文 | 硅谷101
“支付”这件事情，正在变得越来越便捷。无论是网上购物、在线点单，还是线下消费，消费者只需要点击、扫码或碰一碰，就可以完成支付操作。但在这短短几秒钟内，背后实际上运转着一套复杂的支付系统。
在全球范围内，这类交易每分钟发生数百万次，其中不少涉及跨境流动。不同地区的消费者、商户，以及不同的支付方式和金融机构被连接起来，快速完成识别、授权和风险判断。看似简单的付款操作，实际上牵动着整条支付链路。
如今，AI也开始进入这条链路，帮助支付平台实时判断交易应该选择哪条路径、是否放行，甚至能否直接替客户完成交易。支付正在由一个简单的交易动作，逐渐转变为一套智能决策系统。
那么，全球跨境支付的链路如何运作，面临哪些难点？AI的加入正在带来哪些变化，行业巨头又进行了哪些实践？本文将结合蚂蚁国际的实践，探讨跨境支付的运行机制，以及它们如何为AI时代做准备。
从点击付款到最终完成，一笔跨境支付背后需要经过多个环节。
▍跨境支付的复杂性
以购买国际机票为例：在航空公司网站选好机票和支付方式，点击“付款”之后，一笔交易就开始了。
首先，商户向收单机构发起支付请求，并发送交易信息。收单机构是商户进入支付网络的入口，Antom、Stripe、PayPal等支付平台，在实际业务中也会提供收单或相关支付处理服务。它们负责连接不同支付方式、选择合适的支付通道，并识别和控制交易风险。
随后，交易信息进入Visa、Mastercard、银联等卡组织网络。卡组织承担信息转接的角色，将交易信息从商户一侧传递至消费者一侧的发卡机构，也就是消费者银行卡所属的银行。
银行验证账户和支付信息，检查余额或信用额度，同时评估交易风险，决定是否批准交易，并返回授权结果。结果再经由卡组织、收单机构，最终传回商户，这一信息传递和授权过程可能只需要几秒钟。
需要注意的是，此时完成的主要是交易授权，传递的是交易信息，资金尚未真正进入商户账户。实际的资金划转发生在后续的清算和结算环节：资金从发卡机构一侧，经支付网络和收单机构，最终结算至商户账户。如果涉及跨境交易，还可能包含货币兑换和跨境资金流动，整个过程通常需要T+1至T+3天，甚至更久。
因此，一次简单的付款操作背后，商户、收单机构、卡组织和发卡机构已经完成了复杂的信息传递与授权。其中，收单机构是承前启后的关键节点，要确保每笔交易顺利完成并不容易。
▍跨境支付的复杂性
对于希望开展全球业务的商户而言，它面对的并不是一个统一的支付市场。不同国家和地区的支付习惯存在明显差异：北美以信用卡为主，亚洲很多地区更习惯使用电子钱包，另一些市场则较为依赖银行转账，甚至拥有独立的本地支付网络。
这些体系之间并不天然互通，因此，连接不同国家的支付方式、银行与支付网络，本身就是一项复杂的基础设施工程。
另一个难点是跨币种与跨境结算。消费者使用美元付款，而商户最终收到的可能是欧元或人民币，其中需要跨越不同的金融体系，处理汇率、货币兑换和资金结算等问题。
此外，收单机构不仅需要保证交易成功，还必须兼顾成本、风险和消费者的支付体验。例如，一条支付通道的成功率较高，但成本也更高；另一条通道的成本较低，却更容易失败。风险控制同样需要权衡：审核过于宽松可能放过欺诈交易，过于严格又可能误伤正常消费者。
收单机构因此面临复杂的实时决策：选择哪条支付路径、是否放行，以及如何在速度、成本与安全之间取得平衡。当业务从一个国家扩展到几十乃至上百个市场，这些问题的复杂程度还会成倍增加。
过去，这些难题主要依靠支付机构的基础设施、预设规则和人工经验来解决。近几年，AI成为支付链路中的新变量，其影响主要体现在两个层面：
一方面，AI正在优化原有支付链路中的交易决策，提高支付效率与智能化水平；另一方面，随着Agent参与商品搜索、下单，乃至自主完成支付，新的交易方式和支付场景也带来了此前没有遇到的问题。
高度复杂的全球支付网络，长期以来需要大量人工工作来维持运转。进入新市场前，需要调研当地支付方式和监管要求；交易出现异常时，需要人工排查；不同市场的账目不一致，需要财务人员核对；支付渠道发生故障，也需要人工发现、定位并决定处理方式。
随着跨境交易规模增长，这些工作的效率压力不断增加，单纯依赖人力的方式已经难以满足需求。因此，AI成为支付行业降本增效的共同探索方向。头部支付机构开始将AI应用于交易决策、风险控制和运营等环节。
在蚂蚁国际于上海举办的VOYAGE2026上，也展示了支付行业利用AI解决相关问题的具体实践。
Antom是蚂蚁国际旗下主要负责商户跨境支付业务的平台，自2023年开始为全球商户提供支付服务。在此次Antom的演示中，Antom Autopilot智能体体现了AI对支付服务流程的改造：从商户入驻和系统集成，到交易风控、拒付处理、对账及新市场拓展，AI已经参与支付服务的多个环节。
我们通过以下几个典型场景，来说明这个智能体的具体运作方式。
首先是新市场研究。企业进入一个新市场之前，需要先去理解“当地人到底怎么支付”这件事，例如了解当地消费者偏好信用卡还是电子钱包，以及当地的监管制度等。过去，企业自行完成这类调研可能需要约6个月，收单机构提供建议也至少需要2至4周。
通过Antom Autopilot，这一过程可以缩短至5分钟。例如，一家企业准备进入印尼市场，AI会先分析当地支付格局和用户习惯，再根据不同商户类型的业务特点，个性化推荐支付组合。
以前很多系统到“推荐”这一步就结束了，Antom Autopilot还增加了仿真环节：它可以将推荐的支付方案放到结账页和生产环境中提前测试，观察真实交易可能出现的结果，确认后再一键上线。
据Antom称，当市场调研的时间从两周缩短至五分钟，研究成本也随之大幅降低，企业的决策重点因此发生变化。过去，企业需要考虑何时才能负担得起特定市场的调研成本；如今，则可以将更多精力放在判断哪些市场值得立即进入。
第二个难点是安全。在许多海外市场，信用卡仍是普遍的支付方式，而欺诈与拒付，也是信用卡交易中最让商户头疼的问题。
