Employees' willingness to keep using in-house GenAI hinges on managerial authorization: when firms embed AI but don't explicitly permit its use, workers report higher guilt and perceived risk and lower intentions to adopt; formal authorization reverses this effect.

Main Finding

Embeddedness — i.e., integrating generative AI (GenAI) into an organization’s internal software — shapes employees’ psychological responses to GenAI use (guilt and perceived risk) and thereby their intention to continue using GenAI for work. Crucially, the effect of embeddedness depends on whether the organization has authorized the tool: organizational authorization conditions how embeddedness influences guilt and risk, and both guilt (affective) and perceived risk (cognitive) matter for continuance intentions.

Key Points

• Conceptual framing
  • Embeddedness: GenAI integrated into internal infrastructure (e.g., Copilot in Microsoft 365).
  • External GenAI: publicly accessible, outside company infrastructure (e.g., ChatGPT).
  • Authorization: whether organizational permission / formal approval for GenAI use exists. Unauthorized use = Shadow AI.
• Psychological mechanisms
  • Guilt: a self‑conscious emotion reflecting perceived personal responsibility or norm violation; likely to arise when employees are uncertain about appropriateness of GenAI use.
  • Perceived risk: individual assessment of uncertain/adverse outcomes (e.g., compliance, data leaks); shapes behavioral choices.
  • Both guilt and perceived risk are treated as complementary mediators between organizational conditions and continuance intention.
• Hypotheses (as developed)
  • H1: Embeddedness → lower feelings of guilt when using GenAI for work.
  • H2: Embeddedness → lower perceived risk when using GenAI for work.
  • Additional expectations: perceived risk and negative emotions (guilt) reduce continuance intention; authorization moderates the impact of embeddedness (abstract reports that embeddedness effects depend on authorization).
• Practical context
  • Rapid bottom‑up adoption of GenAI (employees adopting non‑authorized tools) creates Shadow AI and ambiguous norms, increasing individual responsibility and potential for emotionally driven avoidance or polarized adoption across employees.

Data & Methods

• Research design: vignette experiment with a 2×2 manipulation of GenAI type (embedded vs external) and organizational authorization (authorized vs not authorized / Shadow AI).
• Outcomes measured: self‑reported guilt (emotion), perceived risk, and intention to continue using GenAI for work (continuance intention).
• Analysis approach: between‑subjects experimental comparisons and mediation/moderation tests linking embeddedness and authorization to guilt and perceived risk, and onward to continuance intention.
• Notes on evidence: the paper is a completed ECIS 2026 research paper; the excerpt reports that results demonstrate an interaction between embeddedness and authorization and confirm the importance of both guilt and perceived risk as mechanisms. Exact sample size and statistical estimates are not provided in the excerpt.

Implications for AI Economics

• For firm-level adoption models
  • Embeddedness is a meaningful firm‑level treatment: embedding AI into workflows changes perceived adoption costs not only through technical integration but through psychological cost channels (reduced risk perceptions and guilt when authorized).
  • Authorization (policy) is a key moderator: rollouts without formal authorization (Shadow AI or implicit embedding via updates) can alter net adoption benefits and may produce hidden externalities.
• Productivity and welfare
  • When embedded and properly authorized, GenAI is more likely to be normalized and adopted, amplifying potential productivity gains; if embedded but unauthorized, psychological frictions (guilt, perceived risk) can dampen uptake or produce uneven access to productivity improvements across workers.
  • Shadow AI use creates unobserved adoption that may alter measured firm productivity, compliance risk, and distributional impacts (e.g., skill or task heterogeneity in who adopts).
• Externalities and regulation
  • Emotional and cognitive responses are channels for adoption externalities (peer effects, norms). Policy and governance (clear authorization, communicated rules, training) can shift these channels and thereby the diffusion and social costs of GenAI.
  • Economists analyzing regulation or incentives should incorporate non‑pecuniary psychological costs and the role of organizational signaling (authorization) when estimating effects of AI policies.
• Research and empirical strategies for economists
  • Natural experiments / staggered rollouts: exploit firm rollouts of embedded GenAI (e.g., Copilot deployments) to estimate causal impacts using difference‑in‑differences on usage logs, productivity, and compliance outcomes.
  • Randomized encouragement / RCTs: randomize authorization communications, training, or in‑product cues to measure effects on perceived risk, guilt (survey proxies), usage, and performance.
  • Combine administrative usage data with survey measures of emotion and risk to estimate mediation (psychological channels) and heterogeneity (by task type, skill).
  • Structural modeling: incorporate psychological costs (guilt, perceived risk) into adoption utility functions to forecast diffusion and welfare under alternative authorization/regulation regimes.
• Policy and managerial takeaways
  • Firms should treat authorization and communication as part of the technology design: embedding GenAI is not sufficient — clear authorization/guidelines reduce psychological frictions and encourage responsible uptake.
  • Monitoring Shadow AI is important for revealing hidden risks and adoption inequality; integrating authorized, vetted GenAI into workflows can reduce perceived risk and moral uncertainty.
  • Economists and decision‑makers should quantify both productivity gains and behavioral/psychological frictions when assessing GenAI interventions, since omission of the latter can bias welfare assessments and policy prescriptions.

If you’d like, I can: - Draft concrete empirical designs (variables, identification strategies, power considerations) for studying the causal impact of embedded GenAI rollouts on productivity and welfare; or - Convert these implications into a short policy memo for firm managers concerned about responsible GenAI deployment.