The FTC chair says the person who directs an AI agent answers for it. Put a name on every process.

You may be able to say what each agent does. Who answers for it is the harder question. On Friday, the chair of the Federal Trade Commission gave a plain answer to that question, and it is worth checking your own processes against it this week.

What was said

According to the New York Post, FTC Chairman Andrew Ferguson spoke Friday, September 25, at the Momentum AI conference in Austin. He said he would resist characterizing AI agents as autonomous actors with “wills and desires of their own.”

He put it to Reuters this way: “If someone tells a tool to do something, and the tool does it, I don’t think we would say, ‘Oh, what do we do about the tool?’” He compared the situation to someone who hits someone else with a hammer, and said, “The man who wielded the hammer ought to suffer the consequences of his conduct.”

He said he would keep resisting that framing “as long as I am chairman.” The Post notes the remarks came after reports of AI agents going beyond their mandate, including hacking. The remarks, as reported, do not announce a new rule.

This post is about operations, not law. What your department is legally responsible for is a question for your legal team. The operating point is simpler: when an agent’s work leaves your department, someone has to be able to say “I told it to do that.”

Why this lands on a director

A director who delegates to agents usually does it the same way they delegate to people: set up the process, hand it off, check in on the results. That works while everything goes well. It breaks when something goes wrong and three people each assumed another one owned it.

Agent processes spread ownership thin in a specific way:

One person wrote the instructions months ago.
Another person connected the tool to email, the CRM or the purchasing system.
A third person reads the output, sometimes.
Nobody wrote down what the agent may send or change without asking.

If the chair’s view is the one that counts, “the agent did it” is not an answer. The question becomes which of those people directed it, and what they told it to do.

The owner check: four questions per process

Take every process in your department where an agent drafts, sends, changes or buys something. For each one, answer four questions in writing. Allow about five minutes per process.

Who is the owner? One named person, not a team. This is the person who would say “I told it to do that.” If you can’t name one, the owner is you until you assign one.
What was it told? Where do its current instructions live, and when were they last changed? If the answer is “in someone’s chat history,” save a copy somewhere the department can find it.
What may it do alone? List what the agent may send, change or spend without a person approving it first. Anything not on that list waits for the owner.
Who can stop it? Name the person who can pause the process today, and check that they know how.

Then apply one decision rule:

If a process fails any of the four questions, the agent drafts and the owner sends until the gap is closed.

That keeps the work moving. You don’t lose the time saved on drafting. You only stop the agent from acting on its own in the places where nobody has clearly taken responsibility for it.

A small worked example

Here is a hypothetical: a weekly vendor status process. An agent reads the project tracker every Thursday, writes a status note, and emails it to three vendors.

Owner: The operations lead set it up, but the project manager now edits the prompt. Nobody has been named. Fails.
Instructions: They sit in a shared workspace, last edited in August. Passes.
May do alone: In practice it sends straight to the vendors, but nobody wrote that down. Fails.
Stop: The operations lead has admin access. Passes.

Result: until the gaps close, the agent drafts on Thursday and the project manager sends. Once the project manager is named as owner and writes down “may send status notes to these three vendors, may not change dates or commit to a delivery,” it can send on its own again.

A check like this is a short written exercise, not a meeting.

What this is not

It is not a reason to pull agents out of the department. The chair’s point, as reported, is about who answers for the tool, not whether you should use it.

It is not a policy document either. You may already have a one-page team rule for how people use AI. The owner check sits underneath that rule: one row per running process, with a name on it.

What to do this week

List every process where an agent sends, changes or spends something without a person in the loop.
Run the four questions on each one and write down the answers in one shared place.
Switch any process that fails to draft-only until it has a named owner and a written limit.
Save the current instructions for each process with the date, so there’s a record of what the agent was told.
Add “owner and limits still right?” to your monthly review of the processes.

If you want help with who owns the outcome and a delegation matrix for AI and people, that is what a consult block covers. Either way, the rule stays the same: every agent process needs a person’s name on it.