Agentic commerce: What changes for merchants when the buyer isn't the decision maker?

For decades, commerce has been built around a familiar relationship. A customer seeks a product or service, decides exactly what they want, chooses where to buy it, and completes the transaction with a merchant.

Technology has repeatedly reshaped that process. The shift from cash to cards changed how merchants accepted payments, ecommerce moved the storefront online, and mobile put the checkout in a customer's pocket. Each shift made buying more convenient for consumers, and merchants adapted the experience around that expectation.

Agentic commerce could change something more fundamental. The customer may still be the person the purchase is for, but an AI agent can increasingly take on the work of discovering products, comparing options, making recommendations and, with the right permissions, clicking buy.

Agentic commerce is gaining momentum, but it’s still at an early stage. Checkout.com's 2026 research found that 89% of merchants are actively preparing for agentic commerce, with 42% already testing it. Yet agentic transactions still account for just 3% of transactions. Meanwhile, Juniper Research forecasts that agentic commerce spend could reach $1.5 trillion globally by 2030, even as trust, infrastructure, and the rules around agent-led purchases are still taking shape.

So what is actually changing for merchants, and where should you be paying attention while the market develops?

To explore these questions, we spoke with payments and AI experts about what happens when an agent takes on more of the buying decision, where the technology still falls short, and what that could mean for merchants as the relationship between businesses and consumers evolves yet again.

From helping customers choose to choosing for them

An agent can increasingly take over the work that leads a customer to a purchase. A traveler, for example, could ask an agent to find a flight under a certain price, compare baggage and cancellation policies, monitor fares, and book it once the right conditions are met.

Matteo Gamba, formerly Head of Product at Wayfair and on the board of the Merchant Risk Council (MRC), now an independent payments and fintech advisor, describes agentic commerce as a “spectrum of delegation.”

“There’s a spectrum of what can be considered agentic commerce, from purely referral traffic (think of it as a replacement for Google or Facebook as a referral source) to fully autonomous agentic shopping,” he says. “But at a higher level, the vision around agentic purchases is basically having a personal shopper always at your disposal that knows you deeply and can make purchases on your behalf.”

Consumers are already showing an appetite for the earlier stages of that spectrum. Visa found that 73% of respondents would be willing to delegate product discovery to an agent and 69% would delegate evaluation. More than half of consumers surveyed by Checkout.com said they would even let AI switch brands if it found a better-value option.

The commercial impact is starting to show too. Ecommerce data and market-intelligence firm ECDB found that AI-sourced shopping traffic already converts 54% higher than non-AI traffic such as paid and organic search, while Shopify's Q1 2026 data shows referral sessions from AI chatbots grew more than eightfold year over year and AI-referred orders grew nearly thirteenfold. 

That changes how merchants have to think about consideration. If more of the buying decision happens before a customer reaches you, your strategy has to account for the agent making that first cut.

But can an agent know what the customer actually wants?

Delegating more of a purchase to an agent assumes it can understand more than the instructions it is given. It has to interpret what the customer means, work out which factors matter most, and make sensible trade-offs between them.

Swapna Kutcharlapati, who leads payment success and monetization at Prime Video, sees this gap between authorization and intent as one of the fundamental challenges agentic commerce has yet to solve. 

She describes asking an AI agent she tested to find toddler-friendly accommodation for a family trip to Hana within a specified budget. The agent followed those instructions and still selected a place with insects and no air conditioning. “The AI agent did exactly what I asked it to do. It did not do what I wanted,” she explains. 

The problem was that “suitable for a toddler” carries reasonable expectations that a human acting on someone's behalf might infer without being explicitly told. That is the gap Swapna draws between authorization and intent. Authorization can establish what the agent was allowed to do, but cannot always establish whether the outcome was what the customer actually wanted. Visa found that 42% of respondents are concerned about purchase accuracy when using AI agents.

For you as a merchant, that changes what it means to understand your customer. The way they frame what they want and the constraints they give their agent may increasingly determine whether your product is considered at all.

What happens when the agent gets it wrong? 

Now let’s say a customer authorizes an agent to make the purchase itself. If the payment is valid but the outcome fails the customer's intent, who deals with the consequences?

Online commerce addressed a similar trust problem through returns, refunds, cancellations, and chargebacks, which helped make the shift to online shopping feel safe for hesitant consumers.

Agentic commerce needs a similar safety net. The difficulty lies in determining where that risk sits in the ecosystem and how much falls to the merchant, agent provider, issuer, or payment network.

Consumers delegate because they want the convenience of handing over the work. But if every mistake comes back to them as a financial or logistical problem, that convenience starts to lose its value. “Someone needs to take the cost away from consumers for agentic commerce to scale,” Swapna says.

