Agentic Commerce in Payments: How AI Agents Could Change Checkout, Risk, and Merchant Control in 2026

Checkout has usually belonged to the merchant. A customer visits the website, compares products, reads the terms, adds an item to the basket, chooses a payment method and completes the purchase inside a journey the merchant can see.

Agentic commerce starts to challenge that pattern.

If an AI agent can help a customer search, compare, select and complete a purchase, part of the commercial journey may move outside the merchant’s usual control zone. That does not make agentic commerce bad. It makes it important.

For merchants, the opportunity is obvious: less friction, more personalised buying journeys and new ways to convert intent into payment. But the control questions are just as important. Who proves customer intent? How are payment credentials protected? What does the merchant see? How are disputes evidenced? Who handles fulfilment, returns and support?

In 2026, the real agentic commerce question is not whether AI can help customers buy. It is whether merchants can keep enough control when AI starts helping customers pay.

Agentic Commerce Moves Checkout Outside the Merchant’s Usual Control Zone

Traditional ecommerce gives merchants a familiar control model. They manage the product page, checkout copy, trust signals, payment options, delivery information, refund policy and customer support prompts. Even if a PSP or wallet sits inside the payment flow, the merchant usually owns the journey around it.

Agentic commerce changes the starting point. Mastercard describes agentic commerce as online or mobile shopping where an AI agent can complete tasks for a user, including searching for items, comparing options and making a purchase with limited or no manual input from that user. Visa’s Intelligent Commerce initiative also focuses on enabling AI agents to transact on behalf of consumers and businesses through APIs, standards and safeguards.

That shift matters because the customer may not begin on the merchant’s website. They may begin inside an AI assistant, a travel planner, a shopping agent, a wallet environment or another interface that filters options before the merchant even appears.

For merchants, this is not only a design issue. It is a control issue.

If the agent displays the wrong availability, misses a shipping condition, misunderstands a refund rule or selects a payment route the merchant would not normally prioritise, the customer may still blame the merchant when the experience breaks.

Agentic commerce may make buying feel simpler for the customer. But it can make visibility more complicated for the merchant.

The Real Payment Question Is Whether the Agent Understands Customer Intent

Payments are built around intent. A customer decides to buy, confirms the amount, accepts the terms, authenticates where required and completes the transaction. Merchants and PSPs use those moments to understand whether the payment is legitimate.

Agentic commerce can make that harder to read.

If an AI agent acts on behalf of the customer, the payment may be valid, but the behaviour may not look like a normal human checkout session. The agent may compare prices at unusual speed, revisit product data repeatedly, use saved preferences or execute instructions based on previous conversations. To a risk system, some of that activity may look closer to automation than ordinary browsing.

A legitimate agent can still look like unusual behaviour

This is where payment teams need a more careful approach. The issue is not that agent-led payments are automatically risky. The issue is that merchants and PSPs need to understand what proves legitimate customer intent when a machine is helping to complete the journey.

OpenAI’s Instant Checkout material says that when someone places an order, ChatGPT sends the necessary details to the merchant backend using the Agentic Commerce Protocol, while the merchant accepts or declines the order, processes the payment through its existing provider and handles fulfilment and customer support. That model keeps the merchant involved, but it also shows why the evidence trail matters.

Merchants may need to know what the customer asked for, what the agent selected, what was shown before payment, and what the customer finally approved. Without that visibility, intent can become harder to defend later.

Merchant Control Becomes Harder When the Interface Changes

The most important agentic commerce issue is not whether AI can complete a purchase. It is whether the merchant still controls the parts of the transaction it remains responsible for.

Even if an AI interface helps the customer buy, the merchant still has to manage price accuracy, stock, fulfilment, returns, customer support, complaints, tax records and dispute evidence. Agentic commerce does not remove those responsibilities. It simply changes the interface through which the customer may reach them.

