What Is Agentic Commerce? A Plain-English Guide

Agentic commerce is when an AI agent shops and pays on behalf of a person or a business. You give the agent a goal. The agent does the steps for you: it searches for products, compares them, and completes the purchase.

This is different from a normal chatbot. A chatbot gives you links and advice. You still do the buying. An agent can finish the whole task, including the payment.

In this guide I explain how it works, why it is happening now, and what is real today versus what is still just talk.

A simple example

You tell your agent: "Order more coffee, the same brand as last time, and keep it under $30."

The agent then:

1. Finds the product in a store that agents can read.
2. Checks the price and stock.
3. Puts it in the cart.
4. Pays, using a method you allowed before.
5. Shows you the receipt.

You did one sentence of work. The agent did the rest.

Why this is happening now

Three things came together at the same time.

First, the AI models got good enough to plan and use tools. An agent can now read a product page, call an API, and follow a set of steps without getting lost.

Second, big companies agreed on shared rules. Google, OpenAI, Stripe, Visa, Mastercard, and others started to build open standards. These standards let an agent and a store talk in the same language.

Third, payments became the focus. Everyone understood the same thing: the hard part is not finding the product. The hard part is paying for it in a safe way.

How an agentic purchase works, step by step

Under the surface, one purchase has a few clear stages.

Discovery. The agent needs to find products it can trust. Stores publish their catalog in a format that agents can read. Today this often uses a protocol called MCP (Model Context Protocol) or a product feed made for agents.

Selection. The agent compares options against your goal — price, brand, delivery time — and picks one.

Checkout. The agent starts a purchase using a commerce protocol. The two most talked about are ACP (from OpenAI and Stripe) and UCP (from Google and Shopify).

Authorization. This is the money-and-trust step, and it is the hardest. The system must prove that you, the human, allowed this purchase. It must also prove the agent is who it says it is.

Payment. The payment is made with a scoped, limited method — for example a single-use token, a network credential from Visa or Mastercard, or in some machine-to-machine cases a stablecoin.

Settlement and record. The money moves, and both sides keep a record for support and disputes.

The part that is actually hard

The AI is not the hard part anymore. The hard part is money and trust.

Think about three questions:

• Who said yes? How do we prove a real human approved this, and not the agent alone?
• Who is the agent? How does the store know this agent is real and not a fake one pretending to be it?
• What if it goes wrong? If an agent buys the wrong thing, who pays for it — the shopper, the merchant, or the platform?

These questions do not have final answers yet. This is why so much work is going into identity, authorization, and spending limits. If you understand this, you understand where the real value is.

What is real today, and what is still hype

It helps to sort news into three groups.

Live. Some things work now. AI assistants can complete real purchases in some stores. Card networks have launched agent programs. A few merchants run full agent checkouts in production.

Announced. Many things are announced but not widely used. A press release is not the same as a shipped feature. Read carefully.

Contributed to a standard. Some protocols were "given to" a standards body like the FIDO Alliance. This is good for the long term, but it does not mean the feature is live today.

A simple rule: when you read a claim, ask if it is live, announced, or just contributed to a standard. This one habit will keep you from being fooled.

Who is involved

Many types of players are building this space: AI companies, commerce protocols, payment protocols, card networks, processors, and standards bodies. Each one owns a different layer.

We map all of them, with names, in a separate post: The agentic commerce landscape: who is building what.

Frequently asked questions

Is agentic commerce the same as a shopping chatbot? No. A chatbot suggests and links. An agent can complete the purchase, including payment.

Do I lose control of my money? You should not. Good systems use limits: a budget, a single-use token, and your approval. You set the rules before the agent acts.

Is this safe? The safety depends on the guardrails, not the AI. Spending limits, agent identity, and clear approval are what make it safe. This is still improving.

Do I need to be technical to use it? No. As a shopper you just give a goal. The technical work is for merchants and builders.

When will this be normal? Some parts are live now. Wider use will grow as identity and payment standards settle over the next year or two.

Key takeaway

Agentic commerce means an AI agent can shop and pay for you. The technology to plan and act is here. The open question is trust: proving who approved a purchase, proving who the agent is, and deciding who pays when something goes wrong. Watch that layer. That is where this space will be won.

Sources and further reading

• Agentic Commerce Protocol (ACP) — OpenAI and Stripe: https://www.agenticcommerce.dev
• Stripe, "Developing an open standard for agentic commerce": https://stripe.com/blog
• Google Developers Blog, "Under the Hood: Universal Commerce Protocol (UCP)" (2026-01-11): https://developers.googleblog.com
• Model Context Protocol (MCP) — specification: https://modelcontextprotocol.io
• Google Cloud, "Announcing Agent Payments Protocol (AP2)" (2025-09-16): https://cloud.google.com/blog
• FIDO Alliance, "FIDO to Develop Standards for Trusted AI Agent Interactions" (2026-04-28): https://fidoalliance.org

Note: this field moves fast. Protocol versions and live/announced status change often — please check the primary sources above for the latest before you rely on a detail.