VGS Scales Agentic Tokens as AI Shoppers Disrupt Merchant Checkouts

AI购物者冲击商家结账，VGS扩展智能体代币应用 代理支付变通方案破坏了争议解决与商家订阅机制，迫使支付行业在自主AI智能体能够安全进行规模化交易之前部署开放代币标准。 Terry Angelos与Alex Rampell探讨了AI智能体从被动对话者转变为配备支付凭据的主动经济参与者的关键演变。他们分析了数字结账的技术摩擦，对比了专有中介解决方案与新兴的开放智能体代币标准，后者可在模型出现幻觉或错误时保留至关重要的卡组织退款权益。讨论详细剖析了在寻求转化率的包容型商家与为保护人类注意力而拦截机器人流量的亚马逊等广告驱动型巨头之间形成的战略分歧。最后，他们概述了VGS等基础设施提供商正如何从单方凭据保险库演变为动态的多方清算所，在银行、商家与自主智能体之间安全地撮合交易。 支付凭据的出现将AI智能体从被动旁观者转变为现实世界的参与者，引发了数字钱包提供商争相与自主智能体整合的竞争。 (0:00) 每年约有三分之一的支付卡因到期、网络切换或欺诈而发生更换，这使得中立代币保险库对接入多家处理机构的企业商家而言必不可少。 (6:07) 在智能体代币采购中使用一次性虚拟卡变通方案破坏了退款等核心消费者保护权益，并阻碍了商家的订阅追加销售，促使人们采用开放智能体代币标准。 (19:42) 商家正分化为两类：优先考虑程序化交易量的包容型平台，以及像亚马逊这样拦截无头智能体以保护广告收入的对抗型平台。 (35:43) 无处不在的AI智能体通过优化刷卡奖励路由以及自动取消未使用的订阅和健身房会员资格，对传统的沉没成本/未兑现（breakage）商业模式构成威胁。 (43:35)

Autonomous artificial intelligence agents are transitioning from passive advisors into economic actors as they gain access to payment credentials, igniting a competitive rush among digital wallets and forcing payment networks to rethink token standards and dispute rules. Giving AI Arms and Legs via Credentials (9:55) According to Terry Angelos and Alex Rampell, artificial intelligence advanced significantly over recent months by shifting from spectators that offer research to actors capable of completing real-world transactions. This transition requires payment credentials, which function as the arms and legs enabling agents like Muse, Instinct, and Town to execute tasks ranging from booking sports lessons to buying retail products. However, capturing and routing credentials exposes developers to severe regulatory and compliance burdens, notably the hundred-plus pages of Payment Card Industry Data Security Standard (PCI DSS) rules. Neutral infrastructure platforms like Very Good Security (VGS) address this bottleneck by taking PCI compliance out of scope for AI labs, vaulting primary account numbers (PANs) and dynamically passing network tokens to merchants. The Emerging Wallet Wars and Walled Gardens (13:43) The rise of agentic transactions has triggered what Angelos describes as a new era of 'wallet wars.' Digital wallet providers such as Stripe Link (approximately 300 million accounts), Shopify's Shop Pay, and PayPal (over 400 million accounts) are competing to link their pre-authenticated user bases directly to agents. Angelos compares Stripe to the America Online (AOL) of agentic commerce—an effective closed garden that works smoothly within its network of roughly 1 million merchants, but struggles on the open web. When an agent purchases outside an intermediary wallet's merchant footprint, platforms resort to workarounds like single-use virtual cards. These synthetic layers break core consumer protections, prevent merchants from establishing ongoing subscription relationships like Walmart+, and complicate dispute handling because issuing banks cannot recognize the true merchant of record. Standardizing Agentic Tokens and Multi-Party Vaults (23:51) To solve the shortcomings of closed ecosystems, payment networks have introduced an open framework of 'agentic tokens.' Under this model, an authenticated agent bounds its spending authority, a specific 16-digit token is minted against the consumer's underlying card on file, and the merchant settles the transaction directly. This structure preserves standard chargeback rights while enabling recurring merchant interactions. Rampell and Angelos emphasize that this shift transforms payment vaults from single-player utilities—which historically held cards for isolated SaaS billing—into multiplayer networks where agnostic vaults supply scoped, interoperable tokens across multiple endpoints, AI models, and downstream payment processors. Transaction Verification and Emerging Liability Gaps (29:45) Rampell points out that legacy transactions depend on three distinct hurdles: verifying mathematical card validity, confirming authorized ownership, and establishing sound mind or intent. While open standards handle syntax and agent registration, determining whether an agent hallucinated or misinterpreted user prompts remains an unsolved challenge for card networks. When an agent makes a booking mistake within programmed price caps, the payment ecosystem currently lacks clear rules defining whether the consumer, the merchant, the AI provider, or the card network absorbs the financial loss. Without explicit liability boundaries, Rampell warns that delegated purchasing risks replicating historical 'friendly fraud' and creating moral hazard if consumers can dispute purchases simply by blaming their software agents. Merchant Resistance and the Erosion of Breakage (35:43) Merchants are bifurcating based on commercial incentives. While conversion-focused retailers welcome programmatic checkouts, advertising-heavy platforms like Amazon have actively blocked agents like Muse to protect high-margin ad impressions that depend on human eyeballs. Beyond merchant blocking, Rampell and Angelos argue that widespread agent adoption threatens legacy business models reliant on consumer friction and 'breakage.' When automated agents optimize purchases, cancel unused recurring subscriptions, and exploit category rewards by routing every charge to the optimal card, companies that profit from unredeemed offers, gym membership inertia, or card-portfolio cross-subsidies will be forced to restructure their terms. While open agentic tokens are resolving structural checkout hurdles across open web retail, the payment industry has yet to establish clear dispute mechanisms and liability frameworks for instances where autonomous software misinterprets user intent or errs during execution.

随着支付凭据的接入，AI 智能体正从被动的信息检索工具跃升为能够自主消费的真实世界行动者；这一演变在重燃数字钱包竞争的同时，也对底层合规托管、多方网络代币化及交易争议的责任划分提出了全新要求。 从观望到执行：支付凭据引发的钱包之战 (12:15) Terry Angelos 指出，过去数月内 AI 智能体完成了从“旁观者”到“行动者”的关键跃迁，其根本原因在于获得了作为“四肢”的支付凭据。以 Muse、Instinct 和 Town 为代表的自主平台不再局限于提供静态建议，而是能跨日程核对并直接完成跨平台交易。然而，若要求消费者在聊天界面中输入原始卡号，不仅破坏体验，还会使 AI 开发者面临长达百余页的 PCI DSS 严格合规审查与潜在法律责任，因此将数据剥离至合规基础设施成为行业刚需。 这一转折迅速引发了各平台争夺智能体默认接口的“钱包之战”。智能体开发者倾向于对接兼具庞大消费者基数与深厚商户网络的钱包，例如拥有约 3 亿消费者账户与 100 万商户的 Stripe Link、规模相当但触达 1.7 亿商户的 Shopify Shop Pay，以及拥有超 4 亿活跃账户与 3000 万商户的 PayPal。Alex Rampell 指出，这一竞争如同当年发卡行斥巨资争夺 Uber 等高频应用内的默认扣款位置，旨在抢占自主交易时代的消费入口。 金库架构重构：从单人订阅到多人代币网络 (23:51) 在传统电商中，年交易额超 10 亿美元的大型零售商通常接入 3 至 5 家收单机构（如 Chase、Stripe 与 Braintree）以实现容灾、成本优化及提高授权率；同时每年约有三分之一的信用卡因到期、丢失或发卡机构重组（如 Capital One 收购 Discover）而失效。若依赖分散的独立金库，商户需面临高昂的同步与失效维护成本。独立中立金库（如管理约百亿凭据的 VGS）不仅通过网络令牌化隔离原始卡号，还能自动捕捉卡号轮换信息，将传统仅用于单一订阅扣费的“单人金库”演进为支持多方协同的“多人网络”。 针对开放互联网的智能体交易，Stripe Link 虽构建了运转顺畅的封闭生态，但在跨商户购买（如在 Walmart 采购）时只能生成一次性虚拟卡代付，不仅导致退单纠纷时服务归属混乱，更割裂了商户与消费者的直接绑定（如无法续订 Walmart+）。为克服该限制，发卡网络推出了“智能体代币”（Agentic Tokens）开放标准：在完成智能体意向验证与金额上限约束后，直接基于用户底层卡生成代币并由商户原生处理，既保留了 Visa 和 Mastercard 核心的退单仲裁保护，也满足了商户沉淀客户关系与二次销售的需求。 交易三要素审查与委托授权下的权责博弈 (29:45) Rampell 分析认为，传统支付体系依赖三重验证：基于 Luhn 校验与银行通信的卡号有效性、持卡人身份真实性，以及交易者的“健全心智与意图”（Sound Mind）。在智能体语境下，前两项技术门槛已能解决，但验证模型是否处于“健全心智”——即是否存在幻觉或曲解意图（例如设定 500 美元限额却预订了错误日期的酒店）——成为全新挑战。由于各卡组织尚未建立针对 AI 幻觉的明确纠纷与退款裁定标准，委托授权的责任归属仍处于未定状态。 这种不确定性极易引发类似早期网络游戏充值中的“友善欺诈”（Friendly Fraud）与道德风险：若消费者在冲动消费或后悔后谎称是智能体误操作，而银行惯于在小额纠纷中向消费者倾斜，现有的退单机制将难以承受。与此相对，零售商对智能体呈现出截然不同的态度：依赖高利润搜索与展示广告的 Amazon 因缺乏“人类眼球”可供变现而选择主动拦截 AI 爬虫；而拥有线下供应链优势的 Walmart 与 Shopify 中小商家，则更倾向于通过原生 API 接纳自动化订单以提振纯交易额。 消除摩擦的长尾需求与损耗商业模式的瓦解 (39:54) Angelos 认为，外界对智能体自主跨行倒资金追求利息并引发银行挤兑的担忧被过度夸大，因为现实中账单支付与账户服务黏性依然坚固；反而是因购买摩擦消除而释放的长尾潜在消费被严重低估，持续倾听需求的智能体能直接转化原本被拖延的消费意图。然而，依靠不对称自动化钻空子的早期红利并不具备可持续性，随着服务商开始屏蔽数据中心 IP 并部署“带立场的 API”，买卖双方的算法将直接在价格与条款上展开博弈。 更深层的颠覆在于对传统“损耗商业模式”（Breakage Models）的冲击。以往健身房闲置会员费、未兑现礼品卡、繁琐折扣券以及遗忘的自动续费，均建立在人类消费者的惰性与健忘之上。当自主智能体能够严格审查条款、自动取消未使用的试用期，并精准按照返现规则在不同信用卡之间切换交易路由时，依赖交叉补贴和消费者遗忘获利的传统商业模式将被迫全面改写条款。 智能体经济的成熟要求支付基础设施从封闭代付转向开放互通的多方凭据网络。然而，自动化采购能否全面普及，核心阻碍不仅在于加密代币的普及，更取决于卡组织如何界定 AI 幻觉导致的法律纠纷责任，以及广告变现平台与履约驱动零售商在流量控制权上的长久博弈。

