What Is Agentic Commerce? AI Shopping and Paying, Explained

Agentic commerce means letting an AI assistant — a chatbot, a browser copilot, or a dedicated shopping agent — actually finish a purchase for you: picking the product, filling in the cart, and authorizing the payment, instead of just suggesting what to buy and leaving the clicking to a human. It's moving from pilot programs to real transactions in 2026, built on a handful of competing-but-overlapping protocols from OpenAI and Stripe, Visa, Mastercard, Google, and PayPal that all try to answer the same question: how does a merchant know a bot is allowed to spend your money, and who's on the hook if it gets it wrong?

What "agentic commerce" actually means

For most of the online-shopping era, "AI shopping help" meant a chatbot recommending products and then handing you a link. Agentic commerce changes the last step: the agent itself completes the transaction. You might tell an assistant like ChatGPT to find a pair of running shoes in your size under $120, and instead of returning a list of links, the agent can show you a cart, let you confirm it, and finish the purchase without you ever opening the retailer's site. The underlying idea is closely related to the broader shift toward AI agents that can take multi-step actions on your behalf, applied specifically to buying things.

The reason this needed new infrastructure, rather than just better chatbots, is that the existing web checkout flow assumes a human is sitting at the keyboard, typing in a card number or tapping a wallet button. An AI agent is not a human and not a browser session a merchant recognizes, so three new problems had to be solved: letting agents see what's for sale in a machine-readable way, letting merchants trust that a given "agent" request is legitimate and not a scraper or fraud bot, and letting the payment network confirm the person behind the agent actually approved the specific purchase.

The three pieces: agent, merchant, and payment rail

Every agentic commerce system being built right now — regardless of which company's name is on it — is solving the same three-part puzzle:

• The agent. This is the AI product the shopper is talking to: ChatGPT, a browser copilot, a shopping-focused assistant, and so on. It needs to understand a catalog, build a cart, and know when to stop and ask the user to confirm.
• The merchant. The retailer has to expose its product feed, pricing, and checkout steps in a structured format an agent can read, and it needs a way to tell a genuine shopping agent apart from a bot trying to scrape inventory or commit fraud.
• The payment rail. Somewhere in the chain, a card network, a payment processor, or a digital wallet has to turn "the agent says buy this" into an actual, authorized charge — ideally without ever handing the agent your raw card number.

The protocols below differ mainly in which of those three legs they focus on, and which companies control that leg.

OpenAI and Stripe: Instant Checkout and the Agentic Commerce Protocol

The most visible agentic-commerce feature so far is Instant Checkout, a feature inside ChatGPT that lets U.S. users buy products — starting with Etsy sellers, with Shopify merchants following — without leaving the chat. According to Stripe's announcement, it's powered by Stripe and runs on the Agentic Commerce Protocol (ACP), an open standard that Stripe, OpenAI, and Meta created together to give merchants one way to plug their catalog and checkout into multiple AI agents instead of building a custom integration for each one.

Technically, ACP breaks the transaction into pieces merchants can implement independently: an "agentic checkout" spec for creating and completing a cart, a feed spec for exposing products, and a "delegated payment" spec that payment providers implement so an agent can hand off payment details safely, as laid out in OpenAI's commerce documentation. The key payment primitive is the Shared Payment Token — Stripe issues this after the buyer selects a payment method, and it's scoped to one merchant and one cart total, so ChatGPT never sees or stores the underlying card number. The merchant can settle that token through Stripe or through its own existing payment provider. The full specification is published openly on GitHub at agenticcommerce.dev, and Stripe says any business already using Stripe can turn on agentic payments with a small code change.

PayPal has also signed on to ACP: PayPal's own newsroom announcement says it will let its wallet (bank transfers, PayPal balance, or cards) work as a funding option inside ChatGPT's Instant Checkout, with merchant catalog integration arriving through 2026, and users keeping PayPal's existing buyer and seller protections.

Visa's answer: Intelligent Commerce and the Trusted Agent Protocol

Visa took a network-level approach with Visa Intelligent Commerce, announced in April 2025 as a way to open its payment network — more than 150 million merchant locations and billions of Visa credentials, by its own count — to the companies building AI agents. Visa's own press release named Anthropic, IBM, Microsoft, Mistral AI, OpenAI, Perplexity, Samsung, and Stripe as early collaborators on the initiative when it launched, though that list reflects companies exploring the broader program rather than specific shipped payment features from each one.

The piece of this that actually addresses the "is this a real shopping agent or a bot" problem is the Trusted Agent Protocol, described on Visa's developer site. It's a cryptographic signature scheme, built on the IETF's RFC 9421 standard, that lets a merchant's server verify a request really came from an approved AI agent: the agent signs its request, and the merchant checks that signature against the agent's public key before deciding whether to treat it as a trusted shopper rather than a scraper to block. Visa publishes the specification and sample code for developers and merchants to implement directly.

