Agentic AI and Liability

Abstract

Agentic AI represents the current frontier of generative AI — and a fundamental shift in what AI systems do. Unlike earlier AI tools that process data or respond to prompts, agentic systems autonomously take actions, make decisions, and increasingly operate with little to no human oversight. The distinction is best illustrated by contrast. Classical software follows explicit rules: "when the price exceeds X, sell." An AI agent given the instruction "optimize procurement costs for Q3" determines for itself which suppliers to contact, what terms to propose, and how aggressively to negotiate. The deployer has delegated not just execution, but judgment. This shift in autonomy carries a corresponding shift in consequence. A conversational AI that drafts a flawed recommendation does not, by itself, act on it. An agentic system can - move money, place orders, delete records, send communications, deploy code - all without requiring human approval at each step. Equipped with web browsers, APIs, and access to live databases, an AI agent operates in an environment too complex to fully constrain through sandboxed permissions alone. Even a well-designed agent, deployed in a sufficiently intricate business environment, will take actions its deployer cannot entirely foresee.