Visa不久前进一步收紧了“收单行监控计划（VAMP）”：按照新的标准，每1000笔交易中，欺诈和争议交易的门槛从过去的22笔降至15笔。一旦超过门槛后，商户和收单机构可能面临罚款，严重时还可能影响支付通道。
与此同时，欺诈攻击手段不断演变，识别和拦截难度也在增加。例如，“盗卡测试”如今已成为普遍问题：攻击者利用机器人持大量盗取的信用卡信息，在商户平台反复尝试支付，以确认哪些卡片可以使用，短时间内便可能发起上万次支付操作。遭遇盗刷的用户向银行申诉后，店铺的投诉率就迅速上升，拒付率甚至可能增长50%。
Leo Li
Antom风控负责人
除传统的欺诈和拒付问题外，攻击者也在不断转向新的业务场景。以AI商户为例，越来越多的企业推出后付费、免费试用期等服务吸引用户，而这类业务过程中也出现了不少新的风险攻击，尤其是“薅羊毛”等滥用行为。
面对这类问题，传统风控通常以事后处理为主：等待拒付报告出现，再由人工定位原因，整个过程可能需要2周到4个月。引入AI后，风控可以转向实时监测与提前预判。
以Antom Autopilot为例，AI持续监测交易，发现异常后自动分析并生成拦截规则。这些规则不会立即上线，而是先通过历史交易“回放”，模拟可能拦截的交易及其对支付成功率的影响。规则上线后，系统每7天自动评估风控效果、生成报告，形成“实时监测—模拟验证—上线—持续复盘”的闭环。
这套系统的效果也非常显著，Antom表示，一家客户的欺诈拒付率从1.18%降至0.02%，从Antom整体商户情况来看，欺诈平均下降约55%。
即便拒付已经发生，AI仍可参与后续争议处理。例如，Antom SmartDispute会提供申诉建议，将平均争议胜诉率从24.5%提升至41.8%。Antom AutoDispute则可以自动整理材料、协助提交申诉，减少人工工作量。
交易完成后，商户还需要面对繁琐的对账工作。跨境交易涉及不同的国家、支付方式和结算规则，例如，在东南亚、欧洲等市场，一次对账可能涉及6种以上货币，不同渠道的结算周期可能分别为T+2或T+7。同一笔交易因结算批次不同，甚至可能出现在不同文件中。传统人工对账一个周期可能需要48至72小时，发现差异后还要继续查找原因。
这也是AI开始发挥作用的地方，Antom除了可以将人工对账自动化，还让AI进一步去解释异常。在找出问题之后，还接着会判断背后的原因是什么，并生成对应的报告，整个流程可以压缩至5分钟。
除上述环节外，支付系统还需要持续稳定运行。跨境支付连接大量渠道与合作机构，任何环节发生故障都可能造成损失。
在东南亚，一些支付渠道每月可能出现1至3次故障。以前这类事情很大程度上，还是靠人工值班和排查，等问题被定位时，可能已经损失了超过5%的交易量。因此，7×24小时的系统运维同样重要。
AI可以持续监测全球支付渠道，发现异常后自动判断影响范围和原因，并进一步提出关停、切换流量、智能路由等处理方案。过去需要数小时甚至数天的检测与处理过程，如今可以缩短至分钟级。
从交易前的市场研究，到交易中的风控，再到交易后的对账以及支付系统的实时运维，AI首先改变的，是支付背后复杂、重复的工作。
但这只是第一步，当AI开始从“帮人把支付做好”，进一步变成“自己参与交易”，支付面对的问题也会完全不一样。
近年来，我们越来越多的听到一个新词：Agentic Commerce（智能体商务）。传统线上购物通常由消费者自行搜索商品、浏览网页、比较价格、领取优惠，最后完成付款。
随着Agent加入，这样的流程就被彻底颠覆了。消费者只需要向AI表达购买意图，搜索、比价乃至付款，便都可以由Agent完成，主流支付平台也正在围绕这一场景加快布局。
这件事带来的变化，并非只是增加了一个AI导购。过去的电商链路以“人找商品”为主，未来可能转向“Agent理解人、再替人完成交易”。当交易参与者发生变化，原有的销售和支付链路也会随之改变。
首先受到影响的是购物入口。过去，独立站能否将流量转化为订单，很大程度上取决于用户能否找到并选定商品，以及商户是否拥有足够的销售与客服人员。
在Agent时代，用户进入网站之后，智能体会主动打招呼，并精准地根据需求去推荐商品、提供建议。接着它还可以继续引导下单、生成订单，并在用户的授权下完成支付。交易完成后，甚至还可以提供售后服务。
但对大量中小企业而言，自行搭建这样的Agent并不容易。Agent要持续运行，背后需要接入模型、技术开发并持续维护，还需要准备云服务器和算力。Agent每次与用户对话、搜索或调用工具，也都会产生Token消耗和推理成本。无论是技术团队、经验还是成本，都可能构成商户的门槛。
因此，主流支付公司都在积极研发一站式的Agent产品。例如，Antom推出的Antom Shopping Agent，可以把访问、导购、下单、支付到售后串成一条完整流程，直接嵌入商户的网站。
除了改变消费者“怎么买”，Agent也在改变商户“怎么卖”。以前的营销通常依靠投放广告、发优惠券，等待用户触达、领取并最终消费，效率存在局限。
研究数据显示，平台获客成本上涨了15%，流量转化却下降了20%。传统优惠券从发放到实际使用，还可能需要超过48小时以上的运营支持，最终核销率仅为1.87%。
随着AI的进入，营销思路开始由“广撒网”，转向“做匹配”，Antom在A+ Rewards上的探索便是一个例子。
A+ Rewards并非单纯的优惠券工具，而是覆盖亚太10大市场、连接超过21个数字支付平台，让商户可以把优惠券、返现、订阅折扣等营销活动，直接放进用户日常使用的钱包场景里。
例如，用户向Agent提出“今天想喝杯咖啡”，AI可以识别意图，再从A+ Rewards中寻找相关优惠并进行推荐，再协助用户完成核销。
这样一来，原本“看到广告、领取优惠、回来消费”的营销链路，由此被压缩为“表达需求、匹配优惠、完成交易”。对商户而言，这既可能提高转化率，也可能触达此前未主动访问过的新用户。
无论是AI进入原有支付链路，还是创造新的交易场景，背后都有一个共同趋势：越来越多的“判断”与“执行”任务正在交由AI，安全问题也因此变得更加复杂。
