JPMorgan Chase refuses to let external AI agents carry out transactions on its retail customers' accounts. Jamie Dimon, its chief executive, confirmed this position on October 13, when presenting the quarterly results. Transfers and bill payments triggered by a third-party autonomous assistant therefore remain blocked. The bank cites the security of its customers, while developing its own tools to automate the management of their money. Here is what this position covers, and why the issue goes beyond the case of a single bank.

What a banking AI agent is

An AI agent is software that does not merely answer questions: it acts autonomously on behalf of its user. Applied to banking, this means an assistant able to check accounts, compare returns, then make a transfer or pay a bill itself. It is this last step, executing a transaction, that JPMorgan denies to agents from outside.

Reading yes, acting no

According to CNBC, cited by Journal du Coin, a JPMorgan spokeswoman confirmed that the bank prohibits third parties from collecting customers' banking credentials to log in on their behalf. No timetable for opening up to transactions by external agents has been announced.

The bank does, however, distinguish executing a transaction from viewing data. It supports secure, customer-authorised sharing through open banking providers such as Plaid. An assistant can thus access certain information on a read-only basis, a balance for example, without being able to move funds.

In substance, the dividing line makes sense: read access exposes information, write access exposes money. And software that logs in with the customer's credentials holds all the keys to the account, without the bank being able to tell the account holder from their agent.

The question of liability

Jamie Dimon also raised the question of liability when a customer delegates actions to an agent. His statements, as reported, do not specify under what conditions a loss would be borne by the customer, by the agent's provider or by the bank. This is probably where the crux of the matter lies: if an agent executes an erroneous transfer, because it misunderstood an instruction or was manipulated, who pays the money back?

Caution does not mean rejecting AI

JPMorgan is not giving up on autonomous AI for all that. The group already uses agents internally and is preparing applications for its customers. Its chief financial officer, Jeremy Barnum, confirmed a targeted pilot of Smart Cash, a tool designed to anticipate cash flows and move funds in order to improve the return on them.

The contrast is worth noting: the bank does not rule out an agent moving its customers' money, provided it is its own. Security is not the only issue, either. According to CNBC, assistants able to compare returns and facilitate transfers could make deposits more mobile. For banks, this could increase competition on the interest paid on savings and raise their funding costs.

Stablecoins, another route to the same questions

Other infrastructures exist for payments made by agents. The x402 protocol, launched by Coinbase in May 2025, allows software and AI agents to pay for digital services in stablecoins directly within their exchanges with a web server, for example to pay for access to an API from a compatible provider.

This mechanism does not, however, give any access to the funds in a bank account: the agent must have a funded wallet and the necessary authorisations. Key security, spending limits and the reliability of the recipient remain decisive. Being technically able to pay therefore does not settle the question of control. Bank account or stablecoins, financial assistants will have to reconcile automation, user consent and the allocation of liability in the event of an error.

And in Switzerland?

The sources consulted say nothing about the position of Swiss banks on this subject. Yet the questions raised by JPMorgan apply to any institution that holds retail customers' accounts. Should an external agent be allowed to read a balance, but not to place an order? Who bears responsibility for an error made by an assistant mandated by the customer? And can a bank develop its own agents while closing the door on those of others, without raising the question of competition?

At least one thing emerges from JPMorgan's position, useful to any user of an AI assistant: authorised read-only data sharing has nothing to do with handing over one's online banking credentials to a third-party service.

Sources

  • Journal du Coin: "JPMorgan's boss judges external AI agents too risky for bank accounts"