A literature review finds generative AI improves business decision-making only conditionally — task-model fit, organizational capabilities, and human oversight determine value; the authors call for integrated multi-level frameworks rather than single-level theories.

Main Finding

No single theoretical framework fully explains how generative AI (GenAI) affects enterprise decision‑making. Individual-level acceptance models (TAM/UTAUT), organizational adoption frameworks (TOE), strategic dynamic‑capabilities theory, the prediction‑machine economic view, and the jagged‑technological‑frontier (capability‑boundary) model each explain different, complementary aspects. The central research prospect is integrating these levels—task, individual, and organizational—into multi‑level models that account for task–model fit, shifting capability boundaries, and governance/explicability constraints.

Key Points

• Frameworks reviewed:
  • Individual adoption: TAM, TAM2, UTAUT — focus on perceived usefulness, effort expectancy, social influence, facilitating conditions.
  • Organizational adoption: TOE — technological, organizational, environmental contexts explain firm‑level readiness.
  • Strategic integration: Dynamic capabilities (sensing, seizing, reconfiguring) — explains why adoption alone may not yield sustained advantage.
  • Economic decomposition: Prediction‑Machine — frames AI as lowering prediction cost and increasing the value of complementary human judgment.
  • Task‑level boundary: Jagged Technological Frontier — capability advances are uneven across tasks; performance depends on task–model fit.
  • Governance/ethics: AI4People, managing‑AI frameworks — stress explicability, accountability, and appropriate human–AI decision structures (full human, hybrid, full automation).
• Empirical evidence (selected, field‑experiment and controlled studies):
  • Brynjolfsson, Li & Raymond: ~14% productivity increase in customer support; gains concentrated among newer/lower‑skilled agents.
  • Dell’Acqua et al. (BCG consultants, GPT‑4): +12.2% tasks completed and 40% higher quality when tasks were within the model’s frontier; performance fell below control for tasks beyond the frontier (false‑confidence/automation bias).
  • Peng et al. (GitHub Copilot): 55.8% faster completion on a standardized programming task.
  • Chui et al. (McKinsey/MGI): large aggregate potential value (trillions) concentrated in language‑intensive functions (customer ops, marketing & sales, software engineering, R&D).
• Observed patterns:
  • Heterogeneous productivity gains by skill level (knowledge leveling effect).
  • Task‑model fit is crucial: GenAI can amplify performance when used within its effective capability frontier and can harm when misapplied.
  • Organizational reconfiguration (dynamic capabilities) is often the bottleneck to converting tool access into sustained decision‑quality gains.
  • Governance and explicability constraints are most salient for high‑stakes, accountable decisions.

Data & Methods

• Study type: Qualitative, descriptive, comparative literature review.
• Scope: Peer‑reviewed information systems, strategic management, and applied economics literature on GenAI in enterprise decision‑making.
• Comparative axes used to synthesize literature:
  • Unit of analysis: individual, organization/firm, or decision task.
  • Theoretical lineage: technology acceptance, organizational adoption, strategic management, or economic.
  • Evidence type: survey construct validation, field experiments, controlled experiments, conceptual/theoretical frameworks.
  • Targeted outcome: adoption, strategic integration, or task‑level capability boundaries.
• Mappings provided in the paper:
  • Frameworks → unit of analysis/central explanatory variable (table).
  • Business functions → most explanatory framework (table).
  • Future challenges → theoretical gaps exposed (table).
• Limitations acknowledged:
  • Qualitative synthesis rather than formal meta‑analysis.
  • Dependence on a selected set of field experiments and cross‑disciplinary conceptual work; possibility of publication/selection bias.
  • Rapidly evolving GenAI capabilities imply findings may shift as models improve and new empirical work appears.

Implications for AI Economics

• Modeling complementarities: The prediction‑machine insight implies GenAI lowers prediction costs but raises the relative economic value of human judgment. Economic models of labor demand and firm value need to explicitly model prediction–judgment decomposition and endogenous task allocation.
• Heterogeneous and conditional productivity gains: Aggregate GDP or sector‑level estimates (e.g., MGI’s trillions) must account for task‑level heterogeneity and task‑model fit (jagged frontier). Macro estimates that ignore capability boundaries risk overstating realized gains.
• Distributional effects: Field evidence of larger gains for lower‑skill or less‑experienced workers suggests short‑to‑medium‑term knowledge leveling, with ambiguous long‑run effects depending on reallocation, upskilling, and dynamic capabilities investment.
• Dynamic uncertainty and investment risk: Firms face uncertainty from capability‑boundary drift (uneven and shifting improvements across tasks). This increases option value of staged investments and raises the importance of real‑options and dynamic investment models in firm valuation.
• Organizational frictions and returns to adoption: TOE and dynamic‑capabilities literatures indicate substantial complementarities between technology and organizational capital (leadership, absorptive capacity, reconfiguration costs). Economic estimates should include these fixed or quasi‑fixed costs and dynamic adjustment lags when predicting adoption timing and welfare gains.
• Governance, liability, and market design: Explicability and accountability constraints affect which decisions firms will automate or assist with GenAI. Regulatory and liability frameworks can materially alter adoption patterns and social welfare outcomes; economic analyses should model policy impacts on adoption incentives and risk allocation.
• Research directions for AI economics:
  • Develop multi‑level empirical models linking individual adoption (TAM/UTAUT), firm readiness (TOE), and task‑level model fit (jagged frontier) to firm productivity and labor demand outcomes.
  • Quantify costs of building dynamic capabilities (sensing/seizing/reconfiguring) and incorporate them into ROI and diffusion models.
  • Measure capability‑boundary drift empirically across task types and incorporate uncertainty into investment/resilience models.
  • Integrate explicability and liability constraints into models of automation choice, insurance markets, and regulatory impact on technology diffusion.
  • Use field experiments and firm‑level longitudinal data to estimate heterogeneous effects by skill, task, and organizational context to improve macro scaling of micro estimates.

Summary conclusion: For economic analysis of GenAI’s enterprise impacts, treat task–model fit, organizational complementarities, and governance constraints as first‑order factors. Predictive gains are necessary but not sufficient — value capture and welfare effects depend on how firms restructure decision processes, allocate judgment, and manage evolving capability boundaries.