## Background 本期视频由著名风投机构 a16z 于 2026 年 10 月 15 日在 YouTube 频道发布。视频聚焦于 AI Agent（智能体）获得消费与支付能力后的全新格局，由 a16z 普通合伙人 Alex Rampell 与支付安全基础设施企业 VGS（Very Good Security）首席战略官 Terry Angelos 展开深度对话，探讨智能体商务（Agentic Commerce）兴起下的钱包战争、代理令牌标准、商户抵抗以及商户生态与权益体系的重塑。 ## TL;DR 对话深入探讨了 AI 智能体如何从被动的“旁观者”进化为掌握支付凭证的“行动者”，以及该质变对底层金融架构带来的巨大冲击。两位嘉宾剖析了现有围墙花园式支付通道（如 Stripe）与新兴开放标准“代理令牌（Agentic Tokens）”的代际交替，并揭示了多收单分片、PCI 合规以及争议拒付（Chargeback）规则的不可替代性。同时，节目深入拆解了商户端分化态度背后的商业算盘（如亚马逊因广告业务而封杀智能体、沃尔玛则看重增量物流转化），并预言智能体的极致套利将终结依赖规则疏忽的传统商业逻辑（如信用卡返现积分）。最后，Terry 阐述了 VGS 如何从单一凭证金库跃升为连接发卡行、商户与 AI 智能体的多人协同清算中介网络。 ## Atmosphere & style 整场对话洋溢着两位相识二十余年的老搭档之间特有的默契与极客式冷幽默。从打趣早年创业公司 TrialPay 到调侃“死牛钱包（实体皮夹）”与“买花给妻子赔罪”的接地气比方，Alex 和 Terry 将晦涩冰冷的支付底层协议（如 PCI DSS、Token 化、Luhn 校验）解构得生动风趣又极为透彻。现场节奏张弛有度，既有对支付网络底层护城河与金融博弈的硬核技术洞见，又充满了老友间轻松随性的调侃与思想碰撞，让观众在会心一笑中清晰窥见未来商业与 AI 交汇的底层技术图景。 ## Bulletin - [12:15](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=735s) AI Agent 正从仅能提供建议的“旁观者”蜕变为能够代客采购的“行动者”，向其安全分发支付凭证（Payment Credentials）已成为催生代理商业（Agentic Commerce）的底层关键。 - [23:51](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=1431s) 现行基于虚拟卡垫付的“围墙花园”代办模式切断了售后追踪与商户订阅链路，卡组织正联合推动“代理令牌”（Agentic Tokens）开放标准，以受控额度和原生卡账户支撑跨商户结算。 - [35:43](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=2143s) 商户对待智能体商务的态度正剧烈分化：亚马逊等依赖消费者“眼球注意力”变现的广告型电商选择封禁购物 Agent，而具备强大履约能力的零售巨头与中小商户则积极拥抱以获取增量订单。 - [43:35](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=2615s) AI 代理对交易条款与信用卡奖励政策的无情套利，正动摇传统商业高度依赖的沉没与疏忽模式（Breakage Model），银行通过非返现场景交叉补贴高额返现的积分生态将面临解体重构。 ## Key takeaways - [12:15](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=735s) AI Agent 通过获取支付凭证实现从“提供推荐”到“代客下单”的质变，全闭环执行能力让智能体实用价值呈指数级增长，并引发了各大钱包接入 Agent 的通道竞争。 - [09:55](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=595s) 接触和存储信用卡卡号面临厚达百页的 PCI DSS 强制安全规范，VGS 等中立安全金库通过隔离和代管敏感数据，帮助 AI 应用和商户在完全免除 PCI 合规负担的前提下完成支付闭环。 - [06:07](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=367s) 大型商户通常将支付流量分片（Sharding）给 3 至 5 家收单机构，加之全网每年约有三分之一的银行卡发生换卡或转网变更，集中式中立金库成为统一管理网络令牌与自动更新卡号的必备设施。 - [13:43](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=823s) 新一轮“钱包之战”在 AI 领域打响，Stripe Link、Shop Pay 和 PayPal 等巨头正凭借各自庞大的存量用户基数和商户覆盖网络，竞相争夺成为 AI Agent 调用的默认支付渠道。 - [19:42](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=1182s) 数字钱包分为“透传凭据型”（如 Apple Pay）与“代办交易型”（如 Stripe Link），Stripe 在非直连商户采用单次虚拟卡代购的“AOL 围墙花园”模式，暴露出割裂商户后续二次扣款与银行拒付客服链条的弊端。 - [22:42](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=1362s) Visa 和万事达卡等卡组织的根本护城河在于一套完备的中立仲裁与拒付（Chargeback）机制，这一信任框架是消费者敢于在陌生网络场景完成支付的基石。 - [23:51](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=1431s) 新兴的“代理令牌”（Agentic Tokens）通过生成带有交易约束（如单笔金额上限）的动态代币，既能让 Agent 在未建联商户处完成支付，又完整保留了用户原生账户的二次订阅扣款权与卡组织风控保护。 - [28:58](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=1738s) 凭据金库的商业模式正在从服务于单商户定期扣款的“单人模式”，演进为将凭证通过开放标准分发给前沿大模型、第三方 Agent 以及多家收单机构调用的“多人协作模式”。 - [29:40](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=1780s) 支付核验包含验证卡片真伪、持有人身份真实性以及付款人心智健全三个步骤；在 Agent 商业中，当智能体产生幻觉订错服务时的责任归属尚处空白，委派授权（Delegated Authority）保障将成为 Agent 平台的核心差异化卖点。 - [35:43](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=2143s) 亚马逊屏蔽采购工具 Muse 反映出平台对“智能体没有眼球、无法变现广告”的防御姿态，这催生了分别服务于“筑墙抵御 AI”和“拥抱接入 AI”的两类截然不同的基础设施商业机会。 - [37:55](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=2275s) 代理商业的发展轨迹类似于开放银行（Open Banking）的历史演进，电商平台未来可能推出带有自身销售偏好和计费限制的专属 API，与代表消费者利益的买方 Agent 展开双向博弈。 - [39:54](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=2394s) AI 代理通过“高频搬家”套取银行活期存款利差的场景被市场高估，而其通过消除操作摩擦力释放出的长尾潜在消费意愿以及主动清理闲置订阅的潜力被严重低估。 - [43:35](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=2615s) AI Agent 对交易规则的极致利用将摧毁依靠消费者疏忽维系的商业模式，如果每笔消费都被 Agent 自动分流到返现力度最大的单张卡上，传统银行信用卡积分的交叉补贴逻辑将彻底瓦解。 - [47:23](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=2843s) 凭证基础设施正经历从“静态金库时代”到“生命周期动态更新时代”，再到连接发卡行与商户两端以实现数据互换与凭证撮合的“网络效应时代”，代理商业正在成为这一网络飞轮的核心催化剂。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=0s) AI Agent 的飞跃与“钱包战争”的序幕 在过去短短几个月内，AI Agent（智能体）经历了一次根本性的质变：它们从只能旁观、给出建议的“旁观者”（spectators），蜕变为了能够真正替用户执行操作的“行动者”（actors）。赋予 Agent 这种“手脚”的核心关键，正是支付凭证（payment credentials）。 ```mermaid flowchart LR accTitle: AI Agent 演进路径 accDescr: 展示 AI Agent 从旁观者演进为行动者，并接入支付凭证与钱包的过程 subgraph Evolution["智能体能力跨越"] direction TB A["早期 Agent<br/>(旁观者 / 仅提供建议)"] -->|"赋予手脚<br/>(获得支付凭证)"| B["新兴 Agent<br/>(行动者 / 执行采购)"] end subgraph Ecosystem["生态格局重塑"] direction TB B --> C{"商户态度分化"} C -->|"阻断屏蔽<br/>(如 Amazon 封禁 Muse)"| D["对抗阵营"] C -->|"开放对接"| E["接纳阵营"] B --> F["钱包战争<br/>(各类钱包竞相接入 Agent)"] end ``` 这种能力的质变立刻在电商和支付领域激起了剧烈反应。例如，Amazon 近期选择直接封禁了采购智能体 Muse。商户对待 Agentic Commerce（智能体商务）的姿态正迅速分化为两派：要么高度敌视并予以对抗，要么积极拥抱与接纳。随着 Muse 和 Instinct 等智能体工具的普及，一旦它们掌握了支付凭证的调用能力，其实用价值便呈指数级攀升。这直接引爆了所谓的“钱包战争”——各大钱包巨头正争先恐后地将自己的支付能力接入主流 AI Agent。而在整个生态的最底层，几乎所有线上消费者的支付凭证都绕不开一家名为 VGS（Very Good Security）的基础设施公司。 - 驱动智能体从旁观者变成行动者的核心是支付凭证（Payment Credentials）。 - Amazon 封禁 Muse 表明部分主流电商平台对智能体自动化采购持防御与对抗态度。 - 当前行业格局呈现出“钱包争相连接 Agent”的竞争白热化态势。 --- ### [00:55](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=55s) VGS 的起源：从 PGP 极客梗到百亿美元级基础设施 a16z 合伙人 Alex Rampell 与 VGS 首席战略官兼董事 Terry Angelos 是一对合作超过二十年的老搭档。早在 2006 年，两人便联合创立了支付技术公司 TrialPay，后来 TrialPay 被 Visa 收购。在 Terry 任职 Visa 高管期间，Visa 投资了由 Alex 早期主导投资的信息安全企业 VGS；如今 Terry 重新回到前线，担任 VGS 首席战略官并与 Alex 并肩坐在董事会中。两人在对谈开场互相打趣，Alex 调侃道：“二十年前 Alex 看起来大不一样，你倒是没怎么变——我觉得你这是在损我，而我是在夸你。” 关于 VGS 这个名字，Alex 透露了一个鲜为人知的极客幽默：许多人没有意识到“Very Good Security”其实是在致敬密码学史上著名的 PGP（Pretty Good Privacy，极佳隐私）。PGP 之所以坚不可摧，是因为利用超大合数进行质因数分解在计算上极其困难。因此，所谓的“Pretty Good”实则是绝对硬核的加密强度，而 VGS 取名“Very Good Security”，正是沿袭了这种工程师式的冷幽默。 Terry 指出，今天的 VGS 已经成长为庞大的支付基础设施支柱： > “我们为全网约三分之一的电商交易额（GMV）保管凭证……最大的在线商户、最大的新型银行（Neobank）以及最顶尖的前沿 AI 实验室（Frontier Labs），全部都在使用 VGS 来保障支付凭证的安全。” 当前围绕 Agentic Commerce 的核心叙事，正是 VGS 在扮演关键枢纽：帮助 Agent 获取凭证、协助钱包把凭证安全移交给 Agent，并在末端帮助履约商户安全接收凭证。 --- ### [03:03](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=183s) 什么是支付凭证？从明文存储到网络 Token 化 为了给普通读者拆解复杂的金融底层，Alex 引导 Terry 用具体用例解释“支付凭证到底是什么”。在普通人眼中，支付凭证就是一张信用卡，但真实的技术实现要复杂得多。以一家大型在线订餐平台为例，当新用户在网站上绑定卡号时，平台本身并不想承担存储和接触敏感卡号的巨大合规风险与泄密风险。VGS 的职责正是在前端直接拦截并截获该卡号，将其存入独立的安全保险库（Vault）。每当订餐平台发起交易时，它只需向 VGS 保险库索取经过安全处理的数据，整个过程平台自身“从未触碰过真实的支付信息”。 Terry 回顾了自 TrialPay 时代以来支付技术的巨大演变。在过去，工程师的做法是将卡片上的 16 位主账号（PAN，Primary Account Number）直接加密保存在数据库中，交易扣款时再解密并发送给 Chase Paymentech 等清算机构。但现今行业规则已经彻底改变： ```mermaid flowchart TD accTitle: 现代网络 Token 化架构 accDescr: 展示用户输入明文 PAN 后，通过 VGS 和卡组织生成商户绑定 Token 的完整流程 subgraph ClientLayer["客户端与保险库"] User["消费者输入 16 位卡号 (PAN)"] --> Vault["VGS 安全保险库"] end subgraph NetworkLayer["卡组织网络"] Vault -->|"请求 Network Token"| Schemes["卡组织网络<br/>(Visa / Mastercard / Amex / Discover)"] Schemes -->|"签发专用 Token"| BoundToken["域名绑定 Token<br/>(Domain-Bound Token)"] end subgraph MerchantProcessing["交易处理"] BoundToken --> Merchant["大型电商平台"] Merchant -->|"提交 Token 扣款"| Processor["收单处理机构<br/>(如 Chase Paymentech)"] end ``` 在现代支付网络中，商户不再存储真实的 16 位卡号，而是向发卡组织（Visa、Mastercard、American Express、Discover）申请一个“网络令牌”（Network Token）。这种令牌具备“域名绑定”（Domain-bound）特性，即它只能在该特定商户的域名下生效。即使黑客攻破了商户数据库，窃取的令牌在其他任何平台都形同废纸。VGS 正是代表商户完成这些凭证保管、并向各大卡组织获取 Network Token 的中枢中介。 - **PAN（Primary Account Number）**：卡面上印制的 16 位主账号明文。 - **Network Token（网络令牌）**：由卡组织签发并限定商户域名的替代标识符，兼顾支付成功率与最高级别安全性。 - **PCI 合规解耦**：商户借助 VGS 实现数据隔离，自身完全跳过敏感支付数据的触碰环节。 --- ### [06:07](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=367s) 多收单机构分片与动态卡号维护 当一家商户的年商品交易总额（GMV）突破 10 亿美元大关时，其底层的支付架构就会变得极其错综复杂。Terry 向 Alex 抛出一个问题：一家 10 亿美元 GMV 的商户，后端通常接入了多少家支付处理机构（Payment Processors）？Terry 透露，行业内的平均配置通常是 3 到 5 家收单机构（例如同时使用 Chase、Stripe、Braintree 等）。 Alex 解释道，商户之所以不把所有鸡蛋放在一个篮子里，是因为不同的处理机构在成本费率、授权通过率（Approval Rates）以及不同国家地区的支持能力上存在显著差异；用计算机科学的术语来说，商户需要将支付流量进行“分片”（Sharding）。但这带来了一个棘手的乘数效应：如果商户在本地保存卡号，拥有 1 亿张卡就意味着必须在 Chase、Stripe 和 Braintree 各自的保险库中分别保存 1 亿份拷贝，并且需要分别向各个网络同步状态更新。 除了通道分发，卡片本身的动态更替也是巨大的业务损耗点。每年自然到期的卡片约占 25%（基于 4 年有效期），而由于卡片被盗、欺诈或发卡机构合并等原因被强制换卡的比例约为 5% 至 8%，这导致全网每年有约三分之一的银行卡会发生变更： > “每年大约有三分之一的银行卡会发生更替变动。” Alex 分享了亲身经历的尴尬痛点：他持有的 Capital One 1.5% 返现 Visa 卡在 Capital One 收购 Discover 网络后被直接强行转换成了 Discover 卡，导致他每周都会收到来自各家签约商户“无法扣款”的扣费失败通知，因为商户手里的老卡号直接作废了。单个独立商户在面对这种跨网络换卡时往往束手无策，而像 Visa、Mastercard、Discover 等卡组织虽然提供了账户自动更新工具（Account Updater Tools），但前提是商户必须拥有集中化的底层设施来承接变更。 Terry 总结道，这正是中立保险库基础设施的核心价值：VGS 作为独立于任何单一收单机构的中立底座，集中维护卡号与 Network Token，并在发卡行切网（如 Capital One 从 Visa 切到 Discover）或补发新卡时，自动通过网络工具拉取最新凭证，使商户既能灵活调度后端 3 到 5 家收单通道，又不会因为卡片状态变更而丢失宝贵的用户订阅与交易。 ### [09:55](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=595s) 代理商业（Agentic Commerce）兴起与 PCI 合规包袱 Alex Rampell 提到，VGS（Very Good Security）原本就是一门优秀的生意，但在过去一段时间里，AI 领域涌现出“代理商业”（Agentic Commerce）的浪潮，许多知名 AI 公司纷纷选择与 VGS 展开合作。背后的核心逻辑很简单：假设用户在聊天界面告诉 AI “帮我给妻子订束花，母亲节到了，不买她会生气，买了她会开心”，此时用户虽然希望 AI 代劳，但无论用户还是 AI 聊天应用本身，都不希望该应用直接存储明文信用卡卡号。即便存了，AI 应用还要解决如何把敏感支付凭证安全传给商户的难题，因此借助一家懂行且深耕已久的第三方基础设施公司才是正解。 ```mermaid flowchart LR accTitle: AI 代理代购鲜花与合规数据流 accDescr: 展示用户通过 AI 代理购买鲜花时，支付敏感凭证由 VGS 安全承接并传递给商户，从而使 AI 应用免于承担 PCI 合规负担。 subgraph Client["用户端"] U["用户指令: 订束花"] end subgraph AgentPlatform["AI 平台"] A["AI 聊天应用"] end subgraph SecureInfra["安全基础设施"] V["VGS<br>(接管 PCI DSS 合规)"] end subgraph MerchantSide["收单与商户"] M["鲜花商户"] end U --> A A -->|"发起购买任务"| V V -->|"安全凭证结算"| M ``` 随后，Alex 和 Terry 探讨了 PCI DSS（支付卡行业数据安全标准）规范带来的重重枷锁。Alex 戏称自己依稀记得这个缩写的含义，并指出只要商户或平台要接触、接收或存储信用卡数据，就必须面对厚达一百多页、密密麻麻的合规条例。无论是 Visa 还是 Mastercard，两家巨头在市场上虽为死对头，但在强制执行 PCI 合规这件事情上立场完全一致——没有合规，任何银行都不会允许你发卡或处理交易。VGS 的核心价值正在于将平台完全“移出 PCI 适用范围”（Take PCI compliance out of scope），让企业能专注于正常业务运作，而不用背负沉重的合规包袱。这种天然的基础设施定位，让 VGS 在面对 AI 代理发起交易的新场景时，迅速占据了核心枢纽的地位。 - PCI DSS（Payment Card Industry Data Security Standard）是一套长达百页的强制性安全规范，涉及所有发卡、收单与交易存储环节。 - 银行业和各大卡组织（Visa、Mastercard）严格要求所有触碰卡号的机构必须达标，否则无法开展支付业务。 - VGS 通过隔离和代管敏感数据，使 AI 厂商不必自行承担合规审核的巨大工程与法律风险。 ### [12:15](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=735s) 从“旁观者”到“行动者”：AI 代理的长出手脚 Terry Angelos 观察到，在过去几个月里，AI 代理经历了一场从“旁观者”（Spectators）向“行动者”（Actors）的质的飞跃。推动这一转变的关键，就是以安全的方式向 AI 交付支付凭证（Payment Credential）。正如 Terry Angelos 所指出的：“I can go from recommending the flowers for your wife to buying the flowers for your wife.”（我能从‘为你妻子推荐花’直接进化为‘为你妻子把花买好’。） 为了具象化这种体验的差异，Terry 分享了自己前一天晚上的亲身经历：他和妻子一直在商量为七岁的儿子请一位钢琴老师。这件事非常繁琐，老师不仅要能上门教学，还要避开孩子每周三天的足球训练时间，两边的日程极难协调。很久以前他问过 ChatGPT，系统仅仅甩给他一堆推荐列表；而在前一天晚上，他把这个任务交给了一个具备执行能力的 AI 代理。AI 代理不仅在全网锁定了符合条件的候选人，询问了他几个确认偏好的问题，对齐了双方日程，还直接完成了预约并锁定了第一节试听课。Terry 感叹，这种全闭环的服务远比在用户门前扔下一堆未完成的待办事项要有价值得多。AI 代理因为获得了支付凭证这一“手脚”，其效用成倍增长；随着 Muse、Instinct 以及 Town 等 AI 代理产品的崛起，接入安全支付能力直接拉开了新一轮“钱包之战”的序幕。 ```mermaid flowchart TD accTitle: AI 代理从信息推荐到执行闭环的演进 accDescr: 展示 AI 从最初被动提供推荐，到获得支付凭证后主动协调日程并完成支付预订的完整行动演进路径。 subgraph Spectator["早期 AI: 旁观者阶段"] S1["接收用户需求"] --> S2["吐出静态推荐列表"] S2 --> S3["把琐碎事务丢回给用户手动操作"] end subgraph Actor["现代 AI Agent: 行动者阶段"] A1["接入支付凭证<br>(长出手脚)"] --> A2["多方比选与日程核对"] A2 --> A3["主动完成交易闭环与预订"] end ``` ### [13:43](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=823s) 新钱包之战打响：网络覆盖与用户底盘的角逐 AI 代理对支付能力的渴求，促使各大钱包生态争相与 AI 平台接轨，展开了激烈的通道争夺。Terry Angelos 指出，任何 AI 代理在选择合作的钱包时，核心考量无外乎两点：第一，该钱包必须沉淀了海量的存量消费者凭证；第二，该钱包必须具备极广的商户覆盖网络（Merchant Reach）。当前在这场角逐中领跑的巨头包括： | 钱包品牌 | 预估消费者基数 | 商户覆盖网络 | |:--- |:--- |:--- | | **Stripe Link** | 约 3 亿消费者账户 | 约 100 万家商户 | | **Shopify Shop Pay** | 规模相仿的消费者基础 | 约 170 万家商户 | | **PayPal** | 逾 4 亿账户（依据不同统计口径） | 约 3000 万家商户 | 这些钱包竞相接入各种 AI 代理，试图成为智能体调用支付时的首选管道。 ### [14:41](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=881s) 从“死牛钱包”到数字钱包的默认卡争夺战 Alex Rampell 回忆起十多年前老东家 Visa 时代的“V.me”项目，调侃钱包之战在支付历史上从来就不是新鲜事。从心理学上的巴甫洛夫条件反射开始，钱包厂商就一直在争夺“默认存储凭证”的地位。为了说明这一点，Alex 掏出自己的物理皮革钱包向镜头展示，将其戏称为“死牛钱包”（Dead cow wallet）。在移动支付普及之前，Capital One 等发卡行不惜斥巨资请影星 Samuel L. Jackson 代言，在广告里反复追问“你的钱包里有什么？（What's in your wallet?）”，目的就是让消费者在 Chipotle 等收银台结账掏出皮革钱包时，下意识抽出他们的卡。 进入数字化时代后，Uber 和 Airbnb 这类平台本身也变成了装有多种凭证的钱包，不仅 Shop Pay 和 PayPal 之间在争夺结账页面的按钮位置，钱包内部的不同卡组织和发卡行同样在展开厮杀。Terry 和 Alex 共同回忆道，在当年推动 Visa 成为 Uber 默认支付卡的项目中，Chase 和 Amex 等发卡行为了抢夺平台上的“置顶卡（Top of wallet）”推荐位，曾经砸下了数千万美元的推广预算。 ### [16:55](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=1015s) 结账界面的“赛车赞助”与凌晨两点的摩擦力 Alex 形象地指出，传统电商的结账页面由于塞满了 PayPal、Google Checkout（即早期版的 Google 支付）等各类支付按钮，五颜六色花里胡哨，看起来就像贴满赞助商车标的“NASCAR 赛车”。从商户端来看，VGS 扮演的角色某种程度上就像是“商户端的钱包”——商户虽然不负责消费付款，但需要安全承接并长期保存用户的资金凭证以便反复扣费，绝不能像老式收银机那样把客人的 16 位卡号和卡背三位验证码随便抄在小纸条上。 而在消费者端，传统数字钱包最早诞生的初衷，其实就是为了解决“摩擦力”（Friction）问题。Alex 打了个生动的比方：在凌晨两点，如果你因为做了错事或者赶上母亲节，想在 1-800-Flowers 上给妻子订一束花赔罪，而你的“死牛钱包”还扔在楼下不知道哪个角落，如果必须下床找卡输入 16 位数字，交易很可能就放弃了；但如果只要登录 PayPal，它就会自动打包调取名下绑定的所有银行卡直接付款。 Alex 指出，如今的代理商业（Agentic Commerce）在底层逻辑上完全是在重演这一幕。如果用户让 AI 代理去买花，AI 却反问：“Terry，你的信用卡号是多少？”这无疑重新制造了巨大的交互阻碍。AI 代理必须像当年 PayPal 包装银行账户一样，拥有安全直接调度资金的凭证权限，才能真正消除摩擦。Terry 总结道，虽然现在“钱包”这个词被过度泛化了，但其本质最基础的版本，依然是将那只实体“死牛钱包”进行数字化封装并无缝调度的载体。 - “NASCAR 效应”：早期结账页面聚集了过多第三方钱包按钮（PayPal、Google Checkout 等），争抢流量入口。 - 商户端痛点：无法也不应直接持有纸质或明文卡号/CVV 进行订阅扣款，必须依靠中立的凭证金库基础设施。 - 消费者端痛点：避免“凌晨找钱包”式的摩擦阻碍，AI 代理只有原生集成免打扰的支付能力，才能将交易摩擦降到最低。 ### [19:10](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=1150s) 钱包模式的分野：传导凭据 vs 代办交易 Terry Angelos 解释了市场上两类截然不同的数字钱包机制。第一类是以 Apple Pay 为代表的“凭据透传型”钱包：当消费者使用 Apple Pay 在商店结账时，苹果所做的本质上是将用户钱包中绑定的 Capital One Visa 卡（经网络标记化后）直接传递给商户；商户接收到该支付凭据后，可以自由决定交由 Chase、Adyen、Stripe 还是 Braintree 来进行清算处理。 与之相对的第二类，是像 Shop Pay、Stripe Link 或 PayPal 这类“交易代办型”钱包。在这种模式下，钱包本身直接向消费者扣除资金，随后再将款项打给商户。随着代理商业（Agentic Commerce）的兴起，这两种模式开始融合混杂。由于这是一套全新的生态系统，Terry Angelos 认为这与早期的互联网格局有着高度相似的发展轨迹。 ```mermaid flowchart LR accTitle: 两种数字钱包交易模式对比 accDescr: 对比Apple Pay凭据透传模式与Shop Pay/Link交易代办模式在资金与凭据流转上的不同路径 subgraph DirectModel["凭据透传模式 (如 Apple Pay)"] UserA["消费者"] -->|"绑定卡片"| WalletA["Apple Pay"] WalletA -->|"透传 Token 凭据"| MerchA["商户"] MerchA -->|"自主选择路由"| ProcA["收单机构 (Adyen/Stripe/Chase)"] end subgraph ManagedModel["代办交易模式 (如 Shop Pay / Link)"] UserB["消费者"] -->|"扣除款项"| WalletB["Link / PayPal"] WalletB -->|"结算资金打款"| MerchB["商户"] end ``` ### [19:42](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=1182s) Stripe：代理商业时代的“AOL 围墙花园” Terry Angelos 提出了一个尖锐的比喻，将 Stripe 称为代理商业时代的“AOL（美国在线）”。他强调这是一种赞美，意味着 Stripe 在早期率先打通了代理交易的闭环，让它能够在一种“围墙花园”式的机制下直接运转起来。然而，这种闭环的局限性十分明显：Stripe Link 的直连能力仅覆盖约 100 万家商户，一旦代理要去网络之外的零售商采购，链路就会受阻。 Terry 分享了一段两周前的亲身经历：他家里要办活动，便让 AI 代理 Instinct 帮他寻找户外灯具。Instinct 最终在库存丰富的沃尔玛（Walmart）找到了商品并加进购物车。然而，在结账那一刻，Link 与沃尔玛之间并没有建立直连。Link 虽存有 Terry 的真实信用卡，却无法直接将该卡丢给 AI 代理再转交给沃尔玛。为了解决这个问题，Stripe 采取了一个非常聪明的“代购”迂回方案： 1. Stripe 自行签发了一张 Visa 单次虚拟卡（Virtual Card），代表 Terry 在沃尔玛完成灯具垫付。 2. Stripe 随后发起第二笔独立扣款，向 Terry 的个人账户收取该笔款项。 这种机制与 DoorDash 派送员持卡替点餐用户向餐厅付款如出一辙。虽然灯具最终顺利送达且金额无误，但这种围墙花园代办机制直接切断了正常的银行服务链路：如果吊灯出现质量问题，用户向自己的开卡行 Capital One 申诉时，银行只会看到一笔 Stripe 的账单记录，根本找不到沃尔玛的交易信息。Terry 笑着提到，在测试阶段有一位用户通过运通（Amex）卡走该流程购买商品出了问题，打电话给运通客服时，运通只认得 Stripe，甚至直接给了一个电话让其联系该“Link 商户”；结果电话直接打到了一位 Stripe 工程师的个人手机上，因为 Stripe 当时根本没有建立起处理退款争议与拒付（Chargeback）的客服运营架构。 - Stripe Link 覆盖的商户量级目前约为 100 万家，在超长尾以及非合作大型商户场景中存在连接断点。 - 代办模式导致商户交易对手方从“最终零售商”变成了“支付中介”，使得持卡人账单与底层消费完全脱节。 ### [22:42](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=1362s) 拒付机制的基石作用：Visa 与万事达卡的护城河 Alex Rampell 针对这一现象指出了传统卡组织的不可替代性。Visa 和万事达卡（Mastercard）之所以拥有牢不可破的商业护城河，核心就在于它们居中调解一切，制定了完备的仲裁规则，而拒付机制是整套商业信任体系的命脉。Alex 强调：“这是一套信任机制。我付钱给你买灯，你不给我发货。哎呀糟糕，我钱没了，我该怎么办？不是这样的，你实际上拥有一套规则，让我可以据此提出投诉。” Alex 进一步将此与加密货币进行了对比：加密货币固然很酷，但它是一种无记名票据（Bearer Instrument）。一旦用加密货币付款，钱被拿走就再也追不回来了；这或许可以被视为一种特性，但在日常消费场景中却是一个致命缺陷。正是因为有 Visa 和万事达卡居中担任仲裁者，消费者才敢放心地从未曾光顾过的商家那里买东西，商户也由此获得了转化率。如果在代理支付中堆叠过多花哨的代办工程手法而破坏了这一机制，整个商业信任基础就会被动摇。 ### [23:51](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=1431s) 代理令牌（Agentic Tokens）：从围墙花园迈向开放标准 为了打破 AOL 式的孤岛，卡组织推出了全新的底层规范——代理令牌（Agentic Tokens）。在这一架构下，诸如 Stripe Link、Shop Pay 或 PayPal 等钱包，能够将支付凭据安全地签发给 AI 代理，AI 代理再将其提交给没有建立直连的商户（如沃尔玛）。此时，底层的交易主体依然是用户本人的银行卡。 Terry 列举了此前沃尔玛代购案例引发的另一个连锁问题：在他成功购买吊灯后，沃尔玛发来一封营销邮件，提议如果他当天订阅 Walmart Plus 会员，之前收取的 6.99 美元运费就可以直接抵扣会员费。然而当 Terry 点击订阅按钮时，扣款立刻报错失败——因为沃尔玛系统里保存的只是一张 Stripe 签发的单次使用虚拟卡。商户花费巨资获取客户，却因为中介隔离而失去了通过交叉销售和订阅维系客户生命周期的权利。 在代理令牌标准下，整个交易逻辑发生了质的飞跃： 1. 钱包在本地保管用户的真实卡片（如 Capital One Visa），但不直接暴露真实卡号。 2. 经过验证的 AI 代理发起购买声明，明确交易约束（例如：“我是已验证的代理，单笔消费上限 100 美元”）。 3. 系统动态生成一个全新的 16 位数字代币（Token），安全传递给代理，代理再交由商户发起扣款。 4. 交易直接绑定用户的原始卡账户，完好保留了商户后续扣款（如订阅会员）的能力，并完整继承了卡组织的拒付与风控规则。 Terry 透露，这项技术在三个月前还完全不存在，而 VGS（Very Good Security）如今已成为代理令牌最大的开通服务商，业务横跨 AI 代理、数字钱包与商户端。他指出：“三个月前这个数字还是零。现在我们每周处理几百万个这样的令牌。”他断言代理商业必将从目前的围墙花园演进为基于开放标准的代理令牌生态。 ```mermaid flowchart TD accTitle: 代理令牌交易与授权流程 accDescr: 展现从用户指令到代理令牌签发、商户处理及卡组织清算的端到端流转 subgraph AuthStage["1. 凭据申请与范围约束"] User["消费者"] -->|"授权代理采购灯具"| Agent["已验证 AI 代理"] Agent -->|"声明约束: 上限$100"| VGS["凭据金库 / Token签发 (VGS)"] Vault["用户真实银行卡"] -.->|"派生隔离凭据"| VGS end subgraph ExecStage["2. 令牌交付与商户清算"] VGS -->|"签发 16位 代理令牌"| Agent Agent -->|"提交代理令牌"| Merchant["商户 (如 Walmart)"] Merchant -->|"发起扣款交易"| Network["卡组织 / 结算网络 (Visa / Mastercard)"] end Network -->|"映射回真实账户完成扣款 / 保留二次扣款与拒付权"| Vault ``` ### [26:41](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=1601s) 凭据金库的跃迁：从单人游戏到多人协同 谈及 VGS 的底层业务逻辑，Alex Rampell 打趣道，从正面角度来看，VGS 做的其实是一门极其“无聊”的生意——本质上就是做一个安全金库（Vault）。在金库业务中，平淡与安全才是最高美德：金库负责把用户的 15 位 Amex 卡号或 16 位 Visa、万事达卡号妥善封存，向外输出可供业务调用的 Token，然后执行诸如“提取并扣款（Retrieve and Charge）”的操作。例如大型汉堡连锁店使用其 App 让数千万消费者点餐时，商户从不碰触真实卡号，只是指示金库将凭据推送到收单通道发起一笔 9 美元的扣款。Alex 推测 VGS 或许已经接触或存储了全美 80% 的银行卡。 Terry Angelos 证实，VGS 平台上的凭据总量已经高达约 100 亿张。他进一步指出，金库业务的本质正在发生剧变——从“单人游戏（Single Player）”全面跨越为“多人协同（Multiplayer）”： - **多收单路由的协同**：甚至在 AI 代理出现之前，大型商户普遍接入了多家收单处理商，VGS 的金库实际上充当了各家收单机构的统一凭据枢纽，将同一套网络令牌动态派发至 Chase、Stripe、Adyen、Braintree 或 Razorpay。 - **AI 代理时代的协同**：在大型前沿大模型（Frontier Model）纷纷下场切入交易场景后，金库不再只是单个企业的静态防火墙，而是演化为连接顶尖模型、第三方代理、各类钱包与全网商户的核心交换基础设施。 - VGS 目前保管与处理的支付凭据规模约为 100 亿张。 - 传统金库模式服务于单一商户与固定收单行，而代理时代的金库必须支持跨多个代理、多级网络令牌及多个处理机构的实时流转。 ### [28:58](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=1738s) 凭据金库的多人化转变 Terry Angelos 指出，支付凭据存储库（Vault）的性质正在发生根本性变革。在过去的模式中，金库本质上是“单人模式”（Single-player），商户或服务商保存用户的卡片信息仅仅是为了在续费订阅时向用户扣款。然而，随着 AI Agent 的出现，Agent 需要使用相同的凭据去代表用户完成购买，这意味着凭据必须通过开放标准授权给第三方 Agent，金库由此演化为“多人模式”（Multiplayer）。 Alex Rampell 对此表示赞同，指出任何从单人模式走向多人模式的商业形态都会变得极具吸引力。这就像密码管理软件 1Password，最初只是用户保存个人凭据的单人工具，后来演化为与家人共享，而如今则进一步需要将访问权限开放给多个不同的实体与 Agent。 ```mermaid flowchart LR accTitle: Single-player vs Multiplayer Credential Vaults accDescr: Diagram contrasting traditional single-player credential vaults with modern multiplayer agentic vaults. subgraph Traditional["传统单人模式 (Single-Player)"] UserA["用户"] -->|"保存支付凭据"| VaultA["商户/平台金库"] VaultA -->|"自动续费扣款"| SubBilling["定期订阅收费"] end subgraph Agentic["智能体多人模式 (Multiplayer)"] UserB["用户"] -->|"委托凭据"| VaultB["开放标准金库"] VaultB -->|"多方分发"| AgentB["AI Agent"] AgentB -->|"代客跨平台下单"| MerchantB["外部各商户"] end ``` ### [29:40](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=1780s) 支付交易的三大核验步骤与现实隐喻 谈及 Visa 和 Mastercard 等卡组织既定的规则和流程时，Alex Rampell 将一笔交易背后的验证逻辑拆解为三个步骤，并对比了现实物理世界与数字世界的巨大差异： 第一步是核实“卡片本身的真实性与有效性”。卡号验证可以通过算法实现（例如 Mod 10 / Luhn 算法校验和），但这只能证明号码符合格式，最终仍需向发卡行确认该卡号是否真实存在。 第二步是核查“交易发起者是否为该卡的合法持有人”。在 25 年前的线下零售场景中，如果去珠宝店购买价值 15,000 美元的商品，店员不仅会拿走实体塑料卡，还会做两件事：核对卡背面的手写签名（Alex 提到 Apple Card 是他见过的第一张彻底取消背面签名栏的物理卡），并要求出示带照片的身份证件（ID）。但在互联网世界，这一步历来极其棘手，面对一个 IP 地址或登录态，系统很难断定屏幕背后的人就是持卡人，而当交互对象变成 AI Agent 时，身份认证变得更加复杂。 第三步是核实“付款人是否具备做出交易决定的健全心智（Sound Mind）”。Alex 以拉斯维加斯的赌场为例：即便前两步都通过了（卡片真实有效、ID 和签名完全匹配），如果用户已经喝了 15 杯酒、连话都说不清楚，赌场柜台是绝不会允许其继续刷卡支付的。 - 步骤一（卡片有效性）：算法校验（Mod 10 Check / Checksum）加发卡行数据库确认。 - 步骤二（持卡人身份真实性）：传统靠签名和实体 ID 校验，线上难以对齐，Agent 场景更成难题。 - 步骤三（交易心智健全）：防止在无意识或精神受损状态下产生高额消费。 ### [32:39](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=1959s) Agent 幻觉、委派授权与拒付责任之争 将上述三步逻辑映射到 Agent 商业（Agentic Commerce）中，风险焦点迅速转移到第二步与第三步。卡片是否有效是一个已经解决的工程问题；第二步中，Agent 必须在 Visa、Mastercard 或 AMEX 等卡组织完成正式注册与备案，卡组织才认可该指令具有合法来源。 然而第三步却悬而未决——Agent 是否“神志清醒”？它是否产生了幻觉（Hallucination）？Terry Angelos 分享了一个亲身经历：他让 Agent 为女儿在圣迭戈举行的体操锦标赛预订酒店，预算限制在 500 美元以内。如果 Agent 订错了日期或产生了幻觉，而卡组织又严格按照标准验证了指令上下文（证明确实是该用户的 Agent 在预订该金额的酒店），那么这笔交易造成的损失究竟该由谁承担？相关的争议解决（Dispute）机制目前仍处于空白状态，Terry 预测未来 Agent 平台可能不得不自行承担一部分过错责任。 Alex Rampell 引入了“委派授权”（Delegated Authority）的概念，并指出这很可能成为 Agent 产品竞争的核心差异化卖点。如果 Muse 和 Instinct 两家 Agent 展开竞争，谁愿意为买错酒店等错误承担赔偿兜底责任，谁就能赢得用户。从商户立场来看，商户根本不在乎指令是来自 AI、人类助理还是持卡人的配偶，商户需要的是终局确定性。 Alex 回忆起他在 TrialPay 工作时经历的“友好欺诈”（Friendly Fraud）现象： > “我花了一万美元买某种愚蠢游戏的虚拟宝石。月底信用卡账单寄到时，我说：‘亲爱的，我从来没在这上面花过一万美元，那不是我干的，我不知道怎么回事。’然后你就拒付这笔交易，不是因为你没得到承诺的服务，而是因为这事说出来太丢人、太尴尬了。” 过去银行对小额争议往往直接偏向消费者，从而制造了道德风险（Moral Hazard）；而在由 AI 代劳的大额消费中，如果委派授权的界限模糊，各方的权责博弈将更加剧烈。 ### [35:43](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=2143s) 亚马逊屏蔽 Muse 与商户对 Agent 的分化态度 Alex Rampell 提到了一个关键事件：亚马逊近期主动屏蔽了购物 Agent 工具 Muse。许多人认为这只是落后势力抗拒新技术，但 Alex 指出这背后有着极其现实的商业逻辑——亚马逊本质上是一家广告公司，最核心的盈利来源是消费者的“眼球注意力”。Agent 替人跑腿并没有眼球，因而无法被广告变现： > “你无法从智能体的眼球上变现，因为智能体根本就没有眼球。” 在对待 Agent 商业的态度上，市场正迅速分化为两大阵营：敌对型（Antagonistic）与接纳型（Receptive）。这催生了截然不同的两种基础设施商业机会： 1. **赋能敌对型商户**：帮助需要保护广告收入或私域流量壁垒的商户建立防火墙，识别并过滤掉 AI 流量。 2. **赋能接纳型商户**：帮助愿意拥抱 AI 的商户无缝对接 ChatGPT、Muse 等前端工具。如果一个网站弹出 15 个验证码（CAPTCHA），逼迫 Agent 叫醒正在睡觉的用户手动解题，用户必然会掉头转向那些对 AI 友好的商户。 Terry Angelos 认为，这是观察 Agent 商业演进中最有趣的博弈阵地。与亚马逊为了保护广告业务而筑墙相反，像沃尔玛（Walmart）这样的巨头至今并未封杀 Agent。对于拥有全美最强大的最后一公里配送物流体系的零售巨头而言，只要商品能够卖出，任何能够带来订单的自动化 Agent 都可能对它们构成长远利好。整个零售与商户生态对 Agent 的态度，注定会演化出形态各异的多样化版本。 ### [37:55](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=2275s) 开放银行的前车之鉴：商家利益、专属 Agent 与 API 博弈 在对待 AI Agent 的态度上，各类商家之间存在着截然相反的动机。Shopify 旗下的海量中小企业（SMB）视 Agent 为维系生存的新销售渠道，属于“别无选择”；沃尔玛等巨头受到追求增量业务的“贪婪”驱动；而亚马逊等平台则充满了被分流和丧失控制权的“恐惧”。Terry Angelos 认为，这种动态关系高度类似于当年开放银行（Open Banking）在金融业落地的曲折过程：核心问题在于，作为消费者的我是否有权将登录与操作权限委托给第三方 Agent？而平台（如亚马逊或大银行）是否有权拒绝？ ```mermaid flowchart TD accTitle: Agentic Commerce vs Open Banking Evolution accDescr: 比较开放银行与智能体商务中平台如何应对第三方接入 subgraph OpenBanking["开放银行历史路径 (Plaid vs Chase)"] A["第三方数据抓取 (Plaid)"] --> B["银行封禁与安全对抗"] B --> C["妥协谈判：专属 API"] C --> D["账户 Token 化 + 限制权限 + 计费"] end subgraph AgenticCommerce["智能体商务演进预测"] E["个人采购 Agent 抓取/代下单"] --> F["亚马逊/大型平台拦截对抗"] F --> G["平台推出专属带立场的 Agent/API"] G --> H["双向 Agent 博弈与最终消费者获益"] end ``` 当年 Plaid 与摩根大通（Chase）等大型银行交锋的最终结果是长期的协商妥协：银行允许第三方获取客户数据，但前提是必须接入银行指定的专属 API，对银行账户进行 Token 化处理，严格限制返回的数据维度，甚至向数据访问方收取高昂费用。类似的情形很可能会在电商领域重化上演。平台本身是有偏好立场的，Terry 借用线下零售打了个生动的比方：去 Best Buy 挑选平板电视时，店内的销售导购表面上是提供帮助，暗中却背负着特定品牌的提成 KPI。未来的线上交易将演变为两股势力的碰撞——拥有海量用户数据、试图引导消费决策的平台端“有偏好 API/Agent”，与代表买家纯粹利益的个人采购 Agent 展开谈判博弈。