That is why merchant-of-record control matters. OpenAI says its Instant Checkout model keeps merchants in control of the customer relationship as a merchant of record, including payments, fulfilment, returns and support. That is a meaningful design point because merchants cannot afford to become invisible at the moment they still carry operational responsibility.

For high-risk and regulated sectors, the control question becomes even sharper. A merchant may need to show market restrictions, age or eligibility information, product terms, refund limitations, risk warnings or licence boundaries before a customer commits to payment.

If those details sit outside the agent’s view, the checkout may feel smooth but become operationally fragile. A cleaner front end does not help if the back end cannot prove what happened.

Agentic commerce works best when it gives customers convenience without turning the merchant’s responsibilities into a black box.

Tokenisation and Delegated Credentials Become More Important

Agent-led payments raise an obvious payment security question: how does an AI agent help complete a purchase without exposing sensitive payment credentials or weakening customer control?

This is why tokenisation, delegated credentials and spend controls are becoming important parts of the agentic commerce discussion. Visa says Intelligent Commerce Connect enables secure payment initiation, tokenisation, spend controls and authentication through a single integration via Visa Acceptance Platform. Mastercard describes Agent Pay as infrastructure for secure, scalable and trusted payments in agentic commerce, with its developer documentation positioning Agent Pay as a remote commerce tokenisation programme designed to support agentic commerce.

The practical point for merchants is simple: agentic commerce cannot rely on a vague idea that the AI “has permission”. Payment permission needs to be structured.

A customer may authorise an agent to buy within a certain budget, from certain merchants, under certain conditions or for a specific purpose. The payment infrastructure then needs to support that permission without handing over raw credentials or making the customer lose control.

Merchants should not assume every agentic payment will use the same model. But they should understand the direction of travel. Trusted agent-led commerce depends on safer credential handling, clearer consent and better controls around what an agent is allowed to do.

Fraud Models Will Need to Rethink What “Normal” Behaviour Looks Like

Fraud systems are trained to spot unusual behaviour. That becomes more complicated when legitimate behaviour starts to look more automated.

An AI agent may act faster than a human, compare more options, interact through APIs, revisit product data, or complete checkout with less visible browsing behaviour. Some of that may be perfectly legitimate. Some of it may also be abused by bad actors trying to test credentials, manipulate listings, scrape offers, exploit refund policies or automate account abuse.

Visa has introduced a Trusted Agent Protocol focused on helping merchants distinguish between malicious bots and legitimate AI agents acting on behalf of consumers. That distinction may become one of the most important risk questions in agentic commerce.

The NIST AI Risk Management Framework is not a payments rulebook, but it gives a useful governance lens because it is designed to help organisations incorporate trustworthiness considerations into AI system design, use and evaluation.

For payments, that trustworthiness lens becomes practical. Agentic commerce needs fraud models that can recognise legitimate delegated action without giving suspicious automation a free pass.

High-Risk Merchants May Face a Slower, More Controlled Adoption Path

Not every merchant category will move into agentic commerce at the same speed. Early use cases may be easier in retail, travel planning, everyday shopping or repeat purchase categories where product rules are relatively clear.

High-risk merchants may face a more controlled path. Sectors such as iGaming, Forex, crypto, adult, CBD, subscriptions and travel often involve additional checks around eligibility, market access, age restrictions, customer identity, refund terms, chargeback evidence and regulatory boundaries.

Sensitive categories will need stronger permission and evidence standards

An AI agent buying a pair of shoes is not the same as an AI agent interacting with a regulated financial product, a gambling wallet, an adult platform or a restricted product category. The merchant may need stronger evidence that the customer was eligible, understood the terms and authorised the action clearly.

That does not mean high-risk sectors should ignore agentic commerce. It means they should watch the infrastructure carefully before assuming it fits their flows.

The first question should not be “Can an agent complete the transaction?” The better question is “Can an agent-led journey satisfy the same permission, compliance, risk and evidence standards that the merchant already needs in its normal checkout?”