前文提到，Antom Autopilot等产品正在提高风控效率，但这一能力首先依赖于对风险的准确识别。为此，Antom推出了Antom 3-in-1 Transformer安全模型。它最大的特点，是将三类不同的数据纳入同一个Transformer模型进行分析。
过去，不少风控模型往往只关注某一种数据维度，例如一笔交易本身的特征或用户行为。但在Antom看来，单一维度很容易看到局部，却可能忽略交易背后的关联。一笔交易有没有风险，往往同时包含多个层面的信息。
这三类数据分别是：
第一，序列数据，用于判断当前交易是否符合用户过去的行为轨迹；
第二，图数据，用于分析设备、账户和资金之间是否存在可疑的关联网络；
第三，表格数据，涵盖上万个专业风险维度，用于判断交易本身是否符合正常支付模式。
它们分别回答三个问题：这个用户过去怎么做？他和谁有关联？这笔交易本身有没有异常？将这些信息结合起来，有助于AI识别更加完整的风险图景。
Leo Li
Antom风控负责人
Antom 3-in-1 Transformer将多种类型的数据更好地结合起来，帮助我们从细微之处更深入地理解风险与支付之间的关系，从而更精准地识别支付过程中存在的风险。
底层模型提供了风险判断，但风险被识别后，还需要落实到防控上。为此，Antom进一步建立了三重防护机制Antom Shield，分别对应AI支付正在面对的三类不同风险。
第一层是支付防护，解决传统的盗卡、欺诈和拒付问题；
第二层是反滥用防护，针对批量注册、设备养号、薅免费额度等新型风险；
第三层是智能体防护，解决Agent开始参与交易之后产生的新风险。
由此，底层模型负责识别风险，Antom Shield负责实施防护，二者结合起来，共同构成AI支付的安全基础。
即便是AI在交易过程中出现了错误，商户的损失也是完全可控的。对此，Antom推出了AgentSafePay风险保障机制。据Antom介绍，使用Antom Shopping Agent及Antom收单服务的商户，如果发生符合保障条件的智能体交易风险，可获得其所称“100%敢付敢赔”的保障。
除了具体支付流程的变化，全球化企业对支付平台的需求也正在扩大。
目前，对于许多全球化企业而言，支付不再只是让客户完成付款。一家企业可能同时在美国、欧洲和东南亚开展业务，每天收到不同币种的款项，随后还需要进行结算、换汇，向供应商和员工付款，并统一管理分散在不同国家和账户中的资金。
因此，除了支付之外，企业还希望资金能够更快、更灵活地在不同账户和市场之间流动。这也是为什么，近几年主流支付平台都在快速行动，开始从单纯处理交易，逐渐延伸到企业账户、跨境付款、资金调拨和流动性管理。
蚂蚁国际的业务布局也反映了这一趋势，目前，整个蚂蚁国际的体系分为四大板块。
其中，Alipay+更靠近消费者一侧，主要连接不同国家和地区的钱包、银行App等支付方式，让消费者可以使用自己熟悉的支付工具，在海外商户完成支付。
Antom则更靠近商户一侧，帮助企业接入不同市场的支付方式，并提供相关数字化服务。目前，Antom覆盖200多个支付市场、300多种支付方式和140多种货币。
交易完成后，资金仍需要继续流动，这便涉及WorldFirst（万里汇）的业务。WorldFirst连接账户、资金和外汇，帮助企业利用多币种账户开展全球收付款及换汇。对于业务更加复杂的企业，还可以对不同实体、账户和市场中的资金进行集中管理与调度。
Carl Huang
万里汇Enterprise总经理
我们的覆盖范围已经超过200个国家，25种货币可以即时到账，85种货币始终可以做到T+0，还可以覆盖60多种数字钱包。
如果将款项支付至员工账户、数字钱包账户，或者以这25种货币进行支付，都可以实现即时到账。
当支付、账户与资金流得到连接后，金融服务也具备了融入企业经营场景的基础。Bettr提供的正是这一层金融服务，包括企业信贷、资金管理、外汇和流动性等。
例如，AliExpress上的中小商户需要补货或短期资金周转时，可以直接在商家后台申请营运资金贷款，无需另行寻找融资渠道。消费者购物时，也可以通过“先买后付”等方式分期付款。
也就是说，金融服务不再是独立存在的产品，可以直接嵌入企业经营和消费者交易之中。从消费者付款、商户收款，到企业资金流动，最后延伸至融资、外汇等服务，原本分散在不同系统中的环节，正在被同一个生态包裹起来。支付也由此成为连接消费者、商户、企业资金和金融服务的一层基础设施。
这些产品与能力进入具体行业后，可以如何发挥作用？航旅业务提供了一个典型案例。
一张跨境机票背后，连接着不同国家、币种和支付方式，也涉及航空公司、代理商等多个参与方。
同时，航旅交易对实时性要求很高，价格、库存和用户决策不断变化。随着AI参与旅行搜索、比价和预订，支付环节的效率问题也更加突出。
Apple Gao
Antom航旅解决方案负责人
行业的结构性变化主要体现在两个层面：一是机票销售结构发生变化；二是最近AI带来的效率提升，颠覆性地改变了整个销售链路。
AI工具给用户带来了很大的便捷性，但如果整个销售分销的结构没有变的话，原本服务的履约是整个B to B销售中最关键的bottleneck（瓶颈），也可能被AI进一步放大。
▍AI比价体验
如今，借助AI，机票比价与购买正在变得更加简单。用户只需提出需求，AI就可以快速比较不同航班与价格，甚至直接给出购买建议。
但AI看到的价格不一定就是最终支付价格，除机票价格本身会变动，跨境支付还会受到汇率波动的影响。AI比价时、实际支付时和后续退款时，汇率都可能不同，同一张机票因此可能出现“三个价格”。
当用户把比价与购买决策交给AI，价格能否保持一致就成为新的问题。支付系统需要尽可能确保AI展示的价格，在后面的交易过程中尽可能保持一致。
Antom采用的一种方式是动态锁汇：通过Falcon TST模型的预测能力，在报价时就锁定当时的汇率，并将该汇率延续至后续支付和退款环节，最长可锁定一年。
