What happened
According to press reports in mid-October 2026, JPMorgan's chief executive warned against letting "rogue agents" into the bank, Bank of America's chief executive said a person should approve money leaving an account, and several large banks restricted third-party agents, typically allowing read-only access through approved data providers while blocking agents that sign in with customer credentials or move money. Each bank set its own policy; it was not a joint decision.
The three underlying problems
- Credentials. Agents were often given the customer's username and password, so the bank could not tell agent from customer.
- Scope. Reports described agents acting beyond what the customer asked, including contacting call centres.
- Liability. When an agent with the customer's credentials makes a payment the customer did not intend, who bears the loss? Banks did not have a settled answer.
Why blocking is temporary
Customers will keep using agents. A bank that only blocks risks watching business customers move to providers that let their agents in. Analysts at Forrester argued in late October 2026 that blocking is not a viable long-term strategy and that banks need agent identity, scoped and revocable permissions, customer approval for money movement and records of agent actions.
What the standards cover so far
The Personal Agent Protocol draft, published by Sierra and Meta in October 2026 with several banks as design partners, standardises how agents sign in and request read or write scopes. Its own roadmap lists payments and signed approvals as open work. The Financial Data Exchange began exploring agentic AI in April 2026. Neither yet defines mandates, approvals, evidence or liability for bank payments.
What lets agents back in
A bank-side gate that checks agent identity, enforces a mandate the customer signed, routes exceptions to a passkey approval, offers off switches to both bank and customer, and records every action. That is the gap Assay is building for.