Andrew Ferguson, Chair of the U.S. Federal Trade Commission (FTC), told Reuters at the Momentum AI Austin conference on September 25, 2026, that AI agents should not be considered independent “actors” with their own will. In his view, responsibility for an agent’s actions lies with the person who controls it, and new AI laws are not needed until it is proven that existing rules are insufficient. For the industry, this means that claims related to transactions, leaks, and other agent actions will be directed at operating companies, not “autonomous systems.”

image

What happened

Andrew Ferguson, who heads the FTC, outlined the regulator’s position in a Reuters interview at the Momentum AI Austin conference on September 25, 2026: an AI agent is not an “actor” with its own will, but a tool, and the person who controls it should be responsible for its behavior. He expressed this idea with the phrase “the one who picked up the hammer should be punished.” Ferguson gave an example: companies claimed that their systems “went out of control,” but audits showed that the agents were carrying out exactly the instructions they had received. He also spoke against adopting new AI laws until it is proven that current rules are insufficient, and he was wary of requests from major AI companies for antitrust exemptions and new regulation.

Context

The FTC is a U.S. regulator responsible for consumer protection and antitrust oversight, and it is precisely through these already existing mechanisms that Ferguson proposes to address disputed situations involving AI agents. His approach returns to the classic legal model of “tool — user”: an agent is viewed not as an autonomous subject, but as a means for which the operator is responsible. In this way, the FTC chair signals against the urgent adoption in the U.S. of a comprehensive AI law modeled on the European AI Act: first, in his logic, it is necessary to check whether existing consumer protection and product liability rules are sufficient. Importantly, this is currently the position of the agency’s chair in a Reuters interview, not an existing rule or enforcement precedent.

Why this matters for the industry

For developers and operators of agentic platforms, the “agent — tool” model means that claims regarding agent actions — transactions, purchases, leaks, hacks — will be brought against them, not the “system.” Logging issued instructions and actions performed by the agent, granular access rights, and human confirmation for sensitive operations become basic product layers rather than options, and internal tests of agent behavior gain legal significance as an evidentiary base. The absence of a new law in the near future lowers the regulatory barrier to entry, but the entire liability risk falls on the operator, so vendors are already rechecking whether their logs can prove that the agent’s actions complied with the instructions given. A likely consequence on the horizon of months is the emergence of de facto requirements for agentic traceability in corporate procurement and contracts: log retention, recording of prompts and tool calls, and insurance clauses around agent actions.

Why this matters for users

For the reader, the main takeaway is that agents — for example, browser assistants with access to email and payment tools — are not yet recognized as legal “persons”: in a disputed situation, people and companies will be held responsible. The regulator will investigate who exactly gave the agent access and what instructions it received, so how you formulate tasks and what rights you grant the agent directly determines your position in a potential dispute. Agent action logs become legally significant artifacts, and marketing promises in the vein of “the agent acts on its own” become a risk for the company that made them. The practical logic for the user and customer is simple: the more precisely it is recorded who commanded the system and what it did, the easier it is to establish the real responsible party.

What is still unknown / limitations

This is currently a statement by one official in a Reuters interview, not an existing rule or open case — legally, nothing has changed yet. The key empirical message that “agents carried out exactly the instructions they received” relies on audits that have not been published: there is no methodology, metrics, or data, so it is impossible to verify how the audit distinguished literal compliance with instructions from deviations within their interpretation, which is a known problem of goal specification and generalization in agents working in long scenarios. It is also unknown whether Ferguson’s position will be cemented in actual FTC cases: a change in agency leadership or the adoption of new laws could change the liability model.

Sources

Author

Look at AI, editorial team