这样，原本可能出现的“三个价格”，可以尽可能变成一个价格。对用户而言，价格更加透明；对商户而言，可以降低汇率波动带来的影响；对AI而言，其提供的报价也更容易在真实交易中兑现。
▍不让支付挡住转化
即使搜索、比价和购买决策已经完成，如果最后卡在支付上，前面的效率提升仍可能落空，航旅交易尤其如此。机票价格和库存实时变化，支付一旦失败，座位可能很快被释放；消费者重新支付时，原有价格和座位也可能已经无法获得。
所以，当AI将购买决策缩短至几秒钟，支付环节也需要兼顾速度与安全，尽可能减少对购买流程的打断。
其中一种方式是智能分层验证：对长期可信的用户减少不必要的验证，对高风险交易则实施更加严格的身份验证。如果发生支付超时或失败，系统还可以进行智能重试，尽可能在价格和库存变化之前完成交易。
Antom表示，在实际应用中，这种方法让整体机票预订支付成功率，提升了3.8个百分点。
▍AI智能风控
当AI开始自动参与风控，大量交易可能会被系统拦截，但航旅支付的特殊之处在于：一笔看似“不正常”的交易，并不一定存在欺诈。
例如，一个人住在A国，却从B国出发，使用C国发行的信用卡购买机票；或者父母替孩子购票、助理为老板订票时，付款人和乘客也并非同一人。
如果AI只分析某一个维度，就容易将这些正常交易误判为风险。因此，AI时代的旅行风控，不能只关注单笔交易，还需要识别交易背后的完整关系，这也正是前文提及的Antom 3-in-1 Transformer的应用场景。
该模型把设备、网络、地理位置、付款人与乘客的关系以及历史行为等信息，放在一起分析，判断这个人、这张卡、这个设备和这次旅行组合起来是否合理，再实时作出风险判断。
据称，在香港航空的实际应用中，该系统实现了99.99%的风险覆盖率，并帮助减少了超过100万美元的欺诈损失。
▍资金的实时流动
支付生态的重要性，也在航旅这个行业中非常凸显。
一张机票售出后，资金还需要在航空公司、OTA和代理商等参与方之间继续流动。传统旅行供应链往往依赖固定授信、保证金和批量结算。而当AI带来的订单越来越实时，原有模式便可能难以适应：订单可能在持续产生，资金却还停留在原来的账期和结算流程里。
Antom的思路是通过企业账户连接不同参与方，使资金随着订单流动，并减少保证金造成的资金占用。
目前，这套账户体系覆盖100多个国家、40多个市场的本地货币账户，并同时提供内置授信能力。这意味着，旅行支付从“先交易、后结算”，逐步转向“资金跟着交易流动”。
Apple Gao
Antom航旅解决方案负责人
随着整个行业发生结构性变化，航旅客户在业务经营中会面临许多挑战。
因此，我们在建设相关能力时，会把围绕航旅客户经营需求的能力单独提取出来。无论是帮助客户降低风险、提升风控效率，还是围绕票价波动提供动态定价（dynamic pricing）能力，这些都会成为我们提供服务时的指引。
整体来看，支付的作用现在已经不仅仅“收款”这么简单了。
一方面，随着AI进入支付体系，支付系统本身变得越来越自主和智能。过去，支付系统主要靠规则驱动：什么交易该拦、走哪条支付渠道、出了问题怎么处理，都需要提前设定规则或者人工介入。现在，AI可以实时分析交易、判断风险、选择更合适的支付路径，甚至在发现问题后自动调整和优化。
另一方面，支付平台的边界也在扩大。以前我们理解“支付”，通常出现在交易的最后一步，即用户选好商品、决定购买并完成付款。如今随着AI开始参与比价、选品、寻找优惠，乃至直接完成购买，支付开始更早地进入交易过程，与导购、营销和转化等环节相连接。交易完成后，支付平台的服务还在向结算、换汇和资金管理延伸。
蚂蚁国际此次VOYAGE2026的主题是“Attention Is All You Need”，其中的“Attention”，对应着AI进入支付领域的两个方向：支付的智能化，以及支付平台角色的重新定位。
目前，Antom围绕这两个方向构建了产品矩阵，包括一个“三合一”支付基础模型Antom 3-in-1 Transformer、一个AI智能体Antom Autopilot、一个AI原生平台，以及一套全栈AI解决方案，覆盖企业的整个支付生命周期。
此外，做全球生意还有一个很现实的问题：本地服务。很多国际支付公司的核心团队位于欧美，而全球跨境电商和出海业务中，有相当一部分商户来自亚洲。
Antom扎根亚洲，在中国及亚洲地区拥有更完整的服务团队，因此能够更及时地响应和处理商户的需求。
Apple Gao
Antom航旅解决方案负责人
Antom是“Ant to merchant”的缩写，由此也可以看出，我们一开始成立之初，就是希望陪伴商户共同成长的。因此，也意味着我们提供的服务不仅限于收单和支付。
我们的目标并不是抢占其他收单机构的市场份额，如果整个市场没有增长，对于商户来说，他们的经营规模和增长空间也不会扩大。所以我们的目标一定是在服务好商户的基础之上，围绕着他的经营的精细化操作和未来的整个收入的增长，去帮他业务变得更大。
所以，支付正在发生的变化，不只是技术的升级。它要解决的问题，正在从“把钱收进来”，到现在需要进一步去回答：“这笔交易，应该怎么发生、怎么管理？”
这意味着，支付的边界也在被重新定义：它正在从交易的最后一步，走向整个交易过程。而这，可能也是AI时代带给支付的全新想象空间。
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Silicate Valley 101
The matter of “payment” is becoming easier and easier. Whether it is online shopping, online billing, or underline consumption, consumers need to click, scan or touch to complete payment operations. But in these short seconds, a complex payment system actually operates behind it.
