What the Agentic Commerce Protocol actually is
The Agentic Commerce Protocol is described on its own spec site as “an open standard for programmatic commerce flows between buyers, AI agents, and businesses”. Stripe and OpenAI co-developed it, it is open source under the Apache 2.0 licence, and ChatGPT was the first AI platform to implement it, with Stripe as the first compatible payment provider.
The specification, maintained on GitHub by OpenAI and Stripe as founding stewards, covers more than most summaries suggest: an agentic checkout API, a delegated payment spec (so an agent can pass a payment credential without exposing the underlying card), product feeds, capability negotiation, and both REST and MCP transport options. It uses date-based versioning; the latest stable release at the time of writing is 2026-04-17, and the project describes itself as being in beta.
On the OpenAI side, the practical entry point today is the product feed. OpenAI’s developer documentation frames ACP as “the connective layer between merchants and ChatGPT users”, enabling ChatGPT to “ingest structured catalog data, understand merchant inventory, and surface relevant products in context”. Merchants supply a structured feed — via daily file upload or an API — with titles, descriptions, images, price and availability, and feed onboarding is currently limited to approved partners who apply via a form. That gating matters: ACP the spec is open to anyone; distribution in ChatGPT is not.
The Instant Checkout arc: shipped, then rethought
Any honest account of ACP has to include what happened to its headline feature. In September 2025, OpenAI launched Instant Checkout: US ChatGPT Free, Plus and Pro users could buy directly from US Etsy sellers inside the chat, single-item purchases only, with “over a million Shopify merchants” announced as coming soon. Merchants remained the merchant of record, handled fulfilment and support through their existing systems, and paid “a small fee on completed purchases” — OpenAI has not published the rate. OpenAI was explicit that Instant Checkout items were not preferred in product rankings.
Roughly six months later, in March 2026, OpenAI changed course. In its “Powering Product Discovery in ChatGPT” announcement, the company wrote: “We’ve found that the initial version of Instant Checkout did not offer the level of flexibility that we aspire to provide, so we’re allowing merchants to use their own checkout experiences while we focus our efforts on product discovery.” ACP itself was not abandoned — OpenAI described it as “expanding” to power discovery — but the completed-purchase-inside-ChatGPT experience was set aside. Shopify merchants’ products now surface via the Shopify Catalog, with purchases completed on the merchant’s own store through an in-app browser, and Walmart launched a dedicated in-ChatGPT app with its own account linking and payments rather than relying on native checkout.
So the current state, verified against OpenAI’s own pages in August 2026: ACP is live as a discovery and feed protocol inside ChatGPT; the agentic checkout and delegated payment specs still exist and are actively versioned in the open-source repo; but first-party checkout inside ChatGPT is no longer the shipping consumer experience. Stripe, for its part, still documents ACP as one of the protocols sellers can use to transact with agents, and lists its agent-side commerce tooling as a private preview with a waitlist.
ACP, Google’s AP2 and the protocol landscape
Google’s parallel effort approaches the problem from the payments layer rather than the shopping surface. The Agent Payments Protocol (AP2), announced in September 2025 with more than 60 payments and technology organisations including Mastercard, PayPal, American Express and Adyen, is “an open protocol… to securely initiate and transact agent-led payments across platforms”. Its core mechanism is cryptographically signed “mandates” — tamper-proof records of what the user authorised an agent to buy, at what price — designed to answer the authorisation, authenticity and accountability questions that arise when a machine spends money. AP2 is at version 0.2, has been donated to the FIDO Alliance for standardisation, and currently supports card payments with wallets, real-time bank transfers and digital currencies on the roadmap. It extends Google’s Agent2Agent (A2A) protocol and complements MCP.
