On September 8, Meta shipped Muse, a personal agent that can read your accounts and act on them. Six weeks later Bank of America’s stock was down 18 percent, Charles Schwab’s 14, Wells Fargo’s 6 and JPMorgan’s 5.
Bank stocks fall for plenty of reasons, and nobody can pin all of that on one product launch. But the fear underneath it is easy to state. If software holds your login and moves your cash wherever the rate is highest, the bank stops being the place you keep money. It becomes a pipe your agent shops between.
On JPMorgan’s third-quarter call, Jamie Dimon gave his answer.
“We’re not going to allow just any outside agents to come in inside the company, because it’s just too risky for clients and for JPMorgan.”
He is right about the specific thing he is blocking. An agent that logs in with a customer’s password looks exactly like the customer. The bank can’t tell what it is allowed to do, can’t cap what it does, and can’t say afterward who is responsible when it moves the wrong money. Every fraud team in the country has spent a decade trying to make borrowed credentials useless, and here is a product category that runs on them.
The rest of the industry, in their own words
The other big banks said versions of the same thing in October, and the variations are worth reading closely.
Brian Moynihan, at Bank of America: “We strongly believe there should be a human in the loop to authorize money moving out of your account.” Rick Wurster, at Schwab: “While people love AI and they’re using it heavily, they’re not ready to turn their financial life over to AI.” Charlie Scharf, at Wells Fargo, on whether agents will drain deposits toward higher rates: “completely overdone, especially for the time period we can see.”
Notice what Moynihan’s sentence permits. A human in the loop to authorize money moving out is a design requirement, and you can build an agent that meets it. Wurster’s sentence is a claim about customers today. Scharf’s has a time limit written into it.
None of them said agents will never be allowed. They said not like this, and not yet.
We have seen this fight before
For most of the 2010s, budgeting apps and aggregators got into bank accounts the same way Muse does now: the customer handed over a password and a server logged in on their behalf. Banks complained about the risk, sometimes blocked the traffic, and the customers kept doing it anyway, because the apps were useful.
The fight ended with banks giving out tokens instead of passwords: scoped, revocable access through an interface the bank controls. JPMorgan went further than most, and in 2025 it reached an agreement under which Plaid pays for the data it pulls from JPMorgan accounts. The bank stopped treating the side door as a threat once it owned the side door and could charge rent on it.
Agents are the same problem with higher stakes, because an aggregator mostly read and an agent will want to act.
Where this probably goes
Blocking works for as long as customers would rather keep the bank than keep the agent. For JPMorgan, with its branches, its brand and its own tools in the works, that may be a long time. JPMorgan is already piloting something called Smart Cash, which predicts a customer’s cash flow and moves funds to earn more yield. That is an agent. It just happens to be the bank’s.
For a bank with less pull, the math runs the other way. If the agents people actually use can’t get into it cleanly, those agents will quietly recommend banks they can get into. Mike Mayo’s line, that the three most important words in banking are trust, trust and trust, cuts both ways here. Trust is the reason to keep strangers out. It is also what a customer will end up giving to whichever agent manages their money well.
So the interesting move is a door: an agent gets its own credential, a defined list of things it may do, limits on how much it can move and how fast, and a tap on the customer’s phone before anything leaves. That satisfies Dimon’s risk, Moynihan’s human and the customer’s convenience at once. Deloitte found 72 percent of bank customers are hesitant to share financial information with generative AI tools, and a bank-controlled door is the most obvious thing that could change that number.
Three things to watch
- A large bank publishes an agent access spec. Permissions, limits, consent, logging. Once one does, the agent builders will build to it first, and the others will be negotiating on its terms.
- An agent moves money between two banks with one tap. Reading balances is the aggregator era. The first sanctioned agent that sweeps cash from one institution to another, with the customer approving on their phone, is the moment the deposit question gets real.
- Someone settles who pays when an agent gets it wrong. Until a regulator, a court or a contract says who eats the loss, every bank has a perfectly good reason to keep saying no.
Dimon can afford to wait for all three. Most banks can’t. They will need to decide what an agent is allowed to do inside their walls before the agents decide which walls are worth visiting.