How AI Is Reshaping Cross-Border Payments: From Transaction Routing and Risk Control to Agentic Commerce

Chen Qian starts from everyday payment: tap once while shopping online, scan or tap in a store, and the payment seems done. In those few seconds, though, the transaction actually travels through a complex global payment system. Consumers, merchants, payment methods, and financial institutions in different regions are connected, completing identification, authorization, and risk decisions at high frequency.

AI is starting to take part in this chain, helping platforms decide which path a transaction should take, whether to let it through, and even whether it can execute a transaction on a customer's behalf. Payment is therefore no longer just an action; it is increasingly an intelligent decision system. This episode is a collaboration with Ant International. Chen Qian uses its business and demos to unpack the difficulties of cross-border payments and the changes AI is bringing.

Take buying a flight home as an example. After the buyer clicks pay, the merchant first sends the transaction information to an acquirer. An acquirer can be thought of as the merchant's entrance into the payment network. The episode cites platforms such as Antom, Stripe, and PayPal to show how acquiring and related processing services connect payment methods, choose rails, and identify and control risk during the transaction.

Next, the transaction information enters card networks such as Visa, Mastercard, and UnionPay, then is passed to the issuer behind the consumer's card. The bank verifies the account and payment details, checks the balance or credit limit, assesses transaction risk, and finally decides whether to approve. The authorization result then travels back to the merchant along the path of issuer, card network, and acquirer.

This round may take only a few seconds, but what it mainly completes is authorization: it transmits transaction information; the money has not yet actually entered the merchant's account. Clearing and settlement still have to follow, and a cross-border transaction may also involve currency conversion and cross-border fund movement. The typical window given in the episode is T+1 to T+3 days, and it can take longer, so a successful payment prompt should not be confused with funds arriving.

Behind a single click, several institutions have already completed a round of collaboration. The acquirer takes on the merchant's needs and also connects to the network behind it, which makes it a key node. Its job is not only to send information out, but to make the transaction complete as smoothly as possible in different markets—and those markets are not one naturally interoperable system.

Different countries and regions have different payment habits: North America commonly uses credit cards; many Asian markets prefer e-wallets; some markets rely heavily on bank transfers or their own local networks. Connecting these methods, banks, and networks is itself a complex infrastructure project. A payment setup from one country cannot simply be copied to another.

Cross-currency settlement adds another layer of difficulty. The consumer pays in dollars; the merchant may ultimately receive euros or renminbi, and conversion, exchange rates, and settlement have to be handled in between. The platform also has to weigh success rate, rail costs, risk, and experience at the same time: the path with a higher success rate may be more expensive, the cheaper path may fail more easily, risk controls that are too loose miss fraud, and controls that are too tight wrongly block ordinary consumers.

These factors turn acquiring into a real-time decision problem: which path the transaction should take, whether to let it through, and how to make it faster and cheaper while keeping it safe. When this expands to dozens or even hundreds of markets, complexity keeps rising. In the past this mainly relied on infrastructure, rules, and human experience. Now AI both optimizes existing decisions and, with the arrival of Agents, creates new transaction scenarios.

Keeping a global payment network running has traditionally taken a lot of people. Entering a new market requires research into payment methods and regulation; unusual transactions need manual review; finance teams have to deal with files that do not match; channel outages also depend on on-call staff to locate. As transaction volume grows, the efficiency limits of manual methods become more obvious, so leading payment institutions have started exploring AI operations together.

The episode visits Ant International's annual VOYAGE conference in Shanghai to observe these explorations. Antom, the platform responsible for merchants' cross-border payments, concentrates several capabilities into Antom Autopilot in its demo. This agent is aimed at the merchant's whole experience of using payment services, not just adding a chat window to a page.

Autopilot's coverage extends from onboarding and system integration to transaction risk, chargeback handling, reconciliation, and expansion into new markets. Chen Qian first looks at market research: a company that wants to enter a country has to understand which payment methods local consumers prefer, and how local rules and business conditions affect a payment plan. That information changes choices about onboarding and operations.

The episode says a company researching a new market on its own may need about half a year, and an acquirer's recommendation may take two to four weeks; Antom's demo compressed this process to about five minutes. Using Indonesia as an example, the AI first analyzes the local landscape and consumer habits, then recommends a payment mix based on the merchant type. These time comparisons come from the company case shown in the episode.

After the recommendation comes a simulation step. The proposed setup is first simulated on the checkout page and in the corresponding environment, so possible transaction outcomes can be observed, and only then is it launched. A conference speaker used this to stress that when research costs fall sharply, a team's question may shift from "when can we afford market research" to "which markets should we enter first now."

