AI is moving from helping people find products to helping them act: comparing options, assembling a basket, choosing delivery and, in some emerging journeys, initiating payment. This shift is often called agentic commerce. It is a significant opportunity for merchants—but it also changes a basic question at checkout: when an AI agent takes an action, how does a business know that the action is legitimate, authorised and within the customer’s instructions?
 
For merchants, the answer cannot simply be “the card worked.” A safe agentic-payment journey needs evidence across the full chain: a recognisable agent, a verified customer mandate, a payment instrument that does not expose raw credentials, transaction limits that the agent cannot exceed, and records that support customer service or dispute handling later.
 
This is why payment authorisation and AI fraud controls are moving to the centre of the conversation. Merchants do not need to rush into untested autonomous checkout. They do need a practical framework for deciding what an AI agent may do, how consent is captured, when a transaction must be stepped up, and when a human should intervene.
 
Quick answer: Agentic commerce is commerce in which an AI agent helps a customer discover, decide on or complete a purchase. Merchant readiness depends on four foundations: verifiable agent identity, explicit and bounded customer consent, secure payment credentials and continuous fraud monitoring—not on a single “approve” moment.

What is agentic commerce—and why is it different from ordinary e-commerce?

An AI shopping assistant can be more active than a search bar or a chatbot. It may interpret a shopper’s preferences, compare multiple merchants, narrow a selection, apply a budget and carry a purchase flow forward. Payment networks and commerce platforms are beginning to build the infrastructure for these experiences. Visa describes the merchant requirement as recognising legitimate agents, allowing them to pay safely and verifying that they act with customer approval. Mastercard similarly highlights registered agents, verifiable intent and authenticated user consent as core elements of agent-led payment journeys.
 
The difference is delegation. In conventional e-commerce, a shopper normally reviews a cart and presses “pay” themselves. In agentic commerce, a customer might permit an agent to take a constrained action on their behalf—for example: “reorder this item every month up to HK$500,” or “book a refundable hotel under this budget.” The merchant must be able to distinguish a genuinely delegated instruction from a bot, an abused account or an agent acting outside its mandate.
 
That makes agentic commerce a payment-design and risk-design issue, not only an AI issue.

Why payment authorisation needs a new model

In an agent-led purchase, authentication and authorisation still matter, but they mean more than a successful login or a valid card. A resilient model should answer five questions:
 
  1. Who is the agent? The merchant or payment ecosystem needs a trustworthy way to recognise the agent or platform making the request.
  2. Who is the customer? The relationship between customer, account and payment credential must remain clear without exposing sensitive details unnecessarily.
  3. What exactly did the customer permit? Record the mandate: merchant or category, maximum amount, currency, time window, frequency, delivery conditions and any exceptions.
  4. Is this transaction inside those limits? The payment request should be evaluated against the mandate in real time—not only at the start of a relationship.
  5. What evidence can be retrieved later? Consent, agent identity, transaction context and decision records should be accessible for service, investigations and dispute handling.
In practice, this points towards scoped and revocable permissions, rather than giving an agent unrestricted access to a customer’s account or payment details. Stripe’s published agentic-commerce approach, for example, describes tokens limited by business, time or amount and revocable by the customer. The precise technical standard will vary as the market evolves; the merchant principle is stable: delegation should be specific, enforceable and auditable.

The AI fraud risks merchants should plan for

AI can make a buying journey more convenient, but it also changes the fraud signals that a merchant sees. Traditional systems often check whether a person and payment credential look legitimate at one point in time. Agentic commerce introduces new questions: is the agent itself legitimate, and is it continuing to behave within the customer’s intent?

1. Agent impersonation and bot camouflage

Not every automated request is a trusted AI agent. Malicious bots may imitate agent traffic, scrape catalogue information, test credentials or bypass a merchant’s intended customer experience. Merchants need a policy for identifying trusted integrations and for applying different controls to unknown or high-risk automated traffic.

