Agentic commerce is a form of online commerce in which autonomous AI agents discover products, compare options, and complete purchases on behalf of human buyers, with the buyer providing only the initial intent and a final authorization. The buyer never sees most product pages, never enters card details, and never navigates a checkout flow.
This is not a chatbot recommending products that a human then buys. Agentic commerce is the AI agent acting as the purchaser, the one constructing the cart, the one passing payment credentials, and the one closing the transaction. The human still authorizes the spend, but the path from intent to checkout runs through software rather than fingertips on a screen.
McKinsey estimates the global agentic commerce opportunity could reach $3 to $5 trillion by 2030. A 2026 IBM study found that 45% of consumers already use AI for at least part of the buying journey. Agent-driven traffic across the open web has grown more than 1,300% in the past nine months. The question for most businesses is no longer whether agentic commerce is coming, but whether their checkout will be discoverable inside it.
This guide explains exactly what agentic commerce is, how it works in practice, the protocols and infrastructure that make it possible, and what merchants need to do to participate.
Agentic commerce is the category of online transactions in which an autonomous AI agent, rather than a human navigating a user interface, performs the discovery, evaluation, and execution of a purchase on behalf of a human buyer.
Three elements are central to that definition:
The category is new enough that the terms get blurred. Three labels worth distinguishing:
| Mode | Who drives the purchase | Where it happens | Example |
|---|---|---|---|
| Traditional ecommerce | Human shopper clicking through a UI | Merchant website or app | Customer browses Amazon, adds to cart, checks out |
| Conversational commerce | Human shopper chatting; agent assists | Chat interface, with merchant UI for checkout | Customer asks Klarna’s bot for a deal; bot links to merchant site |
| Agentic commerce | Autonomous AI agent acting for the human | Agent’s environment (ChatGPT, Gemini, etc.) | Customer tells ChatGPT to buy a gift; ChatGPT discovers, selects, and pays |
The dividing line is who completes the transaction. In traditional ecommerce the human does. In conversational commerce the human still does, with help. In agentic commerce, software completes the transaction inside an environment the merchant does not control.
Every agentic transaction moves through five stages. Understanding the sequence is the first step to understanding what infrastructure has to change to support it.
The buyer expresses what they want to an AI agent. The expression can be sparse (“find me a birthday gift for my mother”) or specific (“a waterproof hiking boot in size 8, under $150, delivered by Friday”). The agent translates this into a structured query.
The agent queries product data from merchants that have made their catalogs accessible through standardized protocols. The agent does not browse storefronts. It does not parse JavaScript-rendered pages. It calls APIs, ingests structured feeds, and reasons across the results. Merchants whose catalogs are not exposed in machine-readable form are invisible at this stage.
The agent narrows the options, sometimes presenting a shortlist to the buyer for approval, sometimes proceeding directly when the buyer has pre-authorized. Once a product is chosen, the agent constructs a cart by calling the merchant’s agentic checkout API.
This is the step that distinguishes agentic commerce from every previous form of automated buying. The agent does not see or transmit raw card details. The buyer authorizes the spend inside the agent’s environment, the agent receives a scoped payment token (limited to one merchant, one currency, a maximum amount, and a short expiration window), and that token is passed to the merchant. The merchant references the token inside a standard authorization request, which is processed through their PSP and acquirer.
From the merchant’s perspective, settlement looks like any other card transaction. The funds flow through the existing acquiring relationships. The agent retains the transaction metadata, tracks fulfillment, and surfaces updates back to the buyer. The merchant captures agent-specific tags so the transaction can be measured separately from human-initiated traffic.
The critical insight is that only steps 1 through 4 are genuinely new. Step 5 runs on the merchant’s existing payment infrastructure, provided that infrastructure can accept tokens from multiple PSPs, route intelligently, and identify agent traffic separately.
Agentic commerce did not become viable through better AI alone. It became viable when two open protocols standardized the interface between AI platforms and merchant systems in late 2025.
ACP is the open standard co-developed by OpenAI, Stripe, and Meta, released under the Apache 2.0 license in September 2025. It powers ChatGPT Instant Checkout and defines four composable building blocks: agentic checkout (creating, updating, and completing sessions), cart and feed (browsing catalogs), delegate payment (passing secure tokens between buyer, agent, and merchant), and delegate authentication (OAuth 2.0 for agents acting on a buyer’s behalf).
Under ACP, the merchant remains the merchant of record. Settlement, compliance, and disputes stay with the merchant and their PSP. The agent is the messenger.
UCP is the open standard co-developed by Google and Shopify, also released in October 2025. It powers commerce inside Google AI Mode and Gemini, and covers the full shopping journey from discovery through fulfillment. UCP has been endorsed by Walmart, Target, Visa, Mastercard, Stripe, and more than 20 other retailers and platforms.
