Citibank, the third largest commercial bank in the United States, changed its terms of service this week regarding third-party AI tools for personal finance. The revised language explicitly tells account holders that if anything goes wrong while using these services, it remains their sole problem, even in the event of an external breach. While AI agents can analyze spending, evaluate investment decisions, offer financial guidance, and handle account transactions or bill payments, lenders are refusing to shoulder the technical liabilities of these automated integrations. The central issue is who absorbs direct financial losses when an autonomous interface errs or exposes access tokens.

Financial institutions have announced that they will not be held responsible for reversing either intentional or unintentional transactions executed by automated assistants. By updating its Online User Agreement, Citibank effectively shifts operational and security exposure away from institutional balance sheets and places it squarely onto consumers and enterprise account users. The policy establishes that customer liability persists even if an autonomous intermediary leaks login credentials to malicious third parties.

"We recently updated our Online User Agreement to provide greater clarity about those risks and to reinforce that customers who choose to grant account access to third-party AI agents are responsible for activities conducted through those agents."

As a Citi spokesperson explained in a statement to TechCrunch, the institution made these revisions to define the risks of sharing online credentials with automated software. When an external financial assistant suffers a security compromise, the bank will not cover the resulting unauthorized transactions.

The Industrywide Wall

This legal shift extends across the banking landscape. JPMorgan and Bank of America, the two largest banks in the country, announced similar policies last month, dismantling expectations that institutional fraud protection would cover automated workflows. PNC, the eighth largest bank, changed its terms today, and banking giant Citizens emailed its customers warning that they are equally on the hook for anything their AI assistant executes. How will autonomous software developers deploy active transaction features when banks refuse to reverse unauthorized transfers?

Developers are forced to retreat into defensive, limited-capability architectures. Anthropic launched a personal finance tool earlier this month that is explicitly billed as read-only. Anthropic designed the software so that it cannot move, manage, or invest money, avoiding execution capabilities entirely to sidestep unhedged liability. AI software companies seek new revenue streams by entering personal finance and capturing transaction fees, yet they encounter severe user hesitation when requiring direct bank credentials.

At the same time, banks view third-party attempts to penetrate their customer relationships as a direct competitive threat. To counteract independent platforms, Citibank explicitly notes that proprietary AI agents provided directly by the bank can be used by customers without worrying about additional liability. Frontier lab developers market an autonomous future where synthetic software manages capital allocations and routine payments, yet legacy institutions have enforced a framework where outside agents remain strictly fenced off as read-only viewers or multi-million-dollar liabilities for anyone running them.

How the picture shifted

US banks used to absorb the losses from unauthorized account activity, but now they leave customers entirely on their own if a third-party automated assistant gets breached or makes an error. At the same time, lenders keep their proprietary automated tools exempt from these harsh liability shifts, guarding direct access to account balances. Handing direct control over payments to an independent agent now means shouldering the entire financial fallout without institutional fraud protections. Contractual barriers from major lenders are effectively confining autonomous tools to passive financial tracking.

AI in FinanceAI AgentsCybersecurityAI RegulationAnthropic