California's rule that a company cannot defend itself by saying its AI acted autonomously has been in force since January, and companies are now stringing agents together so that one hands its work to the next. That combination creates a gap: the law says a human owns the result, but a chain of agents gives nobody an obvious human to pick.
A chain hides where the error began
Take a mid-sized distributor. One agent reads supplier emails about price changes. A second updates the quote spreadsheet. A third sends quotes to customers. A supplier writes that a product now costs a certain amount per box, and the first agent reads it as the price per pallet. The second agent has no reason to doubt the number it was handed, and the third sends a quote that is far too low. A customer accepts it.
No single agent looks reckless. Each did a sensible job with what it received, and the mistake was made at the first step and carried forward untouched. That is why a chain is harder to govern than one tool. Each agent's decision limits what the next one can see, so the later agents cannot catch what the earlier one got wrong.
The customer will not care which agent was at fault. Like the traveller in the Air Canada case, they relied on what your system told them, and the company will be the one answering.
One named person, and a record that lets them act
The answer that works is a named person, not a team or a review process, who owns every output of the workflow before it reaches a customer. This person is picked before the workflow starts. If nobody will put their name to it, the workflow is not ready to run. The person who builds an agent already owns what it does; a chain makes that rule harder to apply, because no one built the whole thing.
The owner can only act fast if each handoff leaves a short note. It should say what the agent was trying to do, what data it could see, how sure it was, and what it did. With those four lines, the distributor's owner can look at the bad quote, see that the first agent was unsure about the unit, and stop the chain in minutes. Without them, someone spends a day rebuilding what happened before they can even decide.
This also means keeping a record of each agent's steps, not only the final output. A log of final outputs shows that a wrong quote went out. A record of each step shows where it went wrong.
The limit will be the number of people willing to sign
If one person can honestly stand behind a handful of workflows, not hundreds, then the number of owners becomes a limit on how fast agents spread. A large company that runs 150,000 agents by 2028 will have them in chains. At five agents to a chain, that is 30,000 workflows, and at five workflows per owner, about 6,000 people whose names are attached. Companies that plan for those people, with time to review and the power to stop a workflow, will be able to run more agents than those that treat ownership as a form to fill in.
If you are asked to be that person, ask for three things before you agree: a way to stop the chain yourself, a step-by-step record you can read, and a limit on what can go out to customers without your review. The ownership question has to be answered before the workflow goes live, because after an incident the court will already have decided who answers.