She thinks accountability should depend partly on how much freedom the customer gave the agent. The tighter the boundaries, the clearer the responsibility when an agent steps outside them.

The industry is already starting to build those controls. Mastercard and Google's Verifiable Intent creates what Mastercard describes as a “tamper-resistant record” of what a customer authorized an agent to do. Swapna has also been exploring similar controls through Patron, an early prototype she built of a consumer trust and authorization management layer in agentic commerce. And American Express is adding its own protection layer through Agent Purchase Protection, designed to cover errors when a registered AI agent deviates from a customer's authenticated purchase intent.

These approaches make delegated authority clearer. The harder part is when the agent stays within those boundaries but the outcome still doesn't reflect what the customer wanted. Amex, for example, excludes subjective or non-verifiable intent such as asking for the “best” option or something “really nice.”

Another consideration on the merchant side is that agents may remove some of the behavioral and payment signals fraud systems rely on to assess whether a transaction is legitimate.

Riskified CMO Jeff Otto told us early agent-led checkout flows are already losing some of those familiar signals. “The result is you're losing somewhere around a third of the signal that most platforms use to make informed decisions, to make sure good customers get approved and fraudsters get denied.”

Riskified saw that play out in its own testing. A fraudulent order using a synthetic identity was approved through an early agentic checkout flow but declined when the same order went through the merchant's traditional checkout.

That broadens the payment-control questions merchants need to solve around agent-led transactions. Merchants need to understand both what happens when an agent makes the wrong purchase and whether they still have enough information to stop a bad transaction before it goes through. 

Is the technology even ready for agentic commerce? 

Even if we get better at managing the risks around agent-led transactions, there is still a more fundamental issue. The technology making the decision has to be reliable in the first place. This is where Georg Zoeller, CTO at NOVI Health and co-founder of the Centre for AI Leadership, brings a useful dose of realism. 

He describes the way large language models (LLMs) represent information as a form of “lossy compression.” Enormous amounts of information are distilled into the model's parameters rather than stored in a form that can always be retrieved exactly. When information is missing or cannot be reliably reconstructed, the model can still generate a plausible answer, which is one reason hallucination remains a major problem.

In a buying context, the difference between a plausible option and the right one for a particular customer can depend on context the model does not reliably preserve or reconstruct. Georg contrasts this with software development, where AI-generated code can be run against tests to see whether it works. A purchasing decision has no equivalent test for whether the agent correctly understood what the customer wanted.

Georg also warns of “prompt injection,” where an agent may encounter instructions embedded in the information it is processing. What an agent reads can then influence its behavior, creating a new attack surface.

Georg points to a recent New York Times report on OpenAI agents attempting to breach government and university websites while carrying out data-collection tasks. “If these models go out and hack foreign governments when asked to perform simple data collection, and the world’s foremost labs can’t seem to prevent it, then how on earth would we be able to trust them to do shopping?” he asks.

“A tiny probability influencing an agent, times huge exposure, can become consequential,” Georg adds. At the scale of millions of purchasing decisions, even a small chance of an agent producing the wrong answer or being influenced in the wrong direction can create significant exposure.

The technology is advancing quickly, and investment is flowing heavily into the space, but reliability and security are still catching up with the ambition. For merchants, that makes understanding the limits of the technology just as important as understanding what agents can do.

What changes for merchants?

For decades, merchants have invested in the places where customers make purchasing decisions. Product pages, merchandising, recommendations, checkout, and the wider shopping experience all give you opportunities to explain your value.

Agentic commerce introduces a new layer between you and that decision. If an agent becomes the first entity evaluating what your business has to offer, you have to think differently about how you earn consideration and create preference. The storefront may still be there, but part of the customer journey could increasingly happen somewhere you don’t control.

That puts more pressure on you to understand what actually drives your customers to buy. Online shopping already pushed businesses to learn from how people behaved through the funnel. Agentic commerce goes further because an agent may now be interpreting those preferences and making trade-offs on your customer's behalf.

It also makes the underlying data more important. Agents need clear, structured information about your business. If they cannot accurately understand your products or policies, you may be less likely to make it into the consideration set at all.

Matteo expects the infrastructure underpinning agentic commerce to become increasingly commoditized, with more of the value shifting toward “data, intelligence, and agentic intelligence.” That makes your data, and how well you understand the customer behind it, increasingly strategic.

Not every purchase will move equally far along this path. Matteo expects predictable, repetitive purchases like groceries to lend themselves more naturally to delegation, while expensive, highly customized, or difficult-to-return products are likely to keep the customer closer to the final decision.

The opportunity will therefore look different for every merchant. What matters is understanding where your customers see value in delegating, what drives those decisions, and whether an agent has enough information to represent your business accurately.

The customer remains at the center, but the way you understand, reach, and win them may have to evolve once again.

‍