2. Mandate and intent mismatch

An agent can technically satisfy a narrow instruction yet still make a purchase the customer did not reasonably expect. A travel agent that selects the cheapest fare but ignores a long stopover, or a shopping agent that combines promotions in an unintended way, creates customer-service and dispute risk. The remedy is not only better model prompting: it is clear transaction rules, confirmation thresholds and a retrievable record of the customer’s mandate.

3. Promotion, inventory and pricing abuse

Agents can monitor stock, vouchers and price changes continuously. That speed may be useful for customers, but it can also magnify promotion stacking, limited-inventory capture and automated testing of discounts. Merchant risk teams should define velocity limits, promotion eligibility controls, inventory rules and anomaly alerts before enabling agent-led purchase paths.

4. Account takeover and credential misuse

An AI layer does not remove familiar payment risks. Stolen credentials, compromised customer accounts and card testing remain relevant. In fact, agents can make bad activity look more “normal” if merchants do not retain enough context about the agent, the customer’s consent and the transaction pattern.

5. Weak evidence when a dispute occurs

A successful payment authorisation alone may not explain whether an agent followed a customer’s instructions. Merchants should plan to retain a privacy-conscious evidence trail showing the mandate, material constraints, agent identity signals, key customer confirmations and the payment decision—subject to applicable law and data-retention policies.

A merchant control framework for safer agentic payments

The following framework is a practical way to prepare. It is intentionally technology-neutral: merchants can use it whether they are assessing a new AI sales channel, designing an internal assistant or speaking with payment and fraud partners.

1. Keep payment credentials out of the agent’s hands

Do not treat an AI agent as a place to store or expose raw card details. Use payment-provider and network-supported credential protection, tokenisation or delegated-payment mechanisms where available, and maintain card-data safeguards. Payment Asia’s is PCI DSS certified and is designed to help merchants process online payments through a secure gateway environment.

2. Verify the agent, not just the customer

Create a clear admission policy for automated actors. Which agent platforms, applications or partners can access product and checkout functions? What technical identity signals or signed requests are required? What rate, scope and catalogue permissions apply? Unknown automation should not receive the same access as a recognised, monitored integration.

3. Make consent explicit, granular and easy to revoke

A customer mandate should be understandable to a person and enforceable by systems. Set meaningful boundaries such as:
  • Maximum order value and currency;
  • Approved merchants, product categories or SKU exclusions;
  • Start and expiry time;
  • One-time, recurring or capped frequency;
  • Delivery, refund or cancellation conditions; and
  • Transactions that require a customer confirmation before payment.
The customer should be able to see and revoke these permissions. If the instruction is vague or the order is materially different, route it to a confirmation step rather than allowing the agent to guess.

4. Monitor behaviour continuously

Agent-driven traffic may behave differently from a human shopper. Look for abnormal request velocity, unusual product views, repeated payment attempts, rapid switching between accounts or inconsistent location/device signals. Combine these with conventional payment signals and business rules. Good fraud controls aim to reduce both losses and unnecessary declines; they must be calibrated to the merchant’s products, customer base and dispute tolerance.

5. Separate discovery, checkout and fulfilment permissions

An agent may need access to a product catalogue without needing permission to purchase. It may be able to build a basket but need confirmation before payment. It may be allowed to create a payment link but not change a delivery address. Breaking a journey into permissions limits the impact of a compromised account or an errant agent.

6. Keep human override points for high-impact actions

High-value orders, cross-border delivery changes, unusual repeat purchases or transactions that conflict with a customer’s established pattern should have a clear escalation route. Human review is not a failure of automation; it is an essential safeguard when the cost of a wrong action is high.

7. Prepare service and dispute evidence from day one

Document the experience you want customers and support teams to have when a question arises. Can the merchant show what was authorised, when it was authorised, which constraints applied and why the order passed or was stopped? Build an evidence model with privacy, security and retention requirements in mind.

