Visa has plugged its payment network into ChatGPT, giving AI agents the power to shop and pay at more than 175 million merchant locations worldwide. The deal marks the moment agentic commerce moved from conference slideware to financial infrastructure. The question now is whether consumers will actually let software spend their money.
By Blumefield | June 12, 2026
The Deal That Puts a Card in ChatGPT's Hand
The most consequential payments announcement of the year arrived on Wednesday with little theatre. At the Visa Payments Forum in San Francisco, Visa and OpenAI announced a strategic collaboration that embeds the world's largest card network directly into ChatGPT, allowing AI agents to complete purchases on a user's behalf. It is the clearest signal yet that agentic commerce, the idea that software agents rather than humans will initiate a meaningful share of online transactions, has graduated from pitch decks to production roadmaps.
The mechanics matter less than the scale. An agent operating inside OpenAI's products will be able to transact at any of the more than 175 million merchant locations that accept Visa, once the user grants permission. Previous experiments in this space were confined to single retailers or curated catalogues. This deal removes the ceiling entirely. If a human with a Visa card can buy it, in principle, so can their agent.
Visa frames the partnership as part of Visa Intelligent Commerce, its initiative for extending payment rails into AI-driven environments. OpenAI gets something it has chased for two years without success: a credible, trusted path from product recommendation to completed transaction.
How the Plumbing Works
Underneath the agentic commerce announcement sits a careful piece of engineering. Transactions will not expose raw card numbers to AI systems. Instead, Visa supplies tokenized credentials that are cryptographically bound to a specific agent and a specific use case, paired with real-time authorization and continuous fraud monitoring. Visa also handles chargebacks and refunds, the unglamorous machinery that makes consumers willing to trust a payment method in the first place.
Control stays nominally with the user. Spending caps, merchant category restrictions, and approval thresholds can all be configured, so an agent might freely reorder household supplies under fifty dollars but require explicit sign-off before booking a flight. "As AI agents become active participants in the economy, Visa's focus is to ensure transactions are trusted, secure and seamless," said Jack Forestell, Visa's chief product and strategy officer, in the official announcement.
The enterprise angle may prove just as significant as the consumer one. The companies said they will explore developer experiences powered by Codex, OpenAI's coding agent, which could eventually purchase inference, APIs, or other developer services on its own within limits a user sets. A coding agent that can provision and pay for its own infrastructure is a small revolution in how software gets built and billed.
What the announcement conspicuously lacks is a date. There is no launch timeline, no pricing, and no user interface to inspect. For now this is infrastructure being poured, not a product being shipped.
A Second Swing After Instant Checkout Missed
OpenAI has been here before, and the first attempt failed. Instant Checkout launched in September 2025, letting ChatGPT users buy from Etsy merchants without leaving the conversation. Adoption never materialised and the feature was quietly retired by March 2026. A follow-up effort, Shopping Research, narrowed its ambition to product discovery, assembling buyers' guides from specifications, reviews, and pricing data.
The lesson OpenAI appears to have drawn is that commerce is not a feature you bolt onto a chatbot. It is a trust problem, and trust in payments is something Visa has spent six decades and more than 300 billion transactions a year accumulating. Marco Mahrus, OpenAI's head of commerce partnerships, acknowledged as much, saying agents will take on a growing role in "tasks that involve money, from purchases and payments to more complex transactions," and that the integration exists to keep those transactions secure, transparent, and under user control.
The behavioural shift is already visible in Visa's own data. Rubail Birwadker, the company's global head of growth, said more than one in five transactions are now influenced by what consumers learn through large language models, a figure he admitted arrived faster than Visa expected. People already ask ChatGPT what to buy. The infrastructure to let it finish the job is the logical, and lucrative, next step.
The Race to Own the Agentic Checkout
Visa is not acting out of generosity. Payment networks earn their economics by sitting in the transaction flow, and agentic commerce threatens to redraw who sits where. If AI platforms become the front door to shopping, the networks that fail to integrate early risk being designed around, whether by stablecoins, account-to-account transfers, or a rival network that moved faster. Notably, the same Visa Payments Forum featured announcements on stablecoin settlement and programmable tokens, a sign the company is hedging every layer of the stack at once.
The competitive stakes explain the urgency across the industry. Analyst forecasts cited around the announcement suggest agentic commerce could account for 15 to 25 percent of e-commerce sales by 2030. As Blumefield has reported through this year's wave of agentic finance launches, from fintech platforms processing agent-initiated payments to brokerages piloting AI-driven trading, the financial system is being rewired for non-human participants at remarkable speed.
For merchants, the calculus is uncomfortable. Most commerce infrastructure was built for a person clicking through a website, not an agent parsing product data and executing a checkout. Retailers that make their catalogues and payment flows legible to agents will capture this traffic. Those that do not may simply become invisible to the fastest-growing class of buyer on the internet.
What Happens When Software Holds the Wallet
The unresolved questions are not technical. Tokenization, spending caps, and fraud models are solved problems in payments. The harder questions are behavioural and regulatory. Who is liable when an agent misreads an instruction and buys the wrong thing at the right price? How do dispute frameworks built around human intent handle a transaction no human directly initiated? Regulators in Brussels and Washington, already scrutinising AI platforms on competition and consumer protection grounds, will find a rich new file to open.
There is also a quieter concern about influence. When an agent chooses among ten retailers selling the same product, the ranking logic becomes commercially decisive in a way that search rankings were, with even less visibility into how decisions are made. The entity that controls the agent controls the shelf.
None of this diminishes the significance of what was announced this week. Agentic commerce now has the backing of the network that processes a third of a trillion transactions a year, integrated into the AI platform with the largest consumer footprint on earth. The product will take time to arrive, and the first versions will be cautious, hedged with approval prompts and low spending limits. But the direction is set. The wallet is becoming an API, and the era of agentic commerce has its first piece of genuine financial plumbing.