Technology

Citi and Wells Fargo Push AI Agent Risk onto the Customer

Citi and Wells Fargo shift AI agent risk onto the customer and leave the bank outside the fraud. What changes for fintechs and operators in LatAm.

Fachada de sucursal bancaria con cajeros automáticos al anochecer

Citibank updated its Online User Agreement in November 2026 so that any transaction executed by a third-party artificial intelligence agent counts as authorized by the customer, even if that agent leaks credentials or is hacked. JPMorgan Chase and Bank of America published equivalent rules in October, and PNC changed its own on November 18, according to reporting by TechCrunch.

The practice is not isolated. Wells Fargo set November 24, 2026 as the effective date for a new version of its Online Access Agreement, which requires the customer to revoke an agent's authorization on their own and to keep evidence of that revocation to hand to the bank when it asks, according to American Banker. Wells Fargo's official page confirms the date and the reservation of the right to restrict or block access from certain third-party AI tools. The institution maintained that it may not have visibility into when the revocation occurred. At least five of the largest U.S. banks already restrict or block third-party agents' access to customer accounts, and the question of who absorbs the loss remains open.

Citibank extended the logic to the hacking scenario. A spokesperson told TechCrunch that the update seeks to clarify the risks of sharing bank credentials and of authorizing those tools, and that the customer who chooses to grant access is responsible for the activity carried out through the agent. Citibank is the third-largest commercial bank in the United States by size. If the third-party tool leaks the credentials, the attacker's transactions also fall on the user's side. On the other side of the counter, Anthropic launched a personal finance tool in early November 2026 labeled as read-only, with no ability to move, manage or invest money.

For a founder integrating agents into a payments product operating in Mexico or Colombia, the data point leaves one question on the table: who absorbs the loss when the agent gets it wrong. That allocation is already in the bank's contract and defines the real cost of the feature.

This article was drafted with the assistance of artificial intelligence from verified sources and reviewed by a human editor before publication.

Sources