Where ACP optimizes for the conversational ChatGPT experience, UCP optimizes for the broader Google ecosystem. The two protocols are complementary, and most enterprise merchants implement both rather than choosing one.
A third protocol, MCP, sits beneath both ACP and UCP. MCP is the framework that enables AI agents to query structured, machine-readable information from external systems including inventory, pricing, and checkout logic. Where ACP and UCP define how an agent transacts, MCP defines how the agent gathers the information it needs to decide what to transact for.
Together, these three protocols form the technical foundation of agentic commerce. For a deeper look at how to implement them, our guide on how to make your checkout AI-agent-ready walks through the specific endpoints, conformance tests, and integration patterns.
The term has been used loosely enough that it is worth saying clearly what falls outside the category.
Agentic commerce is not a recommendation engine. A model that surfaces products on a merchant website is enhancing traditional ecommerce, not replacing it. The human still completes the purchase.
Agentic commerce is not a chatbot with a “buy” button. A conversational interface that hands the buyer off to a merchant checkout page is conversational commerce. The agent did not transact; it routed.
Agentic commerce is not screen-scraping or browser automation. Some early demos of “AI shopping” rely on language models clicking through human-facing pages. This works in controlled environments but breaks at scale, fails fraud screening, and is being deliberately phased out by merchant security systems. The protocol-based approach is what makes agentic commerce a real category rather than a demo trick.
Agentic commerce is not full autonomy without human oversight. Every credible implementation keeps the buyer in the authorization loop. Agents propose, humans dispose. The buyer’s role shifts from clicker to delegator, not from delegator to absent.
Three forces converged in 2024 and 2025 to make agentic commerce viable.
Capable foundation models. Large language models reached the point where they can reason about purchase decisions, handle multi-step planning, and reliably parse structured product data. The reasoning capability is what enables the agent to take a sparse human intent and translate it into a specific transaction.
Standardized protocols. ACP, UCP, and MCP gave merchants a clear specification to build against. Before these protocols existed, every “agentic commerce” project was bespoke and brittle. Now there is a clear interface, and merchants can implement once and serve any compatible agent.
Consumer adoption of AI assistants. ChatGPT alone reached 800 million weekly active users by late 2025, processing roughly 50 million shopping queries every day. Google’s Gemini, Microsoft’s Copilot, Anthropic’s Claude, and a growing ecosystem of specialized agents have made AI assistants a default surface for an entire generation of consumers.
The combination produced the conditions for a category that had been predicted for a decade to finally take off. Cyber Week 2025 was the inflection point: retailers with AI agent integration saw approximately 7× better sales growth than those without, according to Salesforce data. The merchants who waited watched the market move past them in a single quarter.
The category is still young, but several distinct patterns are already in market.
Conversational shopping inside ChatGPT and Gemini. A consumer asks the assistant for a product recommendation. The assistant queries merchant catalogs through ACP or UCP, presents options, accepts a selection, and completes the purchase using a delegated payment token. The buyer never leaves the chat.
Personal shopping agents. Specialized agents that learn a buyer’s preferences and proactively complete recurring purchases. Reordering household goods, restocking pet food, replacing print cartridges, all without prompting from the human.
Travel and ticketing agents. Agents that compare flights, hotels, and tickets across multiple providers, then book the optimal combination on behalf of the buyer. This category is among the earliest to mature because the underlying APIs already existed.
B2B procurement agents. Enterprise software that monitors inventory or supply levels, identifies the optimal supplier, negotiates the purchase, and processes the order through corporate procurement rails. The B2B applications often run with looser human-in-the-loop constraints because corporate spending limits act as the safety mechanism.
Comparison and price-watching agents. Agents that monitor pricing across merchants and execute a purchase the moment a target price is met, often while the buyer is asleep or in a meeting.
The common thread across all five is the same: a human delegated a purchasing decision to software, and software executed the transaction without human navigation of a merchant UI.
For ecommerce businesses, agentic commerce is both an opportunity and an existential question. The opportunity is a new revenue channel growing at unprecedented rates. The existential question is whether your store is even visible inside it.
Discoverability becomes a protocol problem. Search engine optimization made content legible to search engines. Agentic commerce optimization makes commerce legible to agents. The work shifts from keyword strategy and content marketing to structured catalogs, machine-readable APIs, and protocol compliance. Heavy JavaScript frameworks that block agent crawlers, missing product attributes, ambiguous shipping policies, all of these turn into revenue leaks the moment agent traffic becomes meaningful.
Checkout becomes the hardest problem. Most analyst coverage agrees on this point. Discovery is hard but solvable through structured data. Checkout is the layer where agentic commerce breaks most existing infrastructure. Merchants need to accept delegated payment tokens, route agent-initiated transactions intelligently, handle authentication for non-present buyers, and tag agent traffic separately from human traffic. None of this is impossible, but most existing payment stacks do not handle it out of the box.