正如 Terry Angelos 所总结的那样：“In the case of Chase and open banking it was much more a case of like security and control. In the case of Amazon it might be I am an opinionated API. I have an opinion on what you should buy Alex based on many many different sort of you know data points about you. And then my agent may have an opinion about what I want.” ### [39:54](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=2394s) 过度吹捧 vs 被低估的价值：从存款搬家到被压抑的 GDP 讨论到 AI 与智能体商务的技术成熟度，Alex Rampell 指出硅谷与外部世界存在明显的认知落差：硅谷习惯过度炒作（Overhype），每个对话言必称 AI；而外部地区（如爱荷华州）往往低估（Underhype）它的颠覆性，许多人仅因几年前体验不佳便形成刻板印象。Alex 以自己购买特斯拉的经历做类比：2017 年的初代 Model 3 自动驾驶体验糟糕透顶、堪称笑柄，导致他随后换掉；但如今最新版的完全自动驾驶（FSD）已极其出色。AI 的发展同样如此，很多人停留在 2022 年 ChatGPT 只能写诗或帮学生做英语作业的阶段，尚未意识到当前的 Agent 已经能够自主写代码和执行真实消费。 在具体细分领域，Terry 详细拆解了哪些场景被严重夸大，哪些深层潜力被市场忽视： - **被高估的场景（Overhyped）**：利用 Agent 进行极致的高息存款搬家（Bank Run）。市场曾幻想 Agent 会在不同银行间分秒必争地转移资金以最大化存款收益率。然而，现实中的金融摩擦力巨大。Terry 以自身为例，他很喜欢使用 Mercury 个人账户的高收益储蓄功能，同时也是 Chase 的老客户；虽然理论上应把所有流动资金存入 Mercury 并在还款前一刻转出，但现实中 Chase 账单支付系统更完善、生态绑定更深，使得这种追求极致套利的频繁资金划转极难成气候。 - **被低估的场景（Underhyped）**：被唤醒的长尾真实需求与潜在 GDP 释放。人们的购买意愿因流程繁琐被长期压抑。Terry 举例提到自己的骑行 WhatsApp 群聊，好友 Seth 每隔半年就会根据滚阻测试推荐最新款安全轮胎，他每次都心动想买，可实际受制于操作拖延，可能两年才下单一次。如果有一个常驻群聊的 Agent，每逢推荐直接自动匹配并完成下单，Alex 在骑行装备上的消费潜力将被彻底释放。人类的物质欲望是无限的，配合 Agent “永远在线”（Always-on）的主动服务，将创造巨大的增量商业价值；与此同时，Agent 还能主动嗅探并取消被遗忘的扣费订阅，实现一边省钱、一边将结余用于高价值消费的良性循环。 ### [43:35](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=2615s) 信息不对称红利终结：逆向博弈与商业底层规则重塑 当前消费者在 Agent 场景中尝到的甜头，很大程度上来源于早期“信息不对称”的红利。Alex 指出，目前的优势在于极少数尝鲜者拥有尖端工具，而服务商尚不清楚发生了什么；但经济学铁律是“治愈高价的良药就是高价本身”。以早晨 7 点打电话抢订热门餐厅为例，当只有前 5 个人拥有自动抢订脚本时，这是无往不利的利器；可当人手一套工具时，系统就会被彻底挤爆陷入瘫痪。 这种对抗正在演变为客服与技术层面的“猫鼠游戏”： - **反制 Agent 的呼叫中心**：从前是消费者绞尽脑汁试图跳过机器人语音树、按“0”转接人工客服；未来极可能出现反转——Comcast 等运营商在识别出电话对端或网络请求是 Agent 后，会直接拒绝交互并要求“换真人来谈”。Alex 提到，Comcast 已经在网络端屏蔽了 Muse 的 IP 地址，迫使他必须改用 ChatGPT Codex Astra 伪装在本地电脑上登录以代办业务。 - **条款边缘的极致套利（Breakage Model 崩溃）**：过去的消费体系高度依赖消费者对规则的疏忽或妥协（Breakage）。Agent 能把复杂的条款与细则（Terms & Conditions）研究到极致：比如在预订机票时，精准计算出能获得升舱资格的最低边缘票价。 - **信用卡积分奖励体系面临解体**：现行银行信用卡积分模型之所以能持续运作，在于“高额餐饮返现”吸引用户后，用户会顺带在没有返现的加油站等通用场景刷卡消费，银行借此盈利。Alex 和 Terry 强调，若 Agent 全程接管支付，它会无情地将每笔交易分流到返现力度最大的单张卡上，导致发卡行的交叉补贴逻辑彻底失效。面对这种规则的漏洞利用，整个传统商业逻辑将被迫全面重写条款。正如 Terry 评论的那样：“If agents are intermediating all of our transactions and I'm only using the card where I get points, the whole points reward ecosystem falls apart. Anything with a breakage model in it is going to start to fall apart and then we will have to have new terms and conditions.” ### [47:23](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=2843s) VGS 的三个时代：从凭证金库到网络效应 Alex Rampell 指出，由于许多商户使用 VGS 来托管消费者的支付凭证，VGS 在支付链条中占据了极其关键且难以替代的优势地位。随着商户必须存储每一位合作消费者的支付凭据，VGS 的业务正在迅速增长。Alex 借此询问 Terry Angelos，在他们讨论的所有前沿趋势中，VGS 的未来最令人兴奋的方向是什么。 Terry Angelos 解释称，VGS 的发展可以划分为三个重叠并相互叠加复利的时代。第一个时代是“金库时代”（the vault era），这也是 VGS 过去十年深耕的方向，致力于成为全互联网最受信任的支付凭据托管平台。正如 Terry Angelos 所说：“This is 10 years as sort of becoming the most trusted store for credentials across the internet. Very hard business to to execute on, but that's you know that's the sort of privileged position that the company's in.” 第二个时代则是关于金库内外部信息的动态交互与生命周期维护。金库中大约三分之一的凭证需要定期更新，而由于各大卡组织（卡网络）都有自己独立的凭证更新标准，并且会派发多个网络代币（network tokens）来对应金库凭据，这使得确保数据实时刷新的流程变得极其复杂且具高技术壁垒。 ```mermaid flowchart TD accTitle: VGS发展的三大演进阶段 accDescr: 展示VGS从凭证存储金库到动态更新再到多方网络清算撮合的三个演进时代 subgraph Era1["第一时代：金库时代 (Vault Era)"] A["受信任的凭证托管金库<br/>(存储敏感支付凭据与卡号)"] end subgraph Era2["第二时代：动态更新时代 (Lifecycle Era)"] B["跨网络凭证刷新与同步<br/>(更新1/3凭证、对接各卡网络Token)"] end subgraph Era3["第三时代：网络时代 (Network Era)"] C["多方凭证撮合与清算<br/>(规则化中介、Agentic Commerce、银行端Token化)"] end A --> B B --> C ``` - **第一时代（金库时代）**：花费 10 年建立起来的技术壁垒，成为全网商户托管消费者敏感卡信息与凭证的基础设施。 - **第二时代（出入金库时代）**：解决凭证生命周期管理问题，处理约三分之一凭证的失效更新，并与各大支付网络的多重网络代币标准深度对接以实现自动刷新。 - **第三时代（网络时代）**：成为连接多方的凭证中介（broker），在发卡行（Issuers）与商户（Merchants）之间实现合规且增值的凭证撮合与信息置换。 ### [48:50](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=2930s) 代理商业浪潮下的网络效应与老友叙旧 进入第三个阶段——“网络时代”（the network era）后，VGS 的角色演变为在合规规则之下充当多方参与者之间的凭证撮合中介，为整个生态系统的所有参与者创造价值。Terry Angelos 指出，他们目前正在借由代理商业（agentic commerce）落地这一模式。与此同时，VGS 在代币化（tokenization）业务上也开始与银行展开深度合作。这意味着 VGS 既参与发卡端的代币化处理，又负责商户端的卡凭证存储，同时横跨生态的两端。这种双边卡位使 VGS 能够促成发卡行与商户之间深度的凭据与数据互换。Terry Angelos 展望未来的飞轮效应并总结道：“We'll continue to add services and as we add more services on more merchants and more issuers we continue to get this network effect where you can broker credentials you know among the parties and I think agentic commerce is just fuel into that ecosystem.” 在探讨完严肃的商业与技术架构后，对话转向了轻松幽默的叙旧。Alex Rampell 感慨与 Terry Angelos 畅聊往事非常开心，并提到两人很久以前其实录制过一期关于他们早期共同经历（TrialPay 发展史）的播客。Terry Angelos 笑称那期节目很早、录音棚条件也远没有现在好，确实应该找机会重录一次。随后 Terry Angelos 还不忘幽默调侃 Alex 一句：“But I think what's consistent between that first podcast and this one is you should be buying more flowers for for your wife.” Alex Rampell 笑着赞同并表达感谢，节目在老友间的欢笑中圆满收尾。 ## Notable quotes > [35:43](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=2143s) 你无法从智能体的眼球上变现，因为智能体根本就没有眼球。 — Alex Rampell > [06:07](https://www.youtube.com/watch?v=q8cb1XL9sXA&t=367s) 每年大约有三分之一的银行卡会发生更替变动。 — Terry Angelos ## Mentioned Terry Angelos, Alex Rampell, VGS (Very Good Security), Visa, Mastercard, American Express, Discover, Capital One, Chase, Stripe, Stripe Link, Shopify, Shop Pay, PayPal, Apple Pay, Amazon, Walmart, Muse, Instinct, Plaid, Comcast, ChatGPT, DoorDash, TrialPay, 1-800-Flowers, Best Buy ## Video description What happens when AI agents start spending your money? VGS' Terry Angelos argues that agentic commerce will need new payment infrastructure that gives AI the ability to transact without sacrificing the protections consumers expect. Terry joins a16z General Partner Alex Rampell to discuss the emerging wallet wars, how agentic tokens could enable secure purchases, and what happens when an AI buys the wrong thing. They also explore why merchants may resist agentic commerce, how AI could change credit card rewards and subscriptions, and who controls the customer relationship when agents make purchasing decisions on our behalf. Timestamps: 00:00 Intro 00:56 What a payment credential actually is 09:48 When agentic commerce changed the game 13:43 The wallet wars 19:42 Stripe as the AOL of agentic commerce 23:51 The new agentic token standard 29:45 The three checks behind every transaction 35:43 Why Amazon blocked Muse 39:47 Over-hyped vs under-hyped 47:09 Where VGS goes next Resources: Follow Terry Angelos on X: https://x.com/terryangelos Follow Alex Rampell on X: https://x.com/arampell Stay Updated: If you enjoyed this episode, be sure to like, subscribe, and share with your friends! Find a16z on X: https://twitter.com/a16z Find a16z on LinkedIn: https://www.linkedin.com/company/a16z Listen to the a16z Show on Spotify: https://open.spotify.com/show/5bC65RDvs3oxnLyqqvkUYX Listen to the a16z Show on Apple Podcasts: https://podcasts.apple.com/us/podcast/a16z-podcast/id842818711 Follow our host: https://x.com/eriktorenberg Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details please see http://a16z.com/disclosures. [Source](https://www.youtube.com/watch?v=q8cb1XL9sXA)

[35:43] Alex Rampell: 你无法从智能体的眼球上变现，因为智能体根本就没有眼球。 [6:07] Terry Angelos: 每年大约有三分之一的银行卡会发生更替变动。

[0:00] Terry Angelos: In the last 3 months, we've had this huge leap from agents being what I would call spectators to agents being actors because they effectively have arms and legs, which are payment credentials. [39:54] Terry Angelos: I don't think we're going to get a bank run on savings deposits where agents are going to sort of maximize yield and move my money around between different accounts.

