ACP: The Commerce Layer for AI Agents
Customers now buy inside AI conversations. The Agentic Commerce Protocol from Stripe and OpenAI lets an AI agent run your checkout while you stay merchant of record. What it costs to adopt, who is already selling through it, and what a retailer should do now.
Cameron Mukherjee, Director · · Updated
For: Retailers, marketplaces and payments teams evaluating agent-driven checkout
Key points
- The Agentic Commerce Protocol (ACP), co-developed by Stripe and OpenAI, is an open standard that lets AI agents run checkout while the merchant stays merchant of record.
- ACP uses a SharedPaymentToken scoped by merchant, amount and time, so agents never hold raw card credentials.
- Merchants on Stripe can enable agentic checkout with minimal code changes; others implement the ACP endpoints directly.
- ACP is the commerce layer; AP2 provides authorisation proof and x402 provides settlement, and the three are designed to work together.
Part 1 of our three-part series on the agentic commerce stack, written for the directors of businesses that sell online and need to know what changes when the customer is an AI agent. This instalment covers ACP, the checkout layer. Part 2 covers AP2, the authorisation layer. Part 3 covers x402, the settlement layer.
In one paragraph
Since September 2025, people have been buying products inside ChatGPT conversations. The Agentic Commerce Protocol (ACP), an open standard co-developed by Stripe and OpenAI, is what makes that possible: it lets an AI agent run a checkout with a merchant while the merchant stays merchant of record, keeps its customer relationship, and never hands card details to the agent. For a retailer or marketplace, ACP is a new sales channel that can be switched on with modest engineering, and a channel that McKinsey forecasts could carry $3–5 trillion of global retail revenue by 2030.
Why checkout had to change
Online checkout was designed for a person with a browser: pages, forms, card fields in secure frames, redirects. An AI agent cannot click through that. It has no browser session, it does not retain form context, and it must never be handed raw card numbers.
The gap is commercial as much as technical. Bain estimates that 15–25% of US ecommerce will flow through AI agents by 2030. Merchants need a way to sell through agents without rebuilding their commerce stack, and agents need a programmatic checkout that keeps credentials out of their hands. ACP is that interface.
How it works, for a non-engineer
ACP defines four participants and keeps their responsibilities separate:
- The buyer talks to an AI agent, reviews options and authorises the purchase. They never give the agent their card. Instead their payment provider issues a scoped token.
- The agent finds products, opens a checkout with the merchant, shows totals and delivery options, collects the buyer's authorisation and completes the purchase. It never becomes the merchant of record.
- The merchant keeps full control. In Stripe's words: "you maintain your customer relationships as the merchant of record, retaining control over which products can be sold, how they're presented, how transactions are processed, and how orders are fulfilled." Pricing, stock, fraud decisions and fulfilment stay with you.
- The payment provider (initially Stripe) issues a SharedPaymentToken: a token limited to one merchant, one maximum amount and a short time window, used once. It contains no real card data. Stripe's fraud tooling still scores the transaction, so an agent-originated order gets the same accept or decline decision as any other.
The checkout itself runs over a handful of standard web endpoints (create, update, complete, cancel) specified by the protocol, with order updates sent back to the agent afterwards so the buyer can be told when the parcel ships.
What it costs to adopt
For a merchant already on Stripe, agentic checkout can be enabled with minimal code changes. For everyone else, ACP is an open specification: the checkout endpoints can be implemented on any stack, or exposed as a Model Context Protocol (MCP) server for AI-native architectures. Payment providers beyond Stripe can participate through the Delegated Payment Specification; direct integration requires PCI DSS Level 1 or the use of network tokens.
The larger cost is usually not the endpoints. It is product data. An agent can only sell what it can read: accurate titles, attributes, availability, delivery options and returns terms. Businesses whose catalogue lives in several systems that disagree will find that the integration work is the prerequisite.
