The update to Citi's Online User Agreement landed this week, and Citi is not acting alone. PNC rewrote its terms the same day, Citizens emailed customers with a similar warning, and Wells Fargo's new terms take effect on November 24. JPMorgan Chase and Bank of America moved first, in October. In the space of a month, the biggest banks in the United States have gone from blocking AI agents at the door to writing down, in legal language, who pays when an agent goes wrong.

What Citi actually changed#

The new language sits in a section of Citi's Online User Agreement devoted to third-party AI agents. It starts with a warning — customers are cautioned against sharing their credentials with outside AI tools — and then spells out the consequences of doing it anyway. Provide the login or otherwise permit access, and Citi will treat everything the agent does as if the customer did it themselves. That holds even when the agent's actions run against the customer's own instructions.

The second half is the part that has drawn the most attention. If the AI tool shares or leaks the credentials, Citi says later access using them will also be treated as authorized, with an exception only where applicable law says otherwise. A Citi spokesperson told TechCrunch the bank wants customers to understand the risks of sharing banking credentials or authorizing outside tools to reach their accounts, and that the update is meant to give greater clarity about those risks.

A smartphone showing a mobile banking app with billing and pay-bill options on its screen
A mobile banking app on a phone. Photo: Torsten Dettlaff, CC BY-SA 4.0, via Wikimedia Commons.

The industry is moving together#

Citi's move completes a pattern that has formed across the industry in about a month:

BankWhat changedWhen
JPMorgan ChaseSimilar policy on AI agents used with customer credentials; CEO Jamie Dimon raised the risks on the October earnings callOctober
Bank of AmericaSimilar policy on AI tools used with customer credentialsOctober
CitibankCustomers responsible for agent activity, including after a leak or hackThis week
PNCOnly bank-provided connections allowed; no passwords, one-time codes, browser automation or screen scrapingNovember 18
CitizensEmailed customers: on the hook for what their AI assistant doesThis week
Wells FargoTransactions by a customer's delegates, AI agents included, count as authorizedEffective November 24

PNC's version takes a different route from Citi's. Rather than only assigning blame after the fact, it defines the permitted path: connections the bank itself provides, with defined accounts, duration and permissions. Handing an agent a password or a one-time code is prohibited, and so is using browser automation or screen scraping to work around the bank's controls. Break those rules, and PNC says the customer is responsible for the resulting access and activity, to the extent the law permits.

A robotic hand reaching toward a human hand holding a bank card, in front of a glowing bank vault door
Editorial illustration: an AI agent reaching for a customer's bank card. AI-generated image.

The catch: the fine print may not hold up#

There is an important qualification that the scarier summaries leave out: a bank's terms do not automatically erase federal consumer protections. Guidance from the Consumer Financial Protection Bureau under Regulation E says an agreement cannot impose more liability for an unauthorized electronic transfer than the regulation allows. If a third-party service is hacked and a thief moves money without actual authority, that may still count as an unauthorized transfer — and the customer's liability for one is capped, and low if reported quickly.

The hard cases sit in the middle. If a customer intentionally gave an agent authority and the agent then exceeded its instructions, the dispute becomes much harder — the bank's position is that the customer authorized the access, and the regulation's protections for unauthorized transfers may not apply. Business accounts can carry fewer protections than consumer accounts. And as of now, no court or regulator appears to have tested how these clauses hold up against Regulation E, and no US agency has publicly weighed in. Whether the new terms survive a real dispute over a drained account is an open question.

Why banks are drawing the line now#

The timing is not an accident. AI agents are moving into personal finance just as the tools get good enough to be useful — and dangerous. Anthropic launched a personal finance tool earlier this month that is explicitly read-only: it can look at your accounts but, in Anthropic's own words, can't move, manage or invest your money. Google, meanwhile, has made its Gemini Tasks assistant free, with examples that include reaching into a utility account through a password manager, paying bills and using payment cards. Google says sensitive actions need confirmation and the agent runs in an isolated virtual machine. Those are real controls. They do not make handing a bank password to an agent a safe integration.

Banks also have a competitive reason to draw the line. An agent that can shop around for better rates, dodge fees or move money to higher-yield accounts threatens the margins banks earn from inertia — low-interest deposits, unnecessary fees. And the pattern is not unique to banking: AI labs promise agents will change how people engage with every service provider, and the providers keep answering 'hold on a minute.' The same pushback has shown up in ride-hailing and food service. Whoever controls the pathway between the agent and the account controls the relationship — and the revenue.

What to watch#

The next date on the calendar is November 24, when Wells Fargo's updated terms take effect. The bigger questions are whether any regulator weighs in on how Regulation E treats an agent's actions, and whether the first customer who loses money to a compromised agent takes the new clauses to court. On the product side, watch whether agent makers answer with read-only defaults and bank-approved integrations — the model Anthropic chose — or keep asking users for full credentials and hope the banks blink first. Until one of those things happens, the practical rule is simple: don't give an AI agent your banking password, passcode or a shared browser session. Use a bank-approved connection, prefer read-only access, limit the accounts and duration, and check what permissions the agent already holds.

Sources: TechCrunch, “More banks say customers using AI agents are on their own — even if they get hacked”; RaymondTec, “Bank AI Agent Liability, NGINX Flaws, and Java4Shell”; StartupFortune, “Citi, PNC and Citizens tell customers they own any loss an AI agent causes”.