For high-risk merchants, convenience will not be enough. Agentic commerce will need control strong enough to satisfy payment partners, regulators, customers and dispute processes.

Agentic Commerce Could Move Conversion Upstream

Merchants usually think about conversion at the website or checkout level. They optimise page speed, payment method placement, form design, authentication flow, cart recovery and trust signals.

Agentic commerce may move part of that conversion battle earlier.

If an AI agent compares merchants before the customer sees a website, then the merchant may win or lose before the checkout page loads. Product data, stock information, delivery terms, return policy, pricing accuracy, payment compatibility and support reliability may influence whether the agent recommends the merchant at all.

That creates a different kind of optimisation problem. The merchant’s offer must be readable not only to humans, but also to systems that interpret structured information and decide whether it fits the customer’s instruction.

A merchant with unclear pricing, weak product data, vague shipping rules or hard-to-read return terms may be less attractive in an agent-led environment. The customer may never see the merchant’s carefully designed checkout page because the agent filtered it out earlier.

This is why agentic commerce is not only a payment topic. It is also a data quality, product information and trust topic.

The merchant may lose before checkout if the agent cannot understand, verify or trust the offer.

What Merchants Should Review Before Agent-Led Checkout Scales

Merchants do not need to rebuild their payment strategy overnight. Agentic commerce is still developing, and adoption will vary by market, category and provider. But waiting until agent-led checkout becomes mainstream may be too late.

The practical review should begin with control, not excitement.

• Customer intent evidence: Can the business prove what the customer authorised, what the agent selected and what was approved before payment?
• Payment credential handling: Are tokenisation, consent, authentication and spending controls clear enough for delegated transactions?
• Product and pricing data: Can AI systems read accurate availability, pricing, policy, delivery and return information?
• Fraud monitoring: Can risk tools distinguish trusted agent activity from bot abuse, credential testing or suspicious automation?
• Dispute evidence: Can the merchant prove what was shown, selected, confirmed and fulfilled if the customer later challenges the payment?
• Operational ownership: Who handles fulfilment, customer support, refunds, complaints and post-purchase communication?

These are not futuristic questions. They are payment operations questions.

Agentic commerce may change the interface, but merchants will still need to defend the transaction. That means the evidence layer, support process and payment-control model need to be ready before volume arrives.

The strongest merchants will not be the ones that simply add an AI checkout button. They will be the ones that understand where control could leak and fix those points early.

The Real Opportunity Is Not AI Checkout, But Controlled AI Checkout

Agentic commerce could become an important new layer in payments. It may help customers move from intention to purchase faster, reduce unnecessary browsing, improve personalisation and create new routes to conversion.

But the opportunity is not AI checkout by itself. The opportunity is controlled AI checkout.

That means a merchant can still understand the customer’s instruction, rely on secure credential handling, show accurate terms, manage fulfilment, support the customer after purchase and defend the payment if something goes wrong.

For payment teams, this is the difference between a useful new commerce channel and an operational risk.

AI agents may make buying feel more seamless, but seamless is not the same as controlled. Merchants still need to know what happened, why it happened, who authorised it and what evidence exists.

That is where agentic commerce becomes serious. It is not only a front-end innovation. It is a payment-control challenge.

Conclusion

Agentic commerce may change where checkout begins, who shapes product choice and how quickly customer intent turns into payment. That makes it one of the more important payment developments to watch in 2026.

But merchants should not treat AI-agent checkout as a shortcut to effortless conversion. If the customer journey starts outside the merchant website, control becomes more important, not less.

The winners will be merchants that can keep visibility over customer intent, protect payment credentials, manage fulfilment, support customers and produce evidence when disputes arise.

AI agents may help customers buy. Merchants still need to control what they are selling, how it is paid for and how the transaction is defended.

In agentic commerce, the real advantage is not automation alone. It is automation with control.

---

FAQs