Jamie Dimon: “too risky” to let outside AI agents into your bank account

The pitch for AI assistants is simple: tell a bot what you want, let it do the boring work, get your afternoon back. Book the flight, pay the bill, reorder the paper towels. But the places that hold your money and your cart just started slamming doors. Chase won't let outside agents touch accounts. Amazon kicked one off its site. Amex built a whole new system to stop agents from buying the wrong thing. Let me catch you up on what changed and what it means for anyone shipping agent features.

The deep cut

• Access is a permission you have to earn. Chase and Amazon both blocked Meta's Muse before it could finish a task.
• Capture intent before the agent acts, not after. Amex's Create Intent API logs what the customer asked for, then checks the order against it.
• Build brakes so the agent can go fast. Two Sigma's John Feminella routes only the risky calls to a human and lets the rest run.

The doors just closed

The convenience story hit a wall this month. JPMorgan Chase is blocking outside AI agents from logging into retail accounts and acting for customers. On the Oct. 13 earnings call, Jamie Dimon warned about "potentially thousands of rogue agents" and said the bank won't let "just any outside agent" in because it's "too risky." Amazon had already pushed Muse off its site, saying the assistant failed to identify itself and raised concerns about how it handled credentials.

So an agent can research a better savings rate for you, but it can't move the money. It can find the product, but it can't check out. The tedious part, the actual transfer, stays with you.

Note what both companies are doing while they lock the door. Amazon is testing its own shopping agent, Buy for Me. JPMorgan is piloting Smart Cash to move funds and chase yield. They're blocking your agent and building their own.

Intent is the new receipt

The cleaner answer to "did the agent do what I asked" is coming from the card networks. American Express built an Intent Intelligence Framework that captures what a customer actually wanted before the agent buys anything. A Create Intent API logs the request and its limits, the merchant, the budget, the preferences. The customer signs off. Then a token carries that intent through the whole transaction.

Amex VP Hannah Bozian gave the tell: a customer asks for white Nike sneakers in size 9, the agent submits an order for Adidas. With intent logged up front, the bank can stop that transaction before it clears. "It helps mitigate some risk before the risk occurs," she said.

This is the design pattern worth copying. Capture the ask, authenticate it, then check every action against it. Not a consent popup after the fact. A record of the deal, checked in real time.

Brakes, not roadblocks

The firms running agents on real money landed on the same lesson, and it cuts against the instinct to make a human approve everything. At the AI Engineer conference, Two Sigma's John Feminella warned about approval fatigue, where a person rubber-stamping every call slows the work and misses the dangerous one anyway. His team uses classifiers to send only the important requests to a human and lets routine ones run.

Anthropic framed it as a trust layer: who the agent acts as, what it can do without sign-off, and whether it still behaves after you change something. Its trade-break agent proposes order amendments but does not make them. Bloomberg's Andrey Rybka went blunter after agents caused real havoc, including one that read a marketplace review as a stocking instruction: "We want deterministic controls outside the model. Check every call, log it, and keep a kill switch."

Feminella's line is the one to steal for your next review. "Why do cars have brakes? It's so that you can go fast and stop safely."

The data you hand over by accident

The trust problem starts before the first purchase. Meta's Muse asks for bank accounts, passwords, contacts, and chat logs, and it builds hourly profiles of everyone in your life, not just the person who installed it. Signal's Meredith Whittaker called the idea that users made an informed choice "laughable." Privacy law professor Neil Richards said agents should be treated as "an ultra-hazardous product."

Then there's the live demo of why banks are nervous. Elon Musk told followers to photograph their credit cards and drop them in Grok's chat to shop. He posted his own order uncensored, pranksters claimed to cancel it and reorder 400 more, and the stunt backfired on the man selling it.

Not everyone sees a problem. A Y Combinator startup, Outlit, pitches assigning an AI agent to watch every customer's every move. No opt-out. Its founder's defense: "There's nothing that we're doing that's illegal." That is the gap your team has to close. Legal is not the same as trusted, and the banks just made that gap expensive.

Three questions for your team

• If your product calls Chase or Amazon, what happens to the user's task when the agent gets blocked at the door? Have you mapped where your flow dies?
• Do you capture and log what the user actually asked for before the agent acts, the way Amex does, or do you rely on a consent tap after the fact?
• Where are your brakes? Name the specific actions that need a human sign-off and the ones you let run, and make sure you have a kill switch for both.