Agentic Commerce Protocol: who keeps the sale when the buyer is an AI agent

A standard for buying without opening a browser

Stripe and OpenAI maintain a published specification for the Agentic Commerce Protocol (ACP), available at agenticcommerce.dev. It is an open standard, under the Apache 2.0 license, for the purchase flow between a human buyer, an AI agent, and a business. In practice, it defines how an assistant like ChatGPT can build a cart, close an order, and process payment within the conversation itself, without redirecting the user to the store's website.

The protocol is compatible with both REST and MCP (Model Context Protocol), which means a company can expose its checkout in two different ways depending on the stack it already uses. This matters because it doesn't require rewriting commerce infrastructure: the stated idea from Stripe and OpenAI is that ACP works on top of the order backend and payment processor the business already has.

The clause that matters more than the API

The detail that sets ACP apart from a simple shopping plugin is one specific sentence in the documentation: the business remains the merchant of record. That means whoever sells through an agent keeps the relationship with the customer, decides which products enter the agentic catalog, how they are presented, and how the order is fulfilled.

In business terms, this clause addresses the biggest objection any retailer would have before opening its catalog to ChatGPT: becoming just an invisible supplier to a platform that keeps the customer data, the purchase history, and the recurring relationship. With ACP, the sale technically still belongs to the store. The question left open is whether this contractual guarantee holds up once OpenAI concentrates all the discovery demand.

How money changes hands without exposing the card

The payment side solves a concrete technical problem: how an agent processes a transaction without ever having access to the user's sensitive card data. Stripe's answer is the Shared Payment Token, a token that carries the payment authorization without carrying the card number itself.

One possible flow, compatible with this token logic, works like this:

• the buyer authorizes payment within the agent's interface (today, ChatGPT);
• Stripe generates a shared token that represents this authorization;
• the agent passes this token to the business's checkout endpoint;
• the business processes the charge with its own payment provider, without ever seeing the raw card.

This keeps the transaction within PCI compliance scope without forcing the business to blindly trust the agent. Today, Stripe is the only compatible processor cited in the specification, but the protocol was designed to accept any PSP that implements the same token logic.

Who decides who gets into the catalog

The specification makes clear that implementing ACP does not guarantee automatic listing in any agent. Each AI platform manages its own curation process: to sell inside ChatGPT, a business must apply and be accepted. OpenAI itself is, for now, the only AI platform that has implemented the protocol, according to the official FAQ.

In practice, this creates an approval layer that the open standard alone does not remove. The business, in turn, can accept or reject transactions from specific agents, case by case or through a custom rule, which gives some symmetry: rejection is bilateral. But whoever controls the largest volume of AI assistant users today is OpenAI, and it is OpenAI that decides the pace of expansion of the agentic catalog.

What changes in the funnel and in attribution

For those building e-commerce or digital products in Brazil, ACP rewrites a stage of the funnel that was, until now, guaranteed: traffic arrived through the site, and the store controlled the storefront, the price comparison, and the entire checkout. With purchases happening inside an agent, the storefront is now decided by whoever builds the assistant's response, and checkout happens without the store controlling the visual experience or the moment of sale.

This has three direct implications:

1. Revenue attribution moves elsewhere: a sale closed inside ChatGPT doesn't pass through the site's traditional analytics funnel, so measuring agentic commerce performance requires new instrumentation, separate from the usual e-commerce setup.
2. Product SEO gains a sibling: if discovery starts happening inside an agent, optimizing only for search engines no longer covers the whole journey; the competition now includes how the catalog is exposed via ACP.
3. Margin negotiation changes hands: today Stripe charges as a normal processor, but the discovery gatekeeper position OpenAI holds gives it, in the medium term, bargaining power over which businesses appear first.

The counterpoint: an open standard is not guaranteed neutrality

The best argument in favor of ACP is that it is, in fact, open: any business can implement the specification, any PSP can adopt the token model, and nothing stops another AI platform competing with ChatGPT from implementing the same standard tomorrow. This avoids the most feared scenario, a proprietary protocol locked inside a single company.

But openness of specification is not the same as openness of distribution. OpenAI is, today, the only platform that decides who gets into ChatGPT, and ChatGPT is, today, the largest existing agentic commerce channel. A standard can be Apache 2.0 and still have its practical application controlled by whoever holds the largest user base on the agent side. It's the same dynamic apps faced with app stores: the rule is public, but who approves entry is private.

What still can't be measured

So far, the public documentation doesn't include figures for volume transacted via ACP, nor does it list how many businesses have already been approved to sell inside ChatGPT. There is also, so far, no other relevant AI platform that has implemented the protocol besides OpenAI itself, nor any other PSP besides Stripe with the Shared Payment Token in production.

For those deciding to invest engineering effort in exposing a checkout via ACP now, the bet is twofold: believing that agent-driven purchase volume will grow enough to justify the effort, and accepting the risk that curation of who enters the catalog remains concentrated in a single platform for an indefinite time.