By 31 March 2027 we expect no federal court decision and no CFPB, OCC or Federal Reserve text to expressly allocate liability for AI agent initiated transfers. We would be wrong if, by 31 March 2027, at least one federal court decision or CFPB, OCC or Federal Reserve text expressly allocates liability for an AI agent initiated transfer.
Imagine giving an AI agent a bank password and asking it to move $5,000. It misunderstands the instruction, and the money lands in the wrong account. Chase says it can treat the agent's actions as the customer's own. Citi assigns responsibility for the agent's errors to the customer, subject to law. Banks are warning that an automated helper can leave a very human bill.
We would avoid funding money moving agents on an assumption of bank reimbursement. By 31 March 2027 we expect zero federal court decisions or CFPB, OCC or Federal Reserve texts expressly allocating liability for AI agent initiated transfers. The Contract First Gap is the interval in which banks allocate losses in account agreements before federal authorities classify autonomous transfers under Regulation E. One express allocation, whichever party it favors, would change how we price agent risk.
The warning wave began between 13 and 20 October 2026, according to Hunton. Dimon discussed barring outside agents from accounts on 13 October. Chase's revised credential terms took effect on 27 October, 14 days after the 13 October call. Citi's agreement added Section H in November. As of 21 November 2026, Wells Fargo's agent terms are scheduled to start on 24 November, according to Hunton. Banks have stated their position before the law has settled these transfers.
Chase's Section 9 bars outside agents and aggregators from using login credentials. The word 'deemed' does the work: permit an outside agent to use those credentials and Chase treats unintended instructions as the customer's own. Section 17 adds indemnity for actions beyond the agent's authority. Citi's Section H warns, "You are cautioned against using any third-party AI Agent to access your accounts"; it treats activity using supplied credentials as authorized, even after those credentials leak. Citi disclaims a duty to check whether an action matches a prompt, while preserving what applicable law permits. That definition of consent reaches beyond the customer's instruction.
Ally requires a different description. Its public agreement dated 12 December 2025 already disclaims losses when customers share passwords with aggregators or other third parties. An AI agent agreement change at Ally is unconfirmed. The older clause concerns credential risk and gives no evidence of an AI amendment. Chase permits authorized data access through approved APIs while restricting credential sharing. The alternative keeps Chase in charge of the connection.
Regulation E already gave banks much of this position. Its definition of unauthorized excludes a transfer by a person supplied with an access device by the consumer. The CFPB's official interpretation says a consumer is fully liable when that person exceeds the authority granted, until the bank is told the person's authority ended. A human delegate fits more neatly than software. Is an AI agent a person under this language? Does a hostile prompt amount to excessive delegated authority or an unauthorized intrusion? Chase and Citi's wording appears to make the answers favor the bank.
The money is a scenario: $5,000 sent to the wrong recipient. An authorized transfer can leave the customer paying $5,000. For an unauthorized electronic transfer under Regulation E, notice within two business days of discovering lost or stolen access generally caps liability at $50; later notice can raise it to $500, with a separate rule for subsequent transfers after 60 days from the statement. On a $5,000 transfer the customer bears up to $5,000 if the transfer is authorized and $50 to $500 if it is unauthorized, so the classification moves $4,500 to $4,950 of the loss. The classification determines who bears most of the loss.
A password gives an agent the customer's permissions without a distinct audit trail. Chase favors approved APIs, and its 16 September 2025 renewed agreement with Plaid includes a pricing structure. The bank sets API access terms; the login makes an outsider resemble the customer. The warnings give customers a reason to use the controlled route while placing loss on those who choose credentials.
Bank contracts encounter a rather stubborn statute. The Electronic Fund Transfer Act prohibits agreements that waive consumer rights. Its words are, "No writing or other agreement between a consumer and any other person may contain any provision which constitutes a waiver of any right conferred or cause of action created by this subchapter." A customer whose agent is hijacked can argue the attacker acted without actual authority. That would bring the $50 or $500 statutory caps back into play. Citi's own "except as otherwise permitted under applicable law" acknowledges the possibility. The customer needs a favorable classification under existing law.
Britain has already sent much of the bill back to payment firms. The Payment Systems Regulator says firms returned 88% of claimed losses from authorized push payment scams in the nine months to June 2025, up from 66% in the comparable earlier period; GBP 112 million was reimbursed. For covered scam claims, its mandate moves losses from victims to payment firms. A US court could protect a hijacked agent user by March. But Britain adopted a specific reimbursement mandate, while US Regulation E distinguishes authorized transfers from unauthorized ones.
Six earlier disputes provide an uneven reference class. Three of five with outcomes, or 60%, put costs on an intermediary or payment firm: Plaid's $58 million privacy settlement, Block's Cash App redress of at least $75 million plus a $55 million penalty, and Britain's reimbursement mandate. The federal Zelle case was dismissed in March 2025; JPMorgan priced its Plaid access channel, and New York's Zelle case remains open. None involved an AI agent, and US federal enforcement changed in 2025. That record makes reimbursement a cost to consider without predicting when a federal agent ruling arrives.
One federal court or regulator has 130 days, from 21 November 2026 to 31 March 2027, to make our forecast wrong. The consumer case has an existing statute and a $4,950 swing on our $5,000 scenario. That window opens with zero qualifying texts and no scheduled agent decision. Section 1033's 1 April 2026 compliance date passed without effect after a Kentucky court injunction; its reconsideration proposal went to OIRA in August 2026 and remained unpublished in November. That proposal concerns data access rather than agent transfer liability. A federal docket scheduling a Regulation E agent transfer ruling before March would move us toward expecting one allocation. Until then the Contract First Gap leaves banks setting contractual terms.
We would keep customer credentials away from outside agents able to move money and price an erroneous delegated transfer as a customer exposure until a federal allocation changes it. We would be wrong if, by 31 March 2027, at least one federal court decision or CFPB, OCC or Federal Reserve text expressly allocates liability for an AI agent initiated transfer, including whether Regulation E classes it as authorized or unauthorized.
