Let AI read your budget, not hold your wallet: banks are right

Would you give a stranger your online banking password because he promises to find you a better savings rate? Put that way, it sounds absurd. Yet that is roughly the proposition behind a new generation of AI agents, and the banks have just started saying no.

Here is what happened. Meta rolled out Muse, a personal agent that fills in forms, books travel and, per the company's own pages, can help with budgets and money goals. Bank of America and Citigroup have now restricted third-party AI agents, joining JPMorgan Chase and Charles Schwab, according to the Financial Times as relayed by Banking Exchange. Schwab's CEO, Rick Wurster, told Bloomberg TV that the broker stops dozens of agent attempts to reach client money every week, and he named Muse, as Notebookcheck reports.

The real issue is liability, not technology

The banks' stated reasoning is refreshingly plain. Wurster put it as "We can't control it, the client can't control it, something could go wrong and who's liable?" Bank of America, per the same reporting, has limited access while it works out who would be liable for payments an AI tool authorizes. Jamie Dimon said "You don't want rogue agents, and there could be thousands of them."

We think that is the heart of it. The software may well be competent. What is missing is a rulebook saying who eats the loss when an agent misreads an instruction, gets tricked or simply does something you didn't mean. Until that exists, the default answer will probably land on the customer. Dimon himself warned of a liability transfer when customers let outside agents in.

Banks aren't purely noble, and that's fine

Let's not be naive. There is a commercial motive too, and the reporting says as much: lenders worry agents could shift idle cash to higher-yielding accounts. Apollo's Torsten Slok wrote that agents could sweep household cash into accounts paying 3.3% to 5.0%, versus the 0.1% national average on checking, as CNBC reported. Cheap deposits are the banks' raw material.

So two things can be true at once: the banks are protecting their margins, and they are protecting you from a real gap in the rules. Our view is that the second reason is strong enough to stand alone. Notice, too, that they aren't banning AI wholesale. JPMorgan still allows customer-approved data sharing through open banking providers, and Schwab is building connections that let AI see information without moving money. That is the right shape: read access yes, payment authority no.

What we'd do: analyze, don't authorize

The useful part of these tools is on the analysis side. CNBC describes Muse surfacing forgotten recurring charges, a real problem when, per a Mastercard and FT Strategies report, average annual subscription spending among U.S. consumers reached $1,887 in 2025. Spotting a forgotten monthly charge is low-risk and high-value.

Handing over credentials is the opposite trade. Notebookcheck points out that the standard online banking terms of German cooperative banks require customers to keep their PIN and other security credentials secret, and that a bank can treat misuse after sharing them as gross negligence, leaving the customer with the loss. That is Germany; US rules differ, and we haven't seen documentation settling who pays here, which is exactly the point. When the rules are unwritten, it is prudent to assume you're the one holding the bill.

Anthropic's own help page for Claude in Chrome, as cited by Notebookcheck, advises against using the agent to manage financial accounts or investments. If the makers hedge, so should you.

Also, a note of caution on the incident side: Axios, via Newsquawk, reports that some banks are detecting unauthorized agents in accounts, but the sourcing is unattributed, and it's unclear whether this is a breach or legitimate tools misusing credentials. We wouldn't panic, but we wouldn't ignore it either.

The verdict

Banks blocking agents isn't hostility to progress; in our view it is an admission that automation has outrun accountability. Our practical advice, which is general and not personalized: let an AI look at an exported or read-only view of your spending, ask it to find subscriptions and sketch a budget, then make the cancellations and transfers yourself. Don't share your banking login, and don't grant payment authority. When regulators or banks publish clear liability rules, revisit. Until then, the best financial setting for your AI is read-only.