Globally, such transactions occur millions of times a minute, many of which involve cross-border movements. Consumers, merchants and different payment modalities and financial institutions in different regions are connected to quickly complete identification, authorization and risk judgement. It appears to be a simple payment operation, which in effect leads to the entire payment chain.
Today, AI has also begun to enter this chain, helping the payment platform to determine in real time which path the transaction should choose, whether it should be released or even whether it can be done directly for the client. Payments are being gradually transformed into a system of intelligent decision-making by a simple transaction.
So how does the global cross-border payment chain work and what are the difficulties? What changes are being brought about by AI ' s accession and what are the practices of industry giants? This paper will explore operational mechanisms for cross-border payments in the context of ants International practice and how they prepare for the AI era.
A cross-border payment would need to go through multiple links from the point of impact to the time of its finalization.
Complexity of cross-border payments
In the case of the purchase of international air tickets, for example, a transaction began after a “payment” click was made on the airline's website to select the air travel and payment modalities.
First, the business has initiated a payment request from the receiving agency and sent transaction information. The billing agency is the entry point for businesses to payment networks, and payment platforms such as Antom, Stripe, PayPal and, in practice, will also provide receipt or related payment processing services. They are responsible for connecting different modes of payment, selecting the appropriate payment route and identifying and controlling transaction risks.
Subsequently, the transaction information entered the network of card organizations such as Visa, Mastercard, and Union. Cards assume the role of trans-linking information from the business side to the issuing agency on the consumer side, the bank to which the consumer bank card belongs.
Bank validates account and payment information, checks balance or credit quota, assesses the risk of the transaction, decides whether to approve the transaction and returns the results of the authorization. The result is finally sent back to the business through card organizations and billing agencies, a process that may take only a few seconds.
It is important to note that at this time, it is mainly the authorization of transactions, which conveys information on transactions, and that funds are not yet actually entered into the business accounts. The actual transfer took place at the subsequent liquidation and settlement stage: the funds were settled from the issuing agency side, through payment networks and billing agencies, to the commercial account. If cross-border transactions are involved, currency exchange and cross-border financial flows may also be involved, and the process usually takes T+1 to T+3 days, or even longer.
Thus, behind a simple payment exercise, the merchants, billing agencies, card organizations and issuing agencies have completed complex information transmission and authorization. Among them, the receiving agency is the key node of the initial launch, and it is not easy to ensure that each transaction is successfully completed.
Complexity of cross-border payments
For those who wish to do global business, it does not face a single payment market. Disparities in payment practices are evident in different countries and regions: North America is dominated by credit cards, and many parts of Asia are more accustomed to using electronic wallets, while others rely more on bank transfers and even have independent local payment networks.
These systems are not natural interconnectivity, and therefore the payment modalities, banks and payment networks linking different countries are in themselves a complex infrastructure project.
Another difficulty is cross-currency and cross-border settlement. Consumers pay in United States dollars, and business households may eventually receive euros or renminbis, which need to cut across different financial systems to deal with exchange rates, currency exchange and financial settlements.
In addition, billing agencies need to not only ensure the success of transactions, but also balance costs, risks and consumer payment experiences. For example, success rate, a payment channel, is higher but more costly; the other channel is less costly and more likely to fail. Risk control also needs to be weighed: over-lax auditing may spare fraudulent transactions and may be too strict to harm normal consumers.
The receiving agency thus faces complex real-time decision-making: which path to pay, whether to release and how to balance speed, cost and security. The complexity of these issues will multiply as operations expand from one country to dozens or hundreds of markets.
In the past, these difficulties were largely addressed by the infrastructure of payment agencies, pre-set rules and manual experience. In recent years, AI has become a new variable in the payment chain, the impact of which has been mainly at two levels:
On the one hand, AI is optimizing the decision-making of transactions in the original payment chain and improving the efficiency and intelligence of payments; on the other hand, with Agent's involvement in the search for goods, the filing of bills and even the self-completed payment, new trading patterns and payment scenarios also pose problems that have not previously been encountered.
A highly complex global payment network has long required considerable manual work to sustain operations. Before entering the new market, local payment modalities and regulatory requirements need to be investigated; when transactions occur, manual checking is required; the accounts of different markets are inconsistent and require financial staff to reconcile; and the failure of the payment channel requires manual detection, positioning and decision-making.
As the scale of cross-border transactions has grown, the pressure for efficiency in these efforts has increased, and the need has become difficult to meet by relying solely on manpower. As a result, AI became a common explorer for downside efficiency gains in the payment industry. The head payer began to apply AI to the transactional decision-making, risk control and operations.
At Voyage 2026, held in Shanghai by Ants International, concrete practices in the payment industry using AI to address related issues were also demonstrated.
Antom is the main platform for cross-border payments by businesses under the international flag of ants and has been providing payment services to businesses worldwide since 2023. In this demonstration of Antom, Antoine AutopilotZ42Z represents the transformation of AI ' s payment service processes: from business establishment and system integration, to transaction control, non-payment processing, reconciliation and new market expansion, AI has been involved in various aspects of payment services.
We are following several typical scenarios to illustrate how this AI agent works.
The first is new market research. Before an enterprise enters a new market, it needs to understand “how the local people pay”, for example, whether local consumers prefer credit cards or electronic wallets, and local regulatory systems. In the past, it may take about six months for an enterprise to complete such research on its own and at least two to four weeks for the receiving agency to provide advice.
Through Antom Autopilot, the process can be shortened to five minutes. For example, an enterprise is prepared to enter the Indonesian market, and AI will analyse local payment patterns and user habits before recommending a personalized payment portfolio based on business characteristics of different business types.
Many of the previous systems had ended at the point of “recommended” and anatom Autopilot had added a simulation: it could place the recommended payment option on the closing page and in the production environment for pre-testing, observing the likely outcome of a real transaction and confirming that it was then online.
According to Antom, when the time for market research was reduced from two weeks to five minutes, the cost of research was reduced significantly, resulting in a change in the focus of business decision-making. In the past, enterprises needed to consider when they could afford the research costs of particular markets; today, more could be done to determine which markets deserved immediate access.