The second scenario is fraud and chargebacks. The episode mentions Visa's VAMP monitoring requirements, saying the fraud and dispute threshold dropped from 22 to 15 per thousand transactions, to illustrate the constraints merchants and acquirers face. Exceeding the threshold can bring fines, and in serious cases affect rail availability, so risk control directly affects whether a merchant can keep collecting payments smoothly.

Card testing is a typical attack: bots hold large numbers of stolen credit cards and rapidly try payments on a merchant platform, looking for cards that still work. After victims complain to their banks, the merchant's dispute and chargeback metrics can deteriorate suddenly. At the same time, new offerings from AI merchants, such as pay-later and free trials, also attract attacks like bulk registration and quota abuse.

A risk interviewee in the episode stresses that attackers follow changes in the business. Traditional methods often wait for chargeback reports, then rely on people to analyze and locate the problem, which can take a long time. AI's role is to move some of that work forward into continuous monitoring, anomaly analysis, and earlier judgment, so institutions can recognize attack patterns sooner rather than only investigating after losses occur.

After Autopilot finds an anomaly, it can analyze it and generate blocking rules, but new rules are not put into production immediately. It first replays historical transactions to simulate which ones would be blocked and what might happen to the success rate. After launch, the system also evaluates results every 7 days and produces a report, forming a loop of monitoring, validation, launch, and review.

Antom disclosed in the episode that one customer's fraud chargeback rate fell from 1.18% to 0.02%, and that fraud among its merchants overall fell by about 55% on average. After a chargeback occurs, SmartDispute can provide appeal suggestions; the average win rate given in the episode rose from 24.5% to 41.8%. AutoDispute compiles materials and helps with appeals. These are application results as presented by the company.

After a transaction is completed, there is still reconciliation. Different countries, currencies, payment channels, and settlement rules can scatter the same transaction across different batches and files. In one example, a reconciliation involves more than six currencies, and different channels use T+2 or T+7 cycles. Finance teams not only have to find mismatches, but also explain why they happened.

The episode says traditional manual reconciliation may take 48 to 72 hours. Antom not only automates matching, but also tries to have AI analyze the reasons for anomalies and generate reports; processing time in the demo was compressed to about 5 minutes. The value here is not only finding the difference faster, but connecting "what happened" with "why it happened."

Payments are also a system that cannot stop. A cross-border network connects a large number of channels and partners, and a failure in any link can affect transaction volume. Using a Southeast Asian channel as an example, the episode shows that discovering and locating a failure by hand can delay handling. AI can continuously monitor channels, judge the scope and cause of an impact, and then propose options such as shutting a channel down, shifting traffic, or adjusting routing.

From market research to risk, reconciliation, and operations, the first layer of change is making complex, repetitive work more automated, and shortening the distance between finding a problem and dealing with it. But when AI is no longer only assisting people, and instead takes part in transactions itself, the participants, authorization methods, and risks that payment systems face also change.

That is Agentic Commerce. In the past, shopping required consumers themselves to search, browse, compare prices, collect coupons, place orders, and pay. Now a user can first express an intent, such as finding a white dress suitable for hosting a tech summit. An Agent then carries out the later steps. Mainstream payment platforms are therefore starting to build around these new scenarios.

The change is not just adding an AI shopping assistant. It is a shift from "people looking for products" to "an Agent understanding people, then helping complete the transaction." Shopping entry points will change with it. Independent storefronts used to rely on users finding products and making choices themselves, plus human sales and support turning traffic into orders. An Agent may instead ask about needs, recommend, and offer advice.

After recommending, an Agent can also guide checkout, generate an order, pay with the user's authorization, and continue with after-sales service. But small and mid-size merchants usually cannot build and maintain such a system on their own. Continuous operation needs model access, a development team, servers, and compute, and every conversation, search, and tool call also incurs Token and inference costs.

Payment companies are therefore trying to offer one-stop products. The Antom Shopping Agent introduced in the episode can string visiting, guidance, ordering, payment, and after-sales into a flow and embed it in a merchant site. For merchants, the value is lowering the barrier to building and maintaining this. For payment platforms, the service also starts extending to before the traditional payment step.

