Visa and OpenAI Wire ChatGPT Into the Payments Network

Visa said Wednesday it will plug its global payment network into ChatGPT, letting OpenAI’s AI agents search for products, pick one, and pay on a user’s behalf. The deal, announced at the Visa Payments Forum in San Francisco, gives the chatbot a way to charge a Visa card across any merchant that already accepts one.

It is OpenAI’s second try at agentic commerce, after the first ended in March when merchants refused the fee it charged. The party most exposed in the new arrangement is the merchant, and the merchant is the one the announcement does not name. The terms under which that party will accept agent-initiated payments are still undisclosed.

Visa’s Rails, OpenAI’s Front Door

Visa is bringing four pieces of its infrastructure to the deal. The network that clears more than 300 billion transactions a year, the tokenization system that swaps raw card numbers for one-time tokens bound to a specific agent and merchant, the real-time authorization engine, and the fraud-monitoring stack that watches the whole of it. That stack is what the announcement leans on most, with tokenized credentials and real-time authorization listed as the core mechanism for every transaction the new partnership will touch. The four infrastructure pieces in the Visa-OpenAI partnership announcement put the scale of the underlying network at 4.8 billion payment credentials and more than 175 million merchant locations.

Forestell put the moment in generational terms. The network underneath is the same one that pulled e-commerce from stores to websites, mobile apps, and one-click checkout.

AI will transform commerce more profoundly than the internet or mobile technology ever did.

Jack Forestell, Visa’s chief product and strategy officer, said it at the Visa Payments Forum in San Francisco on Wednesday. He returned to the trust theme later in the release, saying the goal is to ensure transactions are trusted, secure and seamless. The agreement reaches past the chatbot, with the two companies pledging to explore embedding payment primitives and trusted agent identity signals into developer experiences built on OpenAI’s coding agent Codex, and into automated business workflows.

• More than 300 billion transactions a year on Visa’s network
• 4.8 billion payment credentials on file
• More than 175 million merchant locations accept Visa

OpenAI Tried This Once Before

This is OpenAI’s second swing at agentic commerce, not its first. The company had launched Instant Checkout in ChatGPT in late 2025, letting the chatbot search the web for a specific item and complete the purchase as a single step.

The product charged merchants 4% of the transaction value, and merchants refused to sign on at scale. OpenAI retired Instant Checkout in March. The Visa version works differently. Where Instant Checkout was an OpenAI-built rail running through a small set of enrolled merchants, this one layers Visa’s existing network under OpenAI’s interface.

Shoppers will be able to link a Visa card to ChatGPT. Merchants that already accept Visa can, in theory, accept agent-initiated transactions without separate enrollment.

Marco Mahrus, OpenAI’s head of partnerships for commerce, framed the new deal as the infrastructure Instant Checkout lacked. He said the goal is to build infrastructure for secure, transparent and user-controlled agentic transactions, helping people do more with AI agents while maintaining confidence that payments are being handled safely and securely. Mahrus was speaking from the same San Francisco stage as Forestell.

Who Is the Silent Party in the Announcement?

Visa and OpenAI did not disclose the financial terms of the collaboration, and did not give details on the fees merchants or customers would pay. No merchant is named in the release, no merchant is quoted in the coverage, and no estimate of integration cost appears in either company’s materials. That 4% was the figure that killed Instant Checkout, and the new version does not name a replacement.

That is the gap the announcement leaves. The same merchants who refused OpenAI’s 4% Instant Checkout fee are now the substrate the new system runs on. The Visa framing leans on guardrails, tokenization, and user-defined permissions, but the underlying economics for the party accepting the transaction are still undisclosed, a point made sharper by the fact that the same merchants killed the first attempt inside six months.

Mastercard Is Building Its Own Rails

Visa is not alone in this. Mastercard said this month that AI agents will be able to procure services on behalf of a business, and in June the company launched an agentic payment push with a 30-plus partner list.

The Mastercard effort has drawn a roster that includes Cloudflare, Coinbase, Stripe, Adyen, and Anchorage Digital. Stephanie Cohen, Cloudflare’s chief strategy officer, framed the partnership as connecting the company’s developer and security platform to payments infrastructure. The design runs on multi-rail settlement, programmable digital dollars, and a layer Mastercard calls Verifiable Intent.

The two efforts are running on different assumptions. Visa is building the consumer-facing surface inside an existing chat assistant and routing agent payments through a card network that already clears more than 300 billion transactions a year. Mastercard is building the protocol layer underneath, betting that agent-driven commerce will move value in directions card networks do not yet reach. The open question is whether the merchant, the silent party in the Visa-OpenAI deal, will face a different set of guardrails, fees, and consent prompts from one rail to the next.

Guardrails and the Human Escape Hatch

Visa’s answer to the trust question is a set of user-defined controls. Transactions under the new partnership will operate inside guardrails the consumer or business sets, including spending limits, required approval thresholds, merchant category restrictions, and per-transaction sign-off. The design keeps the buyer in command even when the agent is the one executing the purchase.

• Spending limits set by the user or business
• Required approval thresholds above a defined amount
• Merchant category restrictions
• Per-transaction sign-off

The system also keeps a human in the loop, at least at the start. Forestell said in an interview that Visa expects most transactions to begin with the agent sending a notification for the consumer to approve the actual purchase, then graduating toward unattended execution as trust builds.

Now, imagine you do that a thousand times over the course of some period of time. And then your agent says, do you want me to just not check?

Forestell, Visa’s chief product and strategy officer, said it in an interview on Wednesday. The harder question is what happens when the agent acts inside the guardrails but the outcome is wrong. Visa will apply its existing dispute framework first, asking whether the consumer really intended the purchase and whether the merchant processed it correctly.

For the case where both sides acted correctly and the agent is the one that erred, the company is modifying its token framework and data capture process through Visa Intelligent Commerce, Forestell said. The deal does not yet say who eats the loss in that scenario.

The S-1 Was Already Two Days Old

The deal lands on the heels of a confidential IPO filing. On Monday, June 8, 2026, OpenAI said it had submitted a confidential S-1 to the SEC, a draft investor prospectus that lets a company begin the IPO process without disclosing detailed financials. The filing landed with a reported valuation of more than $850 billion, with the Wall Street Journal putting the March 2026 figure at about $852 billion after a $122 billion funding round.

Visa did not comment on whether the timing was coordinated, and OpenAI has not said so either. What is on the record is the order of events: a confidential IPO filing on Monday, a payments partnership with the world’s largest payment network outside of China on Wednesday, and a march toward the public market on a clock the deal, if it lands, will help define. ChatGPT’s previous e-commerce build had already been dropped before the new one arrived. Anthropic’s $965 billion IPO filing landed a week earlier, putting two of the most valuable AI companies in the public-offering queue at the same moment.