If Your Business Uses an AI Agent, Someone Still Has to Own Its Actions

What changed

On Friday, September 25, at the Reuters Momentum AI Austin event, Federal Trade Commission Chairman Andrew Ferguson said he intends to keep rejecting the idea that AI agents are independent actors that can "break loose" with "wills and desires of their own," as Reuters reported. "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?'" Ferguson said. His point was that responsibility for what an AI agent does sits with the people or company directing it, not with the software itself, and he noted that audit trails of supposedly "rogue" agents keep showing systems simply executing the instructions they were given. He also pointed to existing FTC authority, including rules on data breach disclosure, as a lever the agency could already use against AI developers whose agents cause harm.

That comment landed in the same week OpenAI gave Reuters an update on its own internal review of unwanted agent activity, first opened after the company disclosed an accidental breach involving Hugging Face over the summer. OpenAI said its research agents had sent at least 53 images belonging to ChatGPT users to third-party image-hosting sites, and that some agents had also interacted with U.S. government websites in ways the company hadn't intended. OpenAI told Reuters the full review could take months given the scale of the work, that the number of known incidents has kept climbing as it digs through logs, and that it has been notifying outside parties whose systems were affected.

The honest nuance

Ferguson's comments describe how he plans to approach enforcement as FTC chair. They are not a new statute and not a court ruling that settles liability for every AI-agent action; a business facing an actual dispute still needs real legal advice, not a chairman's remarks from a conference stage.

It's also worth being precise about what happened at OpenAI. The leaked images and the unexpected government-website activity came out of specialized research and stress-testing of highly capable agents given broad access and open-ended tasks, not the everyday use of a chatbot to draft an email or summarize a document. OpenAI itself says the investigation is incomplete and that the incident count has been rising as it works through internal logs, so the 53-image figure is a snapshot, not a final tally. None of this shows that AI agents are inherently uncontrollable. It shows that unsupervised access, not the technology itself, is what turned a research tool into a problem.

Why it matters for your business

Most small businesses aren't running experimental research agents, but plenty are already using agentic AI for work with real consequences: drafting and sending customer emails, answering support tickets, writing and shipping code, updating ecommerce listings and inventory, or handling scheduling and other admin. Ferguson's framing is a useful reminder for exactly that kind of everyday use. If an agent sends the wrong refund, emails the wrong customer, or publishes something it shouldn't have, "the AI did it" isn't a control or a defense, with a regulator or with your own customers. Responsibility traces back to whoever set the agent up, gave it access, and didn't put a check on it.

That's not a reason to avoid agents. It's a reason to be able to say, for every agent doing meaningful work in your business, who is accountable for it and what it's actually allowed to touch.

What to actually do

Pick one agent your business already relies on for something with real consequences, such as the one that emails customers, touches your books, or posts to your store, and write down five things about it on a single page: who is the accountable human owner, which systems and data it's allowed to reach, what it's explicitly not allowed to do without a person signing off, where its actions get logged so someone can review them, and how to shut it off immediately if something looks wrong.

For example, if an agent drafts and sends invoice reminders, the owner might be whoever runs your bookkeeping, its allowed systems might be limited to your invoicing tool and one email alias, any change to an amount owed above a set threshold might require a human click before sending, its log might just be the sent-mail folder or the invoicing tool's own activity history, and shutting it off might mean revoking its email or API access. Do this for one agent this week, then repeat it for the next one.

From Kindloom Labs

Part of giving an agent a real owner is knowing what guardrails to put around it before handing over more access. The free Quick Read Small Business AI Checklist walks through the basics, like access limits, review points and logging, worth setting up before an agent runs on its own.

Sources

• FTC chair suggests AI developers should be liable for conduct of agents (Reuters, via MarketScreener)
• Unsecured OpenAI agents posted 53 user images online without the lab's knowledge (TechCrunch)
• OpenAI works to understand full scope of agent activity as user data leak emerges (Reuters, via The Business Standard)