The AI industry’s latest attempt to create a business model that actually makes money is framed as “agents” – a convenient marketing term that describes an AI product that executes a process on your behalf without direct human instruction and often requires access to your emails, messages, credit card transactions, location history, and previous chat history with that company. In the last few weeks, Meta and OpenAI rolled out consumer–facing agents, both branded like kids’ toys and obviously targeting kids (but don’t ask Meta about that; that’s very offensive). They are sold as some version of empowerment. Still, they are built on an imbalanced economy, built on tracking and invasive data collection by people who have repeatedly shown a preference for speed and market lock-in over consumer usefulness or safety. These devices are brought to you by the people who think you should accept “some bad things happening” to you so their product can succeed. For years, companies have followed people across websites, apps, and devices to build detailed profiles used for advertising. Those profiles can include sensitive inferences about finances, health, location, interests, and behavior. It’s worse here as they are built on people’s personal interactions with a chatbot – meaning they go beyond signals to make inferences and target you with a product or coupon based on fears, insecurities, or challenges you may have shared with the chatbot product previously. It is an amazing deal for tech companies. 

They get more data, more trust, more access, and even less responsibility (because you asked it to do that thing for you – no matter how they do it, you’re on the hook). In fact, the Muse’s Terms of Service (which you agree to by using it at all) plainly states that “You are also solely responsible for the actions Muse takes or directs in furtherance of those instructions, goals and guidance in connection with your access and use of Muse. This includes, but is not limited to, any contracts, commitments, communications, content or transactions made, whether they are a result of precise instructions, an identified goal, general guidance or otherwise.” And “If, based on your request or instructions, Muse takes actions involving financial transactions, payments, purchases, or transfers, you acknowledge and agree that…you bear all risk of financial loss arising from such actions and your use of Muse…you are solely responsible for setting appropriate limits and safeguards (including transaction limits and category restrictions), monitoring your accounts, and timely disputing errors with the applicable financial institution or payment provider.”

Agents are a dream scenario for surveillance pricing – where prices are tweaked even just a little bit based on your personal data or uncontrollable market factors, sometimes second to second, and can be different to every user from minute to minute. You give power to the companies to make decisions and purchases on your behalf – and all parties involved know you’re using an agent where you have less engagement and oversight, including the service you’re buying from. There is no  reason to think these companies, who have carte blanche access from you with no liability in their terms of service for the “mistakes” it may commit (read: a risk to your safety or wallet), but million-dollar deals with retailers, ad networks, and others, will prioritize your interest over their business partners. In fact, these agreements just expand all the ways they can give you a raw deal without you even realizing it. And when you do – check those terms of service again – you’re out of luck.

In an ideal world, the agentic systems are able to offer seamless navigation of refund processes and negotiate better terms on your behalf – saving you all the friction that companies build to prevent human resolution of an issue. But the trade-offs of giving AI agents access to your account information, your emails, and everything else are real -- and the companies are always a step ahead in making those processes harder. 

To truly have an AI agent exclusively on your side would require an absurd and unacceptable amount of sensitive data to be taken by these big tech companies, which we know do not care about your privacy. AI agents could further intensify this system by demanding access to still more accounts, logins, purchase histories, browsing activity, and payment information—and by generating even more new data ‘on your behalf’ every time they act.

Consider an automated shopping tool asked to find the cheapest groceries or buy a sweater at the best price. To complete the task, it may need access to delivery accounts and stored credit cards. At the same time, the retailer may use its own automated system to change prices or promotions based on location, device, browsing history, past purchases, or infer a level of desperation like a new parent or mourning funeral attendee. This collection of information is augmented by the highly sensitive and personal data that the AI agent has on you from previous interactions. The consumer-facing tool and the business-facing tool then meet in a transaction neither side makes easy to inspect – and with the “convenience” brought to the consumer, creates another vector that is hard to suss out.

This further intensifies the desperate need for strong privacy laws, AI regulations, and specifically ensuring AI companies releasing these products are liable for the bad moves they make. Prohibiting these unconscionable contracts or “take it or leave it” terms are a key provision in Rep. Jayapal’s National AI Charter Act, which CFA endorsed last week. There are many policy solutions here, but it doesn’t have to be this way.

A note on language: The term “AI agent” suggests autonomy, intentionality, and even moral reasoning — attributes that these systems do not possess. This linguistic framing is not benign. It facilitates the avoidance of responsibility, where harms caused by system outputs are blamed on the “algorithm” rather than the developers, designers, or deployers of the technology. It also fosters overreliance, as users may assume these systems are more objective or intelligent than they actually are. For the sake of conceptual precision and public accountability, we should adopt more accurate terminology such as “automated decision-making systems” or “algorithmic tools” when generalizing, and when discussing specific use cases (text generation, ad targeting), than just use those terms. These terms remind us that the systems in question are engineered artifacts — designed by humans and able to be employed in a protective way or not at all.