What is agentic commerce? A plain-English guide to AI agents that shop and pay

Agentic commerce is shopping and payment carried out by an AI agent on someone's behalf. The agent finds products, compares options and completes the purchase within limits the buyer has set, using a checkout interface built for agents and a payment credential scoped to that purchase. What separates it from a shopping chatbot is authority: a chatbot recommends and a person clicks buy, while an agent can complete the purchase itself.

Stripe's documentation describes it in the same terms: agentic commerce lets AI agents transact for the people they serve, so buyers can browse, get recommendations and complete purchases without leaving the conversation. Stripe: agentic commerce

How does agentic commerce work, step by step?

Follow one purchase through six steps:

1. The buyer sets the task and the limits. For example, "reorder printer paper under 40 dollars from our usual supplier." The limits are the authority the agent may use.
2. The agent discovers products. It searches catalogs that merchants have made available to agents, usually as structured product feeds rather than web pages.
3. The agent builds a cart and checks it against the limits. Price, merchant, delivery date and quantity either fit the instruction or the agent stops and asks.
4. The agent starts checkout with the merchant. It calls a checkout interface designed for agents and receives totals, shipping options and terms.
5. A payment credential is presented. Rather than the buyer's raw card details, the merchant receives a credential scoped to this purchase.
6. The order is fulfilled and updated. Confirmation, shipping and refunds flow back to the agent and the buyer, so both see what happened.

Each step has an owner and a failure mode. Agentic commerce: which agent does what maps the roles, and following one order from purchase to refund shows what happens after checkout.

Which protocols are involved?

Open specifications are emerging for the parts that need to work across companies:

Sources: Stripe: Agentic Commerce Protocol, Stripe: Universal Commerce Protocol. The names change faster than the problems they solve, so check what a partner actually implements and which version; MCP, A2A, ACP, AP2 and UCP compares them in more detail.

How do AI agents pay?

The central design question is how to let an agent pay without handing it the buyer's card number. Providers are converging on scoped credentials. Stripe, for example, describes shared payment tokens that give sellers scoped access to a customer's payment method for agent-initiated purchases. Stripe: shared payment tokens Card networks are building their own agent programmes; what Visa and Mastercard are building covers that side.

There is also a machine-to-machine version. Stripe describes machine payments, where a server returns a payment challenge, the agent presents a credential and the payment settles, using protocols such as the Machine Payments Protocol and x402 for paying per API call. Stripe: machine payments

What can go wrong?

The risks are mostly about authority and evidence, not about the AI itself:

• Buying beyond authority. The agent buys something the buyer did not intend. Who is responsible depends on what the buyer authorized and how it was recorded; see an agent can find the product, but may it buy it?.
• Duplicate purchases. Two agents, or one agent retrying, place the same order twice. One durable purchase intention explains the control.
• Fraud and impersonation. A merchant cannot easily tell a legitimate agent from a malicious bot without agent identity. KYA as an identity control plane covers that problem.
• After-sales gaps. Returns, refunds and disputes still need a human-understandable record of what was agreed.

Is agentic commerce real yet?

Parts of it are live, with availability, launch stage and supported countries varying by provider and integration; Stripe's own documentation says as much about its integrations. Most deployments today are narrow: specific merchants, specific product types and buyers who have opted in. For a measured view of what is working, see agentic payments operations: what works, and before launching a pilot, read how to review an agentic commerce pilot.

FAQ

What is agentic commerce in simple terms?

AI agents that can shop and pay for someone, within limits that person has set, rather than only recommending products.

How is agentic commerce different from conversational commerce?

In conversational commerce a person chats with a brand and makes each decision. In agentic commerce the agent can complete steps, including payment, on the person's behalf. See agentic vs conversational vs headless commerce.

Does an AI agent get my card number?

It should not need to. Agentic payment designs use scoped or tokenized credentials so the agent and merchant receive only what the specific purchase requires.

Who is responsible if an AI agent buys the wrong thing?

It depends on what the buyer authorized, how that authority was recorded and the merchant's and provider's terms. Clear mandates and order records are what make the answer possible.

How can a merchant prepare for agentic commerce?

Start with accurate structured product data, clear policies and a checkout your payment provider can expose to agents. See preparing an online store for AI shopping agents.