The second difficulty is security. Credit cards continue to be a widespread form of payment in many overseas markets, and fraud and non-payment are among the most problematic of credit card transactions.
Visa had recently tightened the "VAMP" further: the threshold for fraudulent and controversial transactions had been reduced from 22 to 15 per 1,000 transactions according to the new criteria. Once the threshold has been exceeded, businesses and billing agencies may face fines, which may also affect payment routes in serious cases.
At the same time, the means of fraudulent attacks have evolved and the difficulty of identifying and intercepting has increased. For example, “ticking card testing” has now become a common problem: using robots to hold large amounts of stolen credit card information, the attackers have repeatedly tried to pay on commercial platforms to identify which cards are available and may launch tens of thousands of payment operations in a short period of time. Following complaints to the bank by users of the brushes, the rate of complaints from shops has risen rapidly, and the rate of rejection may even increase by 50 per cent.
Leo Li
Antom Wind Control Officer
In addition to the traditional issues of fraud and non-payment, the attackers are moving towards a new business landscape. By way of example, the introduction of services such as paid and free probationary periods in a growing number of businesses has attracted users, and new risk attacks have emerged in the course of such operations, particularly abuse such as “grab freebies”.
In the face of such problems, traditional wind management is usually based on ex post facto treatment: waiting for a report of non-payment to appear before it is manually identified, the process may take between two weeks and four months. With the introduction of AI, wind controls can shift to real-time monitoring and early prediction.
In the case of Antom Autopilot, for example, AI continuously monitors transactions and automatically analyzes and generates interception rules after detection of anomalies. These rules will not be immediately online, but will first be “released” through historical transactions to simulate transactions that may be intercepted and their impact on the payment of success rate. After the rules go online, the system automatically assesses the wind-controlled effects every seven days, generates reports and forms a closed loop for "real-time monitoring-simulation-up-on-line-continuous disc ".
The effectiveness of the system was also significant, with Antom indicating that the rate of fraudulent rejection by a customer had dropped from 1.18 per cent to 0.02 per cent, and that, in the case of Antom as a whole, fraud had fallen on average by about 55 per cent.
Even if a refusal of payment had occurred, AI could have been involved in subsequent disputes. Antom SmartDispute, for example, will provide recommendations for an increase in the average dispute success rate from 24.5 per cent to 41.8 per cent. Antom AutoDispute, on the other hand, is able to automate materials and assist in the submission of complaints and reduce the manual workload.
When the transaction is completed, the business will also have to face cumbersome reconciliations. Cross-border transactions involve different countries, payment modalities and settlement rules, for example, in markets such as South-East Asia, Europe, where a single reconciliation may involve more than six currencies, and different channels may have a settlement cycle of T+2 or T+7. The same transaction may even appear in different documents due to the fact that the settlement instalment is different. Traditional manual reconciliations may take between 48 and 72 hours for a cycle and continue to search for reasons when discrepancies are identified.
This is also where AI began to function, and besides automating manual reconciliations, Antom allowed AI to further explain anomalies. Once the problem has been identified, it is then decided what is behind it and the corresponding reports are generated, and the process can be condensed to five minutes.
In addition to the above-mentioned elements, the payment system will need to operate steadily and continuously. Cross-border payments are connected to a large number of channels and cooperating agencies, and any chain failure can cause loss.
In South-East Asia, a number of payment channels may fail one or three times a month. In the past, this type of thing had been located to a large extent on manual duty and routing, which may have lost more than 5 per cent of transactions. Therefore, a 7x24-hour system is equally important.
AI can continuously monitor global payment channels, identify anomalies and automatically determine the scope and causes of impact, and further propose options for shutdowns, switching flows, intelligent routing, etc. The detection and treatment process, which used to take hours or even days, can now be reduced to minutes.
From pre-trading market research to the control of transactions, to post-trading reconciliations and real-time operation of payment systems, AI first changed the complex and duplicative work behind payments.
But this is only the first step, and when AI starts from “helping people to make good” to “doing business on their own”, the problems with payment will be completely different.
In recent years, we have heard more and more of a new word: Agency Commerce. Traditional online purchases usually involve consumers searching for commodities, browsing web pages, comparing prices, receiving concessions and finalizing payments.
With the addition of Agent, the process was completely subverted. All consumers need to express to AI their intent to purchase, search, price and even pay can be done by Agent, and mainstream payment platforms are accelerating around the scene.
The change brought about by this is not just the addition of an AI lead purchase. In the past, the electrical chain was dominated by “persons looking for goods” and the future may turn to “Agent understands people and finish deals for others”. When transaction participants change, the original distribution and payment chain changes.
First and foremost, it is the shopping entrance that is affected. In the past, the ability of stand-alone stations to convert traffic into orders depended to a large extent on the ability of users to find and select commodities and the availability of sufficient sales and passenger staff.
In the Age of Agent, when users enter the site, AI agent takes the initiative to greet them, and to recommend and advise them precisely on the basis of demand. It can then continue to direct the orders, generate orders and complete payments with the authorization of the user. After the transaction has been completed, after-sale services may even be provided.
However, for a large number of SMEs, it was not easy to build such an Agent on their own. Agent needs access to models, technology development and continuous maintenance for continuous operation, as well as the preparation of cloud servers and algorithms. Every time Agent talks to users, searchs or calls tools, it also incurs token consumption and reasoning costs. Whether it is a technical team, experience or cost, it may constitute a threshold for business.
As a result, mainstream payment companies are actively developing the all-in-one Agent product. For example, the Antom Shoping Agent, launched by Antom, can provide a complete process of access, direct purchase, billing, and payment for sale, directly embedded in the vendor ' s website.
Apart from changing the consumer's “how to buy”, Agent is also changing the business's “how to sell”. Former marketing usually relies on advertising, coupons, waiting for users to reach, receive and eventually consume, and has limited efficiency.
Research data show that the cost of the Platform ' s take-off has increased by 15 per cent, while the flow transformation has fallen by 20 per cent. Traditional coupons, from issuance to actual use, may also require more than 48 hours of operational support, with a final write-off rate of 1.87 per cent.
With the entry of AI, marketing ideas began to shift from “widespreading” to “matching”, as illustrated by the Antom exploration on A+ Renewals.
