Quantify trade-offs between short-term efficiency gains from algorithmic control and long-term costs in turnover, health, and adaptive capacity.
  • Evaluate how actionable transparency (explanations + contestability + human review) affects worker effort, trust, and platform market outcomes.
  • Estimate how differing allocations of method/decision autonomy shift aggregate productivity and redistribution of surplus between firms and workers.
  • Model externalities from algorithmic governance failures (e.g., sectoral skill attrition, reduced adaptive innovation) and optimal regulation.
• Practical takeaways for economists advising firms or policymakers:
  • When assessing AI managerial tools, explicitly model and measure autonomy facets separately.
  • Prefer designs that preserve method and decision-making autonomy where tasks require adaptation, and emphasize developmental feedback over punitive monitoring to sustain long-term performance.
  • Regulatory evaluation should consider both micro-level welfare (worker stress, fairness) and macro-level productivity dynamics (sustained quality, turnover costs, market competition).

Overall, the review provides a conceptually detailed framework linking specific algorithmic practices to differentiated autonomy effects and downstream economic outcomes. For AI economics, the core implication is to move beyond unitary notions of automation/monitoring and to embed multiple autonomy dimensions, psychological mediators, and temporal trade-offs into empirical and theoretical models.