The Agentic Commerce Protocol decides who keeps the margin when an AI agent sells

The Agentic Commerce Protocol (ACP) is OpenAI and Stripe's attempt to standardize something that still had no standard: what happens the instant an AI agent, not a browser with a human behind it, completes a purchase. The specification is published at agenticcommerce.dev, is open source under the Apache 2.0 license, and already has ChatGPT as the first implementation on the agent side and Stripe as the first compatible processor.

For those building products with embedded commerce, the question is not whether the protocol is technically elegant. It's another one: in this new three-way flow (buyer, agent, business), who controls what, and where the margin lands.

What the ACP solves, and what it deliberately decides

Technically, the ACP defines how an agent passes checkout data to a business and how a secure payment token travels from the buyer to the seller without exposing the original credential. This covers physical goods, digital goods, subscriptions, and asynchronous purchases, and it works both as REST and as MCP, so a merchant doesn't need to rewrite its own payment infrastructure to accept orders coming from an agent.

But the document's most important design decision isn't technical, it's contractual: who is the merchant of record. The specification is direct about this.

With ACP, businesses maintain their customer relationship as the merchant of record, retaining control over which products can be sold, how they're presented, and how orders are fulfilled.

Agentic Commerce Protocol specification, OpenAI and Stripe

This means that ChatGPT, when intermediating the sale, does not take on the role of the store. It doesn't buy the inventory, doesn't issue the invoice, and isn't responsible for the chargeback or the return. The business that implemented the protocol on its own side is the one that keeps that responsibility.

Why the agent doesn't want to be merchant of record

Being the merchant of record is expensive: it involves fraud risk, regulatory KYC obligations, liability for chargebacks, and, in many jurisdictions, a specific license to operate. By designing the protocol so that the end business retains this role, OpenAI builds a high-volume storefront without inheriting the liability of being a retailer.

The documentation itself describes the value of this position for those building agents: "Embed commerce into your application. Let your users discover and transact directly with businesses in your application, without being the merchant of record." In other words: embed commerce into your app, let your users discover and transact directly with businesses, without becoming the store yourself.

For a founder thinking about putting a shopping agent inside their own product, this is the central lesson of the ACP: the model OpenAI tested with ChatGPT is replicable. You can operate a discovery and purchase-intermediation layer, charging by volume or by access, without setting up a retail operation. The trade-off is that you also give up the fine-grained control that a proprietary e-commerce store has over product presentation, dynamic pricing, and conversion data, because, according to the specification, the business on the other side is the one that decides those rules.

Stripe as toll collector: the Shared Payment Token

The second strategic piece of the ACP is Stripe's role. In theory, the specification allows any compatible payment processor to take part in the flow, but Stripe is named as the first compatible PSP, enabling the flow through what the site calls the Shared Payment Token: a token that carries the buyer's payment credential to the agent, and from the agent to the business, without either party seeing the raw card data.

Being the first compatible piece in a standard newly defined by OpenAI has an obvious network effect: every business that implements the ACP to sell through ChatGPT tends to implement it via Stripe first, because that's the documented, tested path with active support. This doesn't stop competitors from building equivalent tokens, but it gives Stripe months, possibly years, of advantage as the default path: every new commerce-agent integration is born testing against Stripe's infrastructure first.

The gatekeeper: why 'open' doesn't guarantee access

Here's the part the protocol's marketing leaves less explicit. The ACP, as a specification, is open: any business can implement it and, in principle, transact with any compatible agent. But the protocol's own FAQ page admits the limit of that openness.

Does implementing the protocol mean my products will automatically be listed through AI agents? No, each AI platform will manage their own process for how businesses can participate. If your business wants to participate in ChatGPT, you'll need to apply.

Agentic Commerce Protocol specification, OpenAI and Stripe

In other words: the protocol is a public road, but the biggest destination on it, ChatGPT, has a toll booth and a gatekeeper. Implementing the ACP correctly is a necessary condition to sell through AI agents, but it isn't a sufficient one: approval from whoever controls distribution (today, OpenAI itself) is still a separate business filter, not a technical one. This reproduces, in a new format, a dynamic that stores already know from marketplaces and app stores: the infrastructure can be open, but the discovery funnel remains the property of whoever holds the largest agent.

The counterpoint: the model also protects the small business

The strongest argument in favor of this design doesn't come from OpenAI or Stripe, it comes from the business itself that implements the protocol. By keeping merchant-of-record status, a small store or a SaaS with embedded commerce doesn't lose the direct relationship with the customer or the transaction data to the agent that brought the traffic. This differs from selling on a traditional marketplace, where the platform often retains the buyer's data and later competes for that same customer base.

If this design holds up, the ACP distributes risk in a way that's more favorable to the merchant than the classic marketplace model: the agent brings qualified demand, but whoever bills the customer, knows the customer, and decides product policy remains the seller. The cost of this advantage is giving up control over the discovery channel, which now depends on third-party approval.

What this changes for those building now

For a Brazilian founder thinking about AI-native commerce, the ACP is not just one more technical integration on the list. It's a signal of where the market is drawing the lines of power: the AI agent becomes the discovery and intent layer, the payment processor becomes the trust toll, and the business that sells remains, structurally, the one that takes on the risk of the sale.

Anyone building a product with embedded checkout has three concrete questions to ask before implementing: if access to the highest-volume channel (today, ChatGPT) depends on applying for and receiving OpenAI's approval, what is the criterion and the timeline for that approval; whether dependence on Stripe's Shared Payment Token creates de facto lock-in, even with the promise of multi-PSP compatibility; and whether giving up control over product presentation inside a chat interface is worth the volume of qualified demand the agent promises to bring.

Anyone who adopts the protocol without answering that is betting on a distribution channel, not implementing a neutral technical standard. OpenAI has already made clear, in the specification's own text, that other AI platforms interested in adopting the ACP can reach out to acp@stripe.com: the standard is open to more agents, which means the power structure described here tends to repeat itself, not remain restricted to ChatGPT.