The Year AI Stopped Assisting and Started Deciding

In March, a logistics firm in Rotterdam signed a $2.4 million freight forwarding contract. The counterparty was an AI agent. Nobody at the firm thought twice about it — the agent had been negotiating terms, adjusting for tariff changes, and routing around port congestion for six months before that signature landed. It had a better track record than the human it replaced.

What’s happening in Rotterdam is happening everywhere, quietly, in procurement systems and trading desks and hospital scheduling platforms. The transition from AI as assistant to AI as decision-maker has not announced itself with a press release. It has arrived invoice by invoice, routing decision by routing decision, one automated approval at a time.

The Accountability Vacuum

Legal frameworks for autonomous systems are almost entirely unprepared for agents that act. Existing law assumes a human in the loop — someone who instructed the system, reviewed the output, and bears responsibility for the consequence. Agentic systems are specifically designed to remove that human from time-sensitive decisions.

“We are operating in a complete regulatory grey zone,” says Dr. Elena Vasquez, professor of AI law at the University of Toronto. “If an AI agent executes a bad trade, or signs a contract under duress, or approves a medical dosage error — current law has no clean answer for who is liable. The vendor? The deployer? The user organisation? We genuinely don’t know.”

The EU AI Act, which entered enforcement in 2025, classifies certain autonomous systems as high-risk and requires human oversight mechanisms. But the definition of “meaningful human oversight” has proven elastic enough to accommodate nearly anything.

What the Companies Say

The five largest enterprise AI vendors have each published agentic AI frameworks in the past 18 months. Each framework includes language about human oversight. None of them define precisely what that oversight must entail before an agent acts.

Where This Goes

The most likely near-term outcome is not regulation but litigation. A significant autonomous system failure will produce a lawsuit that forces courts to assign liability somewhere. That ruling will become the de facto standard until legislatures catch up.

In the meantime, enterprises are deploying agents anyway. The competitive pressure is too great, the efficiency gains too visible. The question of who is responsible when the machine decides wrong is being quietly deferred — right up until it isn’t.