Agentic AI & New Liability Frontiers

The shift from chat interfaces to agents that browse, call tools, move money, and update systems changes the liability map. A wrong answer creates reputational risk. A wrong action creates traditional legal categories — agency, negligence, contract breach, trespass to systems, unauthorized access — with an unfamiliar decision-maker in the loop.

Courts and regulators will follow control and foreseeability. In practice, liability can attach at four layers, often simultaneously:

Classical agency asks whether a human principal authorized an agent to act. With AI agents, the questions become: who set the objective, who granted tool scopes, who could have stopped the action, and who profited from autonomy? Documented authorization boundaries — what the agent may spend, send, delete, or purchase — will matter more than marketing claims about "human-in-the-loop."

Contract terms will push in parallel. Expect customers to demand action logs, spend limits, and indemnities tied to unauthorized agent behavior; expect vendors to demand scoped permissions and proof that the customer configured guards.

Plaintiffs will frame negligence around industry practice: Did you red-team tool use? Did you limit blast radius? Did you monitor for anomalous sequences? Defendants with evals, permission defaults, kill switches, and post-incident reviews will look very different from defendants who shipped autonomy because a demo worked.

The same controls reduce both incident rate and legal exposure. Design them to be producible:

Inventory every agent in production or pilot; classify actions by irreversibility; require approval for the top tier; paper the allocation of liability with customers and tool vendors; and write an incident playbook for "the agent did something we did not intend." The first wave of cases will be won or lost on whether those artifacts existed before the harm.