Hey hey,

How people buy things online is pretty much changing.

For two decades, the checkout flow was roughly the same: browse a website, add items to a cart, enter payment details, and click purchase.

The interface changed, from desktop to mobile, from typed searches to personalised feeds, but the transaction always ended with a human.

Now AI agents are starting to handle that last step.

A user tells ChatGPT, "Find me running shoes under $150 in my size," and the agent searches, compares options, and, with the user's permission, completes the purchase.

The human never visits a product page.

That creates a problem that no payment platform was designed to solve.

How does a business accept a purchase from an AI agent? How does it know the agent is authorised? How does it prevent fraud when the buyer is a bot?

How does it maintain its brand, customer relationships, and control over the transaction when the checkout happens in someone else's AI?

As an answer, Stripe and OpenAI co-developed the Agentic Commerce Protocol (ACP).

It is an open standard, Apache 2.0 licensed, that defines how AI agents, businesses, and payment providers interact during a programmatic purchase.

The first implementation supports Instant Checkout in ChatGPT. US users can now buy from Etsy sellers and, soon, over a million Shopify merchants directly in chat.

Let’s dig in!

Stripe spent a year exploring agentic commerce before building ACP. Three problems surfaced across every conversation with businesses and AI developers.

When an AI agent initiates a transaction on behalf of a buyer, the business needs to verify that the purchase is legitimate.

Traditional fraud signals, such as IP address, fingerprint, and browsing behaviour, don't exist when the buyer is an AI agent making an API call.

Businesses need a way to confirm purchases, securely accept payment credentials, respond to fraud signals specific to bot-initiated transactions, and update their risk models to differentiate good bots from bad ones.

Businesses want to sell through AI agents. It is a new distribution channel. But they don't want to build custom integrations for each agent.

If ChatGPT, Gemini, Claude, and other vertical AI assistants want to allow purchases, a merchant can't maintain a dozen bespoke payment flows.

The ecosystem needs a standard, just as the early web needed HTTP and the early mobile era needed app store payment APIs.

Commerce is not one thing. There are physical goods, digital products, subscriptions, pre-orders, in-store pickups, and dynamic pricing.

A protocol that only handles simple one-time purchases for physical goods would miss most of what businesses actually sell.

ACP needed to support the full complexity of modern commerce from day one, or agents would be limited to the simplest transactions.

The protocol defines four roles in every transaction, each with clear responsibilities and boundaries. The buyer finds a product through an AI surface, such as a chat interface, a voice assistant, or a recommendation agent.

They select what they want, provide or confirm a payment method, and grant permission for the agent to initiate checkout on their behalf.

The buyer remains in control. No purchase happens without consent. The AI agent serves as the interface between the buyer and the business.

It presents products, displays checkout information, and collects payment details. When the buyer confirms, the agent sends a checkout request to the business.

The agent handles the conversation. It does not handle the money. The business receives the checkout request and decides whether to accept it.

This is a critical design choice: the business is always the merchant of record.

It retains control over which products can be sold through agents, how they are presented, how transactions are processed, and how orders are fulfilled.

The business is not disintermediated. It is extended into a new channel.

The payment provider handles the financial mechanics. The AI agent communicates with the business's payment provider to relay payment credentials via a secure token.

The token is programmatically controlled, permissioned, and logged.

The payment provider charges the buyer, settles with the business, and handles disputes, the same role it plays in any online transaction.

The key architectural decision is where credentials flow. The buyer's payment information goes to the payment provider, not to the AI agent.

The AI agent never holds card numbers, bank details, or authentication tokens. It holds a session-scoped reference that the payment provider can resolve.

This separation makes the system trustworthy. The AI agent can initiate a transaction without ever being able to charge the buyer independently.

This mirrors the security pattern Anthropic uses in Managed Agents architecture: credentials are kept structurally unreachable from the code acting on behalf of the user.

In both cases, the insight is the same. Trust is an architecture problem. You do not tell the agent, "Please don't read the credentials." You make the credentials inaccessible.

Stripe could have built a proprietary payment API for AI agents and licensed it. Instead, they made ACP open source under Apache 2.0.

It's because of what happened with HTTP, OAuth, and MCP.

When a new interaction pattern comes, such as web pages, AI tool use, or agent-initiated commerce, the ecosystem fragments if every player builds their own protocol.

Fragmentation raises costs and slows adoption. An open standard means a business deploys ACP once and can accept purchases from any compatible AI agent.

An AI agent develops ACP once and can purchase from any compatible business.

The integration work is O(1) for both sides, regardless of how many agents or businesses enter the ecosystem.

OpenAI's decision to deprecate its proprietary Assistants API in favour of MCP for tool use, with a mid-2026 sunset, suggests that the industry is consolidating around open protocols. ACP follows the same trajectory for commerce specifically.

Also read: Why Your AI Agent Always Forgets Users?

ACP is not theoretical. It is live in production. Instant Checkout in ChatGPT allows US users to buy from Etsy sellers directly in conversation.

Shopify merchants, including Glossier, Vuori, Spanx, and SKIMS, are being onboarded, bringing over a million storefronts into the system.

For businesses already processing payments through Stripe, enabling agentic commerce requires updating as simple as one line of code.

For businesses using other payment providers, Stripe can be added specifically for agentic payments without changing their existing setup.

This low integration cost is intentional. If enabling a new sales channel required a major engineering project, adoption would stall. The early results are shaping Stripe's view.

Transactions are currently simple. It's single-item purchases from a single merchant.

However, ACP is designed for complexity that isn't here yet: multi-merchant carts, "buy for me" background purchases where the agent shops autonomously, usage-based billing, and subscription management.

The protocol is intentionally over-specified for today's use cases, so it does not need to be rewritten when tomorrow's arrive.

If you are building an AI product that involves transactions, any agent that might need to book, purchase, subscribe, or pay on behalf of a user, ACP defines how that works.

If you are building for e-commerce, the question is no longer whether AI agents will be a sales channel because they already are. It is whether your checkout is agent-ready, or you are invisible to the fastest-growing discovery surface in commerce.

Stripe is betting that agentic commerce will follow the mobile commerce pattern: scepticism, then experimentation, then a rapid shift once the infrastructure makes it easy enough.

ACP is the infrastructure play. The pipes that make the shift frictionless. Whether that bet is right depends on whether consumers trust AI agents to buy things for them.

The early signs, from ChatGPT's Instant Checkout rollout, suggest they are willing to try. The protocol is ready. The behaviour is catching up.