[0:00] In the last 3 months, we've had this [0:01] huge leap from agents being what I would [0:04] call spectators to agents being actors [0:06] because they effectively have arms and [0:08] legs, which are payment credentials. [0:09] Amazon recently blocked Muse and there [0:12] are many reasons for this. You can [0:14] either be antagonistic to agent commerce [0:17] or you can be highly receptive. What's [0:19] happened with the rise of, you know, [0:21] Muse and Instinct is once they have [0:23] access to payment credentials, their [0:25] utility increases. So, we sort of have [0:27] kicked off what I kind of call the [0:29] wallet wars where every wallet is now [0:31] rushing to connect to the agents. [0:34] Everybody who's listening to this [0:35] podcast, I almost guarantee you have a [0:37] credential that's stored at VGS. [0:39] >> The largest online merchants, the [0:41] largest neo banks, and the largest [0:42] frontier labs all use VGS to secure [0:46] payment credentials. It's a really [0:47] formidable position to be in. Where do [0:49] you think VGS is going to go? I think we [0:52] are. [0:55] >> Hi everyone. I'm Alex Rampel, partner at [0:57] Andre and Horwitz and I'm here joined by [1:00] Terry Angelos and I'm very excited about [1:02] being joined by Terry because uh Terry [1:04] and I started a company called Trial Pay [1:06] in 2006. [1:08] Um long long time ago, 20 years ago. [1:10] >> 20 years ago. Alex looked very [1:12] different. [1:12] >> Yes. Uh you look the same. So I I think [1:15] you're insulting me and I'm [1:16] complimenting you and hopefully not the [1:18] other way around. Um but today, this is [1:20] the really cool part. Um, I one of my [1:23] first investments here was a company [1:24] called VGS, which stands for very good [1:27] security. And some people might not get [1:29] that that's actually a pun or a joke [1:31] because one of the most famous [1:32] cryptography uh companies was called [1:35] PGP, which is pretty good privacy. And [1:38] obviously, it's very good priv. It's not [1:40] just pretty good privacy. It's [1:41] incredible privacy. It's unbreakable [1:43] privacy because that's how you know it's [1:45] very hard to uh to factor very large [1:47] composite numbers into their primes. And [1:49] pretty good privacy is actually very [1:50] good. So that was kind of the original [1:52] name for very good security. So I I led [1:53] this investment a long time ago. Um [1:56] trial pay got bought by Visa. Visa under [1:59] your leadership invested in VGS and now [2:03] you're you're back and you're the chief [2:05] strategy officer and board member [2:07] alongside me at VGS. [2:09] >> Yeah. Uh it's been a you know very fun [2:12] and you know incredible journey and um I [2:16] think uh if you look at VGS today it's a [2:18] it's a payment infrastructure business. [2:20] Um we secure credentials for about a [2:24] third of all um e-commerce by by GMV. Uh [2:28] and we have you know the largest online [2:30] merchants we have the largest neo banks [2:32] and the largest frontier labs all use [2:35] VGS to secure payment credentials. And I [2:38] think today, you know, part of the [2:40] reason why we want to talk about Agentic [2:42] Commerce is because VGS is uh involved [2:46] in helping agents get those credentials, [2:48] helping wallets hand those credentials [2:51] to agents and also to merchants who have [2:54] to receive these credentials. And so [2:56] we're sort of in the middle of these [2:59] really interesting conversations and and [3:01] and handovers between credentials. [3:03] >> But maybe we should dig into what a [3:05] credential is. is I mean credential is a [3:07] is a fancy way of saying a credit card [3:08] for most you know for for lay people. It [3:11] gets a little more complicated than that [3:12] when you really dig into it. But um you [3:15] work with I I don't know which which [3:17] merchants I'm allowed to say that you [3:18] work with. You work with all the big [3:19] ones, but like why don't we take it why [3:20] don't we take an example merchant um [3:23] that you work with kind of preAI [3:24] pre-agented commerce and just explain [3:26] what you do like why do they use VGS? [3:29] What do they use you for and how [3:31] pervasive is it? Yeah, let's assume [3:33] we're working with like a very large um [3:36] you know online ordering platform and so [3:38] we have um you know that that company [3:40] effectively uh has to store your credit [3:43] card every time that you every time that [3:46] a new customer comes onto that site and [3:49] hands over um you know a credit card [3:52] that company doesn't necessarily want to [3:55] be exposed to that card number and so [3:58] VGS is is involved in capturing that [4:01] card, securely vaulting it and storing [4:06] it. And every time that that the you [4:08] know online ordering service wants to um [4:10] you know transact with one of its [4:13] processors, it goes to the vault and [4:16] retrieves the information uh in such a [4:18] way that the uh the online ordering [4:22] company never actually has to touch the [4:24] payment information. Now, one of the [4:26] things that I think has made VGS a lot [4:29] more interesting since the days that you [4:31] first invested, it's been a decade of [4:33] becoming sort of the most trusted vault [4:36] online [4:37] is that the card networks have also [4:40] introduced a form of tokenization. So [4:43] that in our days back in trial pay, we [4:45] would actually store a 16-digit pen and [4:48] we would like we would, you know, [4:48] encrypt it. [4:49] >> A pen is a primary account number. [4:51] >> A pen is a primary account number, [4:52] >> credit card number. It's the 16-digit [4:54] number that you are, you know, you know, [4:56] that you have on your on your card. We [4:58] would we would encrypt it and vault it [5:01] and then when it came time to charge [5:03] someone, we might unencrypt it and give [5:05] it to I think Chase Payment Tech back in [5:07] those days. Today, that doesn't happen. [5:10] Today, instead, what happens is rather [5:12] than store that that that 16-digit um [5:15] card number and use that, you actually [5:18] go to the network that issued it. So, [5:20] Visa, Mastercard, MX, Discover, and you [5:22] say, I would like another version of [5:25] this card that's actually domain bound [5:28] just to my merchant domain, which is [5:31] safer because if I ever get there's ever [5:33] a problem, the only number that I have [5:35] only works on my merchant domain and I [5:38] am going to use this when I send it to [5:40] Chase Payment. So VGS as a company [5:44] vaults all these credentials on behalf [5:46] of the of the merchant also gets all of [5:49] the network tokens which is the primary [5:52] way that these credentials are passed to [5:55] the payment processor and to the network [5:58] um and that's what we do today. And so [6:00] um because of that complexity of [6:02] interacting with the card networks and [6:05] doing this across multiple processors. [6:07] So actually here's a question for you [6:08] Alex. If you're a merchant with over a [6:10] billion dollars in GMV, how many payment [6:13] processors do you have on the back end? [6:16] >> Well, it depends. Uh there there are [6:18] plenty that just have one, but but you [6:21] are better. I mean, but this goes to the [6:22] use case of of VGS, [6:25] >> which is rather than being locked into [6:27] one payment processor, uh you'd rather [6:30] unbundle that and you'd rather be able [6:33] to pick and choose which payment [6:34] processor you use. But also if you if [6:38] you have a kind of a a shared [6:40] infrastructure that is storing [6:42] credentials, what if one credential goes [6:46] bad? Like what what if my I just had a [6:48] Capital One card. Great example. Capital [6:50] One bought Discover card. Yeah. Or [6:52] Discover Network and then my my fancy [6:54] Capital One cashback rewards 1.5% Visa [6:58] card. It became a Discover card [7:01] >> and every week I get like notifications [7:03] from very sad merchants. we are unable [7:06] to process your card. Why are they [7:08] unable to process my card? Because that [7:09] old card number doesn't exist anymore [7:11] even though I have the same I still have [7:13] a relationship with Capital One. I just [7:15] got a new card number from them. Yeah, [7:17] >> granted it's a little bit complicated. [7:18] It's on a on a new network. Like there's [7:20] also the possibility like merchants [7:22] can't just say, "Oh, we've seen we know [7:24] that this card became that card. One [7:26] merchant cannot do that." [7:27] >> Yeah. [7:28] >> But a large vault can do that. And then [7:31] as as you know the um the actual [7:34] networks themselves like Visa and [7:35] Mastercard and actually Discover too, [7:38] they have these account updating tools. [7:40] Yeah. [7:41] >> So you could say, "Hey, I'm a merchant. [7:42] I have all of these card numbers." And I [7:44] think what is it every year like 10% of [7:46] cards expire? [7:47] >> Yeah. So about a third of all cards get [7:51] turned over every single year. No, even [7:52] more [7:53] >> like like 25% mathematically because [7:55] it's a it's a four-year expiration [7:57] >> and another 5% due to you know in this [8:00] case Capital One buying discover or [8:02] fraud or or something else. So so you [8:04] know back to that question by the way [8:06] for a merchant of about of a billion [8:07] dollars it's about three to five [8:10] processors that that merchant will have. [8:12] So now [8:13] >> they don't have to [8:14] >> uh they don't have to but they should [8:16] point. That that is the average that we [8:17] see you know you know across the [8:19] >> but there's a reason for them to do that [8:21] because you know different costs [8:23] different approval rates different [8:24] regions like you want to shard [8:26] >> correct [8:27] >> in in kind of computer science terms you [8:29] want to you have all of your cards and [8:31] if you just say I use the same processor [8:33] every single time that's just not a good [8:36] idea or it's not as good of an idea as I [8:38] can have multiple processors but in [8:40] order to do that I have to be able to [8:42] securely send the card number to the [8:44] processor that I've chosen. Yeah, that's [8:46] that's right. And in fact, once you have [8:48] multiple processors, then the job of [8:51] vaulting and getting updates is [8:54] multiplied by the number of processors. [8:56] So maybe I have a 100 million cards in [8:58] as as a merchant, but I have to store [9:00] 100 million with Chase and 100 million [9:02] with Stripe and 100 million with [9:04] Brainree. And each of those vaults have [9:06] to be updated at at their networks. And [9:08] so what what VGS does is serve as an [9:11] independent vault neutral infrastructure [9:13] on behalf of the merchants for the [9:15] vaulting and the updates such that the [9:18] merchant can then send a Visa token to [9:23] the merchant via Stripe, [9:25] >> right? [9:25] >> Um back to your kind of Capital One, you [9:28] know, annoying Capital One uh issue. If [9:30] you're a VGS merchant, that should not [9:32] happen. So there are mechanisms when [9:36] Capital One is switching from Visa to [9:38] Mastercard to get sorry Visa to discover [9:42] to get those updates. Um and certainly [9:44] that's the case if there's fraud or [9:46] something else and that's the role that [9:47] that that VGS is playing. [9:48] >> Yep. Yep. Exactly. So so hopefully [9:50] everybody has a good understanding now [9:52] of what very good security does um VGS [9:55] does. So uh out of nowhere and this this [9:58] was a good business years ago. I mean, [10:00] I've I've been a believer for a long [10:02] time and then this whole thing called [10:04] Agentic Commerce showed up. Um, I guess [10:07] how does the world change and you guys [10:09] are at the center of this because again, [10:11] I don't know I don't know what what [10:13] companies that that work with you I'm [10:14] allowed to say or not, but let's just [10:15] say like a lot of AI companies have [10:17] decided to to work with uh with VGS. And [10:20] for what reason? Well, the reason is you [10:22] don't if I say, "Hey, buy me flowers." [10:25] If I go into my my little chat tool and [10:27] say, "Hey, I'd like to buy flowers for [10:29] my wife. It's Mother's Day." She'll get [10:31] mad at me if I don't buy them, and [10:32] she'll be happy if I do buy them. Please [10:34] buy them. I don't want to give my [10:37] credit. Maybe I do want to give my [10:38] credit card number to we'll call it um [10:41] chat app. Um but they don't want to [10:44] store it. And then even if they do store [10:46] it, they have to communicate it to the [10:48] merchant. And how do they do that? And [10:51] like the better option is you have a [10:53] third party that already knows how to do [10:55] this. Yeah. [10:56] >> So you you take out of scope. I mean [10:57] it's interesting like um I don't know if [10:59] most of our audience knows what the PCI [11:02] DSS rules are but that's the payment [11:05] card initiative digital security system [11:08] if I remember correct. Did I get that [11:10] right? Um, so if you've heard of PCI [11:12] compliance that says, well, if you're [11:15] going to be a merchant and you accept [11:16] credit cards, there is a document, I [11:19] think last time I read it, it was over a [11:21] 100 pages of all of the rules that you [11:23] have to do and maintain in order to be [11:26] PCI compliant. And why is being PCI [11:28] compliant very important? Because every [11:30] bank demands it. You cannot process [11:33] credit cards unless you are PCI [11:34] compliant. You cannot issue credit cards [11:36] unless you are PCI compliant. That's why [11:38] it's this payment card initiative. It's [11:39] all the um you know, Visa and Mastercard [11:42] might be enemies, but on this they agree [11:44] you have to be PCI compliant. Uh one of [11:46] the nice things that VGS takes PCI [11:48] compliance out of scope. So you don't [11:51] you basically say I can just run my [11:52] business like a normal business. I don't [11:54] have to follow these 100 pages of rules. [11:56] VGS just focuses on that, right? Which [11:58] kind of goes back to like I want to buy [12:00] flowers for my my beloved wife and I [12:04] want to have my chat application do that [12:06] for me. Turns out VGS is pretty well [12:09] positioned here just by being the [12:11] backend infrastructure, but now can take [12:12] a a pretty vital role in this this whole [12:14] process. [12:15] >> Yeah, I mean I I I think you know in the [12:18] last 3 months we've had this huge leap [12:20] from [12:22] agents being what I would call [12:24] spectators