Marketing is likewise shifting from broadcasting widely to matching intent. The episode cites customer-acquisition costs up 15%, traffic conversion down 20%, and coupon operations taking more than 48 hours with a final redemption rate of only 1.87%, to illustrate the efficiency problems of the traditional "buy ads, send coupons, wait for spending" approach. A+ Rewards explores putting coupons, cashback, and subscription discounts into digital-wallet settings that consumers already use day to day.

The episode says A+ Rewards covers the ten largest markets in Asia-Pacific and more than 21 digital payment platforms. When a user tells an Agent they want coffee, the AI can look for suitable offers, recommend them, and help redeem. The long chain from seeing an ad, collecting a coupon, and coming back to spend is compressed into expressing a need, matching an offer, and completing a transaction. It may also help merchants reach users who had not visited them on their own.

Whether AI is optimizing existing payments or creating new scenarios, more and more judgment and execution are being handed to systems. Higher efficiency depends on the system being able to identify risk accurately. The episode then introduces Antom's 3-in-1 Transformer, which tries to analyze different types of data in a single model, rather than looking at only one side of a transaction or of user behavior.

A single dimension may see a local picture and miss the links behind a transaction. Antom puts three kinds of information together: sequence data observes a user's past behavior, graph data analyzes the network of connections among devices, accounts, and funds, and tabular data contains a large number of specialized risk features used to judge whether the current transaction fits a normal payment pattern.

These three kinds of data answer, respectively, "how the user used to behave," "who they are connected to," and "whether this transaction itself looks unusual." In the episode's explanation, combining them can form a more complete risk picture. A risk interviewee further says that combining multimodal data lets the team understand payment risk from subtle relationships and make more precise judgments.

After identifying risk, controls still have to be executed. The episode introduces Antom Shield's three-layer mechanism: the first layer is payment protection, handling stolen cards, fraud, and chargebacks; the second is anti-abuse, aimed at bulk registration, account farming, and abuse of free quotas; the third is agent protection, handling new risks that appear after Agents join transactions. The underlying models and execution mechanisms play different roles.

The episode also introduces AgentSafePay risk protection. Antom's statement is that merchants who use both Shopping Agent and its acquiring service can receive corresponding compensation when they encounter agent-transaction risks that meet the protection conditions. "Pay with 100% confidence—we'll cover it" is the company's phrasing within this scope, and it should not be understood apart from the required services and the specific conditions of use.

Global merchants need more than collecting money. A company may operate in the United States, Europe, and Southeast Asia at the same time, receive multiple currencies, and still have to settle, convert FX, pay suppliers and employees, and manage funds across countries and accounts in one place. Payment platforms are therefore extending into corporate accounts, cross-border payouts, transfers, and liquidity management, hoping to let funds move more flexibly with the business.

The episode divides Ant International's business into four parts. Alipay+ is closer to consumers, connecting wallets, banking apps, and other payment methods in different countries so users can pay at overseas merchants with tools they already know. Antom is closer to merchants, helping companies connect payment methods and providing related digital services, taking on the work of connecting merchants to the global payment network.

According to the episode, Antom covers more than 200 payment markets, more than 300 payment methods, and more than 140 currencies. After a transaction ends, WorldFirst (万里汇) connects accounts, funds, and FX, helping companies collect and pay through multi-currency accounts, convert currencies, and centrally manage and dispatch funds across different entities and markets.

Speakers in conference clips also described coverage and arrival speed for corporate payouts: coverage of more than 200 countries, 25 currencies that can arrive instantly, 85 currencies that can be T+0, and coverage of more than 60 digital wallets. The episode uses these capabilities to show that once accounts and fund flows are connected, companies do not have to handle global collections, employee payouts, settlement, and conversion in processes that are cut off from one another.

Bettr embeds financial services into operating and spending scenarios, including corporate credit, cash management, FX, and liquidity. The episode gives examples: AliExpress merchants can apply in the backend for operating funds to restock or cover short-term working capital, and consumers may also complete installments through buy now, pay later.

From consumer payment and merchant collection, to corporate fund flows, and on to financing and FX, links that used to be scattered are placed in one ecosystem. Payments are gradually becoming infrastructure that connects consumers, merchants, and corporate funds. To see how these capabilities enter a real industry, the episode turns to travel, where price, inventory, and fulfillment all have to work together in real time.

A cross-border airfare connects different countries, currencies, and payment methods, as well as participants such as airlines and agencies. Prices, inventory, and user choices keep changing, and AI has sped up search, comparison, and booking. A travel interviewee notes that distribution structures are changing on one side, while on the other, AI raising front-end efficiency may also enlarge existing B2B fulfillment bottlenecks.

