Agentic commerce is the set of commercial transactions in which an artificial intelligence agent, acting under a human's standing instructions, completes the discovery, evaluation, authorization, and settlement of a purchase without the human re-confirming each step.

That sentence is the working definition we use in every Major Matters piece on this topic. The rest of this article unpacks each clause and names the structural shifts that follow.

When the buyer is an agent, the assumptions that built modern retail, payments, and loyalty stop holding. Each layer of the commerce stack has to be re-examined for which assumption is breaking and which is not.

Why the definition matters

The term "agentic commerce" has been used loosely for two years to mean everything from a chatbot recommending products to a fully autonomous trading bot moving money on-chain. Loose terms produce loose analysis. We tighten the definition for two reasons.

First, payments networks, regulators, and merchants need a shared vocabulary to make decisions. If a card scheme says "we are launching an agentic commerce product" and a regulator hears "autonomous money movement" and a merchant hears "AI recommendations," every conversation breaks.

Second, the structural questions about agentic commerce only become tractable once the term is precise. Liability, identity, dispute handling, and trust each look different depending on what the agent is and is not allowed to do.

This definition is narrow enough to be useful and broad enough to cover the systems being built.

What counts and what does not

Counts as agentic commerce:

Does not count as agentic commerce:

The dividing line is whether the agent completes a transaction under a human's standing instructions without per-purchase confirmation. Everything else is either retrieval (search) or assistance (chatbot).

The four layers

Every agentic commerce transaction touches four layers. We use this decomposition in every piece on the topic.

1. Discovery

The agent has to find candidate purchases that match the user's stated need. This used to mean ranking a search results page. In agentic commerce, it means the agent reading product data, comparing options, and surfacing one or a small number of candidates.

The discovery layer is where merchants now have to be agent-readable, not just human-readable. Google's llms.txt audit, Anthropic's published agent shopping behavior, and OpenAI's checkout-pivot work are all about this layer. A merchant whose product data is opaque to agents is invisible to agent-driven shoppers.

2. Authorization

The agent has to confirm the purchase falls within the user's mandate. Standing instructions might say "spend up to fifty dollars on household items" or "renew this subscription unless the price increases by more than ten percent." The authorization layer is where the agent decides whether to proceed.

This is where the trickiest commerce questions live. Did the user authorize this specific purchase, or only the general intent? If the agent makes a mistake, who is liable? We call the unsolved part of this question the MM Liability Gap.

3. Payment

The agent has to actually pay. This means selecting a payment method, presenting credentials to a processor, and getting an authorization. The payment layer looks the most like existing commerce because the rails are the same, but the identity question is different: is the cardholder present? In what sense? With what authentication?

Card networks, BNPL providers, and emerging on-chain payment protocols are all racing to solve this layer. Visa has shipped its Trusted Agent Protocol. Mastercard has its agent-aware authentication. Klarna has the agentic-BNPL inside ChatGPT. Coinbase, Stripe, and AWS built the x402 agent payments rail. The work at this layer is the most concrete and the most measurable.

4. Settlement and disputes

Payment is not the end of the transaction. The funds have to settle, the merchant has to deliver, and if something goes wrong, the dispute has to be handled. This is the layer where agentic commerce is most underbuilt. Dispute frameworks were designed for human cardholders calling their bank. When the agent bought the wrong thing on the wrong card, there is no clean dispute primitive.

Together these four layers form what we call the MM Trust Layer Model. Each layer needs a working trust mechanism. A failure at any one breaks the transaction.

What changes when the buyer is an agent

Three things shift, structurally.

The destination economy weakens. Apps, loyalty programs, branded checkout flows, all of these assume the shopper navigates to the merchant. Agents do not navigate. They route. Every owned-destination asset depreciates as agentic share of commerce rises. We have written before about how Klarna picked the agent surface over its own storefront and how this fits the MM Destination Economy Thesis. The same shift is visible in the AWS, Stripe, and Coinbase agent payments rail and across the broader state of the agentic commerce stack.

The identity question becomes harder, not easier. When a cardholder taps their phone, identity is easy: the cardholder is present, the device is bound to them, the biometric matched. When an agent transacts, identity fragments. Was the consumer present? Was the agent acting under fresh consent or standing consent? Was the agent itself authenticated to the merchant and the processor? PYMNTS Intelligence reported in May 2026 that 79 percent of firms verify identity at login, but login is the wrong moment to verify when the next attack vector is an authorized agent acting on a verified identity.

Loyalty and attention economics stop working. As Theodora Lau has argued, the entire post-2010 retail playbook assumes the shopper has attention to capture. Points balances, push notifications, branded apps, retargeted ads, all of these tools influence human buying behavior. Agents have no attention to capture. The whole apparatus needs to be rebuilt around influencing the agent, not the human, and that is a different problem with different solutions.

A note on what we are not saying

We are not predicting that agentic commerce will be most of consumer commerce by any specific date. We have seen too many adoption curves overshoot in early projections and undershoot in late ones to make a useful guess.

We are saying that enough of the four layers are now being built simultaneously that the rest of the stack has to plan for the case where agentic commerce is meaningful, even if its precise share is unknown. As Simon Taylor wrote in his fintech-OS framing, the new commerce stack is vertical rather than horizontal. The vertical stack assumes the buyer is whoever the agent is acting for, not whoever clicks the buy button.

That is a working planning assumption. It does not require believing the most aggressive forecasts.

How to use this definition

Three practical applications.

If you are a merchant: audit your product data, checkout, and dispute paths against the four layers. The discovery layer is the most urgent, the dispute layer is the most unsolved.

If you are a payments network or processor: the authorization and payment layers are where your existing infrastructure either fits or does not. Standing-consent semantics, agent authentication, and per-transaction risk scoring need to be specified.

If you are a regulator or policy analyst: identity-at-authorization, not identity-at-login, is the structural question. Cross-border agent commerce will compound this. The work is harder than the press releases suggest.

Sources

If the four layers all need rebuilding, which one breaks first under load?

Charlie Major is a Product Development Manager at Mastercard. The views and opinions expressed in Major Matters are his own and do not represent those of Mastercard.