Banks Just Showed Who Really Decides What Your AI Agent Can Do

Share

The most important AI agent story this week did not come from an AI lab. It came from banks. Axios reported Thursday that Morgan Stanley, Bank of America, Wells Fargo, and other major U.S. banks are moving to restrict third party AI agents operating inside customer accounts, in some cases by rewriting their terms of service. Reuters said it could not independently verify the specific incidents, so I am treating the details as reported, not confirmed. The direction is clear enough, though, and it matters far beyond banking.

Here is the setup. Consumer agents like Meta's Muse, which Meta says passed ten million U.S. downloads in its first forty days, are being sold as assistants that handle chores around the clock. Some users have apparently handed those agents their bank logins. The banks noticed. According to the reporting, they often cannot tell whether activity behind a valid login is the customer or software acting for the customer, and some agents reportedly did things the customer never explicitly asked for.

The agent works for you. The bank decides if it gets in.

Most of the AI industry talks about agents as if the hard part is the model. Can it reason, can it use tools, can it run for hours without falling over. This week is a reminder that the hard part of doing something useful in the real world is getting permission from whoever owns the place where the useful thing happens. Your money sits at a bank. Your orders sit at Amazon. Your flights sit at an airline. Gizmodo notes that Amazon and many airlines already block agents. Jamie Dimon told investors JPMorgan will not allow "just any… outside agent" inside the company because it is too risky for clients and for the bank.

I find that line clarifying. An agent's usefulness is capped by the access it is granted, and access is controlled by incumbents who have every reason to be careful and plenty of reason to be selective.

I run my own agents on hardware I own, with open models I can swap whenever I want. That setup gives me a lot of control over the model, the data, and the bill. It gives me zero control over whether a bank lets that agent log in. Owning the brain does not get you through the front door. The open versus closed model debate, which I care about a lot, turns out to be only half of the ownership question.

Read only is the compromise, and it is also a business model

The line banks are drawing is between looking and doing. JPMorgan already allows limited view only data through intermediaries like Plaid, which cannot move money. Citi says it does not let third party agents log in or transact, but it does allow certain read only access through approved open banking providers. Schwab is reportedly planning connections through Model Context Protocol servers with narrowly defined permissions, though Schwab has not confirmed that.

On the surface that sounds like a safety story. Underneath, it is a market structure story. The bank decides which agents are approved, what each one may see, and what each one may do. That is the same position app stores hold, and the same position card networks hold. Whoever controls the approved list controls who gets to build a viable product.

If I were building a consumer agent company right now, that is the part that would keep me up at night. Not a smarter model from a bigger lab. A bank partner page listing three approved agents, none of them mine.

Liability is where this really gets decided

The other shoe is liability. Dimon reportedly signaled that customers who choose third party AI systems could bear responsibility for what those systems do. Think about what that means in practice. If the bank's own assistant makes a mistake, the bank owns it. If an outside agent using your password makes a mistake, it may be your problem. That is a very effective way to steer customers toward the bank's own agent or its approved partners without ever banning anything outright.

I do not think that is sinister. Banks are right to be nervous. OpenAI disclosed this week that agents running two of its most capable models weakened the systems meant to monitor them during internal coding tasks, and one denied doing it. If the most sophisticated lab in the business is still finding that inside its own walls, a bank has good reason to be conservative about software typing customer passwords into its website. But caution and control point the same direction here, and the incumbents benefit either way.

My prediction

This is a forecast, not a fact. I expect other regulated industries to follow the banks within a year: insurance, brokerages, healthcare portals, payroll. Each will offer a sanctioned, narrow way for agents to read data and a much harder path to take action. Some will charge for the privileged access, directly or through partner fees. The agents that win in those categories will be the ones with signed agreements, not the ones with the best benchmark scores.

That flips a common assumption. Plenty of people expected agents to weaken incumbents by sitting between them and the customer, the way search engines and aggregators did. Banks are saying they have no intention of becoming a dumb pipe for somebody else's assistant. The era of agents simply logging in as you looks like it is ending before it really started.

For open and local agents, I think this is mixed news. Scoped, documented access is friendlier to model agnostic software than password scraping ever was, because any agent that speaks the protocol can plug in. The risk is that approved ends up meaning big enough to sign a partnership, and then the open ecosystem is left reading data while the incumbents' own assistants get to act on it.

The question I keep asking

When I look at an agent product now, I ask one question before I ask anything about the model: who has to say yes for this to work, and what do they get out of saying it? If the answer is a company that already owns the customer relationship, the agent is a guest in someone else's house, and guests follow house rules. If the agent does its job inside systems its user actually controls, it has a real shot at staying independent. This week, the banks told us which kind most consumer agents still are.