Adyen: 'The AI Was Never the Hard Part' — Agentic Commerce Is Still at 0.5, and Payment Infrastructure Decides the Winners

'We Are at Version 0.5'

Adyen on Agentic Commerce: The AI Was Never the Hard Part

Adyen's Karan Katyal tells Karen Webster the industry is barely off zero. Here's what gets built first, and why AI was never the hard part.

www.pymnts.com

Ask how far along agentic commerce is and you usually get an adjective. Early. Nascent. Emerging. PYMNTS CEO Karen Webster wanted a number instead.

So she handed Adyen's global head of agentic commerce, Karan Katyal, a scale. Zero means nowhere. Five means agentic commerce is just an embedded part of how buying happens, as ordinary as tapping a card. Where are we now?

The answer was "We are at version 0.5." Not a one. A half. On the most-hyped shift in commerce since the smartphone, the person running the agentic business put us at one-tenth of the way to the midpoint. When Webster pushed him out a year, to next June, Katyal said 1.5.

From 0.5 to 1.5 is a single point on the scale, which sounds like a rounding error. But it is also a tripling: in twelve months we will do three times everything built to date, and still land 70% short of the halfway mark. The slope is steep and the summit is far. Both are true at once, and that tension captures the whole shape of this moment.

'The AI Was Never the Hard Part'

This is the heart of the conversation. Katyal's take is that the intelligence already works. People lean on large language models to research products, compare options, and shorten a list of twenty down to three. Discovery feels effortless. It is everything after discovery that falls apart.

"

Infrastructure is a much bigger block than folks thought about, perhaps.

Source: Karan Katyal, Global Head of Agentic Commerce, Adyen

Here is what that means in practice. A traditional product listing needs maybe a dozen attributes to look good on a webpage. An AI agent trying to tell two similar products apart needs far more, roughly three times as much: detailed specs, reviews, context, the stuff a human reads between the lines. Most catalogs simply do not carry it. Making a catalog legible to machines is not a tweak; it is a project.

And the theme Katyal kept returning to was trust. Trust, risk, and fraud are not bolted on at the end; they have to sit at the center of the foundation of whatever you build. Because the moment an agent acts on your behalf, the questions get sharper. Who is liable when the agent books the wrong flight? When it misreads what you asked for and buys it anyway? A botched checkout is annoying. A botched delegation becomes a dispute about responsibility.

This diagnosis is not unique to Adyen. Mastercard's digital leader Sherri Haymond described the same gap, the absence of anyone reading the terms once spending is delegated to an agent, as the industry's biggest opportunity and biggest vulnerability at once. The players at the front line of payments are all pointing at the same spot.

This Isn't One Thing, and That's Why It's Slow

Part of why we are stuck at 0.5 is that agentic commerce is not a single product anyone can ship. Katyal described several models maturing at the same time, each with its own technical demands.

In some, the human stays in the loop the whole way. In others, the agent buys autonomously once you set the rules. Then there is machine-to-machine, where software transacts with software, and business procurement, which is its own animal entirely. Four directions, four sets of standards, and no referee.

"

Fragmentation will continue for quite some time. Agentic commerce is really almost like a direction of travel.

Source: Karan Katyal, Global Head of Agentic Commerce, Adyen

A heading you point toward, not a destination you arrive at, while competing protocols, onboarding flows, checkout requirements, and payment methods sort themselves out underneath. The standards landscape alone makes the fragmentation plain. OpenAI and Stripe's ACP (Agentic Commerce Protocol) powers ChatGPT Instant Checkout; Google and Shopify's UCP covers discovery and cart; AP2, Google-initiated and now FIDO-governed, proves who authorized the payment. Visa Intelligent Commerce and Mastercard Agent Pay each add their own entry point for tokenization and authentication. The layers may be complementary, but from a merchant's seat they are an unreconciled crowd.

That is why Katyal is blunt about the money. Agent commerce is almost a series of experiments, and if the thing you need to justify it is hard numbers, now is not the time yet.

Don't Read '0.5' as 'Come Back Later'

Here is where Webster pushed back, and where merchants should pay attention. Low maturity is not permission to wait.

She has watched this movie before. Businesses that dragged their feet on the shift from stores to digital, and again from digital to mobile, spent years clawing back ground they did not have to lose. Waiting for certainty has almost never been the winning move. The consumers are already, in her words, knocking at the door.

The catch is that they cannot get through it. "Consumers just can't close the loop," Webster said. The interest is real and the intent is real, but the path from intent to purchase is broken. The merchants who fix their half of that path now, during the experimentation window, are the ones who will be ready when the loop finally closes.

So what gets fixed first? Katyal expects low-drama purchases to lead: repetitive household restocks, low-consideration buys where handing off the task removes friction without making anyone nervous about losing control. Webster added a twist. AI may prove just as valuable at the top of the consideration curve, on big researched purchases, where the real payoff is the gathering and comparing rather than the click that ends it.

Both agreed there will be no single universal way to shop. It bends to the person and the moment. Sometimes you want the agent to handle it; sometimes you want to drive. The infrastructure has to support both.

Adyen's Bet: Build the Connectors, Not Another Platform

Against that backdrop, Adyen rolled out Adyen Agentic on June 16, 2026. The suite bundles discovery, checkout orchestration, and payments into modular pieces meant to cut the integration headache across the various AI commerce protocols.

It splits into three layers. Agentic Feed is a structured layer that distributes catalog, pricing, and availability data in real time across conversational commerce environments. Agentic Cart is the orchestration layer that connects the checkout, tax, fulfillment, and order management systems merchants already run to conversational platforms. And Agentic Payments is the payments and fraud layer, delivering authentication, token portability, merchant-of-record preservation, and risk management across protocols that are still evolving.

The logic tracks with everything above. Help merchants make their product data machine-readable, let them keep the checkout and fulfillment systems they already run, and extend the authentication, tokenization, and fraud controls they trust into agentic environments, without forcing a rebuild for every new platform that shows up. Calling it a "universal translator" is Adyen's way of signaling that connector character.

The early lineup is telling. American Express, Mastercard, and Visa join as payment networks; Salesforce as a platform; and ESW, Scheels, Sézane, and SharkNinja as merchants. Card networks that normally compete sitting on the same translation layer makes sense in a world stuck at 0.5, where nobody knows which protocol wins: betting on one is not the smart play. Katyal's conclusion was clean. The smart move is not to bet on one, but to be able to plug into all of them. Adyen Agentic launches in limited availability for enterprise merchants in the U.S., with global expansion to follow.

Conclusion

The line "the AI was never the hard part" drags the agentic commerce debate back to solid ground. Discovery already works. What gets tested from here is the plumbing, machine-readable catalogs, authentication, trust, and clear liability, hard to see yet indispensable to everyone.

From 0.5 to 1.5: tripling and still short of the midpoint does not mean there is time to wait, it means there is no room to wait for certainty. Before guessing which protocol wins, make your product data machine-readable and open your payment acceptance to multiple standards. That is what decides whether you are on the ready side the day the loop finally closes.