Citi Puts Liability for AI Agent Activity on the Customer

Citibank has updated its Online User Agreement so that customers who give third-party AI agents access to their accounts are responsible for the activity those agents carry out. The update came in the week of November 18, 2026, according to TechCrunch.
The liability extends to compromise. If an AI tool leaks the credentials it holds or is hacked, the customer is responsible for what attackers do with the account. Under the terms, that exposure sits with the customer, not the bank.
Citi's agreement separates outside tools from its own. The bank's terms say that AI agents provided by Citi can be used without additional liability concern. The dividing line is who supplies the agent, not the authentication method or the scope of access the agent receives.
Citi is not alone. JPMorgan and Bank of America announced similar policies in October 2026, holding customers liable for AI-agent activity. PNC changed its terms on customer use of AI agents on November 18, 2026, and Citizens emailed customers the same week warning that they are liable for anything their AI assistant does. Across these four large U.S. retail banks, the changes place the risk of agent-driven access on the account holder.
For practitioners, the central question is the delegation chain. A consumer agent that logs into a bank on a user's behalf typically holds a session token, a short-lived credential that proves a login is valid, or stored credentials. The agent vendor, the user's device and the bank each sit at a different point in that chain. The agreements assign the bank's side of the chain to the customer whenever the agent is outside the bank's control, much as a homeowner who hands a spare key to a contractor carries the consequences of what that contractor does. The reported terms do not specify how a customer would scope or revoke an agent's access, meaning limit what it may do and switch it off. They also do not describe what logging a bank would provide to show which actions an agent took rather than the user.
The broader context here is a shift in where liability sits. Banks have long held the fraud-loss exposure for unauthorized transactions under many consumer frameworks. These terms move part of that burden onto the customer when the unauthorized party is an agent the customer chose to connect. The contract language appears to be the mechanism. The incentive it creates is also visible: customers who want agent convenience with bank-level protection are steered toward bank-provided agents, where the terms say no additional liability concern applies.
In my view, the practical test for these policies will be how cleanly banks can tell an agent-initiated transaction from a user-initiated one, and how quickly a customer can revoke access. Those are engineering questions as much as legal ones, and the agreements do not answer them on their own.
There is a constructive reading as well. Agentic tools that reconcile accounts, flag subscriptions or move money on a schedule could become routine if banks define a clear authorization model with scoped permissions, auditable actions and revocation that works. The updated terms suggest that banks see a path where sanctioned agents operate inside a defined liability boundary. Whether that path delivers the convenience its proponents describe will depend on the controls built around it.


