Liability as a Requirement: Allocating Civil Responsibility for Agentic AI Systems

Abstract

The proliferation of agentic artificial intelligence (AI) systems autonomous, goal-directed agents capable of acting with minimal human oversight, has exposed fundamental inadequacies in existing civil liability frameworks. Traditional tort, product liability, and custodian regimes struggle to accommodate systems whose emergent behaviour arises from complex, opaque, and dynamically adaptive processes. This paper argues that civil liability allocation should be reconceived as a legal requirements specification problem. Drawing on comparative civil law analysis, the least-cost avoider principle from law and economics, and the Systems-Theoretic Accident Model and Processes (STAMP), we argue that the deployer occupies the primary, though not exclusive, locus of civil accountability. We propose five categories of legal requirements distributed across the AI supply chain, with deployer-centred obligations complemented by developer coliability in layered architectures. A worked case study demonstrates how the framework applies to a realistic multi-party deployment scenario and addresses the concern that existing causal doctrines are structurally inadequate for agentic AI harm.