A+Rewards is not a mere coupon tool, but rather covers the 10 major markets in Asia and the Pacific, connecting more than 21 digital payment platforms, allowing businesses to place vouchers, refunds, subscriptions, etc., directly in the wallet scene that users use daily.
For example, the user proposed to Agent “to have a cup of coffee today”, and AI could identify the intent, then look for and recommend the benefits from A+ Renewals and then assist the user in completing the write-off.
As a result, the marketing chain, which “sees advertising, receiving concessions, returning to consumption”, was reduced to “expressing needs, matching preferences, and completing transactions”. For businesses, this may increase conversion rates or reach new users who have not previously been actively visited.
Whether AI enters the old payment chain or creates new trading scenes, there is a common trend behind it: more and more “judgment” and “execution” tasks are being handed over to AI and security issues are becoming more complex.
As mentioned earlier, products such as Antom Autopilot are improving the efficiency of wind control, but this capacity depends first and foremost on the accurate identification of risks. To that end, Antom launched the Antom 3-in-1 Transformer security model. Its greatest feature is the integration of three different types of data into the analysis of the same Transformer model.
In the past, a number of wind-control models tended to focus on a particular data dimension, such as the characteristics of a transaction itself or user behaviour. In Antom's view, however, a single dimension can easily be seen in local areas, while the connection behind the transaction may be ignored. A transaction is not risky and often contains information at multiple levels.
The three types of data are:
First, serial data used to determine whether current transactions correspond to the user ' s past course of behaviour;
Second, figure data to analyse whether there is a suspicious connection between equipment, accounts and funds;
Third, tabular data, covering tens of thousands of professional risk dimensions, are used to determine whether the transaction itself is in conformity with normal payment patterns.
They answered three questions: what did the user do in the past? Who's he related to? Is there anything unusual about the deal? Combining this information helps AI to identify a more complete risk profile.
Leo Li
Antom Wind Control Officer
Antom 3-in-1 Transformer better combines a variety of types of data, helping us to understand the relationship between risk and payment in a more nuanced way, and thus to identify more precisely the risks in the payment process.
The bottom model provides a risk judgement, but once the risk has been identified, it needs to be applied to prevention. To this end, Antom further established a triple protection mechanism, Antom Shield, which corresponds to three different types of risks that AI is facing.
The first level is payment protection against traditional card theft, fraud and non-payment;
The second level is anti-abuse protection, which addresses new types of risks such as mass registration, equipment numbering, 薅 free quota;
The third level is AI agent protection, which addresses new risks arising from Agent ' s involvement in the transaction.
As a result, the bottom model is responsible for risk identification and Antom Shield for protection, which together form the safety basis for AI payments.
Even if AI had made a mistake in the course of the transaction, the business loss was entirely manageable. In response, Antom introduced the AgentSafePay risk safeguard mechanism. According to Antom, a business using Antom Shoping Agent and Antom's billing services is entitled to what it claims to be “100 per cent of the risk of a AI agent transaction that meets the safeguards”.
In addition to changes in specific payment processes, the demand for payment platforms by globalized enterprises is expanding.
At present, for many globalized enterprises, payments are no longer simply to be made by customers. An enterprise may operate simultaneously in the United States, Europe and South-East Asia and receive funds in different currencies on a daily basis, followed by settlements, swaps, payments to suppliers and employees and the harmonization of the management of funds dispersed across countries and accounts.
Thus, in addition to payments, enterprises also want to see faster and more flexible financial flows between different accounts and markets. This is why, in recent years, mainstream payment platforms have been moving quickly, starting from the mere processing of transactions to the business accounts, cross-border payments, fund transfers and liquidity management.
The operational layout of ants International also reflects this trend, and at present the whole ants international system is divided into four main blocks.
Among them, Alipay+, which is closer to the consumer side, is linked mainly to different countries and regions ' wallets, bank Apps, etc., so that consumers can use their familiar means of payment to make payments abroad.
Antom, on the other hand, is closer to the business side, helping businesses to access different markets and to provide related digital services. At present, Antom covers over 200 payment markets, over 300 payment modes and over 140 currencies.
Once the transaction has been completed, there is still a need to continue the flow of funds, which involves the operations of WorldFirst. WorldFirst connects accounts, funds and foreign exchange to help businesses use multi-currency accounts for global payments and exchange of remittances. For more complex businesses, central management and movement of funds in different entities, accounts and markets is also possible.
Carl Huang
General Manager of Enterprise.
Our coverage has exceeded 200 countries, 25 currencies are on call, 85 are always on T+0 and over 60 digital wallets.
Immediate payments can be made if they are made to an employee ' s account, a digital wallet account or in these 25 currencies.
When payments, accounts and financial flows are linked, financial services also have the basis for integration into the business landscape. It is at this level of financial services provided by Bettr, including business credit, financial management, foreign exchange and liquidity.
For example, when small and medium-sized businesses on AliExpress need replenishment or short-term liquidity, they can apply for operating capital loans directly from the back office without having to find alternative sources of financing. Consumers may also be paid in instalments, such as “buy first” or “pay first”.
In other words, financial services are no longer independent products and can be embedded directly in business and consumer transactions. From consumer payments, business receipts, to business capital flows, to services such as finance, foreign exchange and so forth, which were scattered in different systems and are being wrapped up in the same ecology. Payments thus become an infrastructure linking consumers, businesses, business capital and financial services.
How can these products and capacities work when they enter specific industries? The flight service provides a typical case.
A cross-border airline ticket is linked to different countries, currencies and modes of payment and involves multiple participants, such as airlines, agents and others.
At the same time, the real-time nature of airline transactions is high and prices, inventories and user decisions are changing. The efficiency of the payment chain has become more prominent as AI has been involved in travel search, price and booking.
Apple Gao
Antom Flight Solutions Manager
The structural changes in the industry are mainly at two levels: first, a change in the structure of air ticket sales; and second, a recent increase in efficiency brought about by AI, which has revolutionized the whole distribution chain.
The AI tool provides users with a great deal of ease, but if the overall distribution structure remains the same, performance of the original service is the most critical bottleneck (bottleneck) in the overall B to B sales and may be further amplified by AI.
It's a price-matching experience.
Today, through AI, the price of air tickets is becoming simpler to purchase. Users need only submit their needs so that AI can quickly compare different flights with prices and even direct purchase advice.
