An AI agent makes a call that costs real money, or harms a real customer, and something predictable happens in the room afterward. It fills with candidates for blame and empties of owners. The vendor built the model. IT deployed it. The business asked for it. A human was somewhere near the loop. And the agent, the thing that actually acted, is software, so it can't be sued, fired, or made to explain itself. Everyone is adjacent to the decision. No one made it.
Here's what I want you to take from this issue, stated plainly because the law now states it plainly: that room is arguing about a question that has already been answered. When an autonomous agent makes the call, the deploying company answers for it. The open question, the only one you still control, is who inside the company that means. And you only control it before the call.
Who's accountable for an AI agent? The law has answered
Look at how uniformly the answer has come in. When Air Canada's chatbot invented a refund policy, the airline argued the bot was, in effect, a separate entity responsible for its own statements. The tribunal called the argument remarkable and rejected it: the company owned what its bot said. California made it statute: a law effective January 1, 2026 bars anyone who developed, modified, or used an AI system from arguing the AI autonomously caused the harm. A Munich court this year held Google directly liable for false statements in its AI Overviews, as Google's own content. And the EU AI Act builds the same assumption into its structure: obligations attach to providers and deployers, named legal persons, never to the system.
Different jurisdictions, different legal tools, one direction: accountability runs to the humans and entities behind the agent. "The AI decided" is not a defense anywhere it has been tried. Delegating the action to software never delegated the answering.
Anyone who has operated automation for a living already knew this in their bones. For 25 years, when a system I built made a call, "the system decided" was never an answer I could give a customer. The law is not inventing a new duty here. It's refusing to let the new technology dissolve an old one.
So the interesting question was never whether your company answers. It's the one the statutes can't settle for you: which person. And this is where most organizations are genuinely exposed, because they've treated the ambiguity as a kind of comfort. Nobody owns the agent, so nobody is on the hook, so nobody has the uncomfortable conversation. That comfort is an illusion with a timer on it. Accountability is assigned in advance or it's litigated afterward, internally if not in court, and ambiguity doesn't protect anyone. It just means the decision gets made later, under pressure, by people angrier than you, with the record you happened to keep.
The fix is unglamorous. Every agent gets a named owner: a person, not a department, because a department can't be accountable, only a person can, and "IT" ran the servers, not the decision. The owner of the agent is the owner of the business decision it automates. Write down what the agent may do, because the boundary you set is your foreseeability: an agent with unbounded authority makes every failure foreseeable, and "nobody could have predicted it" collapses the moment someone asks why there was no bound. And give whoever reviews the agent real authority and a record of using it. I made this argument about proof in Issue 006, and it lands harder here: a reviewer who couldn't actually evaluate or intervene isn't oversight, and when the record is examined, it shows.
None of this is about fear of the technology. It's the opposite. An agent with a named owner, a written boundary, and real oversight is an agent you can defend, expand, and let run harder, precisely because when someone asks who answers for it, you have an answer that isn't a shrug.
The agent made the call. Your company answers for it. The only thing still undecided is which person answers, and whether that gets decided by you, calmly, this quarter, or by an incident, loudly, later.