A conceptual framework warns that AI investments rarely deliver strategic value on their own: firms convert generative and agentic AI into transformation and innovation only when culture, skills and governance absorb those capabilities, with employee AI literacy and innovation capacity mediating outcomes and regulatory context shaping effectiveness.

Main Finding

The paper develops a conceptual, theory-driven framework linking AI capability inputs (generative, predictive, and agentic AI) to organizational outputs (digital transformation maturity, innovation performance, and decision quality). It argues that AI capabilities only produce business value when converted through organizational absorption processes (culture, skills, governance), and that this conversion is mediated by innovation capability and employee AI literacy and moderated by the AI governance/regulatory environment and industry context. The work highlights an "AI adoption–value paradox"—substantial AI investment often fails to deliver strategic value when governance and organizational structures are not adapted—and proposes a mixed-methods empirical design (survey + interviews; PLS‑SEM + thematic analysis) to test the framework. The article is primarily conceptual and methodological: it proposes hypotheses, measurement constructs, and an analysis plan rather than reporting final empirical results.

Key Points

• Theoretical grounding: extends Dynamic Capabilities Theory to the era of agentic AI and integrates TOE (Technology–Organization–Environment), Diffusion of Innovation, and Decision Intelligence / Bounded Rationality perspectives.
• Novelty: explicitly includes agentic AI (systems that plan, use tools, and act autonomously) alongside generative AI as distinct inputs, focusing on their implications for strategy, governance, and organizational reconfiguration.
• Core logic: input → process → output. AI capability inputs must be absorbed (culture, skills, governance) to produce transformation, innovation, and decision-quality outcomes.
• Mediators and moderators:
  • Mediators: organizational innovation capability; employee AI literacy.
  • Moderators: external AI governance/regulatory environment; industry type.
• Managerial problem: without updated governance, decision-right allocation, and capability reconfiguration, AI pilots often fail to scale or generate strategic returns.
• Empirical plan: mixed-methods explanatory sequential design—structured survey of mid-to-large firms in finance, manufacturing, retail, and professional services; follow-up semi-structured interviews with C-suite and senior IT executives.
• Measurement: proposed multi-item Likert scales for constructs (AI adoption, digital maturity, innovation capability, decision quality, employee AI literacy, governance environment).
• Hypotheses (examples): H1—AI adoption positively affects digital transformation maturity; H2—innovation capability mediates AI adoption → performance; H3—AI governance quality moderates the AI adoption → performance link.
• Analytical approach: PLS‑SEM (variance-based SEM) with bootstrapping for path significance, mediation via product-of-coefficients with bootstrapped CIs, moderation via interaction terms; thematic analysis for qualitative data with inter-coder reconciliation.

Data & Methods

• Population/sampling: mid-sized to large enterprises that have operational/strategic AI use in finance, manufacturing, retail, and professional/business services. Survey respondents: managers/technical staff with AI decision access. Interviewees: C-suite and senior IT leaders.
• Instruments:
  • Structured survey: multi-item reflective scales (6–8 items for AI adoption and digital maturity; 5–7 items for innovation capability and decision quality; 4–6 items for governance environment; 4–5 items for employee AI literacy). Typical scales: 7‑point Likert for core constructs, 5‑point for mediators/moderators.
  • Semi-structured interview guide: governance models, organizational absorption barriers, examples of scaling or failed pilots.
• Analysis:
  • Quantitative: Partial least squares structural equation modeling (PLS‑SEM); composite reliability, AVE, HTMT for validity; SRMR, Q², PLSpredict for fit/predictive relevance; bootstrapped significance tests (e.g., 5,000 resamples).
  • Mediation: bootstrapped indirect effects (product of paths); assess partial vs full mediation.
  • Moderation: two-stage interaction approach; probe simple slopes at ±1 SD of moderator.
  • Qualitative: thematic analysis by multiple coders with reconciliation to ensure inter-coder reliability; used to explain governance practices and absorption dynamics not captured by survey items.
• Ethical/validity procedures: pilot testing, Cronbach’s alpha/internal consistency checks, anonymization, informed consent, and standard institutional ethics protections.
• Scope/limits of study: organizational/decision-level focus (not algorithmic performance); excludes consumer-only AI startups; limited to selected industries and firm sizes; the paper proposes empirical validation rather than reporting completed large‑scale empirical findings.

Implications for AI Economics

• Returns to AI investments are heterogeneous and conditional: economic models estimating returns to AI must account for complementarities (governance, skills, culture). Merely measuring technical adoption underestimates required organizational inputs to generate value.
• Complementary investments matter: human capital and governance (training, AI literacy, decision-right allocation) are likely to be key mediators of productivity gains—policy and firm-level cost–benefit analyses should include these investments.
• Adoption–value paradox and misallocation risk: observed low realized returns in some firms may reflect organizational frictions, not only technological immaturity; this affects macro estimates of AI-driven productivity and may produce persistent cross‑firm dispersion.
• Agentic AI amplifies dynamic-capability effects: as AI becomes more autonomous, firms’ ability to sense, seize, and reconfigure assets will determine comparative advantage. Models of technological diffusion should incorporate firms’ dynamic capability endowments.
• Regulatory design has economic consequences: the governance/regulatory environment moderates AI value capture. Tighter regulation can reduce certain risks but may also change the cost–benefit calculus—affecting diffusion rates, investments in on‑premise/“sovereign” models, and incumbent vs entrant dynamics.
• Platform and ecosystem dynamics: agentic and generative AI integrated into platforms may generate stronger network effects and winner-take-most industry structures; competition models should allow for increased returns to scale via embedded AI capabilities.
• Labor and task reallocation: agentic AI’s capacity for multi-step autonomy changes the nature of task complementarities; economists should model shifts in task portfolios, demand for supervision/AI governance roles, and potential reductions in some middle-skill tasks but increases in governance and higher-skill roles.
• Measurement implications: standard productivity metrics may not capture innovations in decision quality or reduced decision latency; new metrics (decision quality, innovation throughput) may be needed for evaluating AI’s economic impact.
• Policy levers: targeted subsidies or training programs that lower the cost of organizational absorption (e.g., AI literacy programs, governance toolkits) may raise aggregate returns to AI and reduce heterogeneity in realized benefits.
• Empirical research agenda: microdata linking firm-level AI adoption to outcomes should include variables capturing governance maturity, AI literacy, and innovation capability to avoid omitted variable bias. Natural experiments from regulation (e.g., phased AI rules) could be exploited to estimate causal effects.
• Welfare considerations: while agentic AI could raise firm-level productivity and innovation, distributional effects (industry concentration, labor displacement) and systemic risk from autonomous decision systems warrant explicit economic analysis and possibly redistribution/mitigation policies.

Limitations to note for economists using this work: the paper proposes a framework and empirical plan but does not yet report completed empirical estimates. Any economic modeling or policy recommendations should treat the framework's causal claims as hypotheses pending empirical validation.