AI Just Learned to Finish the Job

Visa, Mastercard, and Amex just gave AI agents the ability to actually buy something — not just recommend it.

For the last two years, AI could help you shop and then it handed you back the checkout page. It could compare headphones, find a flight, draft the perfect gift list, and then it stopped right at the part where money actually changes hands. That gap just closed, and I think it's one of the more underappreciated shifts happening in the economy right now.

On June 10, Visa announced it's integrating its payment network directly into OpenAI's agent systems, letting ChatGPT complete purchases at any Visa-accepting merchant. Mastercard rolled out a parallel framework the same day, built for high-frequency, machine-to-machine payments. American Express followed in April with its own agentic commerce kit and purchase protection built specifically for AI-initiated transactions. Google has its own protocol. Stripe has one too. In the span of about a year, every major payment rail in the country decided, independently, that agents are about to become real participants in the economy.

This isn't AI helping you shop anymore. It's AI finishing the purchase.

What actually works today is more capable than most people realize. Link a card, set your limits, and an agent can shop across any Visa-accepting merchant, complete the checkout, and hand you a confirmation, no re-typing your card number, no bouncing between ten browser tabs. The same rails are being built to handle more than retail too: business invoice payments, an AI coding agent buying its own compute or API access, even recurring machine-to-machine purchases that happen continuously in the background with no human in the loop at all. That's a genuinely different category of convenience than "AI helps you shop." It's AI that finishes the job.

What's notable isn't just that this is happening, it's how fast the model changed. OpenAI's first attempt at this, an in-chat "Instant Checkout," got quietly retired after less than a year. It charged merchants a fee, and adoption never really took off. The version that replaced it flips the model: agents discover and negotiate inside the chat, but the actual transaction routes through the merchant's own site using the card infrastructure that already exists. That's a smarter design. It doesn't ask merchants to trust a new system, it just lets AI plug into the one they already trust.

The obvious question is liability. If an agent buys the wrong size, the wrong flight, or ten of something instead of one, who's on the hook? I don't think that question gets taken seriously enough in most of the AI hype cycle, but the payments industry has actually been building real answers to it, not just promises. Every one of these frameworks is layering on the same core protections: spend caps you set yourself, merchant restrictions, approval prompts for anything above your comfort threshold, and a way to identify and verify the agent itself before it's ever allowed to touch your card. American Express has gone a step further, backing registered agent purchases with dedicated purchase protection, the same instinct that's made their card trusted for decades, now extended to transactions a human never actually clicked "buy" on. Visa's version leans on the fraud detection, chargeback, and refund infrastructure it already runs for billions of human transactions a day, just extended to cover a new kind of buyer. None of that makes the system foolproof. But it tells me the industry isn't racing ahead of the safety question, it's building the guardrails and the capability at the same time, which is exactly the order I'd want it done in.

McKinsey and the International Council of Shopping Centers have put out a joint estimate of roughly $1 trillion in AI-orchestrated US retail revenue by 2030, and $3 to 5 trillion globally. I'd treat that as a forecast, not a measured outcome, the kind of number that tends to get revised in both directions before it's real. But directionally, it points at something I've believed for a while: the biggest winners in this next wave of AI aren't necessarily the flashiest model companies. They're the infrastructure layers that get to sit in the middle of every transaction and take a toll for making it trustworthy.

That's the same instinct behind a lot of how we think about opportunity at Gravette Capital, less about betting on the shiniest new interface, more about identifying which rails become indispensable once the novelty wears off. Payments have always been one of the best businesses in the world precisely because they're boring, durable, and hard to replace. Agentic commerce doesn't change that. It just adds a new kind of customer to the rail, one that never sleeps, never gets distracted, and transacts at machine speed.

We're early. Most merchants still aren't built to be "read" by an agent the way a human reads a storefront. But the infrastructure race already has a winner's circle forming, and it's worth watching who's standing in it a year from now.

"The agent doesn't need convincing. It needs infrastructure it can trust. Build that, and you've built something durable."

— Paul Gravette, Founder & CEO, Gravette Capital