AI Agents Can Now Pay for Things: Inside the New 'Agentic Commerce' Protocols

For most of the past few years, "AI agent" meant something that could research, draft, or click around a website on your behalf, then hand the last step — entering a card number, hitting "place order" — back to you. That last step is now getting automated too. Over the past year, Visa, Mastercard, Google, OpenAI, and Stripe have each shipped a real, working protocol for letting an AI agent complete a purchase directly, and the approaches differ enough that it's worth understanding what's actually happening before you let a chatbot near your payment method.

What's live right now

The most concrete example is inside ChatGPT. OpenAI and Stripe built Instant Checkout, which lets US ChatGPT users buy directly from Etsy sellers in the chat window, with a large number of Shopify merchants (Stripe's announcement named brands like Glossier, SKIMS, Spanx, and Vuori) rolling out as well. It runs on the Agentic Commerce Protocol (ACP), a standard OpenAI and Stripe co-developed and published openly so other merchants and platforms can adopt it. The pitch to merchants is that it's a small integration change if they already use Stripe; the pitch to shoppers is that the price isn't marked up and ChatGPT's product recommendations aren't influenced by who's paying for placement, according to OpenAI's own announcement.

That's one working implementation, but it isn't the only standard in play, which is exactly where this gets more complicated.

The card networks built their own verification layers

Visa and Mastercard, whose business is verifying and settling payments rather than building chat interfaces, each shipped something different: a way for a merchant to confirm that an incoming request genuinely comes from an authorized agent acting for a real customer, rather than a scraper or a bad actor pretending to be one.

Visa's version, the Trusted Agent Protocol, was built with Cloudflare and works by attaching a cryptographically signed identity to the agent's web requests. A merchant's server can check that signature against a Visa-run directory and treat the request as coming from a known, accountable agent instead of an anonymous bot — which matters given that retailers have reported enormous increases in AI-driven traffic hitting their sites over the past year, much of it from crawlers and agents with no way to verify who or what they were.

Mastercard's Agent Pay, announced with Microsoft and IBM as early partners, takes a similar goal and applies it to the token itself: it issues an "Agentic Token" scoped to one agent, one merchant, and one consent policy, so an agent can complete a checkout without ever handling the actual card number. Mastercard has since extended the idea further with Agent Pay for Machines, aimed at agent-to-agent payments — think one AI system paying another automatically for API usage or data, potentially in amounts as small as a fraction of a cent.

Google's approach: an open standard, not a network

Rather than building this on top of its own payment rails, Google published AP2 (Agent Payments Protocol) as an open specification that other companies can build on, and got dozens of partners — including PayPal, Mastercard, American Express, and Coinbase — to sign on at launch. AP2's core idea is a pair of signed digital records: an "Intent Mandate" that captures what you actually asked the agent to do, and a "Cart Mandate" that captures the specific purchase the agent is proposing, both cryptographically tied back to you. The idea is that if an agent buys the wrong thing, there's a verifiable record of exactly what it was authorized to do and what it actually did — useful for resolving disputes about whether the agent or the merchant made the mistake. Google later handed governance of the protocol to the FIDO Alliance, the same industry body that manages passkey standards, which suggests it's aiming for AP2 to become shared infrastructure rather than a Google product.

Why this fragmentation is normal, and mostly fine, for now

Having Visa, Mastercard, Google, and OpenAI each ship something different sounds chaotic, but it maps onto a familiar pattern: card networks build verification and settlement, platforms build the shopping experience, and an open protocol tries to standardize the parts everyone needs (proof of consent, a clear record of what was authorized). Several of these efforts are explicitly designed to interoperate rather than compete — Mastercard has said its framework is meant to work alongside AP2 and similar protocols, not replace them.

What actually matters if you try this

The useful question isn't "which protocol is best" — you won't get to choose that as a user, the merchant and platform decide it for you. The useful question is what you're actually authorizing. Look for whether a purchase requires your confirmation before it completes, whether the amount and merchant are scoped in advance rather than left open-ended, and whether you can revoke the agent's access without canceling your entire card. Every protocol described above was designed with some version of scoped, revocable permission rather than handing an agent your raw card number and unlimited discretion — but "designed with" and "implemented carefully by every merchant" aren't the same guarantee. Treat an AI shopping agent the way you'd treat a very literal-minded assistant: fine for a specific, bounded task you've described clearly, and worth watching closely the first several times before you trust it with anything you'd be annoyed to have gotten wrong.