AI agents in client accounts: block, tolerate or govern, the choice now facing banks
On 22 October, Axios revealed that large US banks were finding AI agents inside their clients' accounts, logged in with the clients' own credentials. Several executives called the situation "horrifying". For a Swiss or European bank, the question is no longer whether these agents will arrive, but what answer to give them.
What happened in October
The facts fit in a few lines. Meta launched Muse in the United States in early September: more than 10 million downloads in 40 days, according to its 21 October announcement. OpenAI followed with its Dots in late September, and Anthropic opened the Duet preview on 21 October.
These agents book, negotiate and fill in forms. And, increasingly, they log in to banks with the password the client hands them, which the terms of almost every institution forbid.
The responses were fast and scattered. Morgan Stanley is blocking third-party agents on its platforms. JPMorgan, Bank of America and Citi restrict them and push towards read-only access. Charles Schwab says it blocks unauthorised agents several dozen times a week.
On 13 October, Jamie Dimon described these accesses as "simply too risky for clients and for JPMorgan". Two days later, PNC chief William Demchak called the transactions "incredibly dangerous" and asked for legislation on the transfer of liability.
The underlying problem is not fraud, it is ambiguity. One banking executive put it plainly: it has become hard to tell whether an agent is acting on the client's instruction or has slipped out of their control.
Why blocking will not hold for long in Europe
Muse is not yet available in Europe, and Meta has announced no date. That breathing space is useful, but short, and the European legal context makes the wall strategy more fragile than across the Atlantic.
On 15 October, in case C-70/25, the Court of Justice of the European Union ruled that a bank must refund an unauthorised transaction without delay, even where it alleges gross negligence by the client. The argument about fault comes afterwards, not before.
Applied to agents, this means a bank that lets an agent in without being able to tell it apart from fraud will carry both the burden of proof and the cash. A Swiss bank is not bound by this ruling, but its European clients and its EU subsidiaries are.
On the Swiss side, FINMA published Guidance 02/2026 on digital fraud in April. Of 19 institutions surveyed, 8 had no dedicated policy, 7 had no standard response plan, and most monitored transactions with fixed thresholds rather than scenarios.
To monitoring systems, an agent logging in with a client's credentials looks like a money mule, or a takeover. Or like nothing at all. That is precisely the grey zone FINMA is asking institutions to shrink.
Three paths, and what each one costs
We see three possible answers for a chief risk officer or a head of digital. They are not mutually exclusive, but they do not share the same horizon.
Block: detect agents by their technical fingerprint and refuse the session. Fast and reassuring, but fragile: agents imitate a human browser better every month, and clients find workarounds.
Tolerate in read-only mode: allow balances and transactions to be viewed, forbid any order. This is the JPMorgan and Citi line. It calms the transfer risk, not the data leakage risk.
Govern through delegated access: give the agent its own identity, limited rights, ceilings, a log and one-click revocation. This is the only path that turns the risk into a service.
The third path has a precedent: open banking. The PSD2 APIs showed that a third party can act for a client without ever seeing the password. The draft FIDA regulation, under discussion in Brussels, extends that logic to savings and insurance.
The market is already organising itself on that side. Walmart has opened an official gateway for shopping agents, with approval of each purchase or a spending cap. Anthropic's Duet asks permission before paying or signing and never sees passwords.
A bank that waits for regulation will arrive after its competitors and after its clients. The useful question is not "should we let agents in", but "on what terms, for which actions, with what audit trail".
The example of a Geneva private bank
Take a Geneva private bank with CHF 25 billion under management, 300 staff and a European and Middle Eastern clientele. Its e-banking already sees a few dozen oddly shaped sessions every week: regular hours, very fast navigation, no hesitation.
Step one, over 30 days: measure. The security and data teams cross-check login logs against the known signatures of the market's agents. The result: under 1% of sessions, but concentrated on wealthy, active clients, exactly the ones the bank wants to keep.
Step two, over 60 days: decide and write it down. The risk committee opts for immediate read-only mode for any detected agent, and a delegated-access workstream for three actions: view, prepare a transfer to a registered beneficiary, download a statement. The terms and conditions are rewritten accordingly.
Step three, over 90 days: build. The agent receives an identifier distinct from the client, rights per action, a daily ceiling and a log the relationship manager can read. Every order prepared by the agent waits for human validation in the client's mobile app.
The figures to track fit on one page: share of agent sessions identified, time to revoke, number of orders validated or refused, incidents reported. Within six months, the bank knows whether agents are a service channel or a door to close.
What we take away
October moved agents out of demos and into accounts. US banks reacted by blocking, then by negotiating read-only access. Swiss and European banks have a few months' head start, and a legal framework that obliges them to do better.
Our conviction: the right level of decision is the risk committee, not the IT department alone. It takes a written position on agents, a measured detection setup, and a delegated-access architecture built on what open banking has already proven.
If you want to see where your institution stands on this, the 39 Advisory AI Barometer gives a first reference point in a few minutes. And if you would rather talk it through, get in touch.





Comments