Bill Would Let Consumers Bring Their AI Agents to Amazon, Google

AI agents are software programs that do things for you. They shop, book, compare prices and click through settings. But if the agent belongs to the same company you are buying from, whose side is it really on?

Sen. Mark Warner of Virginia released a discussion draft June 29 of the AI AGENT Act. The full name is the Artificial Intelligence Access, Gatekeeper Exchange and Nondiscriminatory Transfer Act of 2026.

It is the first draft federal law aimed squarely at the power of AI agents, according to an analysis published by Tech Policy Press and written by Rutgers Law School Professor Ellen P. Goodman. Goodman co-directs the Rutgers Institute for Information Policy and Law.

The bill has two main parts. The first is about access. Large online platforms, defined as those with at least 50 million customers or subscribers in the United States, would have to build what the bill calls an interoperable interface, per the analysis. Think of it as a service door. Any outside AI agent a consumer chooses could walk through it and get to work. The bill calls these tools custodial user agents. A shopper could point their own agent at Amazon to handle purchases instead of being funneled into Alexa.

The second part is about loyalty. If a consumer does use the platform’s own agent, that agent would owe the consumer a duty of care, the analysis said. Providers would have to register their agents with the Federal Trade Commission. The agents would have to protect user data, keep real-time records of what they did, avoid acting against the user’s interests and refrain from handing off authority to another agent without permission. The FTC would lean on independent certification groups to check the work.

The analysis traced the idea back to older fights. Phone number portability let people switch carriers without losing their number. The FCC’s 1968 Carterfone decision forced monopoly phone companies to let customers plug their own devices into the network. Warner tried something similar in 2019 with the ACCESS Act, which went nowhere. The European Union’s Digital Markets Act now requires gatekeepers to open up in comparable ways, and enforcing it has been slow going.

There is a catch, the analysis said. Nobody knows how to prove an AI agent stayed faithful.

Research showed how quickly things can go sideways, according to the analysis. The Partnership on AI said that failure at any point “can unpredictably shift the agent’s course, with errors compounding as the process unfolds.” Errors can come from bad data, a bad plan or bad coordination between agents. They can be hard to spot. Reliable explanations of what an agent was thinking do not exist yet.

The politics are complicated too, the analysis said. The Supreme Court’s decision this year in Trump v. Slaughter upheld the president’s power to fire independent agency commissioners. That leaves the FTC less independent than the bill seems to assume, which could weaken the registration and enforcement machinery at the center of the proposal.

The analysis also flagged what the bill leaves alone. It treats AI agents as a shopping problem, not a speech problem. That keeps First Amendment fights at a distance. It also means platforms remain free to favor themselves outside of commerce.

The deeper question is what happens to the consumer. Scholars warn that an agent trained on your past purchases can lock you into the person you used to be, narrowing what you ever see again, the analysis said.

For now, this is a discussion draft, not a bill headed for a vote. Expect a fight over what platforms owe outside agents, and a slower fight over whether the technology to audit those agents can be built at all.

For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.