OpenAI has introduced a feature called ChatGPT Wallet. This feature allows AI agents operating within ChatGPT to select products and complete the payment step on a user's behalf. American Express, Mastercard and Visa have partnered on this infrastructure, with the card networks developing mechanisms that grant agents payment authority under specific conditions.

What happened?

The feature OpenAI announced makes it possible for agents running inside ChatGPT to access a user's card information and make purchases. The simultaneous participation of three major card networks signals that this isn't an experimental integration, but a payment channel directly backed by OpenAI.

What does it mean for AI agents to go shopping?

OpenAI's arrival at this point is no accident. ChatGPT first gained the ability to compare and recommend products, then came checkout integrations with partners like Shopify that moved the purchasing flow directly into the conversation. Wallet is the latest link in that chain: the agent no longer just makes recommendations, it now carries payment identity as well. This turns OpenAI from a conversational product into a payment intermediary — an unusual role for an AI company to take on.

The card networks' rapid alignment with this has its own logic. For Visa, Mastercard and Amex, the real risk is that the agent economy grows outside their reach — for instance, through stablecoin-based rails or closed in-platform wallets. Making their cards the default payment method for agents is the most direct way to head off that risk. Google's AP2 (Agent Payments Protocol), announced last year together with Mastercard, PayPal and other payment providers, was born of the same motivation: establishing a shared standard that cryptographically verifies on whose behalf an agent is transacting and within what limits.

Who's liable: how will fraud risk be managed?

The real open question here isn't technical, it's legal. Card networks' chargeback systems are built on the assumption that the person approving a transaction is human. If an agent buys a product the user didn't want, orders the wrong quantity, or is manipulated by a prompt injection attack into paying a fraudulent merchant, who bears the loss — OpenAI, the card network, or the user? For now there's no clear answer to this question; OpenAI's statements appear to offer limited detail on how liability would be shared.

The race for the agent economy: OpenAI's closed model or Google's open protocol?

FeatureChatGPT Wallet (OpenAI)AP2 (led by Google)
StructureClosed, exclusive to the OpenAI ecosystemOpen protocol, multiple providers
Card partnersAmex, Mastercard, VisaMastercard, PayPal, and other payment providers
How it worksPayment happens directly in the ChatGPT interfaceVerifiable authorization standard between agents
Scope of useOnly agents within ChatGPTShared infrastructure any platform can adopt

This table actually captures the main competitive axis of the period ahead: OpenAI is moving fast within its own walled garden, while Google and its partners are trying to build an industry-wide standard. The closed model can deliver a faster user experience in the short term; the open protocol carries the potential for broader acceptance among merchants and banks in the long run.

What's next?

Three things are worth watching: which countries and merchants Wallet rolls out to, what policies issuing banks adopt for fraud and chargeback cases, and how Amazon and Google accelerate their own agent-payment solutions in response to this competition. It's not yet clear when or to what extent Turkey's card network and banking infrastructure will support this system.

Ultimately, this step is a concrete sign of the shift toward an era where agents move not just information but money as well. It's a technically sensible evolution, but the trust and liability framework is advancing more slowly than the payment authority itself — and that gap will remain the real limiting factor.