Who is already doing it
ACP launched on 29 September 2025 with Instant Checkout in ChatGPT for US users, starting with Etsy merchants and expanding to Shopify. Payment details are, as OpenAI puts it, "processed securely by trusted payments partners, and not stored by OpenAI."
| When | What |
|---|---|
| April 2025 | Mastercard Agent Pay and PayPal Agent Toolkit launch |
| 29 September 2025 | ACP launches with ChatGPT Instant Checkout |
| 14 October 2025 | Salesforce Agentforce Commerce integration announced |
| 28 October 2025 | PayPal becomes an ACP-compatible payment provider, rolling out to its merchants from 2026 |
Direct early adopters include URBN (Anthropologie, Free People, Urban Outfitters), Coach, Kate Spade, Revolve and Ashley Furniture. Platform partners include Etsy, Shopify, Wix, WooCommerce, BigCommerce, Squarespace and commercetools. PwC has a consulting partnership to help larger enterprises structure product data and pilot integrations.
Where ACP sits in the stack
ACP is the commerce layer of three complementary protocols. As Orium's analysis puts it, "ACP, AP2, and x402 are not direct competitors so much as layers in an emerging agentic commerce stack."
| Protocol | Developed by | Answers |
|---|---|---|
| ACP | Stripe and OpenAI | How does an agent complete a purchase? |
| AP2 | Google and 60+ partners | Who authorised this transaction? |
| x402 | Coinbase and Cloudflare | How do the funds move? |
ACP is deliberately the path of least resistance: it adapts existing card infrastructure rather than replacing it, which is why it was deployable to Stripe's merchants on day one.
What this means for your business
- A new channel with measurable attribution. Orders arriving through an agent are tagged as such, so you can see what the channel is worth before investing further.
- Control stays with you. Which products, at what price, on what terms, with your fraud rules. Nothing about ACP obliges you to discount or to hand the customer relationship to a platform.
- Product data becomes a revenue asset. Businesses with clean, complete catalogue data will be recommended by agents; those without will be invisible to them.
- Payment risk does not increase. Scoped tokens and existing fraud scoring mean an agent order carries no more credential exposure than a web order.
What to do now
- Ask your platform or payment provider when agentic checkout is available to you and what it requires.
- Audit your product data for completeness and consistency across every system that holds it.
- Decide the commercial terms for the channel: which ranges, which prices, which delivery options.
- Instrument attribution so agent-originated orders are visible in your reporting from the first one.
How Hexploits helps
We build the integrations that give an agent one truthful view of stock, price and order status, and the agentic workflows and governance that keep an automated channel under control. For ecommerce and retail businesses we have written up the pressures we see most often on our sector page.
Request a proposal or talk to an engineer. You will have a written scope and an indicative price within two working days.
Continue to Part 2: AP2, the authorisation layer, or Part 3: x402, the settlement layer.
Questions this raises
What is the Agentic Commerce Protocol?
Does the AI agent see the customer’s card details?
What does a merchant need to do to support agentic checkout?
How Hexploits helps
- Agentic AI workflows with governance
An agentic workflow is an AI system that carries out multi-step tasks by calling your systems and tools, rather than only answering questions.
- Systems integration and API development
Systems integration connects the software a business already runs, such as finance, CRM, ERP, HR and industry systems, so data is entered once and flows automatically.
More insights
EU AI Act High-Risk Series, Part 5: Critical Infrastructure
11 September 2026
AI managing power grids, water supply or road traffic is high-risk under the EU AI Act only where it is a genuine safety component, and part of the scope test is still draft guidance. How utilities, operators and their vendors should classify, and what to build if they are in scope.
EU AI Act High-Risk Series, Part 4: Healthcare & MedTech
4 September 2026
Healthcare AI has two EU AI Act deadlines eight months apart, and which applies depends on whether your product is legally a medical device. How to classify correctly, what goes in the technical file, and why health-system procurement is already asking.
EU AI Act High-Risk Series, Part 3: Education & EdTech
28 August 2026
Admissions, grading and exam-monitoring tools are high-risk under the EU AI Act, and emotion recognition in proctoring is banned outright. What universities, exam boards and EdTech vendors have to build by December 2027, and why institutional buyers are already asking.
Next step
Have a question this raised?
Tell us about the system and the sector. A named engineer replies within one working day. A written scope and an indicative price within two working days of a short scoping call.