It is easier for users to have AI compare flights and give purchase advice, but if sales structure and fulfillment capacity do not keep up, a faster front end will not remove back-end problems. Air tickets are especially exposed to exchange rates: one price when comparing, another when actually paying, and possibly a different one again at refund, so the same ticket ends up with "three prices."

When consumers hand comparison and purchase decisions to AI, whether a quoted price can be honored in later steps becomes a new requirement. The episode splits the problem into the ticket's own volatility and cross-border exchange-rate volatility. The payment system needs to keep quotes, payment, and refunds as consistent as possible, reducing the gap between the price a user sees and the amount that finally occurs.

Antom introduces dynamic FX lock in the episode, using the forecasting ability of the Falcon TST model to lock the exchange rate at quote time and carry it through payment and refund, with a maximum duration of one year given by the company. The aim is to turn "three prices" into a consistent price as far as possible, lowering FX impact for merchants and raising the chance that a price given by AI is honored in a real transaction.

If search and decision-making are already done, but the transaction fails at the last payment step, time saved at the front end can still be wasted. Airfare prices and seat inventory keep changing; if payment fails and the user has to start over, the original seat or price may already be gone. Payment has to be safe and also reduce unnecessary interruptions, so finer judgments are needed between risk and conversion.

The episode introduces intelligent layered verification: reduce unnecessary checks for long-trusted users, and strengthen identity checks for high-risk transactions; when a payment times out or fails, intelligent retries then try to finish before prices and inventory change. Antom says this application raised the payment success rate for flight bookings by 3.8 percentage points; the figure still belongs to a scenario result disclosed by the company.

In travel, a combination that looks unusual is not necessarily fraud. Someone may live in country A, depart from country B, and pay with a credit card from country C. Parents buy tickets for children, assistants book for bosses, and the payer and passenger can be different people. Looking at only one dimension makes it easy to block by mistake; risk judgment has to take in the full set of relationships behind the trip.

The episode describes the system combining device, network, location, the relationship between payer and passenger, historical behavior, and other information to judge whether the combination of person, card, device, and journey is reasonable. Antom claims, with a Hong Kong Airlines case, 99.99% risk coverage and a reduction of more than one million dollars in fraud losses. These metrics should be understood according to the interviewed company's case definition.

After a ticket is sold, funds still have to flow among participants such as airlines, OTAs, and agencies. Traditional supply chains rely on fixed credit lines, deposits, and batch settlement, but orders are becoming more real-time while funds may sit in the old settlement cycle. Antom's approach is to connect participants with corporate accounts so funds move with orders, reducing how much working capital deposits tie up.

The account system introduced in the episode covers more than 100 countries and local-currency accounts in more than 40 markets, and includes built-in credit capability. A travel interviewee says service has to be built around challenges in customers' operations, including risk, fare volatility, and dynamic pricing. Payment services are no longer only responsible for settlement after a transaction is done; they have to work in sync with orders and actual operations.

Returning to the larger picture, AI is moving payment systems from pre-set rules and manual handling toward more real-time analysis and adjustment. A system can judge risk, choose rails, and optimize after it finds problems. At the same time, the boundary of the payment platform is also expanding: it no longer appears only after a consumer has decided what to buy, and is starting to take part in earlier discovery, marketing, and conversion.

Before a transaction, the platform connects comparison, product selection, and offers; after a transaction, it continues into settlement, FX, and cash management. The episode uses this to explain the meaning of VOYAGE's theme, Attention Is All You Need: on one side, payment decisions becoming more intelligent; on the other, a repositioning of the payment platform's role in transactions and operations.

Antom organizes the related capabilities as foundation models, Autopilot, an AI-native platform, and full-stack solutions, covering the payment lifecycle. Chen Qian also raises a practical question: global merchants need local service, and Asia is also an important source of cross-border e-commerce and overseas expansion. Whether the company has a complete service team on the ground affects whether merchants can get a timely response when problems arise.

An interviewee in the episode explains that Antom is short for Ant to Merchant, and that the starting point is accompanying merchants as they grow; the service is not limited to acquiring and payouts. Its stated goal is also not only to take share from other acquirers, but to help merchants expand their business and revenue through finer operating capabilities, so the overall transaction market keeps growing.

Chen Qian concludes that the problem payments face is expanding from "getting the money in" to "how a transaction should happen and how it should be managed." What AI brings is not only a technical upgrade; it also moves payments from the last step of a transaction into the whole process. The episode closes on this change, keeping the links among merchant operations, platform services, and transaction decisions.