Behavioral frictions—miscalibrated trust, cognitive overload, and weak governance—help explain why many corporate AI pilots never reach production, with coding assistants a high‑visibility example. The paper offers a practical diagnostic framework for managers to detect disengagement and guide institutionalization, alongside a research agenda for mixed‑methods validation.

Main Finding

Successful scaling of corporate AI pilots depends not only on initial user acceptance but on a two-level interaction between behavioral mediators (trust calibration, perceived usefulness, cognitive load, disengagement triggers, organizational influence) and concrete organizational enabling conditions (executive sponsorship, data readiness, operationalization capability, business-use-case fit, governance). Without alignment between these layers, pilots—even those that look promising—tend to remain isolated or be abandoned during transition to production. The paper synthesizes TAM and TOE frameworks with human–AI trust research and data-cascade evidence to produce a scaling-oriented conceptual model, illustrated with AI coding assistants as a focal case.

Key Points

• Scaling is a behavioral systems problem: sustained use requires routinization under uncertainty, accountability, and social evaluation rather than a one-time acceptance decision.
• Two-level model:
  • Level 1 (behavioral mediators): perceived usefulness, trust calibration (role-differentiated and evidence-based), interaction/cognitive costs, organizational influence (norms, policy), and disengagement triggers (accountability, reputational risk, emotional strain).
  • Level 2 (organizational enablers): executive sponsorship and continuity, data readiness (quality, lineage, accessibility), operationalization capability (MLOps/AIOps, integration, monitoring), business-use-case fit, and firm AI capability/maturity.
• Trust calibration is central: calibrated trust (matching expectations to actual competence and limits) enables explicit reliance policies, reducing oscillation between overreliance and rejection.
• Cognitive load and verification costs drive silent abandonment: even productive tools are discarded if they impose ongoing mental overhead without institutional support to reduce it.
• Data cascades and invisible data work create delayed downstream failures that pilots often mask; data readiness is therefore a primary scaling determinant.
• Governance and accountability framing matter: clear allowed-use policies, documentation norms, and managerial signaling externalize risk and lower individual disincentives to adopt.
• Well-being and autonomy influence behavioral sustainability: perceived surveillance or autonomy loss can convert “assistance” into perceived managerial monitoring, reducing long-term engagement.
• Practical measurement: the paper maps constructs to indicators (e.g., self-reported productivity, reliance frequency, verification intensity, data coverage, MLOps maturity, sponsor continuity) to support diagnostics and empirical testing.

Data & Methods

• Approach: structured literature synthesis + comparative conceptual analysis + inductive–deductive framework building.
• Evidence base: 21 peer-reviewed, scholarly, and institutional sources (primarily within the last five years), including TAM/TOE extensions, trust-in-AI research, NIST AI RMF, studies on pilot-to-production failures, and analyses of data cascades and MLOps practices.
• Case focus: AI coding assistants used as a representative, high-visibility class of workplace AI that concentrates trust, workload, and governance tensions.
• Outputs: a conceptual scaling model linking behavioral mediators to organizational enablers, a harm-aware governance framing (adapted from NIST), and a construct-to-indicator mapping (Table 1 in the paper) for operational measurement and future mixed-method validation.

Implications for AI Economics

• Investment vs. realization gap: economic returns to AI pilots are contingent on organizational complements (data infrastructure, operational capabilities, governance). Capital invested in models without these complements risks low or delayed returns—i.e., stranded R&D.
• Valuation of AI initiatives should incorporate scaling probability: when performing cost-benefit or ROI analyses, firms and evaluators should discount pilot gains by the probability of successful operationalization (which depends on the two-level factors).
• Complementarity and productivity measurement: AI is an organizational-general-purpose technology whose productivity effects depend on complementary assets (managerial practices, worker skills, data assets). Cross-firm differences in these complements explain heterogeneous productivity gains and diffusion patterns.
• Labor and wage effects: behavioral frictions (autonomy loss, well-being impacts, cognitive load) affect adoption and hence the realized labor-augmenting effects of AI. Models predicting displacement or augmentation should include adoption frictions and incentives for use.
• Policy and governance economics: regulators and corporate policymakers aiming to maximize social value from AI should incentivize investments in data quality, operationalization capacity, and governance (e.g., standards, auditability) rather than only model development. Liability and accountability regimes shape individual use incentives and therefore the diffusion of AI within firms.
• Firm strategy and competitive advantage: building firm-level AI capability (data pipelines, MLOps, governance, sponsorship) is an asset that compounds with time; early movers with strong complements can capture larger value from similar AI tools than laggards who only run isolated pilots.
• Empirical research agenda for AI economics:
  • Quantify the impact of organizational enablers on probability and timing of scale-up (panel studies across firms/sectors).
  • Measure economic losses from pilot-to-production failures (stranded projects, repeated experiments).
  • Field experiments or quasi-experimental studies testing interventions (e.g., explicit governance, MLOps investments, training) on adoption, productivity, and well-being.
  • Incorporate behavioral measures (trust calibration, cognitive load, perceived accountability) into structural productivity models to explain heterogeneity in AI returns.
  • Model macro-level diffusion incorporating firm heterogeneity in organizational readiness to better predict aggregate productivity effects of AI.

Takeaway: Economists, firm managers, and policymakers assessing the value of AI must treat scaling as a joint behavioral–organizational problem. Evaluations, investment decisions, and policies that ignore trust dynamics, data cascades, and operational readiness will systematically overstate the near-term economic benefits of pilot-stage AI.