The risk of vendor lock-in increases. The fastest path to ACP support today is a single-line update inside one specific PSP integration. The fastest path to UCP is being on one specific commerce platform. Both options ship quickly because they bundle protocol compliance with a particular vendor’s product. The trade-off is exactly the kind of single-vendor concentration that enterprise merchants have spent the last decade trying to eliminate.
Analytics need a separate lens. Without specific tagging, agent-driven transactions look identical to human checkouts in merchant analytics. This is how merchants quietly lose money for quarters. Authorization rates that drop on agent traffic but stay flat overall are invisible without instrumentation. Conversion gaps that emerge from agent-specific friction points are invisible without instrumentation. Tagging from day one is the cheapest insurance policy in this stack.
For merchants thinking about how to participate, our companion guide on making your checkout AI-agent-ready covers the four capabilities your stack actually needs and a week-by-week implementation roadmap.
The consumer experience changes more quietly than the merchant experience, but it changes substantively.
The job of shopping shrinks. Consumers no longer evaluate dozens of options across tabs. They describe what they want and let the agent do the comparison work. The cognitive load of choice falls dramatically.
Brand loyalty becomes more fragile. When the agent is doing the picking, the brands that win are the ones that are easiest to transact with, not necessarily the ones with the strongest brand affinity. Discoverability, clear policies, and frictionless agent-side checkout matter more than logo recognition.
Trust shifts from merchant to agent. The consumer is trusting the agent to act in their interest. This raises new questions about how agents are paid (do they take a commission? are they sponsored?), how they prioritize options, and how disputes are resolved when an agent buys the wrong thing.
Privacy and authorization patterns change. The agent needs to know enough about the buyer to act on their behalf, which means more data flows between humans, agents, and merchants. The protocols build in scoped tokens and short-lived authorizations to constrain this, but the consumer surface area is genuinely larger than in traditional e-commerce.
A handful of confusions show up consistently in early coverage of the category.
Myth: Agentic commerce will replace human shopping. It will not. Agentic commerce is best suited to repeat purchases, well-defined queries, and tasks with measurable success criteria. Discretionary shopping, browse-for-fun experiences, and emotionally driven purchases remain human-driven. The two modes coexist.
Myth: You need to be on a specific platform or PSP to participate. ACP, UCP, and MCP are all open standards. Any merchant can implement them against any commerce backend and any payment provider. The bundled paths from Stripe or Shopify ship fastest, but they are not the only paths. Merchants using payment orchestration can implement the protocols once and route agent-initiated transactions across multiple PSPs.
Myth: Agentic commerce is just a fancy word for chatbots. Chatbots respond to prompts inside a conversational interface. Agents plan, act, and transact across systems. The distinction is the difference between assistance and autonomy.
Myth: Agentic commerce is insecure. The protocols are designed with security in mind from the start. Payment tokens are scoped to a single merchant, a single currency, a maximum amount, and minutes-long expirations. Raw card data never reaches the agent. The compliance burden sits with the merchant and PSP, which is the same place it sat in traditional ecommerce. A well-implemented agentic flow is at least as secure as a standard card-not-present transaction.
Myth: There is plenty of time to prepare. Agent-driven traffic has grown 1,300% in nine months. The retailers who treated this as a 2027 problem in 2024 are already losing market share to competitors who treated it as a 2025 one. The lead time for protocol implementation, merchant program approvals, and conformance testing is measured in months, not weeks. Starting in 2026 is not early.
For merchants beginning to plan their approach, the work falls into four capability areas:
The order matters. Discoverability without checkout is a half-built solution. Checkout without visibility is a black box. Visibility without routing is a measurement system with nothing to optimize.
The most defensible architecture is one in which protocol compliance lives in a layer separate from payment processing. The merchant implements ACP and UCP endpoints once, those endpoints sit inside an orchestration platform that connects to multiple PSPs, and each agent-initiated transaction is routed in real time to whichever PSP is most likely to approve it. This preserves flexibility as the protocols and provider options continue to evolve through 2026 and 2027.
This is exactly the problem Gr4vy’s Agentic Development Kit (ADK) was designed to solve. The ADK gives merchants the infrastructure and step-by-step guidance to enable their storefront inside AI environments like ChatGPT, orchestrate the resulting transactions across more than 400 PSPs and payment methods, and maintain control over routing, security, and performance, all without replatforming and without locking themselves into a single PSP.
Agentic commerce is online shopping where an AI agent does the buying instead of a human clicking through a website. The human tells the agent what they want, authorizes the spend, and the agent handles discovery, selection, payment, and tracking. It is different from a chatbot recommending products: the agent is the purchaser, not the recommender.
