Autonomous AI shifts decision authority away from humans and fragments responsibility, creating an accountability gap that corporate governance must close. The Dynamic Authority Delegation Model assigns clear roles to strategic decision-makers, algorithmic execution, and institutional oversight to make accountability for agentic systems auditable and practicable.

Main Finding

The paper argues that agentic, autonomous AI transforms IT from a passive tool into an "algorithmic proxy" that can make adaptive, high-impact decisions. Traditional, human-centered accountability models therefore break down. To fill the resulting accountability gap, the author develops the Dynamic Authority Delegation Model (DADM), which distributes responsibility across three layers—human strategic intent, algorithmic operational execution, and institutional oversight—and proposes concrete governance mechanisms (goal specification, boundary-setting, auditing, and institutional liability) to make corporate accountability practicable for autonomous systems.

Key Points

• Conceptual distinction
  • Automation vs autonomy: automation follows human-defined rules; autonomy selects strategies in uncertain environments and adapts, producing opacity and unpredictability.
  • Algorithmic proxy: when organizations delegate objectives to agentic AI, the system becomes a relational agent (principal–agent framing) that executes strategies on behalf of principals but lacks moral agency.
• Accountability taxonomy
  • Moral responsibility: requires intent and ethical understanding (only humans).
  • Causal responsibility: refers to which actor/system functionally produced an outcome (algorithms can be causally responsible).
  • Governance responsibility: institutional duty to anticipate, constrain, monitor, audit, and correct autonomous decision-making (organizational obligation).
• Two mechanisms that generate the accountability gap
  • Black-box attribution problem: technical opacity and complex training/data interactions make tracing causal chains hard.
  • Problem of many hands: many actors (developers, deployers, vendors, operators, managers) jointly contribute to outcomes, diffusing individual responsibility.
• Moral Crumple Zone: when accidents occur, responsibility is often unfairly concentrated on the weakest or most visible human actor (e.g., safety driver), masking organizational governance failures.
• Dynamic Authority Delegation Model (DADM)
  • Distributes responsibility across three layers:
  • Strategic intent (human/organization sets goals and incentives).
  • Operational execution (algorithmic agent implements strategy).
  • Institutional oversight (governance, audit, constraints, liability).
  • Practical governance actions include explicit goal definition, bounded autonomy, continuous monitoring and auditing, and retaining institutional liability for high-risk deployments.
• Contributions
  • Reframes algorithmic accountability as an IT management and corporate governance problem rather than only a philosophical or legal one.
  • Provides a theory-building, actionable framework for auditors, regulators, and managers.
• Limitations
  • Conceptual and qualitative design; uses purposive, illustrative case analysis (Uber AV, 2010 Flash Crash, COMPAS) and published reports rather than primary empirical testing.

Data & Methods

• Research design: qualitative, conceptual, theory-building.
• Literature synthesis: narrative review of AI governance, ethics, and information systems literatures (search protocol inspired by PRISMA-ScR; sources include Web of Science, Scopus, ACM, IEEE, Google Scholar).
• Case analysis: purposive cross-case comparison of three public failure cases chosen to represent physical (Uber autonomous vehicle accident), financial (2010 Flash Crash), and social (COMPAS recidivism scoring) domains.
• Evidence base: official investigation reports and secondary scholarly analyses; cases used illustratively to derive mechanisms and validate the DADM construct rather than for formal hypothesis testing.

Implications for AI Economics

Policy and regulation - Liability design matters: allocation of governance responsibility to organizations (per DADM) supports regulatory approaches that hold deployers/providers institutionally accountable (aligns with elements of the EU AI Act). Clear liability rules will shape firms' incentives to invest in governance vs. to externalize risk. - Regulatory arbitrage risks: without harmonized standards, firms may relocate deployment to jurisdictions with weaker governance or looser liability, affecting global diffusion of agentic AI. Firm behavior and organizational economics - Governance as a production input: managing algorithmic initiative becomes an explicit, potentially costly input (monitoring, audits, compliance, human oversight). Firms will trade off governance costs against gains from autonomy; this affects adoption timing and scale. - Principal–agent framing: DADM suggests new contract and incentive structures—contracts must specify goals, enforce constraints, and internalize downstream harm; performance metrics should account for externalities and ethical constraints. - Moral hazard and risk-shifting: if institutional oversight is weak or liability is ill-specified, firms may under-invest in safe deployment; conversely, strong governance obligations increase internalization of risk and may reduce hazard-taking. Markets for governance services - Demand for third-party auditing, certification, algorithmic insurance, and compliance tools will grow. These markets will influence cost structures of AI adoption and create new firms/industries (auditors, certifiers, insurers). - Insurance implications: actuaries will need to price risks arising from autonomous agents and interacting algorithms (systemic risks like market flashes require new underwriting models). Systemic and aggregate effects - Interaction externalities and systemic risk: cases like the Flash Crash show how interacting autonomous agents can generate large, sudden aggregate consequences. Regulators and economists should model endogenous risk from multi-agent interactions. - Competition and market power: firms with superior governance capabilities (or lower governance costs) may gain advantages; conversely, small firms may be crowded out by high compliance costs, altering market structure. Research agenda for AI economics - Formalize DADM in principal–agent and contracting models to quantify how governance costs, liability rules, and incentive structures affect adoption, investment, and welfare. - Empirically estimate governance cost elasticities: how much oversight, auditing, and constraint reduce accident probability and how those costs scale with system autonomy. - Insurance and pricing: develop models for underwriting algorithmic risk and for pooling systemic risks from interacting agents. - Macro-level modeling: analyze how liability regimes and governance standards influence industry concentration, innovation rates, and social welfare when agentic AI is deployed at scale. Practical recommendations (for economists advising firms/regulators) - Incorporate governance cost channels into technology adoption models rather than treating AI as a pure productivity shock. - Design liability and incentive mechanisms that internalize externalities and encourage investments in monitoring/auditing infrastructure. - Support development of standardized audit metrics and certification processes to reduce information asymmetries in markets for algorithmic governance services.

––––– If you want, I can: (a) produce a 1-page slide-ready summary for policymakers; (b) sketch a simple principal–agent model that formalizes DADM for economists, or (c) extract policy prescriptions tied specifically to the EU AI Act. Which would you prefer?