@sytaylor - Banks Versus AI Agents

Open original

@sytaylor on X Banks are starting to block AI agents that log in with customer passwords, while Walmart and others build official front doors for them. The post maps four layers (entry, identity, tools, intent) and argues that existing protocols like Visa's TAP, Mastercard's Verifiable Intent, and Google's AP2 can be stitched together so agents act only within what a customer approved, with an audit trail, though banks will likely charge for that access.

  • Banks like JPMorgan are blocking shared-credential agents, since a shared login doesn't reveal which agent is acting, what the customer asked for, or whether the action matches the instruction.
  • The four-layer map (Entry, Identity, Tools, Intent) separates how an agent gets in, proves who is behind it, performs actions, and checks the action against what the customer approved.
  • Intent is the hard part: a "freeze my card" instruction shouldn't permit a balance transfer, and a hotel booking has to be checked against price, fees, and amenities after the fact.
  • The author expects "permissioned" access to mean "paid," pointing to the years-long JPMorgan and Plaid data-access deal as a sign banks will charge for it.

Banks are starting to block AI agents that log in with customer-supplied credentials. JPMorgan says it is blocking shared-credential services that move money while it works on a proper alternative, and the author frames this as a schism: some firms, like Walmart with its MCP-based Agent Gateway and Agent Profiles due next month with spending limits, are building front doors for agents, while banks are guarding the back. The core problem is that a handed-over password tells the bank nothing about which agent is acting, what the customer asked it to do, or whether the action matches that request. The author argues that David Sacks is right that banks should support access customers can grant, limit, and revoke, but the hard part is the mechanics. The author proposes a four-layer map. Entry covers how an agent discovers and signs into a service: Personal Agent Protocol (now called Poppy), with Bank of America and Wells Fargo as design partners, plus Visa's Trusted Agent Protocol, which lets a shopping agent present signed claims about who it is and how the customer will pay, and Web Bot Auth, which checks the request signature against the operator's published key. Identity covers proving the person behind the agent: Proof's x401 asks for evidence, OpenID4VP delivers a credential from a wallet, and FIDO passkeys let the user authenticate on-device without giving the agent a password. Tools covers what the agent can do once inside, through MCP, WebMCP, APIs, and Agent2Agent for task handoff. The author stresses that discovering a "transfer money" tool does not grant permission to use it. Intent is the layer that matters most for money: the agent's action must match the customer's approved terms. Mastercard's Verifiable Intent, Google's AP2 mandates, and Visa Intelligent Commerce all try to bind a signed customer approval to the actual transaction, and the author's hotel example shows why: a $190 room with $60 in fees breaks a $200 nightly cap even though it looks compliant at a glance. The author's own risk line is a purchase under a couple hundred dollars can proceed freely, while anything above that should require the agent to check in and leave an audit trail. The catch, which the author admits, is that "permissioned" often means "paid." The JPMorgan and Plaid data-access agreement already includes pricing and took nearly a decade to settle, so banks may well hold out for payment from companies like Meta. The protocols largely exist; the open question is who gets paid to stitch them together.

Ask about this article