Somewhere in a European ecommerce backend this morning, a payment was approved. The buyer never opened the product page. The buyer was software.
Agentic commerce has stopped being a conference slide. Google’s Native Checkout in AI Mode, OpenAI’s Instant Checkout inside ChatGPT and Amazon’s delegated purchasing all run on live payments infrastructure today. The International Monetary Fund describes them not as experiments but as emerging design patterns in payment initiation, which is a sober framing from an institution that does not chase trends.
For merchants, agentic commerce raises an uncomfortable question. Payments systems spent fifteen years learning to detect automation and block it on sight. Agentic commerce now asks those same payments systems to welcome automation, verify it and settle for it.
Vladyslav Kolodistyi, who works on payments infrastructure at PayAdmit, thinks most merchants are staring at the wrong layer of the problem.
“Everyone treats agentic commerce as a marketing channel,” says Vladyslav Kolodistyi. “It is not. Agentic commerce is an authorisation problem wearing a shopping interface. The moment the initiator of a payment is not a human being, every assumption inside your payments risk stack is quietly wrong.”
Agentic payments break the assumptions the checkout was built on
Traditional payments risk models read behaviour. Mouse movement, typing cadence, time on page, device fingerprint. Those signals exist because a human generates them. AI agents generate none of them, and the payments engine reads that absence as fraud.
The result is not a security incident. It is silent revenue loss. Agentic payments that should have cleared get declined, and nobody files a support ticket, because from the merchant’s side the payment simply never happened.
Vladyslav Kolodistyi has watched this pattern before in a different form. “We went through this with mobile payments and again with tokenised payments,” he says. “A new initiation method arrives, legacy risk rules read it as anomalous, approval rates drop for eighteen months until somebody retunes the model. Agentic commerce will follow the same curve unless payments teams get ahead of it.”
The card networks reached that conclusion earlier than most merchants did. Visa’s Intelligent Commerce programme and Mastercard’s Agent Suite, both live in 2026, converge on one idea: Know Your Agent. Registration, cryptographic signatures and network tokens that let payments infrastructure separate legitimate AI agents from a scraper with good manners.
The distinction matters commercially, not only technically. Under a Know Your Agent model, AI agents carry verifiable identity into the payments flow, and the merchant can price risk rather than guess at it. Without it, AI agents look like every other unwanted automated visitor, and the payments stack treats them accordingly.
There is a scale argument underneath all of this. Boston Consulting Group has estimated that agent-led shopping could account for more than a quarter of ecommerce spending within several years, a figure reported by eMarketer in its payments outlook. Whether agentic commerce hits that mark early or late, no payments organisation can treat a quarter of a channel as an edge case.
Four questions define whether a payments stack is ready for agentic commerce:
- Can the system identify which of the AI agents initiated a payment, and prove the shopper authorised it to act?
- Does the fraud model score AI agents separately, or does it inherit rules written to stop bots outright?
- When AI agents buy the wrong item, who owns the dispute: the shopper, the agent platform, or the merchant?
- Do product data feeds carry enough structure for AI agents to evaluate the offer before the checkout is ever reached?
The fourth question is the least glamorous and the most often skipped. Agentic commerce runs on machine-readable catalogue data. Payments only become relevant once AI agents have already decided to buy, and AI agents cannot decide anything from a marketing page written for human eyes.
Designing an AI agent checkout that does not decline its best customers
The phrase AI agent checkout suggests a new page, a new button, a new interface. Vladyslav Kolodistyi argues the opposite.
“There is no new page,” he says. “The AI agent checkout is your existing checkout, invoked by a caller you cannot see. The visible layer is unchanged. The payments authorisation layer carries all the difference, and that is where payments teams should spend the budget.”
This is why he places agentic commerce readiness inside the risk and screening function rather than the front end. The practical question is not whether payments arrive from AI agents. It is whether the payments platform can express the difference between authorised and unauthorised automation as a scoring decision at the checkout.
According to Vladyslav Kolodistyi, the framing many merchants use makes the work harder than it needs to be. “Human or bot is the wrong axis,” he says. “Authorised or unauthorised is the right one. AI agents acting inside a mandate the customer signed produce good payments. A scraper hammering your endpoint does not. Both are automation. Your payments logic has to hold both ideas at once.”
The economics deserve attention too. Agent-led conversions on the Stripe-powered Agentic Commerce Protocol carry a platform fee of around four percent, a figure the IMF cites in its own analysis. Layered on standard processing costs, agentic commerce arrives with a margin structure that looks nothing like organic traffic. That is a commercial decision hiding inside what many payments teams file as a technical integration.
There is a fragmentation problem worth naming. Several competing agentic commerce protocols are live at once, and no merchant can integrate all of them cleanly. Visa has responded with a protocol-agnostic on-ramp so merchants can accept agent-initiated payments without replacing their processor. Vladyslav Kolodistyi reads that as a signal about sequencing.
“When the networks build an abstraction layer over agentic commerce, they are telling you the standards war is unresolved,” he says. “Do not pick a winner. Build the payments plumbing so switching protocols is a configuration change, not a rewrite. That is the discipline that saved merchants during the 3D Secure transitions.”
His practical advice for payments leaders is narrower than the hype suggests. Audit checkout decline logic before peak season and find out how the payments risk engine currently scores non-human initiators. Structure the product feed so AI agents can parse it. Decide in advance where agent disputes route. None of that requires betting on which agentic commerce protocol survives.
For merchants outside the enterprise tier, Vladyslav Kolodistyi is direct about timing. “Most businesses do not need an agentic commerce strategy this quarter,” he says. “They need to know their payments stack will not silently reject agentic commerce traffic when it arrives. Those are different projects, and only one of them is urgent.”
What makes agentic commerce genuinely new is not that machines can buy things. Automated payments have existed for decades as standing orders and subscription billing. What is new is discretion. AI agents choose, compare and commit inside limits a person set earlier, and the payments industry has no settled vocabulary for delegated intent of that kind. Mastercard’s Verifiable Intent work and the wider Know Your Agent effort are early attempts to build one, and both sit squarely in the payments authorisation layer rather than the checkout interface.
There is a data question that sits before payments entirely. Agentic commerce depends on merchants exposing structured product information that AI agents can read, and Vladyslav Kolodistyi points out that most catalogues fail that test today.
“Agentic commerce does not reward the best marketing copy,” says Vladyslav Kolodistyi. “Agentic commerce rewards the cleanest data. AI agents compare on attributes, and if your attributes are trapped inside an image, AI agents skip you before payments are ever involved.”
Fulfilment logic shifts too. When AI agents place orders at machine speed, returns behave differently and payments reconciliation has to cope with volume patterns no human buyer would produce. Vladyslav Kolodistyi describes agentic commerce demand as spikier than seasonal traffic and harder to forecast, which matters for any payments team sizing capacity.
Consent is the other unsettled question. A mandate signed once can authorise many payments later, and Vladyslav Kolodistyi argues that payments infrastructure needs a durable record of that mandate rather than a checkbox buried in an onboarding flow.
“Agentic commerce will produce disputes about scope,” he says. “AI agents will buy something the customer technically permitted and did not actually want. Payments systems have to store the mandate, not only the transaction, or agentic commerce disputes become unwinnable.”
Merchants who want more of Vladyslav Kolodistyi’s commentary on agentic commerce and payments infrastructure can follow his LinkedIn profile, where he writes on the operational side of AI agents in payments.
The businesses that handle agentic commerce well will not be the ones with the loudest launch. They will be the ones whose payments infrastructure treated AI agents as an authorisation question early, quietly, and before the checkout volume showed up.
