Agentic commerce: What it is, how it works, and what comes next

Agentic commerce is when an AI agent acts on a shopper's behalf, searching, comparing, and building a cart based on a natural language request rather than a traditional search bar or category menu. The agent does the legwork but the shopper still approves the purchase.

Agentic commerce is expected to span three evolutionary stages:

1. Assisted discovery: An agent helps a shopper find and compare products, then routes them to a merchant to complete the purchase themselves.

2. Human-in-the-loop purchasing: An agent builds a cart with live pricing, tax, and availability, and the shopper reviews and approves the payment.

3. Autonomous purchasing: An agent completes a transaction under a pre-agreed mandate, with no confirmation needed for each individual purchase.

Most live activity today sits in the first two stages. Autonomous purchasing exists in early pilots and is likely years off in markets like Europe, where regulation still assumes a person is present to authorize each transaction.

How does agentic commerce work?

The purchase flow for a typical human in the loop agentic commerce transaction follows three steps:

• Discovery: In response to a request from the shopper, the agent queries structured product data including catalog, pricing, and availability.

• Cart and checkout: The shopper makes a selection and the agent builds a dynamic cart, recalculating pricing, tax, and delivery in real time.

• Payment and authorization: The shopper reviews and approves the payment, and the transaction is screened for fraud before it processes.

Agentic commerce vs conversational commerce

Conversational commerce is a subset of agentic commerce. AI tools answer questions and make recommendations, but the shopper still handles discovery and checkout themselves. This is closer to the assisted discovery than human-in-the-loop purchasing where the agent takes on part of the discovery and checkout process on the shopper's behalf.

Agentic commerce vs traditional ecommerce

Agentic commerce is structured differently from a traditional ecommerce checkout. A traditional flow hands off sequentially between systems, giving each one time to catch up before the next step starts. Agent sessions don't work that way. They query pricing, tax, and inventory repeatedly and simultaneously throughout a single interaction, which is part of why agentic commerce places different demands on the systems behind it than a standard checkout does.

The below table shows the differences between agentic commerce and traditional ecommerce:

A protocol is the set of rules governing how an AI agent is authorized to act, how payment gets executed, how consent is captured, and who's liable if something goes wrong.

Two protocols are shaping the space right now:

• Agentic Commerce Protocol (ACP) associated with OpenAI

• Universal Commerce Protocol (UCP) associated with Google

These are focused on solving the human-in-the-loop agentic transaction. Meanwhile, other protocols are emerging for different parts of the agentic commerce landscape. For example, x402, backed by the Linux Foundation, is built for machine-to-machine payments, where software pays software with no consumer in the transaction flow at all.

In all cases, it’s still early days, so if you're making infrastructure decisions, you’ll want to avoid betting too early on one protocol as you might have to overhaul your systems later.

Fraud prevention has traditionally asked two questions: 

1. Is this person who they claim to be?

2. Are they authorized to make this transaction? 

Agentic commerce adds a new dimension. With an agent sitting between the shopper and the merchant, you also need to identify legitimate agents versus malicious ones, and confirm that the agent is acting in line with what the customer intended.

This can be complicated to get right. Fraud prevention has long assumed bots are bad by default. But with agentic commerce, a legitimate shopping agent can look identical to a scraper or exploit bot. And even a fully authorized agent can drift from what the shopper wanted, swapping in a cheaper product or picking the slowest shipping option to hit a price target.

Two things keep this in check for now. First, a verifiable, auditable mandate that records what an agent was actually approved to do. Second, the shopper confirming each payment. However, as fully autonomous, agent-executed payments become more common, that second backstop disappears.

Agentic commerce also makes the fraud you’re already dealing with (promotion abuse and inventory-window exploitation) happen faster and at greater scale. So your focus needs to shift from a single check at the point of payment to a continuous layer of control across the whole interaction. That’s why 30% of US enterprise merchants say AI platform trust scoring is the most critical new signal for agentic commerce, ahead of other emerging risk indicators. 

Agentic commerce compliance

Regulation adds another layer of complexity, particularly in Europe. Frameworks like SCA and 3DS2 assume a real person is confirming the purchase, which is why human-in-the-loop purchasing is likely to remain the standard in the region for some time. In the US, there's no equivalent real-time authentication mandate, which is part of why agentic pilots are moving faster there.

No ecommerce platform or payment provider can guarantee compliance. The right infrastructure gets you thinking about agent trust, consent, and liability earlier in the process, before problems reach checkout.

