Banks block AI agents capable of moving money
Citigroup, Bank of America, Schwab, and JPMorgan restrict AI agents able to open sessions and transfer funds. Underlying concerns include security, legal liability, and threats to bank deposits.
Major American banks are tightening their rules against AI agents capable of connecting to client accounts and moving money without human intervention. Citigroup and Bank of America have joined Charles Schwab and JPMorgan Chase in imposing restrictions, citing security, legal responsibility, and a banking model unwilling to see deposits move on their own.
A human before the transfer, not after
At Bank of America, the message is clear: a human must approve before an account is debited. Its CEO Brian Moynihan stated about automated transfers: "We feel very strongly that there should be a human involved in authorizing before money moves out of your account," adding that unresolved liability issues make these operations impossible to manage properly. The same stance is held by Charles Schwab, which says it blocks unauthorized agents ten to twenty times a week, according to its CEO Rick Wurster. "We can’t control it, the client can’t control it, something could go wrong and who’s liable?" he summarized. The core issue is this: who is responsible when software acts on behalf of a client?
These safeguards mainly target a new generation of personal assistants, like Meta Platforms' Muse, designed to perform daily tasks on behalf of the user. The problem is not theoretical. Giving your credentials to an external bot is like letting a machine enter the vault with a diagonally printed badge, and banks clearly do not want to test what happens next.
The real risk: deposits that leave
The concern is not only about cybersecurity. AI agents programmed to optimize clients' returns could move dormant money from deposit accounts to higher-yield investments, which would weaken banks' funding base. According to economist Torsten Slok from Apollo, about 1.6 trillion dollars of non-interest-bearing deposits held by JPMorgan Chase, Bank of America, and Wells Fargo could be threatened by these automated reallocations. The same calculation mentions 47 billion dollars in additional annual interest costs for the sector if banks had to compensate for this deposit outflow.
For institutions, the issue goes beyond simply locking passwords. The traditional banking model relies on a large volume of low-cost deposits, which then fuel lending. If AI agents start continuously arbitraging these funds, the relationship between client, bank, and savings product becomes more unstable. The debate is no longer just about using a new interface, but about how software autonomy can shift value from one balance sheet to another.
Read-only access, not hands-free
Banks are not completely closing the door to AI tools. Schwab and JPMorgan are working on secure connections, notably via Model Context Protocol (MCP) servers, which allow an AI tool to check balances or conduct financial research without gaining the power to transfer money or pay bills. At the same time, Schwab is promoting its in-house assistant Charley, a recently announced internal AI tool, to keep the experience within a controlled framework.
This design choice is quite telling: reading yes, execution no. Institutions want to benefit from the productivity and research gains offered by agents while preventing uses that affect treasury, payments, or implicit authorizations. This reflex is shared by other major digital players. Amazon recently technically blocked integrations of Meta Muse and asked Perplexity AI to review the integration of its Comet browser. The anti-agent front is expanding, slowly but surely.
Why regulation will have to catch up
So far, US federal regulators have generally supported these safeguards, but they acknowledge that more comprehensive legislation will be needed to set authorization standards and clarify liability when autonomous software acts on behalf of a client. Legal uncertainty is currently one of the main barriers to adopting agents in finance, as it touches on fraud, consent, and the chain of responsibility.
In Europe, the approach is already more defined with the AI Act, which pushes companies to document access, traceability, and human supervision. The text does not target "AI agents" as an autonomous category, but it imposes, for certain high-risk systems, stricter control over what is authorized to act on their behalf. According to the Okta Global CISO Insights 2026 report, 81% of security leaders surveyed say they worry about excessive AI access, while only 47% say they can identify all AI agents present in their environment. In other words, governance often starts with an overlooked obvious fact: knowing who has the right to open the door.
What this means for businesses
For banks as well as other organizations, the issue now goes beyond mere compliance. An AI agent must be identified, linked to a responsible party, limited in its rights, and monitored over time. Without this identity infrastructure, it is difficult to know which actions were authorized, by whom, and within what scope. And as long as a company does not know exactly what an agent can do, it does not truly control it.
The real shift is here: the question is no longer just whether AI can respond quickly, but whether it can act without misusing your access. American banks have already answered with caution.
Key points
- Citigroup and Bank of America have tightened their rules on AI agents.
- Charles Schwab blocks unauthorized agents ten to twenty times per week.
- 1.6 trillion dollars of deposits are cited by Apollo as threatened.
- 47 billion dollars in additional interest costs are mentioned by Apollo.
- The AI Act strengthens human supervision for certain high-risk uses.
- Okta cites 81% of security leaders worried about excessive AI access.
By the numbers
- 10 to 20 times per week — attempts blocked at Charles Schwab, according to Rick Wurster.
- 1.6 trillion dollars — non-interest-bearing deposits at JPMorgan, Bank of America, and Wells Fargo, according to Apollo.
- 47 billion dollars — estimated annual additional interest cost for the sector, according to Apollo.
- 81% — security leaders concerned about excessive AI access, according to Okta Global CISO Insights 2026.
- 47% — leaders able to identify all AI agents in their environment, according to Okta.