OpenAI and Visa’s new agentic AI payment collaboration means ChatGPT-style agents may be able to shop and pay for you with Visa credentials, under spending limits, merchant rules, and approval controls you set. Announced on June 10, 2026, at Visa Payments Forum in San Francisco, the deal brings Visa’s tokenization, authorization, and fraud monitoring into OpenAI’s agentic commerce plans.
What the OpenAI-Visa agentic AI payment deal actually changes
The simplest reading is this: OpenAI wants agents that can complete purchases, and Visa wants those purchases to run through a trusted card network rather than a messy patchwork of one-off checkout integrations. Visa said the collaboration will support secure Visa payments inside “agentic commerce” across OpenAI experiences.
AP reported on June 10, 2026, that Visa had embedded its payment network inside ChatGPT, allowing agents to shop and complete transactions on a user’s behalf, potentially at merchants that accept Visa. That matters because Visa acceptance is far broader than a closed pilot with a single retailer.
An agentic AI payment is different from a saved-card checkout. You’re not just clicking “buy now.” You’re giving an AI agent permission to act within a defined boundary, such as “buy printer paper under $25 from an office-supply merchant” or “ask me before any purchase over $100.”
Visa says transactions will operate under user-defined permissions, policies, and controls, including spending limits, merchant categories, and required approvals. OpenAI’s head of partnerships for commerce, Marco Mahrus, described the goal as “secure, transparent and user-controlled agentic transactions.” That phrasing is careful, and it should be.
How Visa is trying to make AI shopping less risky
The payment plumbing is the real story. Visa said the OpenAI integration will use tokenized Visa credentials, real-time authorization, and fraud monitoring. In plain English, the AI agent shouldn’t be handling your raw card number like a sloppy browser extension from 2014.
Tokenization replaces the sensitive card number with a payment token that can be scoped, monitored, and revoked. Real-time authorization checks whether a transaction fits the rules at the moment it happens. Fraud monitoring looks for patterns that don’t match normal card behavior.
Banks have a reason to care. AP reported on June 10, 2026, that banks have been concerned about fraud claims when an AI agent uses a customer’s credit or debit card. If a bot orders the wrong product, buys from the wrong seller, or gets manipulated by a malicious listing, who eats the dispute?
My view: the controls are more important than the AI model. A brilliant shopping agent with vague permissioning is a liability. A merely competent one that can’t spend outside its lane is far easier to trust.
The 2025 groundwork: OpenAI, Stripe, and Visa Intelligent Commerce
The Visa-OpenAI announcement didn’t appear out of nowhere. In September 2025, OpenAI announced Instant Checkout in ChatGPT and the Agentic Commerce Protocol with Stripe. OpenAI said merchants using other payment processors could accept agentic payments through Stripe’s Shared Payment Token API or the protocol’s Delegated Payments Spec.
Visa had also been building toward this. In April 2025, the company launched Visa Intelligent Commerce with AI industry partners including OpenAI, Anthropic, IBM, Microsoft, Mistral AI, Perplexity, Samsung, and Stripe. By December 18, 2025, Visa said it and partners had completed hundreds of secure, agent-initiated transactions, with more than 100 partners working across the agentic commerce ecosystem.
Those dates matter because they show a shift from “AI can recommend products” to “AI can initiate a financial action.” If you follow agent design more broadly, that same move from suggestion to execution is visible in software work too; our guide to AI loops that build, test, and improve outputs explains why autonomy gets powerful only when feedback and constraints are built in.
Visa also said on December 18, 2025, that 47% of U.S. shoppers had used AI tools for at least one shopping task. The company predicted millions of consumers would use AI agents to complete purchases by the 2026 holiday season. Predictions are not adoption, but the direction is clear.
| Milestone | Date | What it added | Why it matters |
|---|---|---|---|
| Visa Intelligent Commerce launch | April 2025 | AI commerce partner network including OpenAI, Anthropic, Microsoft, Stripe, and others | Created the commercial framework for agent-led payments |
| OpenAI Instant Checkout and Agentic Commerce Protocol | September 2025 | ChatGPT checkout with Stripe and delegated payment specifications | Gave merchants a path to accept agentic payments |
| Visa agent transaction update | December 18, 2025 | Hundreds of secure agent-initiated transactions completed with partners | Showed early testing beyond slideware |
| OpenAI-Visa collaboration | June 10, 2026 | Visa payments in OpenAI agentic commerce experiences | Connects agent shopping to Visa’s payment network and controls |
A practical spending-limit example
Here’s the calculation most announcements skip. Say you authorize an AI assistant to reorder household supplies with a $75 monthly cap, a $25 per-transaction cap, and approval required for marketplace sellers without a 4-star rating. If the agent finds detergent for $18, trash bags for $16, and paper towels for $22, it can spend $56 without bothering you.
Now add a $24 replacement coffee filter. The individual purchase is under $25, but the monthly total would hit $80. A well-designed agentic AI payment flow should stop and ask because the aggregate limit is breached. That’s a tiny edge case, but it’s where trust is won or lost.
Merchant category controls create another useful boundary. You might allow groceries, pharmacies, and pet supplies while blocking electronics or travel. Honestly, that option only makes sense if categories are enforced by the payment system, not guessed from a product description written by a seller.