Conversational commerce uses chat as an interface to help a human shopper complete a purchase, but the human still completes the purchase. Agentic commerce uses an autonomous agent to actually complete the purchase on the human’s behalf. The dividing line is who closes the transaction.
The category involves three main types of participants. AI platforms (OpenAI, Google, Anthropic, Microsoft, Meta) provide the agents and the consumer-facing environments. Protocol bodies (the ACP and UCP working groups) define the standards. Payment infrastructure providers (PSPs, acquirers, orchestration platforms) handle the financial transactions. Merchants sit at the center as the parties whose catalogs are being purchased from.
Yes, when implemented correctly. The protocols use scoped payment tokens that are limited to one merchant, one currency, one amount, and a minutes-long expiration window. Raw card data never reaches the agent. The merchant’s PSP handles authorization and fraud screening with the same protections used for any card-not-present transaction. The compliance bar is PCI DSS Level 1, the same standard that applies to traditional ecommerce.
McKinsey estimates the global opportunity at $3 to $5 trillion by 2030. Other analysts project the addressable market will grow from approximately $135 billion in 2025 to $1.7 trillion by 2030. Most forecasts converge on 15 to 25% of ecommerce transactions being agent-driven by the end of the decade.
ACP (Agentic Commerce Protocol) is the open standard co-developed by OpenAI, Stripe, and Meta that powers ChatGPT Instant Checkout. UCP (Universal Commerce Protocol) is the open standard co-developed by Google and Shopify that powers commerce inside Google AI Mode and Gemini. MCP (Model Context Protocol) is the underlying framework that enables AI agents to query structured information from external systems. Together, these three protocols form the technical foundation of agentic commerce.
No. ACP and UCP are open standards released under permissive licenses. Any merchant can implement them against any commerce backend and any PSP. Shopify and Stripe offer the fastest paths to specific implementations, but the protocols themselves are infrastructure-agnostic.
No. Agentic commerce is an additional channel that grows alongside traditional ecommerce rather than replacing it. Discretionary shopping, browsing, and emotionally driven purchases remain human-driven. Repeat purchases, well-defined queries, and tasks with clear success criteria shift toward agents. Most merchants will operate both channels in parallel.
The buyer authorizes the spend inside the agent’s environment. The agent receives a scoped payment token (rather than a raw card number) and passes it to the merchant. The merchant references the token inside a standard authorization request, which is processed through their PSP and acquirer. From the merchant’s payment infrastructure perspective, the transaction looks similar to any card-not-present payment, but the credential is a delegated token instead of a stored card.
A consumer tells ChatGPT they need a birthday gift for their mother under $75, delivered by Friday. ChatGPT queries merchant catalogs through ACP, identifies several candidates, presents three options, accepts the buyer’s selection, receives a delegated payment token after the buyer authorizes the spend, and completes the purchase with the merchant. The buyer never visited the merchant’s website. The merchant received a fully authorized card transaction with metadata identifying it as agent-initiated.
Four capabilities: a machine-readable product catalog, agentic checkout API endpoints that implement ACP and UCP, delegated payment token support inside a PCI DSS Level 1 environment, and analytics that tag and measure agent traffic separately. Most enterprise merchants find that the fastest path to all four is a payment orchestration layer that sits above their existing PSPs and handles the protocol implementation centrally.
Agentic commerce is the next architectural shift in ecommerce, comparable in scale to the move from desktop to mobile a decade ago. The protocols that make it work are open, the consumer adoption is already meaningful, and the early data from Cyber Week 2025 shows that participating merchants outperform non-participating ones by a wide margin.
The merchants who emerge from this shift in the strongest position will share three characteristics. Their catalogs will be machine-readable and current. Their checkout will support both ACP and UCP without being locked to a single PSP. Their payment infrastructure will route agent-driven transactions intelligently and surface separate analytics for them.
For most enterprise stacks, getting there does not require starting over. It requires a clear inventory of where the four capabilities above already exist, where the gaps are, and the sequence of work that closes them in weeks rather than quarters.
If you’re ready to audit your stack against the four capabilities or map out a practical implementation path, contact our team or explore the Agentic Development Kit for a step-by-step framework.
Agent-driven traffic across the web has grown more than 1,300% in the past nine months.…
Most businesses use the terms payment gateway, payment processor, and payment orchestration platform as if…
The average online checkout loses seven out of every ten shoppers. Across all industries, the…
For years, platforms treated payments as a necessary layer to enable transactions. Something to plug…
Most merchants focus on the transaction rate. By the time interchange, scheme fees, chargebacks, and…
A subscription business with fifty thousand active subscribers will process roughly six hundred thousand recurring…