However, prices observed by AI may not necessarily be the final payment, and cross-border payments will be affected by exchange rate fluctuations, in addition to the price of the airfare itself. The exchange rate may vary between AI prices, actual payments and subsequent refunds, resulting in a “three price” for the same ticket.
The question of consistency of prices becomes new when users turn to AI for price and purchase decisions. The payment system needs to ensure, to the extent possible, that the prices displayed by AI are as consistent as possible in the subsequent course of transactions.
Antom uses a dynamic lock-in approach: through the forecasting capability of the Falcon TST model, the exchange rate at the time was locked at the time of the offer and the exchange rate continued to the subsequent payment and refund chain for up to one year.
In this way, the “three prices” that would otherwise have been possible could have been converted to one price as far as possible. For users, prices are more transparent; for merchants, the impact of exchange rate fluctuations can be reduced; and for AI, the offers it provides are easier to cash in real transactions.
I don't want the money blocking the conversion.
Even when search, price and purchase decisions have been completed, previous efficiency gains, especially in airline transactions, may be frustrated if the final card is on payment. Air ticket prices and stocks change in real time, and the seat may be released soon if the payment fails; the original price and seat may be unavailable when the consumer is repaid.
As a result, when AI reduces purchasing decisions to a few seconds, the payment chain also needs to balance speed and security and minimize disruptions to the purchase process.
One such approach is smart stratification: less unnecessary authentication for long-term, credible users, and more stringent identification for high-risk transactions. In the event of payment of timeout or failure, the system can also perform smart retry, where possible, to complete the transaction before the price and inventory changes.
Antom states that, in practical application, this method resulted in an increase of 3.8 percentage points in the amount of success rate for the overall airfare.
▍Ai Smart Wind Control
When AI begins to participate automatically in wind control, a large number of transactions may be blocked by the system, but the special feature of the airline payments is that a transaction that appears to be “unusual” does not necessarily involve fraud.
For example, a person who lives in State A but from State B uses a credit card issued by State C to purchase a ticket; or a parent who buys a ticket for the child and an assistant orders a ticket for the boss is not the same person, either.
If AI analyses only one dimension, it is easy to miscalculate these normal transactions as risk. As a result, the AI-era travel control cannot focus only on single transactions, but also on identifying the whole relationship behind them, which is also the application of Antoine 3-in-1 Transformer mentioned above.
The model brings together information on equipment, networks, geographical location, payer-passenger relations and historical behaviour to determine whether the person, the card, the equipment and the trip together are reasonable and make a real-time risk judgement.
It was stated that in the practical application of Hong Kong aviation, the system achieved 99.99 per cent risk coverage and helped to reduce fraud losses in excess of $1 million.
Real-time flows of funds
The importance of the ecology of payments has also been highlighted in the tourism industry.
Following the sale of an airline ticket, funds will also need to continue to flow between the participants, such as airlines, OTA and agents. Traditional travel supply chains often rely on fixed letters of credit, bonds and bulk settlements. As the orders brought by AI become more and more real time, it may be difficult to adapt to the old model: the orders may continue to arise, but the funds remain in the original account period and settlement process.
The idea of Antom is to connect the different participants through the business account, to allow the funds to flow with the order and to reduce the amount of money spent on the bond.
Currently, the system covers local currency accounts in more than 100 countries and 40 markets, while providing built-in creditworthiness. This means that travel payments move from “dealing first, settling later” to “money following transaction flows”.
Apple Gao
Antom Flight Solutions Manager
As the industry as a whole undergoes structural changes, airline customers face many challenges in their operations.
As a result, when we build the relevant capacity, we will be able to single out the capacity around the business needs of the airline ' s clients. Whether it helps clients reduce risks, improves wind efficiency, or provides dynamic pricing around ticket price fluctuations, these will guide our service delivery.
Overall, the effect of payments is now more than simply “receipts”.
On the one hand, as AI enters the payment system, the payment system itself becomes increasingly autonomous and intelligent. In the past, payment systems had been largely rule-driven: what transactions should be stopped, which channels of payment should be pursued, and how problems should be dealt with required pre-set rules or manual intervention. Now, AI can analyse transactions in real time, determine risks, choose a more appropriate payment path, and even automatically adjust and optimize when problems are identified.
On the other hand, the boundaries of payment platforms are also expanding. Previously, we understood “payment”, usually at the last step of the transaction, where the user chose the commodity, decided to buy and complete the payment. Now, as AI begins to participate in price-price, selection, search for a preference, or even directly completes the purchase, the payment begins to enter the transaction process earlier, linked to the chain of conduct, marketing and conversion. After the completion of the transaction, the services of the payment platform are being extended to settlement, exchange and fund management.
The theme of Voyage2026 this time is “Attention Is All You Need”, which corresponds to two directions for AI to enter the payment field: the smartness of payments and the repositioning of payment platform roles.
At present, Antom has built its product matrix around these two directions, including a three-in-one payment base model, Antom 3-in-1 Transformer, an AZAI agentAntom Autopilot, an AAI original platform, and a full-house AI solution covering the entire payment life cycle of the enterprise.
In addition, there is a real problem with global business: local services. The core team of many international payment companies is located in Europe and the United States, while a significant proportion of global cross-border electricians and offshore businesses come from Asia.
Antom has taken root in Asia, with more complete service teams in China and Asia, and is thus able to respond and respond to business needs in a more timely manner.
Apple Gao
Antom Flight Solutions Manager
Antom is the abbreviation of "Ant to Mercant", which also shows that we wanted to be with the business community from the very beginning. Therefore, it also means that our services are not limited to receipt and payment.
Our goal is not to capture the market share of other billing institutions, and if the entire market does not grow, the size of their operations and the space for growth will not expand for the business community. So our goal must be to make his business bigger, on the basis of good business, around the finer operation of his business and the future growth of his entire income.
So, the changes that are taking place are not just technological upgrades. The problem it is going to solve is from “receiving money” to the need for a further answer: “How should this deal happen and how should it be managed?”
This means that the boundary of payment is also being redefined: It is moving from the last step of the transaction to the entire course of the transaction. And this is probably the new imagination that the AI era brought to pay.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.