A tougher case is substitutions. What if the agent is told to buy “the cheapest USB-C charger from a reputable brand” and picks a model that’s technically cheap but unsafe or underpowered? Payment controls can stop suspicious spending, but they don’t guarantee product quality. AI commerce still needs product intelligence, return handling, and clear receipts.
What merchants and developers should watch next
For merchants, the interesting question is whether agent traffic behaves like search traffic, affiliate traffic, or something stranger. An AI agent may compare price, shipping time, return policy, seller reputation, and card reward eligibility before a human ever sees the cart.
Online retailers already plan for sudden demand spikes from product drops and flash sales, but agent-driven buying could create a different kind of pressure: many small automated comparisons and purchases happening at once. If that sounds familiar, our piece on how retailers manage traffic surges during flash sales is a useful parallel.
Visa’s broader Payments Forum announcement named Agent Score, Agentic Directory, and a Large Transaction Model among new AI commerce capabilities. Reported as part of the June 10, 2026, update, those tools point toward identity, risk scoring, and transaction analysis for software agents rather than only human shoppers.
Developers should also watch Codex. Visa said it will explore embedding payment primitives and trusted agent identity signals into developer-focused experiences powered by OpenAI Codex. If that grows into real tooling, payment permissions could become something developers define as carefully as API scopes.
There’s a security angle too. PCI DSS and browser-side script controls already create headaches for checkout teams, and agentic commerce adds another actor to the chain. Teams modernizing payment pages may find useful context in this report on AI-assisted PCI DSS script authorization.
Set guardrails before you let an AI agent pay
The first consumer versions will probably feel conservative, and that’s a good thing. A payment agent should earn trust purchase by purchase, not start with an open wallet and a vague instruction to “get the best deal.”
Before enabling any agentic AI payment feature, you should check the controls rather than the marketing language. The questions are boring. They’re also the difference between convenience and a customer-service mess.
- Set both per-purchase and monthly spending limits, because one small transaction can still push a budget over the edge.
- Limit merchant categories to purchases you’re comfortable automating, such as groceries or recurring supplies.
- Require approval for substitutions, unfamiliar merchants, used goods, international orders, or delivery fees above a fixed amount.
- Confirm how receipts, returns, chargebacks, and warranty claims are handled when the agent initiated the purchase.
- Use a card or credential you can freeze quickly if the agent behaves oddly or a merchant dispute gets messy.
People often focus on whether the AI “understands” what they want. I’d focus first on reversibility. If you can cancel, return, dispute, or revoke access without a scavenger hunt through settings, the system is much safer.
Mobile shopping will be part of this too. If OpenAI experiences and banking apps make agent permissions manageable on a phone, adoption gets easier; our broader look at mobile applications in 2026 explains why the app layer still shapes consumer behavior more than most infrastructure announcements do.
Why Amazon, banks, and card networks may not agree
Agentic commerce sounds neat until it threatens someone’s control point. Amazon, Walmart, Shopify merchants, banks, card networks, processors, and AI companies all want to be the trusted layer between intent and purchase.
AP reported that Visa’s prior AI-commerce efforts were limited to a single retailer or a small group of enrolled merchants. The OpenAI relationship is different because ChatGPT could become a general shopping interface, not just a retailer-specific assistant. That creates tension with platforms that prefer to keep discovery, checkout, and customer data inside their own walls.
Amazon is the obvious pressure point. If a ChatGPT agent can compare merchants and buy wherever Visa is accepted, Amazon has to decide whether to welcome the traffic, restrict bots, or build a better agent of its own. We’ve covered that strategic fork in Amazon’s choice between resisting and embracing AI shopping bots.
Card issuers face a narrower but serious question: when an approved AI agent makes a bad purchase, is that fraud, user error, merchant misconduct, or product misrepresentation? The answer will shape liability rules. It may also shape how aggressively banks allow agentic AI payment activity on debit cards versus credit cards.
A counter-argument deserves airtime. Some consumers won’t want this at all. They’ll use AI for comparison, then pay themselves, because shopping is partly about judgment and partly about control. That’s not backward. For expensive goods, medical products, or anything with safety implications, manual checkout may remain the wiser default.
FAQ
What is an agentic AI payment?
An agentic AI payment is a transaction initiated by an AI agent under rules you set, such as spending limits, merchant categories, and approval requirements. The agent acts for you, but it shouldn’t have unlimited authority.
Can ChatGPT buy things with Visa now?
Visa and OpenAI announced their collaboration on June 10, 2026, to enable Visa payments in OpenAI agentic commerce experiences. AP reported that ChatGPT agents could shop and complete transactions for users, potentially at merchants that accept Visa.
How does Visa protect card details in AI transactions?
Visa said the payment capability will use tokenized Visa credentials, real-time authorization, and fraud monitoring. Tokenization helps avoid exposing the raw card number to the agent or merchant workflow.
Is this the same as OpenAI Instant Checkout?
No. OpenAI announced Instant Checkout in ChatGPT and the Agentic Commerce Protocol with Stripe in September 2025. The Visa collaboration, announced in June 2026, brings Visa’s network, credentialing, and security infrastructure into OpenAI’s agentic commerce plans.
What should I limit before letting an AI agent shop?
Start with per-purchase caps, monthly caps, allowed merchant categories, and approval rules for unfamiliar sellers or substitutions. The safest setup is narrow, reversible, and easy to pause.
EN 
FR 
ES