Five large US banks move to block customers' AI agents as a new benchmark shows end-state scoring overstates agent success

According to Axios, JPMorgan Chase, Morgan Stanley, Bank of America, Wells Fargo and Citi are each limiting third-party agent access to customer accounts. Customers had given their own login credentials to autonomous assistants, which then carried out online-banking operations on their behalf. The banks can see which credentials logged in. They often cannot tell whether a human or an agent directed the session, or whether the customer authorized a specific action. Some banks detected agents attempting transactions that breached policy.

JPMorgan allows limited, view-only data retrieval through aggregators such as Plaid. That route does not let agents move funds or pay bills. Jamie Dimon, JPMorgan's chief executive, told investors the bank would not grant "just any" outside agent broad access to its systems.

New research on Indian retail banking shows why procedure matters as much as outcome. Krishanu Adhikary and Prashant Devadiga of the National Payments Corporation of India (NPCI) published TauIndianBankBench, a set of 1,000 tasks in which an assistant serves a simulated customer under bank rules. An example rule is that the assistant must obtain a closure quote before it closes a deposit. Many required steps leave the final database state unchanged. The benchmark therefore checks that the agent took every required step and made no unrequested change. Across the models tested, 6.5 to 12.0 percent of sessions passed an end-state-only check even though the agent skipped a required step or made an unrequested change. Tau-style benchmarks that grade only final records would count those sessions as successes.