Payments are the execution layer of agentic commerce. Once a shopper approves, an agent needs a way to pay that a merchant can authorize, process, and screen for fraud, just like any other transaction. That’s why tokenization is such a powerful tool. Instead of handing raw card details to an agent, you can assign a secure token that allows the agent to act on the shopper’s behalf (with their approval). This allows you to recognize the same customer across different agent sessions and platforms rather than treating each one as a new buyer.

According to OpenAI's B2B Partnerships lead speaking at NRF 2026, high-intent traffic driven by AI agents is converting two to three times better than typical search traffic. Moving up the funnel, agentic commerce also opens up a new discovery surface for customers who’d rather describe what they want than browse for it.

But, while these early indicators look good, agentic commerce also comes with tradeoffs. When a third party controls discovery and checkout, you risk becoming an interchangeable back-end supplier rather than a valued brand. There's also a data question: if the platform holds the mandate, sets the limits, and keeps the proof of purchase, it could end up owning more of the customer relationship than you do.

You can lower these risks by retaining control of the payment token, rather than letting the AI platform issue and hold it. That way, you're the one who recognizes the shopper regardless of which AI platform they’re using.

The most common use cases today fall into two areas:

• Retail discovery-to-checkout: A shopper asks an agent to find and compare a product against criteria like style, price, sustainability, or delivery timing. They make a choice, the agent builds the cart, and the shopper approves. 

• Travel booking: An agent compares flights or hotel options against a shopper's stated preferences and completes the booking once approved.

Within retail, businesses are adopting agentic commerce at different rates. For example, apparel, fast fashion, and sports and outdoor are early movers because these tend to be lower-consideration purchases where a shopper is comfortable letting an agent narrow the options down. However, we’re seeing slower adoption in luxury where purchases tend to involve more consideration, research and reassurance, often with a human touchpoint. 

Within all categories, businesses are taking different approaches. Some are using proprietary platforms that let them move fast on new protocols. Others are waiting to see what their commerce platforms build before committing engineering resources. For more details on how different businesses are approaching agentic commerce see: Insights from five leading retailers.

LLMs have reached 1 billion users in roughly three years, which is the fastest adoption of any technology on record. Behind the scenes, platforms and card networks are actively building the infrastructure to support agent-led transactions at scale. But many hard questions, like which protocols will dominate, how liability gets shared, and how authorization works across platforms, are still unanswered.

A big consideration for businesses is whether you can work with every AI platform without rebuilding your setup each time a new one launches. The future of agentic commerce won't be about ripping and replacing infrastructure; it will be about plugging new channels into the systems merchants already trust.

Since agentic commerce is still in the pilot phase, you don’t need to have a complete agentic strategy locked in but here are a few things you can do right now to prepare:

Ensure your product data is machine-readable

Standard product feeds were built to serve humans and search engines. AI is different. It requires access to a variety of product information in real-time which it will query several times during a single session. 

Ensure all your data is accessible in one place, not distributed across several departments. For example, one brand found their product weight lived in a completely different system from the rest of their catalog, with no reliable link between them. A gap like that is manageable when the front end is your own website or app. But for an AI platform, that’s a dead end. Learn more in: Agentic commerce and product feeds: a guide for retailers.

Keep your infrastructure as open as possible

With no dominant protocol yet, building for one platform's requirements risks having to rebuild for the next. Favor an infrastructure that lets you adapt as standards evolve, rather than locking into a single integration too early.

Ensure clear consent and mandate handling

Make sure you’re clear on what an agent is authorized to do on a shopper's behalf, what limits apply, and how that authorization is recorded and audited.

Align your teams

Agentic purchases touch fraud, risk, product, support, and marketing all at once. So it’s important to get these teams talking to each other early, rather than treating it as a single team's problem to solve.

We built Adyen Agentic as a universal translator connecting businesses to multiple AI platforms and protocols via a single integration. It follows the same three-stage structure covered earlier: discovery, checkout, and payment and security and is underpinned by four principles:

1. Agents can only act under an explicit, verifiable shopper mandate. 

2. Businesses keep control of their own payment methods and tokenization. 

3. A universal payment token lets businesses recognize shoppers across agentic channels.

4. Businesses retain ownership of their transaction data and the customer relationship, before, during, and after the purchase.

We're also active in shaping the standards this all depends on, including membership in the Agentic AI Foundation, collaboration with Google on the Agent Payments Protocol, and founding membership of the x402 Foundation.

Agentic commerce is moving quickly, but for most businesses, right now, the shopper still needs to make the final purchase decision. This gives you time to get the fundamentals right: structured data, flexible infrastructure, clear consent handling, and a payments setup that recognizes your shoppers wherever they show up.

Stay up to date on agentic commerce in our latest articles >