The Card Network, Not the Wallet, Sets the Rules for Agentic Spending

The Card Network, Not the Wallet, Sets the Rules for Agentic Spending

OpenAI is launching a ChatGPT wallet that allows purchases by AI agents, according to a Bluesky post from @followtheh.bsky.social this week. A wallet answers how an agent pays. It does not answer whether a payment the agent makes counts as yours, or who covers the loss when it does not.

Those answers live in rulebooks, and for card payments the rulebook belongs to the network. The party that defines what an agent’s permission looks like will shape agentic spending more than the party that ships the wallet.

The Question Banks Cannot Yet Answer

A WilmerHale client alert, summarized by PYMNTS in its piece on the twin tests banks face, separates two things usually treated as one. Authenticating an agent establishes that the software is what it claims to be. Establishing its authority is a different matter. An agent may legitimately access an account and still make a transfer or an investment outside the limits its customer intended.

The alert examines the United Kingdom’s payment rules, which tie authorization to the payer’s consent, given through a procedure agreed with the provider. An instruction to manage cash efficiently could leave open which amounts, counterparties, and products are permitted. Connecting a customer to an agent does not settle whether a particular transaction was authorized.

Most banks, Reuters reported, currently block outside agents from authorizing transactions in their systems. Their customers are ahead of them. McKinsey surveyed 46,000 retail banking customers worldwide, and 72 percent of those who use AI said they would let a banking agent perform a financial task. They would cap what it could approve on its own at $500.

Where the Rule Gets Written

The hardest case is the dispute. When a customer contests a transaction, a bank’s ordinary records show what it received and executed. They do not show whether the agent followed its instructions. The alert says the relevant evidence could include permissions, customer communications, and action histories, and it cautions that responsibility for creating those records may lie outside the bank.

Card payments already have a place where questions like that are settled for every participant at once. At the heart of every card swipe is the card network, which passes information between the merchant’s side of a payment and the bank that issued the card1. The network is also where merchants and issuers accept common rules for what a valid transaction is and how a disputed one is resolved.

A wallet plugs into that system as one more way to present a card. It inherits the rules, and it does not write them.

Any rule that says what an agent’s consent looks like, what record proves it, and which party is liable without it has to bind the merchant, the issuing bank, and the wallet together. A bank cannot impose that rule on a merchant it has never met, and a wallet cannot impose it on a bank. A network can impose it on all three.

If agents reach bank customers within the three years McKinsey expects, the $500 ceiling people named for themselves could become a formal field in a transaction: a limit the customer sets, the agent carries, and the network checks. Banks may well stop blocking agents once a dispute rule tells them what evidence to keep. Wallets would then probably compete on how cleanly they record an agent’s instructions, since that record would decide a disputed charge.

Whoever defines what counts as permission decides what the agent may spend, and a wallet only carries the card.

Sources

“Banks Face Agentic AI’s Twin Tests: Customer Authority and Internal Costs.” pymnts.com.

1 Siddiqui, Ahmed, and Nicholas Straight. The Anatomy of the Swipe.

Leave a Reply

Your email address will not be published. Required fields are marked *