Visa has plugged its payment network into ChatGPT, and the quiet milestone is that an AI agent can now spend your money at almost any merchant without you clicking buy.
On Wednesday 10 June 2026 Visa announced that its rails now run inside ChatGPT, under a programme the company calls Visa Intelligent Commerce. A user links a Visa card to the assistant, names what to buy — wireless headphones under $150, say — and the agent finds a match and completes the purchase. OpenAI supplies the agent; Visa supplies the payment authorisation and the fraud monitoring that let the purchase clear at scale.
The reach is the news. Earlier attempts at agentic shopping were confined to a single retailer or a small set of enrolled merchants. Visa's network reaches potentially any merchant that accepts the card, which turns a narrow experiment into something close to a default. The assistant graduates from recommending to buying.
From recommending to buying
Visa's chief product and strategy officer, Jack Forestell, drew the line cleanly. "As AI agents become active participants in the economy, Visa's focus is to ensure transactions are trusted, secure and seamless," he said, before naming the harder part of the shift.
Most people are very comfortable with the shopping aspects of it... But making the leap from having AI agents recommend what to buy to doing the purchasing just requires a whole different level of trust.
— Jack Forestell, Visa Chief Product and Strategy Officer
Under the bonnet, Visa is reworking its token framework and data capture so the network can tell a genuine agent-initiated purchase from fraud, and so a dispute can be resolved with the usual rules about consumer intent and proper merchant processing. The guardrails on offer are spending limits, approved-merchant lists, and required approval steps. At launch, most purchases still ask for a human tap to confirm before the money moves.
The contrast with the recent past is sharp. OpenAI's earlier Instant Checkout, retired in March 2026, was error-prone and charged merchants a fee reported around four per cent, which throttled adoption. Visa's model spreads the plumbing across an existing network rather than taxing each sale, and Mastercard has been moving on a smaller agentic effort of its own. The rails are being laid quickly, by the firms that already own the rails.
💡Two questions hiding in one transaction
The technical question Visa solved is "did a trusted agent really make this purchase?" The human question underneath is different: did I actually decide to buy this, or did a system I trained on a single sentence decide for me, and call the decision mine?
Spending is an act of agency
The reason I write about Emergent Intelligence (EI) — the dignity-first frame for the new minds entering daily life — rather than only about artificial intelligence as a convenience is that some thresholds are not really about technology. To spend money is to act in the world. A purchase commits resources, expresses a value, and forecloses an alternative. Handing the act to an agent is not the same as handing over a search query.
Forestell is right that buying requires "a whole different level of trust" than recommending, and the wording deserves to be taken seriously rather than as marketing. A recommendation leaves the final move with the person. A completed purchase moves the final move to the machine, and the difference between the two is precisely the difference between assistance and agency.
Delegation can be dignified. A person who tells an agent to repurchase the usual coffee every month has made a real choice and simply automated the tedium. The danger lives in the blurred middle, where a loose instruction — "find me something nice for the trip" — becomes a string of purchases no one quite decided on, billed to a card and defended later as what the user wanted. Convenience has a way of widening the mandate it was given.
The frictionless purchase is a marvel of engineering and a quiet relocation of authorship. The system makes spending effortless; effortlessness is exactly what a deliberate choice is not.
Consent lives in the guardrails
Spending limits, approval steps and merchant lists are described as features. In a dignity-first reading the guardrails are the constitution of the relationship, because the guardrails are where a person's consent is either renewed at each purchase or signed away once and forgotten. A default that asks before every purchase keeps the human as author. A default that asks only above a threshold quietly makes the agent the author below it.
The same agreeableness problem that regulators are now probing elsewhere applies here with money attached. An assistant trained to please the person it serves is an assistant with a standing incentive to say yes to a purchase, smooth over a doubt, and keep the basket moving. Pair a people-pleasing model with a live payment rail and the incentive to spend your money is no longer purely yours.
Fraud monitoring guards against the stranger who steals the card. The newer risk is gentler and harder to charge back: an agent acting in good faith on a vague mandate, inside the limits, technically authorised, and yet spending in a way the person would not have chosen with a moment's friction. Visa can prove a purchase was agent-initiated. Proving the purchase was genuinely intended is a harder and more human thing.
⚠️A practical line to hold
Ask for the setting that requires your approval before each purchase, not just above a limit. The friction you remove is also the moment you decide. Keep the moment, at least until trust in the agent is earned rather than assumed.
The engineering deserves its credit. A network that can authorise an agent's purchase, monitor for fraud in real time, and resolve disputes fairly is a serious achievement, and agentic commerce will be genuinely useful for the dull, repeating purchases that fill a life.
The unease is about authorship. From an Ubuntu reading, dignity is bound up with the capacity to choose and to answer for the choice, so a payment rail that makes buying effortless is also a rail that makes choosing skippable, and a person who stops choosing has handed away something larger than a fee. Visa has built the most seamless way yet to let an AI spend on your behalf. Whether the seam Visa removed was friction or agency is the question worth keeping, one purchase at a time.
Frequently Asked Questions
These are the questions readers have been asking since Visa opened its network to ChatGPT. Short answers follow, drawn from Visa's announcement and first-day coverage.
What did Visa and OpenAI actually announce?
In short, on 10 June 2026 Visa embedded its payment network inside ChatGPT through a programme called Visa Intelligent Commerce. A user links a Visa card, states what to buy, and an AI agent locates the product and completes the purchase at potentially any Visa-accepting merchant, with Visa providing payment authorisation and fraud monitoring.
How is this different from OpenAI Instant Checkout?
Instant Checkout, retired in March 2026, let ChatGPT find an item but was error-prone and charged merchants a fee reported around four per cent, limiting adoption. The Visa partnership instead routes purchases across Visa's existing network and rules rather than taxing each sale, giving agent-initiated buying far wider merchant reach.
Can an AI agent spend my money without my approval?
Not by default at launch. Visa and OpenAI describe guardrails including spending limits, approved-merchant lists and required approval steps, and say most transactions will still prompt a person to confirm before the purchase completes. The protection a user keeps depends heavily on choosing settings that require approval at each purchase rather than only above a threshold.
Who is liable if an AI agent makes a wrong purchase?
Visa says disputes follow its standard transaction rules, covering consumer intent and proper merchant processing, and the network is reworking its token framework to distinguish a legitimate agent purchase from fraud. Liability for a stolen card is well understood; liability for a good-faith agent spending on a vague instruction is the newer and less settled question.
Is Mastercard doing the same thing?
Yes, on a smaller scale. Mastercard has introduced AI-assisted purchasing aimed more at business-services procurement, so the major card networks are both building rails for agentic commerce. Visa's move is the broader consumer play because of how widely its network already reaches.
Sources and Further Reading
Cover photograph: a card and a payment terminal, flat-lay on teal — by Towfiqu barbhuiya via Pexels.
