Visa plugs its payment network into ChatGPT for AI agents

Betting that people will soon trust artificial intelligence agents to do their shopping, Visa said on June 10, 2026 that it has embedded its payment network inside ChatGPT, letting the chatbot complete purchases on a user's behalf at any merchant that accepts Visa. It is the clearest signal yet that the next phase of agentic AI is moving from research demos into the rails that already move consumer money.

The move is structurally different from OpenAI's last commerce attempt. When OpenAI launched Instant Checkout in late 2025, the feature worked only with a small set of enrolled merchants and charged them a fee that retailers broadly rejected. OpenAI retired Instant Checkout in March. Visa's collaboration is different. Users will be able to link a Visa card directly to ChatGPT, and any merchant on Visa's network can be reached without a separate integration. The economic shape of the deal is the same kind of deal Visa has had with banks and merchants for decades, just with a new buyer at the keyboard.

The announcement came at the Visa Payments Forum in San Francisco on June 10, 2026, and AP News covered the rollout the same day, with reporting that names the executive, the timing, and the differences from the failed Instant Checkout product. For OpenAI, the deal is also a reversal of strategy. Instant Checkout was OpenAI's commerce bet as a merchant relationship. The Visa deal is OpenAI's commerce bet as a card-network relationship, and it is a bet that the network will reach more merchants faster than OpenAI ever could on its own.

Why the payments rail will decide who wins

Agentic commerce has spent the last two years as a product story. Demo videos of AI agents booking travel, ordering groceries, and refilling household items have circulated since early 2024. Almost none of them processed real transactions at scale. The reason was plumbing. Recommending a product is a text problem that any large language model can solve. Moving money, dealing with chargebacks, matching the right card to the right consumer intent, and binding the buyer's authorization to a specific action is infrastructure that takes decades to build and a regulatory regime to operate inside.

Visa is offering OpenAI exactly that infrastructure through a system Visa is calling Visa Intelligent Commerce. The press materials, which AP News covered on June 10, describe a token framework and data capture process that reuses the same dispute and authorization rules Visa applies to any other card transaction. Jack Forestell, Visa's chief product and strategy officer, told AP the goal is to make agentic transactions look the same to banks, merchants, and consumers as any other card transaction, even though the buyer is software.

The phrasing matters. Banks, the most regulated participants in the chain, have been the loudest skeptics of agentic payments. The concern is liability. If an AI agent buys something a cardholder did not want, who pays for the chargeback. If an agent overshoots a budget the cardholder never set, does the bank refund the difference. Forestell told AP the framework will have spending limits, required approval steps, and approved merchant lists. That language is the same language banks and card networks have been using for two decades to authorize corporate cards and recurring payments. It is the first sign that Visa is folding agentic payments into a category the financial system already knows how to underwrite.

There is also a question about who carries the operational risk when an agent misfires. Card networks have spent forty years building processes to handle contested charges, stolen cards, and merchant disputes. AI agents introduce a new failure mode: the agent acts on stale information, hallucinates a product that does not exist, or misreads a return policy. Visa's claim is that the new token framework will record enough detail about the agent's authorization to resolve disputes the way any other transaction would be resolved, but that claim is unproven at scale and is the part of the deal that regulators, banks, and consumer advocates will be watching most carefully.

What merchants, banks, and Mastercard are now racing to set

For merchants, the question is whether an agent-driven purchase is more or less valuable than a human-driven one. The answer is unsettled. On the merchant side, the fee OpenAI charged for Instant Checkout was 4% of the transaction. Merchants refused it because it ate the margin on items that were already competitive on price. Visa did not disclose the fee structure of the new arrangement, and Forestell was explicit in his AP interview that terms were not being shared. Without numbers, the merchant economics are still the central open question.

What merchants do know is the distribution. Visa's network reaches tens of millions of merchants globally, which is a different order of magnitude from the curated retailer list OpenAI assembled for Instant Checkout. A merchant that already accepts Visa does not have to do additional integration work to receive an agent-driven transaction, which removes the largest adoption friction that held Instant Checkout back. The trade-off is that merchants do not get a direct relationship with the agent's user, the way they would have under OpenAI's prior product, and they will not have first-party data about which agent made the purchase. For most merchants, the volume is worth the data loss. For a small number of large retailers that have built direct-to-consumer relationships through programs like Amazon Prime, the calculus is more complicated.

