Agentic commerce

This is made possible by large language models in assistants such as ChatGPT, Gemini or specialised shopping agents, combined with open protocols through which agent, merchant and payment service provider can talk to each other securely. For retail, this shifts the decisive question: no longer just "How do I convince a person?" but also "How does my range become findable and buyable for a machine?"

How agentic commerce works

A purchase by an agent takes place in several steps: the agent interprets the user's intention, searches for suitable products via interfaces or product feeds, evaluates options on the basis of price, availability and reviews, and finally triggers the order, including payment. The trickiest step is payment: a merchant must be able to be sure that an agent really is acting on behalf of an authorised customer and that a specific purchase intention has been authorised. Several open standards that connect agents, shops and payment systems were created for exactly this purpose in 2025 and 2026.

The competing protocols in 2026

The Agentic Commerce Protocol (ACP) was developed and published by OpenAI together with Stripe. Among other things, it powers OpenAI's "Instant Checkout", through which users can buy from connected merchants directly in ChatGPT; shops that already use Stripe can activate agentic payments with little effort. In parallel, Google is pursuing its own approach with the Agent Payments Protocol (AP2): it relies on cryptographically signed "mandates", digital proofs with which a user verifiably authorises a specific transaction intention; in 2026 AP2 was handed over to the FIDO Alliance and thus transferred to vendor-neutral standardisation. In addition, Shopify and Google presented the Universal Commerce Protocol (UCP) at the beginning of 2026. The standards partly compete and partly complement each other; a final market convention is not yet in sight.

Table: the most important agentic commerce standards

What agentic commerce means for retailers

When agents, and not just people, do the shopping, machine-readable, structured product data becomes a competitive factor. An agent does not evaluate a pretty banner but clean attributes: precise titles, complete specifications, correct prices, reliable availability and structured markup. Anyone who maintains their data centrally and provides it via open interfaces has the advantage; the closeness to omnichannel and clean product data management is obvious. At the same time, a strategic question arises: if the agent makes the product selection, classic brand management at the storefront loses influence, and factors such as data quality, price structure and reviews gain it. Retailers who create a technically clean e-commerce foundation early keep the option of connecting to agents open.

Prerequisites: how a shop becomes agent-ready

For an agent to be able to use a product range at all, it needs a technical foundation that goes beyond the classic shop designed for people. At its core are four things: structured product data with unambiguous attributes, prices and availability; a machine-readable interface or a well-maintained product feed through which an agent can query the range; structured markup (Schema.org) that makes content and meaning explicit; and a reliable checkout interface that accepts an order together with an authorised payment. Much of this contributes to existing goals at the same time: clean product data also improves classic conversion and visibility. Agent readiness is therefore rarely a project in its own right but the logical continuation of good product data management and an open, API-capable architecture.

Risks, law and the customer relationship

As convenient as agent-led purchases are, they raise just as many questions. Legally, it has to be clarified who concludes a contract when a machine acts on someone's behalf, how withdrawal and warranty apply and what effective consent to payment looks like; AP2's cryptographic mandates are an attempt to make exactly this verifiable. Strategically, retailers risk losing the direct customer relationship: if an agent stands between brand and buyer, the storefront, brand image and upselling lose influence, while objective signals such as data quality, price and reviews tip the balance. And operationally, there is the question of fraud prevention when automated actors make requests in large numbers. These points are no reason to wait, but they are a reason to design the connection deliberately, and in doing so to decide which parts of the customer relationship you want to keep for yourself even in an agentic world.

A practical example

A customer asks his AI assistant to reorder printer cartridges for a particular model, "preferably where I last bought them, otherwise somewhere cheaper". The agent identifies the right product via the product interfaces of several merchants, checks prices and delivery times, recognises the preferred shop and triggers the order there via an agentic checkout protocol. The payment is authorised via a signed mandate that covers exactly this purchase: amount, merchant, time. The customer receives a confirmation in the chat. For the merchant, there was no person in the funnel; what tipped the balance was data quality, availability and the technical connection to the protocol.

Opportunities and open questions

Agentic commerce promises convenience and can greatly simplify repeat purchases. However, central questions remain open: how do retailers keep the customer relationship when an agent stands in between? How are liability and withdrawal regulated when a machine buys? And which standard will prevail? For companies, the pragmatic answer is not to connect every platform immediately, but to keep their own data foundation and interfaces clean enough that a connection remains possible as soon as the market consolidates. Anyone who needs advice on this will find it in our AI consulting.

Conversational, social and agentic commerce

Agentic commerce is one of a series of new forms of retail with which it is easily confused. Conversational commerce means selling in dialogue, via chatbots, messengers or voice assistants, but in the end it is usually the person who makes and confirms the purchase decision. Social commerce moves the purchase into social networks, where discovery and transaction merge, but it too is controlled by people. Agentic commerce is the next step: here the person delegates the execution to an agent that selects and buys independently, provided it is authorised to do so. The boundaries are fluid (a voice assistant can begin conversationally and end agentically), but the decisive difference lies in the degree of autonomy. For retailers, this development means that they have to prepare their range not only for human buyers in social and conversational channels, but increasingly also make it machine-readable and buyable for autonomous actors.

Frequently asked questions about agentic commerce

What distinguishes agentic commerce from conversational commerce?

Conversational commerce conducts a dialogue in which the person usually makes the purchase at the end. Agentic commerce goes further: the agent carries out the purchase independently, including selection and payment, provided it is authorised to do so.

Do I as a retailer have to connect a protocol now?

Not necessarily. It makes more sense to first ensure structured, machine-readable product data and open interfaces. This keeps a connection to ACP, AP2 or UCP possible as soon as a standard prevails.

What role do reviews and data quality play?

A central one. Since the agent does not make an emotional purchase decision, objective signals such as complete attributes, correct prices, availability and reviews gain considerably in weight.

Background on the open standard: Stripe on the Agentic Commerce Protocol and the ACP specification on GitHub.