AI Accountability: Who Owns Agent Work?

A customer receives a contract summary with a material error. A regulator asks why a filing contained an outdated figure. A vendor got a commitment nobody remembers authorizing. In each case, the trail leads back to work performed by an AI agent, and in each case, the same question arrives wearing a lawyer's face: who is accountable?

This question is going to define a decade of corporate governance, and it deserves better than the two lazy answers currently circulating. The first lazy answer is panic: nobody's accountable, it's a machine, we're headed for a responsibility vacuum. The second is dismissal: obviously the company's accountable, same as always, nothing new here. The panic is wrong because the vacuum only exists if you design one. The dismissal is wrong because "the company" is not an answer any regulator, judge, or wronged customer has ever accepted for long. Accountability that belongs to everyone belongs to no one, and organizations know this, which is why they invented org charts, sign-offs, and named owners in the first place.

Start with the one principle that should be non-negotiable, stated plainly enough to put in a policy: an agent can perform work, but it cannot hold accountability. Not partially, not "shared," not as a mitigating factor. When something goes wrong, "the AI did it" must be treated, inside the company, as an incomplete sentence, the same way "the spreadsheet did it" or "the intern did it" is incomplete. Somewhere behind every agent action is a chain of human decisions: someone delegated the task, someone configured the boundaries, someone decided the output could ship with or without review. Accountability lives in that chain. The whole game is making the chain explicit before the incident instead of forensically reconstructing it after.

An agent can perform work, but it cannot hold accountability. When something goes wrong, "the AI did it" must be treated as an incomplete sentence.

And the chain has more links than people expect, which is where the genuinely new thinking is required. Consider the contract summary with the error. Candidate one: the employee who delegated the task. Candidate two: the person who reviewed the output, or was supposed to. Candidate three: whoever configured the agent and decided what sources it draws from, because the error came from an outdated document it was pointed at. Candidate four: whoever decided this category of work could ship with sampling review instead of full review. Candidate five, in the cases that will make case law: the agent was part of a chain, one agent drafted, another checked, and the handoff between them is where the error slipped through, so who owns a seam between two machines?

A traditional org chart answers none of this, because it maps people, not delegations. Which points to the actual work in front of every company adopting agents seriously: building what amounts to an accountability map for non-human work. Not a bureaucratic monster. Three boring artifacts, mostly.

First, every agent has a named human owner, the way every employee has a manager. The owner answers for the agent's configuration, its access, and its standing behavior. When the owner leaves, ownership transfers explicitly or the agent stops. This single rule prevents the most common failure pattern already appearing in the wild, the orphan agent still doing work that no living employee is responsible for.

Second, every delegation of consequence has a record: who assigned what, under what instructions, with what review requirement. This sounds heavy and isn't; most of it can be captured where the work already happens. Its value is precisely proportional to how badly things went. On a good day nobody reads it. On the bad day, it's the difference between an incident review and a blame lottery.

Third, and hardest, review obligations are explicit rather than vibes-based. "A human checks important stuff" is not a policy; it's a hope. Real policy names which categories of output require human sign-off before shipping, which get sampled, and which are trusted, and it puts a name next to each sign-off. This is where companies discover the uncomfortable economics of the whole transition: review is the new scarce resource, and pretending otherwise, letting review requirements exist on paper while volume makes them impossible in practice, creates the worst of all worlds, accountability assigned to people who never had realistic means to exercise it. Assigning someone responsibility for reviewing a thousand outputs a day is not accountability. It's a designated scapegoat with extra steps, and the first serious lawsuits in this space will feast on exactly that gap between nominal and actual review.

It's worth saying what all this scaffolding is for, because it isn't primarily about blame. Blame is the failure mode. The purpose of clear accountability is that it changes behavior before anything goes wrong. The employee who knows their name stands behind the agent's output reviews differently than the one who assumes someone else is watching. The owner who knows the configuration is theirs updates the sources when the pricing changes. Ambiguity, not machinery, is what actually produces recklessness; every aviation and medical safety regime learned this generations ago. Clear ownership is not the enemy of adopting agents aggressively. It's the precondition, because the alternative to designed accountability is not freedom, it's the freeze that follows the first bad incident, when a company with no map overcorrects into banning everything.

One more prediction, offered with reasonable confidence. Within a few years, "human in the loop" will stop being a slogan companies say and become a design specification regulators and courts inspect. Which loop. Which human. With what actual capacity to intervene. The companies that treated the phrase as decoration will discover, in a deposition, the difference between a loop and a circle drawn around the org chart. The companies that did the boring work, named owners, recorded delegations, honest review economics, will discover they accidentally built something more valuable than compliance: an organization that can hand real work to machines and still, always, produce a human who can say "that was mine."

That sentence, "that was mine," is the entire subject. Everything worth building here exists so that when the question arrives, someone can say it, and mean it, and have known it all along.