What Is Agentic Commerce?
Agentic commerce is shopping in which AI agents discover, compare, and buy products for people. Definition, scope, players, and open questions.
In short: Agentic commerce is commerce in which an AI agent, acting on a person’s or business’s behalf, carries out some or all of the steps of a purchase: understanding the need, finding and comparing products, checking out, paying, and handling what happens after. It differs from earlier e-commerce in that the “shopper” doing the clicking is software, which forces new answers to old questions about product data, identity, payment authorization, and liability.
Definition
Agentic commerce is the buying and selling of goods and services where an AI agent takes actions in the purchase process, not just gives advice. An agent here means software, usually built on a large language model (LLM), that can interpret a goal stated in natural language, plan steps, call tools or websites, and act with some degree of independence.
A working definition used across this wiki:
Agentic commerce is any transaction in which an AI agent performs at least one consequential commercial action (selecting, carting, checking out, or paying) on behalf of a principal, within limits the principal has set.
Three elements matter:
- A principal. A consumer, employee, or business on whose behalf the agent acts.
- Delegated action. The agent does something with commercial consequence, not just summarizes reviews.
- Bounded authority. The agent operates within explicit or implicit limits: a budget, a merchant list, a requirement to confirm before paying. How those limits are expressed and verified is the subject of levels of autonomy and payments and trust.
The term covers a spectrum, from a chatbot that builds a cart and hands the shopper a checkout link, to an agent that reorders household goods automatically when a price threshold is met. The how it works page walks through each stage.
How it differs from related ideas
| Concept | What the software does | Who completes the purchase | Example |
|---|---|---|---|
| Traditional e-commerce | Hosts a catalog, search, and checkout for humans | The human, clicking through a website or app | A shopper browses a retailer’s site and pays |
| Recommendation engines | Ranks or suggests products based on behavior | The human | “Customers who bought this also bought” |
| Conversational commerce | Answers questions or takes orders through chat or voice, usually on one merchant’s channel | The human, often with a scripted flow | A brand’s support bot, ordering through a messaging app |
| AI-assisted shopping | Researches and compares across sources in natural language | The human, after clicking out | Asking an AI assistant for the best running shoes under a set price |
| Agentic commerce | Plans and executes purchase steps across merchants | The agent, with human approval or within a pre-authorized mandate | An assistant that buys an item when it drops below a target price |
The distinctions are about who acts and across what boundary. Recommendation engines and most conversational commerce live inside one merchant’s walls. Agentic commerce typically involves a third-party agent (built by an AI company, a browser, a bank, or a marketplace) interacting with many merchants that did not build it. That third-party position is why the field needs shared protocols, agent identity schemes, and new payment credentials.
Why now
Several developments converged between late 2024 and 2026 (see the full timeline):
- Capable models with tool use. LLMs became reliable enough to follow multi-step instructions and call external tools. Browser-using agents such as OpenAI’s Operator (January 2025) showed agents could navigate ordinary websites.
- Shared integration standards. The Model Context Protocol (MCP, November 2024) gave agents a common way to connect to tools and data. Commerce-specific protocols followed: the Agentic Commerce Protocol (ACP, September 2025), the Agent Payments Protocol (AP2, September 2025), and the Universal Commerce Protocol (UCP, January 2026).
- Network and processor investment. Visa (Intelligent Commerce) and Mastercard (Agent Pay) both announced agent-specific tokenized credentials in April 2025. Stripe, PayPal, Adyen, and American Express followed with agent payment products.
- Consumer behavior. Shoppers started using AI assistants for product research at scale. Amazon has said its Rufus assistant helped more than 300 million customers in 2025, and OpenAI launched a dedicated shopping research mode in ChatGPT in November 2025.
- Platform defaults. As of October 2026, large commerce platforms have begun making stores agent-purchasable by default. Shopify enabled agent checkout by default for eligible US merchants in its Meta channel in September 2026.