to agents being actors. And [12:27] how do they act on my behalf? Well, if [12:29] you give them a payment credential in a [12:32] safe way, [12:33] >> right, [12:34] >> I can go from recommending the flowers [12:37] for your wife to buying the flowers for [12:39] your wife. So, like I'll give you an [12:40] example. Like just last night, my wife [12:43] and I have been talking about getting a [12:44] piano teacher for our seven-year-old [12:46] son. And it's complicated because we you [12:49] have to know our schedule and he has [12:51] soccer three days a week and we want [12:52] someone who can come to the house. We [12:53] have to know the schedule of the piano [12:55] teacher. And you know, I asked chat GPT [12:58] about this a while ago and I got some [12:59] recommendations, but just last night I [13:02] asked an agent and it did the entire [13:05] transaction task for me. It found a [13:08] bunch of people, asked me a few [13:09] questions, checked the schedule and [13:12] booked someone for, you know, for the [13:14] very first lesson. So that is infinitely [13:17] more useful than dumping yet another [13:19] task on my doorstep. So, so agents have [13:22] all of a sudden become significantly [13:24] more helpful and useful because they [13:27] effectively have arms and legs which are [13:29] payment credentials. And so what what's [13:31] happened with the rise of you know Muse [13:33] and Instinct uh and Town and other AI [13:36] agents is once they have access to [13:40] payment credentials their utility [13:42] increases. And so we sort of have kicked [13:44] off what I kind of call the wallet wars [13:47] where every wallet is now rushing to [13:50] connect to the agents. And of course as [13:53] an agent I want a wallet that has sort [13:54] of two things. One the the most number [13:57] of consumers that already have a [13:59] credential at that wallet and a wallet [14:02] that has a very broad merchant reach. [14:04] And so you know right now this is sort [14:07] of companies like Stripe Link. They have [14:09] about 300 million consumers have an [14:11] account at link. About a million [14:13] merchants use link. You have Shopify [14:15] with Shopay similarish um you know [14:18] consumer size probably more like 1.7 [14:21] million merchants [14:22] >> right [14:22] >> then you have PayPal which has got you [14:24] know over 400 million accounts depends [14:28] on how you kind of you know sort of you [14:30] know h how that's actually measured but [14:33] they have on the order of 30 million [14:35] merchants. So these wallets are [14:37] connecting to the agents and I think are [14:39] providing a lot more utility. [14:41] >> So it's interesting though because [14:42] wallet wars have been around for a [14:44] while. So remember uh not not to speak [14:46] ill of our of our old employer, remember [14:48] Vime? [14:49] >> Yes. [14:50] >> Right. VME. [14:51] >> VM me. But it was something before that. [14:52] I forgot. They changed the name to Vime. [14:54] But there have been many attempts like [14:56] why why were there wallet wars 10 or 15 [14:58] years ago? Because it turns out that [15:00] being the default stored credential for [15:02] a consumer, right? Like if I can if I [15:05] can get trained in a Pavlovian way where [15:07] it's like oh I see flowers I use PayPal. [15:09] Well PayPal is better off for that. [15:11] >> That's correct. [15:12] >> Um or one of the things that we worked [15:14] on at Trial Pay was how does a Visa It's [15:17] very good for Visa. It's very good for [15:18] Visa's issuing banks if the default card [15:21] in your Uber wallet because Uber is a [15:23] wallet too. Yeah. [15:24] >> You can store eight different cards or [15:26] credentials within Uber. being the top [15:29] of wallet is very very important because [15:31] when digital wallets came out like I I [15:34] kind of like to think before digital [15:35] wallets you had what I like to call dead [15:37] cow wallets you know an actual I think I [15:39] have one here give a demo to our [15:40] audience this is called a dead cow [15:42] wallet um has lots of lots of [15:44] credentials in there um but before [15:47] Capital One would pay Samuel LJ Jackson [15:50] to say what's in your wallet why would [15:52] they do that because when you go to [15:53] Chipotle they want you to pull out your [15:55] dead cow wallet and say like I'm going [15:57] to pay with my Capital one card. [15:58] >> Yeah. [15:58] >> Right. And that was the whereas for when [16:01] digital wallets emerged, it's like I'm [16:03] shop pay, I'm PayPal, I want to be your [16:07] default. And then actually what the [16:08] interesting thing is like within that [16:09] wallet, there were other credentials and [16:11] those guys were competing to be the [16:13] number one. Like if if you recall at at [16:15] at Visa when we were dealing with Uber [16:18] and Airbnb, there was tens of millions [16:21] of dollars paid by the issuers. So like [16:24] Chase and Ammex to prefer their cards on [16:28] those platforms. So that's one war which [16:30] is like [16:31] >> this is my card. I want you know I want [16:33] to sort of you know promote this to the [16:36] to the to the top of wallet. You're [16:38] talking about the actual network war [16:39] where Visa and and we helped Visa become [16:42] a default card at at at Uber and now you [16:45] have this other layer, this abstraction [16:47] layer of IMA wallet that actually [16:50] processes the transaction, the shop [16:52] pays, the links and the and the PayPals. [16:55] And so we talked about how a checkout [16:57] looks like a bit like NASCAR because you [16:59] have multiple different buttons. [17:01] >> Pay with PayPal, pay with Google check. [17:03] remember Google checkout came out along [17:05] >> they were yet another but there were [17:06] like there are consumer wallets and then [17:08] there are merchant because like the [17:10] interesting thing is that I I kind of [17:11] think of VGS as a merchant wallet to an [17:14] extent I mean like it's it's a strange [17:16] metaphor to use because a merchant isn't [17:18] paying for things a merchant is [17:19] receiving payments but like once I [17:22] receive the payment like where where do [17:23] I put it especially if I'm billing this [17:25] customer again and again and again and [17:27] again for PCI reasons and whatnot I [17:30] don't want to store scraps of paper with [17:32] like your se that that secret [17:34] three-digit code that of course is not [17:36] secret because it's on the back of your [17:37] card, but I need it to process it. I [17:39] don't want to store your 16digit, you [17:41] know, credential and just keep it in [17:43] scraps of paper and like my my my uh [17:46] original um cash register. Like I need [17:48] something to go receive that for me. On [17:50] the consumer side, it's like I don't [17:53] want to remember like what was the [17:54] original purpose of a consumer wallet? [17:56] It's that it's 2:00 in the morning. I [17:58] want to buy flowers for my wife. My [18:00] wallet's downstairs. like I don't know [18:02] where my wallet is. I want to go pay on [18:04] 1800 Flowers right now. And if I just [18:06] enter my PayPal credentials, they were a [18:09] rapper and link to all of my my kind of [18:12] all the credentials in my dead cow [18:14] wallet. Like that that was the purpose [18:15] back then. But what I find interesting [18:17] with all of aentic commerce is it's kind [18:19] of the same thing. [18:20] >> It's like, you know, you're you're it's [18:22] still 2 o'clock in the morning. You [18:23] know, you you still want to buy flowers [18:25] for for your wife or something bad that [18:27] you did or, you know, Mother's Day or [18:29] whatever. [18:29] >> Something good, Alex. [18:30] >> Something good, right? you know, always [18:31] always time to buy flowers. I mean, make [18:33] sure we're sponsored by 1800 Flowers, so [18:35] make sure to buy there. Um, so you want [18:37] to go buy and your agent doesn't have [18:40] access. Like it it's the same problem, [18:42] right? It's like the agent could say, [18:43] "Hey, Terry, where's your credit card [18:45] number?" And that's more work and more [18:47] friction versus it has access to it in [18:49] the same way that you just gave access [18:51] to all of your cards, all of your bank [18:53] accounts to PayPal to reduce the [18:55] friction when you were doing things [18:56] manually 15 years ago. [18:58] >> Yeah. And look, I think the word wallet [19:00] is is a bit overused. There's sort of [19:03] different versions of it. One is the the [19:05] digital version of your of your dead cow [19:08] wallet, which is that it just stores the [19:10] payment credential. And when it go to [19:13] when you go to a merchant, you hand that [19:14] credential to the merchant. All right? [19:16] That's kind of what Apple Pay does, [19:17] right? when when I shop at Apple Pay or [19:20] so when I shop at a store with Apple [19:21] Pay, Apple is effectively giving my [19:24] Capital One Visa card effectively [19:26] tokenized in the wallet to the merchant. [19:29] The merchant receives this credential [19:32] and it decides whether it processes it [19:34] through Chase or Adin or Stripe or [19:36] Brainree. [19:38] Then you have wallets that actually take [19:40] care of the transaction for you. So [19:42] think about shop pay or link or PayPal [19:45] where the wallet is in fact transacting [19:48] with me collecting the money from me and [19:51] remitting that back to the merchant. And [19:52] what's happening with Agent Gente [19:54] commerce is we're sort of getting this [19:56] this the sort of you know blend of the [19:58] two and it's such a new ecosystem that [20:02] it actually has parallels for me back to [20:04] the early days of the internet. And uh [20:07] for those people who were around in the [20:08] sort of AOL era, Stripe in many ways is [20:12] like the AOL of Agentic Commerce. Um, [20:15] and I I mean that as a as a compliment [20:17] in the sense that they've figured out [20:20] how to make agentic commerce just work [20:24] in a way that's a very kind of, you [20:25] know, uh, sort of, you know, world [20:28] garden approach. So, you know, a few [20:29] examples here would be um, if I use [20:32] stripe link, links only available at a [20:35] million merchants. So, it's great if in [20:37] fact the item that I want is available [20:39] at one of those million merchants, but [20:41] very often it's available at some place [20:44] where there isn't a Stripe link [20:46] connection. So, a good example was I [20:48] think two weeks ago I had this event at [20:50] my house. I wanted to buy some outdoor [20:52] lights and I asked Instinct to go and [20:55] help me find these lights and I found [20:57] the lights at Walmart which makes a ton [20:59] of sense. Walmart has lots of inventory. [21:02] And the agent went loaded the cart and [21:05] when it came time to pay, link didn't [21:08] have a connection to Walmart. [21:11] It link has my card, but it can't just [21:13] give my card over to the agent and over [21:15] to Walmart. So, Stripe did something [21:18] pretty interesting. They purchased the [21:21] chandeliers on my behalf. They issued [21:24] their own uh actually this is like a a [21:26] Visa virtual card, one time use card. [21:29] made the purchase and then they charged [21:30] me Terry as at at a second transaction. [21:34] This is a little bit how like Door Dash [21:35] works today. You know, you like spend a [21:37] hundred bucks at Door Dash and then the [21:38] Dasher goes and and and pays the [21:42] restaurant. Um, but in the case of this [21:44] Walmart transaction, this is the point [21:47] about it sort of being the AOL, it [21:49] worked just great. Like I actually like [21:50] the the the item showed up, the [21:53] transaction went through, I was charged [21:54] the right amount. But if there's a [21:57] problem with the chandelier, typically I [21:59] call up my bank, Capital One, and I say, [22:02] "Hey, I had this transaction with [22:03] Walmart, right? [22:04] >> Something, you know, went wrong." And [22:06] they say, "Well, we don't see a Walmart [22:08] transaction here. [22:09] >> We see a stripe. [22:09] >> We just see a stripe transaction." And [22:11] in fact, you will find this pretty [22:12] pretty funny. I know someone who uh in [22:15] the sort of testing phase uh did this [22:17] exact flow on their on their AMX card [22:21] called up AMX and said, "Hey, I I you [22:23] know, this transaction with the merchant [22:25] didn't go through, but Amx only saw [22:27] Stripe." So, they gave them a phone [22:30] number to call whoever this link [22:33] merchant is. the person called them, it [22:36] went straight to the cell phone of a [22:37] Stripe engineer because Stripe's not set [22:40] up to run a chargeback, you know, sort [22:42] of [22:42] >> Well, this is the main like one of the [22:44] reasons why Visa and Mastercard have [22:46] such impregnable positions is because [22:48] they intermediate everything and they [22:50] set the rules and chargebacks are very [22:52] important. Yeah. [22:53] >> Right. Like what if I I paid for [22:55] something? It's a trust mechanism. I pay [22:57] you for something. I pay you for lights. [22:59] You don't deliver the lights. Like, oh [23:01] shoot, I don't have money anymore. What [23:03] do I do? No, no, no. You actually have [23:04] rules by which I can kind of lodge a [23:08] complaint [23:08] >> and then there's a central intermediary [23:11] that helps um you know adjudicate those [23:14] complaints in a way that is you know [23:15] more more or less fair that both parties [23:17] agree on. I mean it's part of the [23:18] problem like this is why crypto cool but [23:21] like it's kind it's a bearer instrument. [23:23] I pay you with crypto. I'm never getting [23:24] that back. Which again can be a bug or a [23:26] feature but for your lights example like [23:28] what if the lights don't work? you want [23:30] to have Visa and Mastercard in the [23:32] middle. That's that's high high utility [23:34] actually even for the merchant because [23:36] now I am I'm much more willing to buy [23:38] with the merchant knowing that if they [23:40] don't actually deliver a product that [23:42] works I can get my money back. [23:43] >> Yeah. [23:44] >> So that trust mechanism is very [23:45] important but again if you if you add [23:47] all these kind of um interesting and [23:49] very clever engineering steps it just [23:51] the the the purpose it doesn't work as [23:54] well. [23:55] >> Yeah. So in fact this is like I think [23:56] the right sort of direction of travel [23:59] from the AOL ecosystem to the open [24:02] standards ecosystem. [24:04] The card networks have