A third layer has since emerged: the Universal Commerce Protocol (UCP), an Apache-2.0 standard developed by Google with industry partners covering catalogue search, cart, checkout and order management, which uses AP2 for the payment leg. Stripe’s seller documentation now lists both UCP and ACP as supported paths — a useful signal that payment infrastructure providers are hedging across protocols rather than picking a winner.
| Dimension | ACP (Agentic Commerce Protocol) | AP2 (Agent Payments Protocol) |
|---|---|---|
| Stewards | OpenAI and Stripe (open source, Apache 2.0) | Google with 60+ payments/tech partners; donated to FIDO Alliance |
| Layer standardised | Full commerce flow: product feeds, checkout API, delegated payment | Payment authorisation: signed intent/cart mandates, credentials |
| First surface | ChatGPT (Instant Checkout, Sept 2025; discovery-centric since March 2026) | No consumer surface of its own; designed for any agent platform |
| Status (Aug 2026) | Beta; latest spec release 2026-04-17; feeds live with approved partners | v0.2 specification; standardisation in progress at FIDO |
| Trust model | Merchant stays merchant of record; user confirms each step | Cryptographic mandates create a non-repudiable audit trail |
| Relationship to MCP | Offers MCP as a transport option | Positioned as complementary extension alongside A2A and MCP |
The B2B SaaS translation — clearly labelled inference
Everything above is verifiable fact. What follows is inference: ACP today is consumer-retail-centric — physical and digital goods, subscriptions, product feeds with prices and availability. Nothing in OpenAI’s or Stripe’s documentation describes a B2B software sales flow. But B2B teams would be careless to ignore it, because the pattern generalises.
B2B has no “add to cart” — but it has a checkout-equivalent
In B2B SaaS, the conversion event isn’t a card payment; it’s a booked demo, a trial signup, or a partnership application. If AI agents become a meaningful research layer for buyers — and agentic buyers are already hitting B2B sites — then the ACP arc predicts what they’ll need from you: first, structured data about what you sell; second, a machine-completable next step. An agent that can parse your offer but can’t book the demo delivers a worse handoff. An agent that can do neither recommends your competitor.
What the retail arc actually teaches
The most instructive fact in this whole story is which half of ACP survived contact with reality. Checkout-in-chat was retired within six months; the structured product feed became the centre of the protocol. Discovery data outlived transaction plumbing. The B2B read: the durable, low-regret investment is making your offer machine-readable — clear entity data, honest comparison content, pages written so agents can parse and quote them accurately — not building speculative transaction endpoints for protocols that are still reshaping themselves quarterly.
What to do now versus what’s premature
Worth doing now:
- Structured offer data. Schema.org markup for your products, plans and FAQs; consistent naming of what you sell across every page. This is the B2B analogue of the ACP product feed, and it pays off in AI search visibility today regardless of how commerce protocols shake out.
- Agent-completable next steps. Audit your demo-booking and application forms: labelled fields, no unnecessary steps, a clear machine-discoverable path from “what is this product” to “book a conversation”. If an agent (or a hurried human) can’t complete your form, that’s revenue leaking now.
- An MCP-shaped view of your capabilities. ACP supports MCP as a transport and AP2 positions itself alongside MCP, which suggests exposing structured actions to buying agents via MCP is the most protocol-neutral bet a B2B team can place.
- Instant follow-up on agent-referred leads. A buyer whose agent researched five vendors and completed your form expects a response in minutes, not days. This is where autonomous outreach earns its keep: Zian’s SmartReach AI™ orchestrates message, channel and timing per lead with intelligent follow-up pacing, so the handoff from a buyer’s agent lands with a seller that responds like software but converses like a person.
Premature:
- Building a bespoke ACP checkout endpoint for a B2B product. The spec’s checkout flow assumes cart-style purchases; OpenAI’s own consumer implementation of it has already been withdrawn once.
- Betting exclusively on one protocol. With ACP, AP2 and UCP all under active development and Stripe supporting multiple paths, lock-in is the main risk.