For banks, the integration changes how fraud detection has to work. A cardholder buying the same product every month looks identical to a cardholder's agent restocking the same product. Banks will need signals that distinguish human intent from agent intent, and the card networks will have to share those signals in a way that does not break the dispute framework. Forestell said the company is modifying its whole token framework and data capture process to handle that gap, but the practical effect will only be visible once the system has been running for several months at real volume.

There is also a regulator angle. The Consumer Financial Protection Bureau and its international counterparts have not yet issued clear guidance on agent-initiated payments. Most consumer protection frameworks were written assuming a human cardholder and a human merchant, with the card network as the intermediary. When one of the three actors is software, the framework has to be reinterpreted. Visa, as the operator of the network, is in a position to set the de facto standard for what counts as authorized agent behavior. That is significant power, and it is one of the reasons Mastercard has been pushing its own agentic commerce protocols in parallel, including an "agent suite" for deploying agentic AI and a set of agent-to-agent commerce protocols Mastercard has been talking about publicly since early 2026.

Mastercard is the obvious competitive reference point. Mastercard's framing, as covered in Axios in January 2026, is that agents will increasingly procure services on behalf of businesses, not just consumers. The Visa-OpenAI deal is a consumer-first version of the same race. The two networks are now competing to define the standards that any AI agent will have to use to touch a card, and that is the prize worth more than any individual transaction. The bigger signal is structural. For most of 2024 and 2025, the agentic AI conversation centered on what models can do. With the Visa-OpenAI deal, the conversation is moving to what rails they will use, who controls them, and how much of the fee stack gets shared with the model provider. The model is the visible product. The payment network is the margin, and the payment network now has a permanent seat at the table.

Trust is the real product, and it is the bottleneck

Forestell's quote in the AP interview is the most honest framing in the announcement. "We're generally at a place where most people are very comfortable with the shopping aspects of it and have discovered this as a superior discovery experience," he said. "But making the leap from having AI agents recommend what to buy to doing the purchasing just requires a whole different level of trust."

That trust has to be earned, and it will be earned one transaction at a time. Visa's own plan is to start with humans in the loop. Most early transactions will require explicit approval, with the agent sending a notification before the purchase goes through. Over time, as consumers see the agent get it right, the company expects approval prompts to fade for routine categories. That is the same trust curve that took card-not-present transactions from a niche in 2005 to a majority of e-commerce volume by 2015, and it is the curve that took direct debit and automatic bill pay from opt-in afterthoughts to default settings in personal finance apps.

The risk is asymmetric. A consumer who lets an AI agent order a hundred dollars of the wrong groceries is annoyed. A consumer who lets an AI agent handle a thousand dollars of subscription renewals and gets billed for items they do not recognize is angry and is the kind of customer who closes the card. Banks see chargeback rates at this level as a real cost, not a rounding error, and they will pressure Visa to add friction at exactly the point consumers are hoping to remove it. The next 12 months of agentic commerce product work will be less about model capability and more about how the dispute framework handles the first wave of contested agent-driven charges.

There is also the question of habit. Most consumers do not know what their bank dispute process looks like in practice, and a meaningful share of contested charges are resolved by the cardholder simply giving up and accepting the loss. That works when the loss is ten dollars. It does not work when the loss is recurring and the consumer is not sure they authorized it in the first place. The first quarter of agentic payment volume at scale will surface a population of contested transactions that the existing framework was not designed to handle, and the resolution process for those cases will set the public's mental model of how safe agentic payments really are.

For enterprises building internal AI agents, the Visa-OpenAI deal is a useful data point even if the company never touches consumer commerce. The same authorization, audit, and tokenization framework that Visa is building for consumer card transactions will eventually show up in enterprise procurement workflows, where AI agents are already being asked to buy cloud capacity, software seats, and ad inventory on behalf of their companies. Our earlier piece on linking a bank account to ChatGPT covered the consumer-side foundation; the Visa deal is the network-level version of the same idea. If you are mapping the broader picture of how AI is moving from chat interfaces to systems that actually do work in the world, our Enterprise AI in 2026: Use Cases, Governance, and Rollout page walks through the categories of internal AI programs that are working today and the governance patterns that hold them up. Agentic payments are not in that resource yet, but the same trust and audit story is the through-line, and it is the most useful lens for thinking about what this deal will actually mean over the next year.