Governing Autonomous Agents: Accountability Beyond Capability
Digital Dignity Institute Research Team
Abstract
As AI systems gain the ability to act, communicate, and make decisions with reduced human supervision, governance frameworks must address not only what these systems can do but who remains responsible for their actions. This paper proposes a responsibility allocation framework for agentic AI across public, commercial, and personal contexts.
The Accountability Gap in Agentic AI
As AI systems gain the ability to act, communicate, and make decisions with reduced human supervision — booking appointments, managing finances, drafting communications, negotiating contracts — the question of who is responsible for their actions becomes both more urgent and more difficult to answer. Existing accountability frameworks assume a human actor who can be identified, held responsible, and required to make amends. Agentic AI systems disrupt this assumption: they act on behalf of principals who may not have anticipated or authorised specific actions, through chains of delegation that obscure the locus of responsibility.
A Responsibility Allocation Framework
We propose a responsibility allocation framework for agentic AI organised around three principles. First, responsibility should follow the chain of delegation: the party that authorised an agent to act in a given domain bears primary responsibility for the consequences of that action, subject to the agent's compliance with its authorisation. Second, platform operators bear secondary responsibility for harms that result from foreseeable misuse or malfunction of their systems. Third, where responsibility cannot be clearly allocated through the chain of delegation, a residual liability regime should ensure that affected parties have access to redress.
Public, Commercial, and Personal Contexts
The framework is applied across three contexts in which agentic AI is already deployed or imminently likely to be deployed: public administration (automated decision-making in government services); commercial deployment (AI agents acting on behalf of businesses in commercial transactions); and personal use (AI agents acting on behalf of individuals in personal and financial matters). Each context presents distinctive accountability challenges, and the framework is adapted accordingly.
Governance Recommendations
The paper concludes with governance recommendations for legislators, regulators, and platform operators. For legislators, we propose a statutory framework for agentic AI liability that codifies the responsibility allocation principles and establishes a residual liability regime. For regulators, we propose mandatory registration and audit requirements for high-stakes agentic AI deployments. For platform operators, we propose design requirements — including mandatory audit trails, override mechanisms, and scope limitations — that make the chain of delegation legible and contestable.
Key Findings
- 01
Existing accountability frameworks are structurally inadequate for agentic AI: they assume a human actor at the point of decision.
- 02
Responsibility should follow the chain of delegation, with platform operators bearing secondary responsibility for foreseeable harms.
- 03
A residual liability regime is necessary to ensure that affected parties have access to redress when responsibility cannot be clearly allocated.
- 04
Mandatory audit trails and override mechanisms are essential design requirements for high-stakes agentic AI deployments.
- 05
The governance challenge is not primarily technical but legal and institutional: existing frameworks must be adapted, not merely applied.
Methodology
Legal and philosophical analysis of accountability frameworks; comparative analysis of liability regimes across jurisdictions; case study analysis of agentic AI deployments in public administration, commercial, and personal contexts; expert consultation with AI researchers, legal scholars, and regulatory practitioners.