Citi, the third-largest commercial bank in the US, updated its Online User Agreement this week. The wording is blunt: if you let a third-party AI tool into your bank account, you're on the hook for everything that goes through it. JPMorgan and Bank of America — the country's two biggest banks — announced the same thing a month earlier, and PNC (eighth-largest) amended its terms the same day TechCrunch published its story. Citizens also sent customers a letter to that effect this week — the publication has seen it.
The mechanics are simple, and that's what makes them unpleasant. You hand an agent your login, password, or access token — it analyzes your spending, offers investment advice, pays your bills. Citi goes further than its competitors: if the tool leaks your credentials or gets hacked itself, the customer is liable for the hackers' actions too. Quote from a bank spokesperson to TechCrunch: "we want customers to understand the risks involved in sharing banking credentials or authorizing such tools." The terms also spell out one exception: agents provided by Citi itself can be used without any added liability.
Our read: this isn't about security — it's about defending the franchise. Anthropic shipped a financial tool explicitly labeled "read-only," with a promise that it "cannot move, manage, or invest your money" — and the banks got moving anyway. The moment an agent sits between the customer and the account, it takes over the very role the bank monetizes: picking a savings product, spotting a junk fee, moving money around.
What this means for someone with a phone. Disputing a fraudulent transaction with your bank used to follow roughly one script: prove it wasn't you. Now there's a new category — "you gave access yourself" — and it overrides even a hack. We think this creates a hole the size of an entire business model: the average user can't tell "my agent glitched" from "my agent got hacked," and in both scenarios the bank stays out of it. And yes, "read-only" is a vendor's marketing promise, not a guarantee issued by the bank.
Who wins: the banks — they keep control of the channel and offload the cost of incident investigations. Who loses: personal-finance startups. A good product is no longer enough; now they also have to convince the customer that the risk buried in their agreement is worth it. It looks like a pattern we've already seen in ride-hailing and delivery: the labs promise to reinvent the service, and the service says "hold on a second."
We haven't seen the full texts of the agreements — only the Citi press office quote and a summary. There isn't a single number: how many agent-related incidents have already happened, how much customers have lost, whether any bank has actually denied a refund. The technical side is unclear too: how does a bank distinguish an agent session from a regular customer login when the token has been handed to a vendor? The Citizens story rests on a letter only the TechCrunch team has seen. And most importantly — the regulators are silent, and it's on them whether these clauses actually hold up in a dispute.
Forecast: within a quarter, at least one public case will emerge where a customer of a US bank disputes charges made by an AI agent, and the bank will have to explain its position publicly — not just in the fine print.
Disclaimer: not officially confirmed: the full texts of the new Citi, JPMorgan, Bank of America, and PNC agreements haven't been published — there's only the Citi press office quote and TechCrunch's summary, and the Citizens story rests on a single letter that only the publication's team has seen.
Source: TechCrunch
Source: techcrunch.com · post in Telegram