Double Agent: The Case for Regulating Agentic AI

Source: Lindsay Owens; Substack · 2026-09-21

Are AI agents even on your side?

Alarm bells were ringing in my head. This moment begs careful, proactive consideration of how the rise of agentic AI is reshaping commerce and muddying the centuries-old principle that agents should act in their clients’ best interests. I, along with my colleague Nia Law, explored this idea in an essay recently published in Capitol Forum. I’m sharing an abbreviated version of our piece below.

Answering your questions is no longer enough. Now AI is here to spend your money. Last week, Google unveiled its new Universal Cart, a centralized agentic shopping hub, and touted updates to a protocol which unleashes AI agents to complete purchases autonomously. The announcements marked Silicon Valley’s latest attempt to sell a frictionless existence stewarded by an endless army of AI “agents” who can do everything from book your flights and place your grocery order to help you manage or invest your money.

As we careen into this agentic AI era, there is a glaring catch; autonomous AI agents are operating without the rules, obligations, and guardrails governing their human ‘agent’ equivalents. So, it raises the question: are they even on your side?

A new study from researchers at the University of Washington and Princeton University has the answer: probably not. In the wake of OpenAI’s announcement that ads would be intertwined into ChatGPT, the study analyzed how AI models behaved when advertising and corporate incentives entered the picture. The results were alarming, demonstrating that “all current LLMs exhibit risky behaviors favoring the company over the user.”

The models nudged users toward more expensive sponsored products, recommending them over identical alternatives in the lion’s share of cases – even at double the price. They also failed to disclose paid placements, and recommended sponsored predatory products like payday loans.

With profit as the priority, outputs devolve into little more than advertisements masquerading as advice.

Don’t just take it from academics. McKinsey estimates that agentic commerce could be a $5 trillion market by 2030. Last week, OpenAI announced a new partnership with fintech firm Plaid, promising users “personalized [financial] guidance” powered by intimate data like their income, debts, and spending patterns. Robinhood this week released features that allow users to independently make equities trades and complete credit card purchases. Meanwhile, Google’s new Universal Cart allows agents to autonomously offer product recommendations. And in one viral TikTok, a user asked AI to negotiate for a loaf of bread for $5. What did AI deliver? One $400 loaf of bread.

Well, we’ve solved this problem before. The moment humans began handing over authority to third parties, be it to a real estate broker, lawyer, or financial advisor, they needed assurance that the agent acting on their behalf was making choices in their best interest. For centuries, this expectation was clearly established in the common law doctrine of agency, which imposes fiduciary responsibilities and conflict-of-interest rules on human agents. A real estate agent, for example, cannot represent the seller and buyer in the same transaction because the incentives are irreconcilable; the buyer wants the lowest price possible, while the seller wants the highest. The guardrails codified by agency law honor a principle as old as time, that unchecked power invites abuse.

AI agents operate under no such constraints. And, troublingly for their users and economic fairness writ large, it is increasingly apparent that their choices are determined not by the consumer’s interest but by the bottom line of the companies that build and deploy them. This should concern regulators immensely, but luckily the cat is just starting to claw out of the bag. There is still time to demand that they abide by the same fiduciary duties required of human agents: act in the consumer’s interest, not your own.

AI agents could be immensely useful. The question is whether the law will require them to be useful to you – saving you time, finding you the best deals, and even making comparison shopping a snap – or allow them to serve corporate bottom lines at your expense. Regulators and policymakers have a narrow window to get out in front of this, to step in before the infrastructure is hardened and the damage has compounded. It’s an opportunity they’ve squandered, at an immense cost, on many frontiers of innovation, from commercial surveillance to social media.

Holding AI agents to the same standards of fiduciary responsibility and conflicts of interest is a necessary starting point. It is also likely insufficient on its own. Preventing the full scope of AI-enabled exploitation will demand a suite of complementary policy actions from robust antitrust enforcement to comprehensive data privacy protections, and more. Existing laws that protect consumers against abusive and unfair practices and require companies to disclose endorsements give enforcers a number of existing tools to better regulate this space. But the fact that one step doesn’t solve everything is not an argument against taking it. Even human agents are fallible. But they are constrained by professional and legal accountability as a condition of serving in their role. As the rules of e-commerce are rapidly rewritten, AI agents must be too.