Tech companies want AI agents, not just chatbots, to manage shopping, inboxes and subscriptions. Whether to trust them with money and personal data is the question NPR tech reporter John Ruwitch and personal-finance reporter Stephan Bisaha took up.
Unlike chatbots that answer questions, agents are designed to do real-world tasks: manage email, fill out forms, or watch for a price drop and present buying options. Ruwitch's analogy was that chatbots are GPS, while these agents are more like a self-driving car that you still supervise. Products already on the market include OpenAI's Dots, xAI's Grok Bot, and Meta's Muse, which the analytics firm Sensor Tower has ranked as one of the fastest-growing apps ever, NPR reported. Dots costs about $100 a month. Grok Bot starts at $30. Muse has paid and free versions; users Bisaha spoke with said the free tier seemed to do more than they needed.
The practical payoff is often unglamorous. Scott Burka, a Muse user in Nashville, said the agent helped him cancel forgotten charges — "things my kids had subscribed to on their iPads or, like, a chess account" — worth about $100 a month. Agents can also compare car insurance rates at renewal without being asked. That is the "low-hanging fruit," Bisaha said: tedious tasks people forget.
The catch is data, and dumb mistakes
The tools work only if they can reach accounts, email, calendars, contacts or cloud files. "The more data you give these things, the more useful they are as a tool, and, in turn, the more data the companies have on you," Ruwitch said. Meta told NPR that Muse is "highly secure," that each agent is designed not to be linked to other agents, that its internet access is limited, and that people control permissions, including whether the app can see passwords and whether it must check before sending an email or buying something. Muse can look into a bank account if a user allows it, Ruwitch said, but cannot move money; purchases usually go through third-party checkout, as in ordinary e-commerce, so the app is not holding the credit card in the way a wallet would.
Critics say the agents also fail in ordinary ways. Jiaxin Pei, an assistant professor of information science at the University of Texas at Austin, told NPR they "are not as reliable as human beings. They make stupid mistakes all the time." Bisaha offered his own example: he wanted grill inserts to use a new gas grill like a griddle, and Muse "confidently" recommended parts that were the wrong size. "These agents, they can find you deals you'd otherwise miss, but it pays to double check their work," he said. He added that these consumer tools are more limited than the "rogue" systems in some recent security headlines.
Many banks are warning that, without clearer guidelines, AI agents could raise fraud and scam risks, and that it is not clear who is liable when something goes wrong — the bank, the AI company or the customer, NPR reported. The reporters' advice was practical, not a ban: use agents for forgettable chores, limit permissions, require approval for sensitive steps, and verify recommendations before paying.