Mastercard's Agent Pay

Mastercard's version, Agent Pay, launched the same week in April 2025 and builds on its existing tokenization technology — the same system behind contactless taps, stored cards, and Payment Passkeys — by extending it into "Agentic Tokens" for AI-initiated purchases. Mastercard's own announcement names Microsoft (via Azure OpenAI Service and Copilot Studio) and IBM (via watsonx Orchestrate) as platform partners, with Braintree and Checkout.com enhancing tokenization on the acquiring side.

Mastercard has since layered on an Agent Pay Acceptance Framework, which requires an AI agent to register and be verified on Mastercard's network before it can transact, and it describes device-level biometrics as part of how a consumer's agent can be strongly authenticated. The company has been rolling the program out market by market — including Latin America and the Caribbean and, in early 2026, Australia's first authenticated agentic transactions — through its own regional newsroom announcements.

Google's Agent Payments Protocol (AP2)

Where Visa and Mastercard work from the card-network side, Google approached agentic commerce as a protocol problem that doesn't assume any particular payment method. Its Agent Payments Protocol (AP2), announced on Google Cloud's blog in September 2025, is built around "Mandates" — tamper-proof, cryptographically signed records, based on verifiable credentials, that create an audit trail from a user's instruction to the final charge. An Intent Mandate captures what the user actually asked for (including any rules, like a price ceiling, for purchases the agent makes later without the user present); a Cart Mandate locks in the exact items and price once the user approves a specific cart, so it can't be altered afterward.

AP2 is explicitly payment-method-agnostic — Google says it covers cards, real-time bank transfers, and even stablecoins, and it can work as an extension of Google's own Agent2Agent (A2A) protocol alongside the Model Context Protocol. Google says it developed AP2 with more than 60 payments and technology companies, including American Express, Coinbase, Etsy, Intuit, Mastercard, PayPal, Salesforce, and ServiceNow, and has published the full specification and a reference implementation as open source.

Who pays if the agent gets it wrong?

Every one of these programs leans on the same basic trick to limit risk: the agent itself never gets your raw card number. It gets a token, a mandate, or a signed credential that's scoped to one purchase, one merchant, and (often) one price ceiling. That's the same tokenization logic that already protects contactless and card-on-file payments — Visa and Mastercard are explicitly extending infrastructure they already use for fraud screening and authorization, rather than inventing a parallel payment system.

What's genuinely less settled is liability for the messier cases that aren't simple fraud — an agent buying the wrong size, misreading a price, or acting slightly outside what a user actually intended. The protocols create an audit trail (Google's chain of Intent, Cart, and Payment mandates is explicitly designed for this; Visa's signed agent requests serve a similar purpose), so there's a record of what the agent was authorized to do and what it actually did. But none of the companies involved has published a single, universal rule for who eats the cost when an agent's purchase goes wrong without fraud being involved — that's still being worked out deal by deal, and card-network rules on confirmed fraud are the clearest piece so far.

What it means for shoppers

In practice, none of these systems currently let an AI assistant spend money with zero human confirmation for one-off purchases. Instant Checkout, for example, still shows you the cart before it charges you. The bigger shift is in standing permission: AP2's Intent Mandate and similar approaches are built for the case where you tell an agent "buy this jacket in blue if it drops below $80" and walk away, trusting the agent to act later within rules you set up front. That kind of delegated, unattended purchase is the part worth paying attention to, because it's where you're trusting the mandate's price and quantity limits to actually hold.

For now, opting in is specific rather than global — you connect a payment method (a card, a PayPal wallet, etc.) to the agent or chat app you're using, and purchases happen only through merchants and catalogs that have integrated with that particular protocol. If you've tried an AI browser with agentic shopping features, you've likely already seen an early version of this same pattern: the assistant drafts the cart, and you still click to confirm before anything is charged.

What it means for merchants

For retailers, the pitch is discovery and conversion without having to build a custom integration for every AI platform: expose a product feed once, in a standard format, and show up inside ChatGPT, a Visa-connected agent, or any other participating assistant. The cost is usually a transaction fee on completed agent-driven sales, similar to existing payment-processing fees, and the requirement to implement whichever checkout and authentication spec a given protocol demands — plus, for the ones built around bot-detection like Visa's Trusted Agent Protocol, the work of distinguishing a legitimate shopping agent from scraper traffic that merchants have spent years trying to block outright.

What's next

None of these protocols is finished. ACP's specification is still being revised through its open GitHub repository, Visa's Trusted Agent Protocol pages describe the product as still rolling out rather than globally available, and Mastercard and Google are both still adding partners and regional launches month by month. The practical question for 2027 isn't whether agents will be able to check out — several already can, in limited pilots — it's whether a shopper's agent of choice, a merchant's payment processor, and the card network behind it all will actually agree on the same signals of trust, or whether buying something through an AI assistant will depend on which specific combination of apps, banks, and retailers you happen to be using.

Written byTheo Park

Theo Park runs the AI desk at Pandromeda. He follows model launches from the frontier labs and the open-weight community, tracks the assistants and developer tools built on them, and explains what each release changes on pricing, capability and safety. His reporting leans on primary sources: model cards, technical reports, API documentation and the companies' own announcements.

More from Theo ParkAll AI coverage