- Expecting agents to sign contracts or complete negotiated enterprise purchases. Nothing shipped today does this; complex B2B deals still end in a human conversation — which is why the sensible preparation is faster, better-instrumented human-facing conversion, not imaginary machine-to-machine closing.
Sources & ownership
| Claim | Owner | Owner URL |
|---|---|---|
| ACP is “an open standard for programmatic commerce flows between buyers, AI agents, and businesses”; developed by Stripe and OpenAI; Apache 2.0; ChatGPT first implementer | OpenAI & Stripe (ACP project) | agenticcommerce.dev |
| Instant Checkout launch: US users buying from US Etsy sellers, single-item, Shopify “coming soon”; merchant of record stays with merchant; “small fee on completed purchases”; results not preferred | OpenAI | openai.com/index/buy-it-in-chatgpt |
| March 2026 pivot: “initial version of Instant Checkout did not offer the level of flexibility…”; ACP expanded to product discovery; Walmart in-ChatGPT app; Shopify purchases completed on merchant’s store | OpenAI | openai.com/index/powering-product-discovery-in-chatgpt |
| Product feed requirement (file upload or API), feed onboarding limited to approved partners; ACP as “connective layer” for catalogue ingestion | OpenAI | developers.openai.com/commerce |
| ACP spec contents (agentic checkout, delegated payment, feeds, MCP), beta status, date-based versioning, latest release 2026-04-17 | OpenAI & Stripe (GitHub) | github.com/agentic-commerce-protocol |
| Stripe supports sellers via UCP or ACP; Shared Payment Tokens; agent-side tooling in private preview | Stripe | docs.stripe.com/agentic-commerce |
| AP2 announcement: 60+ partner organisations; signed mandates; extension of A2A, complements MCP | cloud.google.com (AP2 announcement) | |
| AP2 v0.2; donation to FIDO Alliance; card payments supported now, other rails on roadmap | AP2 project (Google) | ap2-protocol.org |
| UCP scope (catalogue, cart, checkout, orders) and integration with AP2/A2A/MCP | UCP project (Google & partners) | ucp.dev |
FAQ
Is the Agentic Commerce Protocol open source?
Yes. The protocol’s own site, agenticcommerce.dev, states that ACP is open source under the Apache 2.0 licence, that Stripe and OpenAI developed it “to define a common language for how agents and businesses transact”, and that ChatGPT was the first AI platform to implement it. The versioned specification is maintained publicly on GitHub.
Can people still buy things inside ChatGPT with Instant Checkout?
Not in its original form. OpenAI announced in March 2026 that the initial version of Instant Checkout “did not offer the level of flexibility” it aspired to, so merchants now use their own checkout experiences while ChatGPT focuses on product discovery. Purchases are typically completed on the merchant’s own site, and some retailers, such as Walmart, have built dedicated in-ChatGPT apps instead.
What is the difference between ACP and Google’s AP2?
ACP (OpenAI and Stripe) standardises the whole commerce flow between an agent and a business: product feeds, checkout and delegated payment. AP2 (Google, with 60+ payments partners) standardises only the payment-authorisation layer, using cryptographically signed mandates to prove what a user authorised an agent to buy. They sit at different layers and are not direct substitutes; both interoperate with MCP.
Does ACP apply to B2B SaaS today?
Not directly. ACP’s shipped implementations are consumer-retail-centric: product feeds with prices and availability, and card-based checkout. B2B SaaS has no add-to-cart moment, so any B2B application is inference. The transferable lesson is structural: agents need machine-readable offer data and a completable next step, which in B2B means structured content and friction-free demo or application forms.
What should a B2B SaaS team do about agentic commerce right now?
Three things are low-regret: add structured data (schema.org) describing what you sell; make your demo-booking or application form simple enough for an agent or a hurried human to complete; and instrument fast follow-up so agent-referred leads get a response in minutes. Building bespoke checkout endpoints for any single protocol is premature while ACP, AP2 and UCP are all still moving.