A readiness checklist for merchants

Before opening an agent-led checkout or integrating with a new AI commerce channel, ask these questions:
  • Do we know which automated agents can access our catalogue and checkout?
  • Can we distinguish customer authentication from an agent’s delegated authority?
  • Can a customer set, view and revoke limits for price, frequency, merchant and time?
  • Are payment credentials tokenised or otherwise protected from unnecessary exposure?
  • What happens if the order differs from the customer’s instruction?
  • Do fraud rules evaluate agent-specific behaviour as well as conventional transaction risk?
  • Are promotions, vouchers, inventory and delivery changes protected from automated abuse?
  • Can support teams retrieve an appropriate audit trail if a customer disputes an agent-led order?
  • Who owns the escalation path across commerce, payments, fraud, compliance and customer service?
If any answer is unclear, start with a limited use case rather than a fully autonomous buying experience. For example, an agent can begin by helping customers discover products or prepare a cart, while the customer still confirms payment. Measure approval quality, false declines, complaints, chargebacks and fulfilment exceptions before expanding its authority.

Build the payment foundation before the autonomous journey

Agentic commerce is emerging quickly, but merchants can act now without making exaggerated technology claims. The first step is to make checkout, payment acceptance, merchant operations and fraud governance more reliable for every digital customer—human or AI-assisted.
 
Payment Asia supports this foundation through:
 
  • : online payment acceptance for Visa, Mastercard, UnionPay, Alipay, WeChat Pay, FPS, Octopus and more, with API documentation, a merchant back office and flexible settlement options.
  • : local collection and disbursement options designed for supported markets across Asia, Africa and Latin America—useful when cross-border customers expect familiar local payment experiences.
  • : secure, shareable payment links for merchant-defined amounts, customer-defined amounts and recurring payments, with real-time notifications.
These products should be assessed against a merchant’s own use case, technical architecture, compliance obligations and risk appetite. A payment gateway is a critical part of a secure commerce stack; it does not by itself make a merchant “agentic-commerce ready.” The control design around consent, identity, fraud operations and customer support remains essential.

The next move: controlled, customer-first experimentation

The merchants most likely to benefit from agentic commerce will not be those that simply automate the most steps. They will be the ones that earn trust: making customer intent clear, using secure payment mechanisms, setting guardrails that are visible to customers and responding well when something does not look right.
If your business is improving online checkout, expanding payment methods or preparing a safer digital-payment foundation for future AI-assisted commerce, start with a merchant setup conversation.
Apply for a Payment Asia Merchant Account

For a specific agentic-commerce integration, Payment Asia’s technical and compliance teams should validate scope, supported capabilities and required controls before launch.

Frequently asked questions

What is agentic commerce?

Agentic commerce is a buying journey in which an AI agent helps a customer discover, compare, decide on or, where authorised, complete a purchase. The customer’s intent and control should remain central even when the agent performs actions.

They can be designed more safely when merchants combine verified agent identity, customer authentication, explicit delegated consent, scoped payment credentials, real-time risk monitoring and clear dispute processes. No single control is enough on its own.

Authentication establishes who is accessing an account. Payment authorisation determines whether a specific transaction is permitted. In an agentic journey, merchants also need to verify whether the agent has authority to act and whether the requested payment stays inside the customer’s mandate.

The traditional risks remain, including account takeover and stolen credentials. Agentic commerce adds risks such as agent impersonation, mandate mismatch, promotion abuse and limited evidence of what a customer actually permitted. Merchant controls must therefore evaluate behaviour and delegated authority as well as the payment instrument.

PCI DSS is an important security foundation for protecting cardholder data, but it is only one part of readiness. Merchants also need governance for agent identity, customer consent, transaction boundaries, fraud monitoring and customer support.

Payment Asia provides online payment, alternative payment-method and payment-link solutions that can support secure digital commerce foundations. Any proposed agentic-payment use case should be reviewed with Payment Asia’s technical and compliance teams to confirm scope, supported capabilities, security controls and launch requirements before implementation.

Editorial sources and further reading