Who is involved
| Role | What they do | Examples (non-exhaustive) |
|---|---|---|
| Consumers and businesses (principals) | Set goals and limits, approve or delegate purchases | Individual shoppers, procurement teams |
| Shopping agents | Interpret intent, search, compare, and act | ChatGPT, Google AI Mode and Gemini, Microsoft Copilot, Perplexity, Alexa for Shopping, Meta’s Muse |
| Merchants and brands | Expose products, accept agent orders, fulfill | Retailers, DTC brands, marketplaces |
| Commerce platforms | Make merchant catalogs and checkouts agent-accessible | Shopify, BigCommerce, Salesforce Commerce Cloud, commercetools, WooCommerce |
| Product data providers | Supply structured, current product information to agents | Merchant feeds, catalog aggregators, product APIs |
| Payments providers and networks | Issue agent credentials, process payments, manage risk | Visa, Mastercard, American Express, Stripe, PayPal, Adyen |
| Trust and identity infrastructure | Verify agents and signals of user intent | Cloudflare (Web Bot Auth), network agent registries |
| Standards bodies and foundations | Govern open protocols | Linux Foundation (Agentic AI Foundation, A2A project, x402 Foundation), FIDO Alliance (AP2), IETF (Web Bot Auth) |
The market landscape page covers these categories in more depth.
What changes for each side
- For shoppers: less time on search and form-filling, but new questions about whether an agent’s choices are neutral and who fixes mistakes.
- For merchants and brands: the agent, not a human eye, increasingly reads the product page. Accurate, structured product data and machine-readable policies matter more than page design. See agent optimization and for brands.
- For developers: new integration surfaces (protocol endpoints, MCP servers, payment token APIs). See for developers.
- For payments: card-not-present transactions where the “customer” is software, requiring new ways to prove the human authorized the purchase. See for payments providers.
- For affiliates and agencies: traditional click-based attribution breaks when no human clicks. See attribution and affiliate.
Key open questions
As of October 2026, several basic questions remain unsettled:
- Where does checkout happen? OpenAI launched in-chat Instant Checkout in September 2025 and pulled back in March 2026, routing most purchases to merchants’ own checkouts or retailer apps inside ChatGPT. Google, Microsoft, and Meta continue to offer in-assistant checkout. Whether “discover in the agent, buy on the merchant’s site” or “buy in the agent” wins is open.
- Which protocols win? ACP, UCP, AP2, network-specific frameworks, and x402 overlap. Some vendors (for example Adyen, with its June 2026 Adyen Agentic product) explicitly support several at once. See protocols overview.
- Can merchants refuse agents? Amazon blocked Meta’s Muse agent from purchasing on its site in September 2026, and sued Perplexity over its Comet browser agent in November 2025. In August 2026 the US Ninth Circuit vacated a preliminary injunction Amazon had won in that case, holding that the user, not the tool maker, accesses the site under federal anti-hacking law. Terms-of-service and technical blocking remain available to merchants. See agent identity and fraud.
- Who is liable when an agent errs? Card networks and issuers are building rules for agent-initiated disputes. American Express announced protection for purchases by registered agents in April 2026. Regulation is still forming. See consumer protection and liability.
- How neutral are agents? Agents that rank products may also carry advertising or take transaction fees. How disclosure and ranking fairness will work is unresolved.
- How does attribution work? When an agent chooses the product, the usual signals (clicks, sessions, last-touch) fade. See attribution and affiliate.
- Will consumers trust it? Surveys through 2026 consistently report that many consumers are uncomfortable letting agents pay autonomously, even as AI-assisted research becomes common.
Where to go next
- New to the topic: How it works, then Levels of autonomy.
- Want the standards: Protocols overview.
- Need terms defined: Glossary.
- Quick answers: FAQ.
Sources
- Introducing the Model Context Protocol — Anthropic
- Agentic Commerce Protocol — agenticcommerce.dev
- Universal Commerce Protocol — ucp.dev
- New tech and tools for retailers to succeed in an agentic shopping era — Google
- Find and Buy with AI: Visa Unveils New Era of Commerce — Visa
- Mastercard Unveils Agent Pay — Mastercard
- Meet Alexa for Shopping — Amazon
- OpenAI Ends Instant Checkout in ChatGPT — The Keyword
- Shopify opens checkout to browser-based AI agents — TechCrunch
- Amazon Blocks Meta’s Muse AI Agent From Shopping — Dataconomy
- Ninth Circuit Vacates CFAA Injunction Against Perplexity’s Comet AI Agent — Jones Day
- American Express Debuts ACE Developer Kit and Agent Purchase Protection — American Express
- Adyen’s universal translator for agentic commerce — Finextra