Generative AI acts as a quasi-agent inside firms, reshaping evidence and framing across decision phases through knowledge coupling, decoupling and hiding. This shift makes hidden knowledge — not just hidden human actions — the central agency risk and calls for phase-sensitive safeguards for traceability, reconstructability, contestability and human justificatory ownership.

Main Finding

Generative AI (GenAI) creates a form of "quasi-agency" in data-driven decision-making (DDDM): it actively shapes framing, evidence construction, prioritization, justification, and organizational memory without formal authority or accountability. This reshapes organizational knowledge across the decision cycle via three mechanisms—knowledge coupling, knowledge decoupling, and knowledge hiding—and shifts principal–agent risk from hidden human actions toward transformations of hidden knowledge. Effective governance therefore requires phase-sensitive safeguards for traceability, reconstructability, contestability, and human justificatory ownership.

Key Points

• Concept introduced: GenAI quasi-agency — the capacity of GenAI systems to materially influence decision-relevant knowledge without formal intent or responsibility.
• Three recurring mechanisms by which GenAI transforms knowledge:
  • Knowledge coupling: linking previously separate pieces of information or stages in the decision cycle.
  • Knowledge decoupling: breaking or abstracting links between data, evidence, and decisions (reducing direct traceability).
  • Knowledge hiding: producing transformations or intermediations of knowledge that are opaque to humans or formal records.
• These mechanisms operate with varying intensity across decision phases (e.g., framing, evidence generation, synthesis, justification, retention).
• Governance needs are phase-sensitive; requirements differ across the DDDM cycle:
  • Traceability: being able to follow how evidence and claims were generated.
  • Reconstructability: being able to recreate the reasoning/evidence pathway.
  • Contestability: providing grounds for challenging AI-influenced outputs.
  • Human justificatory ownership: ensuring humans retain responsibility and ability to justify decisions.
• Agency risk shifts: rather than primarily hidden human action being the problem, the transformation and concealment of knowledge by GenAI become central sources of risk.
• The study extends principal–agent theory by foregrounding hidden knowledge transformations as an agency problem.

Data & Methods

• Design: Abductive qualitative research.
• Sample: 120 managers from Italian small and medium-sized enterprises (SMEs) who actively use GenAI in recurring DDDM tasks.
• Instrument: Ten open-ended interview questions, aligned to stages of the decision cycle (framing, evidence construction, synthesis, justification, retention).
• Analysis:
  • Reflexive thematic analysis to identify recurring patterns.
  • Gioia-informed coding procedures to build higher-order concepts and a phase-sensitive framework.
• Outputs: A phase-sensitive model of GenAI quasi-agency mapping how the three mechanisms vary in intensity across decision phases and what governance requirements they imply.
• Limitations (implicit from method): qualitative and context-specific (Italian SMEs), so findings are interpretive and may require quantitative validation or wider-sector replication to generalize.

Implications for AI Economics

• Information asymmetries and market efficiency:
  • GenAI-induced knowledge hiding increases information asymmetry between principals and agents, potentially degrading market efficiency, pricing, and contract performance where transparent evidence is critical.
• Principal–agent contracts and incentives:
  • Contract design must account for risks from opaque knowledge transformations (not just hidden effort/actions). New clauses or monitoring mechanisms may be needed for traceability and reconstructability of AI-augmented outputs.
• Liability, regulation, and auditability:
  • Regulators and auditors should prioritize standards and tools for provenance, explainability, and reconstructability of GenAI outputs across decision phases.
  • Economic models of liability should incorporate quasi-agency: allocation of responsibility where AI materially shapes decisions but lacks legal personhood.
• Organizational investment and factor allocation:
  • Firms may need to reallocate resources toward epistemic governance (provenance tooling, logging, human review capacity), shifting investment from pure productivity gains to governance and compliance.
  • Returns to such investments affect comparative advantage across firms and sectors: those that manage GenAI governance well may capture more value.
• Labor, skills, and task reconfiguration:
  • Demand will rise for roles that can contest and justify AI-influenced decisions (interpretive skills, auditing, epistemic oversight), affecting wage structures and human capital formation.
• Innovation and adoption dynamics:
  • Phase-sensitive governance costs may shape adoption: tasks/phases where GenAI causes more knowledge hiding or decoupling will face higher governance costs, slowing uptake or limiting use to less sensitive decision phases.
• Empirical research directions:
  • Quantify economic impacts of knowledge-hiding mechanisms on decision quality, error rates, and transaction costs.
  • Model equilibria where firms choose levels of GenAI use and governance investment given regulatory regimes and market penalties for opaque decisions.

If you want, I can (a) map the three mechanisms to specific decision phases with examples, (b) sketch a simple principal–agent contract modification that accounts for quasi-agency, or (c) propose empirical designs to quantify the economic effects described above.