Definition
Agentic commerce is buying in which an AI agent, rather than a person clicking through a website, does the work: finding products, comparing them, and completing checkout on the buyer's behalf. The buyer states an intent — "reorder the filter that fits my furnace, under $40, arriving before Friday" — and the agent resolves it against real catalogues and pays.
What makes this a distinct term rather than a marketing phrase for chatbots is that it cannot be done with the existing web. A checkout flow is designed around a human in a browser session: form fields, a cart cookie, a card number typed into an iframe, a CAPTCHA. An agent has none of that reliably, and screen-scraping a checkout is fragile, unauthorised, and uninsurable. So agentic commerce required new protocols, and by 2026 there are two competing ones, which is the practical situation anyone reading this needs to understand.
A note on the acronym before going further. In this field, ACP means the Agentic Commerce Protocol, from OpenAI and Stripe. Elsewhere in AI, ACP meant IBM's Agent Communication Protocol, a spec for agents talking to each other that merged into A2A in August 2025. They are unrelated. This page uses ACP for the commerce protocol throughout.
How It Works
Three layers have to exist before an agent can buy something, and the protocols differ in how much of the stack they claim.
Discovery. The agent needs a machine-readable view of what a merchant sells, in stock, at what price, shippable where. This is a product feed problem, and it is the part most merchants already half-solve for existing shopping channels.
Checkout. The agent needs to construct a cart and submit an order without a browser session. This is where the protocols do their real work, because it is the step that has no existing standard.
Payment. Someone must authorise a charge that the cardholder is not personally clicking "confirm" on. This is the hardest layer, and it is a liability question before it is a technical one: if an agent buys the wrong thing, who eats it?
ACP's answer to the payment layer is a shared payment token rather than a raw credential. Stripe's implementation hands the merchant a token that authorises one transaction, so the card number never reaches the agent or the merchant's systems, and PCI scope does not expand. Critically, the merchant remains merchant of record — the agent brokers, but the commercial relationship and the liability stay with the seller. That design choice is why merchants were willing to adopt it at all.
ACP is published under Apache-2.0 and supports both REST and MCP integration patterns, which matters because a merchant who already exposes an MCP server for other purposes has less new surface to build.
Google split the problem differently. AP2, the Agent Payments Protocol, announced 17 September 2025 with more than 60 partners including Mastercard, PayPal and Coinbase, addresses payment authorisation on its own. UCP, launched 11 January 2026, covers the full journey — discovery, purchase, and post-purchase support.
Real-World Applications
This is live, not prospective, and the deployments divide cleanly along protocol lines.
ChatGPT, via ACP. OpenAI shipped Instant Checkout on 29 September 2025, the same day ACP was published, initially limited to US buyers purchasing from US-based Etsy sellers, with Shopify merchants following. By March 2026 OpenAI had already revised the approach — moving away from the original Instant Checkout flow toward product discovery plus merchant-controlled checkout. That reversal inside six months is worth noting: the first shape of agentic checkout did not survive contact with real merchants.
Google AI Mode and Gemini, via UCP. Google co-developed UCP with Shopify, Etsy, Wayfair, Target and Walmart — that is, with the merchants rather than only with a payment processor.
Chinese platforms, on their own rails. Alibaba, Tencent and ByteDance are building agentic commerce inside their existing super-apps, where the payment layer, the catalogue and the assistant already belong to one company. The structural advantage is real: none of the protocol negotiation described above is necessary when one firm owns every layer.
The most informative single event so far is an alliance change. Stripe co-authored ACP with OpenAI in September 2025, and on 24 April 2026 joined the Universal Commerce Protocol Tech Council — alongside Amazon, Meta, Microsoft and Salesforce. A protocol's co-author hedging onto the rival standard seven months later says more about where this is heading than any adoption statistic.
Challenges
Two protocols is not a transitional state you can wait out. The protocols map to buyer populations, not to technical preferences: ACP reaches people asking ChatGPT, UCP reaches people asking Google Gemini or AI Mode. A merchant who integrates one is not "adopting agentic commerce" — they are adopting half of it and are invisible to the other half. The arithmetic is unforgiving: two integrations, two sets of feed mappings, two checkout endpoints, two certification processes, and two ongoing maintenance obligations against specifications that are both younger than a year and both still changing. For a large retailer that is a real programme; for a small one it is the difference between doing this and not.
Liability is unresolved at the edges. The merchant-of-record design settles the common case. It does not settle what happens when an agent misreads a size chart, buys a non-returnable item, or acts on a stale price. Chargeback frameworks were written for a human who can be asked what they intended.
Discovery becomes a ranking problem you do not control. When an agent picks one of five comparable products, the criteria are the model's, not the shopper's stated preferences alone. This is Generative Engine Optimization with a purchase attached, and merchants have even less visibility into it than into search ranking.
Prompt injection acquires a payment method. An agent that reads product pages, reviews and merchant-supplied content, and that also holds authorisation to spend, is a new class of target. A malicious listing that manipulates an agent into a purchase is not a hypothetical attack shape; it is the obvious one.
The first design already failed once. OpenAI's retreat from Instant Checkout v1 within six months should temper any assumption that current specifications are settled. Integration work done against a specification this young carries rework risk.
Future Trends
The consolidation question is the one worth watching, and the April 2026 Tech Council expansion is the strongest signal available: when Amazon, Meta, Microsoft, Salesforce and Stripe all join one standard's governance body, the other standard's position depends increasingly on a single distribution channel. ACP is not weak — ChatGPT is an enormous surface, and Apache-2.0 licensing means nobody can withdraw it — but a protocol backed by one platform against a protocol backed by a consortium is a familiar shape, and it usually resolves the same way.
The less-discussed trajectory is that the merchant-of-record design may not hold. It exists because merchants would not have adopted anything else in 2025. As agent-originated volume grows, the pressure to move the commercial relationship toward whoever owns the agent — which is where the customer relationship actually lives — will grow with it. Watch for that in specification changes rather than in announcements.
Note: this term was last reviewed in August 2026. Protocol launch dates, governance membership and platform behaviour are stated as of that review, and in a field where a flagship checkout flow was redesigned within six months of launch, these are the parts of this page most likely to age.