actually come out [24:06] with a new type of standard. They're [24:09] called Agentic tokens that will allow a [24:13] wallet like a Stripe link or a Shopee or [24:15] a PayPal to hand over a credential to [24:18] the agent in a safe way and then have [24:21] the agent hand it to the merchant in [24:23] such a way that when that Walmart [24:25] transaction occurred, it's my card. In [24:27] fact, like one sort of, you know, um [24:30] addendum to the Walmart transaction [24:32] after I bought the chandelier, [24:34] everything was fine. I got an email from [24:36] Walmart to say, "Hey, Terry, you paid [24:38] for delivery." It was like $6.99. Why [24:41] don't you sign up to Walmart Plus and [24:44] here's an offer. If you sign up today, [24:46] we will credit the delivery toward your [24:49] Walmart Plus subscription. So, I said, [24:51] "Great." I hit the button and of course [24:53] the transaction failed because the card [24:56] that they had was the Stripe one time [24:58] use card. And part of the, you know, you [25:01] mentioned like rules and norms. Part of [25:03] the norms of e-commerce is that Walmart [25:05] spends a lot of money to get me as a [25:07] customer and then they have the right to [25:11] try to have a relationship with me, [25:13] upsell me, cross-ell me. And in this [25:15] case, because we were using the sort of [25:17] AOL version, they weren't able to easily [25:21] uh allow me to subscribe to all, you [25:24] know, to to this to this new service. [25:25] With the new Agentic standards, what [25:28] should happen is if any of the wallets [25:31] don't have a relationship with the [25:32] merchant, but they have my card on file, [25:35] my Capital One uh Visa card, that should [25:38] be uh um tokenized with the agent. [25:41] There's a whole process by which the [25:43] agent says, "I'm about to make a [25:45] transaction. I'm a verified agent. I'm [25:47] going to spend no more than $100." [25:50] That um the token is nothing more than a [25:53] new 16-digit number. Right? [25:54] >> We give that safely to the agent. Agent [25:57] gives it to Walmart and then Walmart [26:00] runs the transaction, but it's my card [26:02] that is being charged and that preserves [26:05] this ability to have a follow- on [26:07] transaction. It preserves all of the [26:09] rules around chargebacks. And so, you [26:11] know, 3 months ago, this was at zero. It [26:13] did not exist. VGS is by far the largest [26:18] uh provisioner of these agentic tokens. [26:20] We do it for the for the agents. We do [26:22] it for the wallets. We do it for some of [26:24] the merchants and this was, you know, [26:26] zero three months ago. Now we're doing a [26:29] few million of these tokens per week. So [26:32] I expect that as we make this transition [26:34] from the from the walled garden to the [26:36] sort of open standards, this will be the [26:38] way that that Agent Commerce evolves. [26:41] >> Yeah, I mean it's it's one of the [26:42] reasons why this like kind of long I I [26:45] mean this as a positive, but like VGS [26:47] has a very boring business of why is it [26:50] boring? Because we have a vault. I mean, [26:52] anybody who's in a vaulting business is [26:53] kind it's not a it's not an exciting [26:55] business. Like your job is to be not [26:56] exciting and be safe and be secure and [26:59] it's like things go in the vault and [27:00] then every now and then things go out of [27:01] the vault, right? It's not like a [27:03] one-way door because this the way that I [27:04] always thought about VGS is um there are [27:07] these credentials that you store. So [27:09] it's like you get these 15digit AMX [27:11] numbers or 16digit, you know, Visa, [27:12] Mastercard Discover numbers, you store [27:15] them, you give back like a, you know, a [27:17] token. This is not a crypto token. This [27:19] is just like here here's something that [27:20] you can use and then there are verbs [27:22] that can be performed on those [27:24] credentials, [27:25] >> right? [27:25] >> So if I am, you know, a big burger chain [27:29] or something hypothetically, uh we don't [27:31] know which one, but like you work with a [27:33] big burger chain, right? Like uh I I I I [27:37] have a consumer app that has, you know, [27:39] tens of millions of users. I know that [27:42] whenever that user says, I want to buy a [27:44] hamburger, like VGS plays a key role. [27:47] It's a very boring role, but I retrieve [27:49] and charge, right? So, I retrieve it [27:51] with the purpose of sending that to a [27:53] payment processor and saying, "Charge [27:55] Terry $9 right now, but I never have to [27:58] touch that credential again." You know, [27:59] ve very, very boring business. But [28:01] you've gotten to I mean, what what [28:02] percentage of cards do you think you've [28:04] seen or touch or store? It's probably [28:06] what is it 80% of cards in America? [28:09] >> I mean, we have like 10 billion [28:10] credentials. Um, it's a lot of [28:11] credentials. Um, I I think you you're [28:14] touching on something really [28:15] interesting, which is vaults are going [28:17] from single player to multiplayer, and [28:20] that makes them a lot more valuable. So, [28:21] for example, put aside agents, the fact [28:25] that merchants have multiple processors [28:28] means that the vault at VGS is [28:31] effectively the the vault for each of [28:33] those processors, right? because the the [28:35] the network tokens stored in that vault [28:38] now have to be sent to Chase, Stripe, [28:40] Adion, Brainree, Razer Pay, [28:42] >> right? [28:43] >> Uh that's multiplayer with agents. [28:47] Uh if you have a vault, let's assume [28:49] you're like a large, you know, frontier [28:51] model. [28:53] Here's a good example. Your large [28:55] frontier model, you're revolting cards [28:58] in order to charge for your [29:00] subscription. [29:01] But now the agent wants to use that same [29:04] transa that same credential to make a [29:07] payment. And so the vault used to be [29:10] single player. It used to be just to [29:11] make sure that you could pay for the [29:13] subscription. But now you can [29:17] effectively give that credential [29:19] according to these open standards to the [29:21] agent. It's now multiplayer, [29:22] >> right? [29:23] >> And so I think you know any business [29:25] that is going from single player to [29:27] multiplayer inherently becomes more [29:28] interesting. I mean even one password, [29:30] right? like it's a used to be a single [29:32] player vault [29:33] >> and maybe I share my credentials with my [29:35] wife on a few things, but now it's [29:37] multiplayer cuz I'm trying to give it [29:39] access to multiple different parties. [29:40] >> Well, this is a really complicated thing [29:42] going back to like the rules and [29:44] procedures of what Visa and Mastercard [29:45] have done, which is normally I just have [29:48] to there there are almost like three [29:50] steps to a transaction. I kind of think [29:52] of number one is is this identity real, [29:56] right? Like in general, like is this a [29:57] real card number? [29:59] >> Right? That's important. But wait a [30:00] minute, I can have a real card number [30:02] and you're a thief. It's still a real [30:04] card number. So, kind of step two is [30:06] once I validate because there actually [30:08] there are two ways of validating a card [30:09] number. There's one there is a um [30:12] there's like a uh cyclicality check, [30:14] right? You kind of add up the Yeah, it's [30:16] check some, right? Um but it's this mod [30:18] 10 check of the credit card number. Is [30:20] it valid? But that doesn't mean it's a [30:21] real card, but it means that it's like a [30:23] valid it's in the valid universe of [30:25] cards. But I can go check with the bank. [30:27] Is this an actual card? That's step one. [30:29] Step two is, is the person that I'm [30:32] interacting with the valid owner of this [30:34] card? So, like if you if you went to a [30:36] jewelry store 25 years ago and you [30:38] wanted to buy something for $15,000, [30:41] they'll take the credit card number, [30:42] they'll take your credit card, your [30:43] piece of plastic, but they'll do two [30:45] things as well. Remember, it used to be [30:46] a you used to have to sign the signature [30:48] of the back of the card. Like I thought [30:50] the first card that I ever saw that just [30:51] got rid of that alto together was Apple. [30:54] So the Apple card, I mean, even though [30:55] they want you to use Apple Pay, if you [30:57] get the physical Apple metal card, [30:59] there's no place to sign on the back. [31:01] None. But they would also say, "Can I [31:04] see your ID?" Like, I remember I bought [31:05] uh I bought something very expensive one [31:07] time. Uh might have been jewelry for my [31:08] wife that I sometimes buy flowers for as [31:10] well. This was probably 25 years ago. [31:12] And they wanted to see my ID. They [31:14] checked the signature. Why? Because they [31:16] knew that the card was valid, but they [31:17] want to make sure that I am the actual [31:19] owner of the card. And then number [31:21] three, and you probably only see this in [31:23] Vegas, um, but it's like, are you of [31:26] sound mind to make this transaction? So [31:28] like if if you went to a to a casino in [31:30] Vegas and you pass, you know, number [31:32] one, card's valid. Number two, it's [31:33] like, yep, here's your ID, here's your [31:35] signature, it matches entirely. But [31:37] there's step three, it's like if you've [31:39] had 15 drinks and you can't even speak [31:42] properly, they're probably going to not [31:44] let you pay, right? So this was kind of [31:46] like the old world. And if you bring [31:48] that metaphor to the new world, well, [31:51] it's like, well, is the card valid? [31:52] Yeah. Does the person [31:54] >> agent? [31:55] >> Well, but this is tricky, right? It's [31:56] like the card is going to be valid. Like [31:58] that that part solved problem. [31:59] >> But the second part, it's like what does [32:01] that mean, right? Like is the person [32:03] that I am interacting with, is my [32:05] interlocutor? Um, I love that word. Is [32:08] is is the person that I'm interacting [32:10] with, are they the valid owner of this [32:12] card? It was always a hard problem to [32:14] figure out on the internet because I [32:15] can't like signature picture like what [32:17] does that mean? You're an IP address and [32:19] maybe I've seen you before and I've [32:21] checked your login, but now I have an [32:22] agent. And then that part three is [32:25] pretty important as well from a [32:26] liability perspective because you do see [32:28] these anecdotes on on X and whatnot of [32:30] people that are on instinct. They're [32:32] like, "What do you mean you spent $900, [32:34] you idiot?" And like who where does the [32:37] chargeback come from? So, so I mean this [32:39] is exactly what these open standards are [32:41] are are solving for. So, let's let's [32:43] take an example where I instruct my [32:46] agent. In fact, this is a re real [32:47] example. I was looking for a hotel [32:50] uh for a trip we're making for my [32:52] daughter's gymnastics tournament and I [32:54] said go and find um a hotel uh in San [32:58] Diego and and don't spend more than you [33:01] know $500. So the first question when [33:03] that when that agent is provisioned is [33:06] to your question is like is this a real [33:07] card? That's a solve problem. The second [33:09] one is is the request for the agent a [33:12] valid request and the way that this [33:14] works with the networks is that agent [33:17] actually has to be registered with Visa [33:20] or Mastercard or AMX. And so there's a [33:22] whole sort of mechanism there. [33:24] >> The third question is like is the agent [33:27] of sound mind like did it get it right? [33:29] >> Right. [33:30] >> And was it drunk? you know, exactly. Did [33:32] it hallucinate? Um, and so, um, that [33:36] part I don't think we fully understand [33:38] yet because if there's a mistake that [33:40] the agent makes and the the cart network [33:44] did everything right. They validated [33:46] that this is your agent and your [33:47] instruction. In fact, part of the way [33:49] these agent these these new standards [33:51] work is you give it context. You say, [33:53] "I'm buying a this is a hotel [33:54] reservation at this amount. It's hard [33:57] for me to know the if you got the if you [33:59] got the actual dates right. If there's [34:02] an error there then that dispute [34:04] mechanism is sort of unresolved. But I [34:07] think we will we will get to that point [34:09] and probably the agents will have to [34:11] take on some responsibility. [34:12] >> Well, it's this idea of delegated [34:13] authority. [34:14] >> Yeah. [34:15] >> Um but it's how do you how do you like [34:17] who is liable? How does this change? I [34:19] mean it might be something where if two [34:22] agents are competing for your business [34:23] like hey Terry please use me. I'm Muse [34:25] or please use me. I'm Instinct. What is [34:28] one of the value ads that I'm going to [34:29] give you? I will take on because if if [34:32] the merchants are smart, they're not [34:33] going to be like I don't give a crap [34:35] that like you got the wrong hotel room [34:38] or you did this. It's like you told you [34:40] gave your card, you delegated authority [34:43] to either an AI agent or your human [34:45] assistant or your wife or whatever. Like [34:48] I don't care. I need to know if I'm [34:50] accepting payments for something that [34:52] there's no excuse at the end. I mean, [34:54] this was called for uh for for those [34:56] that know this term friendly fraud. Once [34:58] upon a time, there's this concept of I [35:00] can buy something. We used to see this [35:02] at trial pay. [35:02] >> All right. [35:03] >> I spent $10,000 on virtual gems. My wife [35:07] and I share the I didn't actually do [35:08] this. Just for disclosure, but let's [35:10] just say I did. I spent $10,000 on [35:12] virtual gems for some stupid game. The [35:14] credit card bill comes due at the end of [35:16] the month. I'm like, "Honey, I never [35:17] spent $10,000 on that wasn't me. I don't [35:20] know." like and you you refuse the [35:22] transaction not because you didn't get [35:24] the service that was offered but because [35:25] it's so embarrassingly terrible. [35:27] >> And merchants had to deal with this all [35:30] the time. [35:31] >> Yeah. [35:31] >> And in many cases like the the the banks [35:34] would say, "Well, if it's a $10 [35:36] transaction, it's not even worth [35:37] fighting it. We'll just give it to the [35:38] consumer every time." Which created [35:40] moral hazard. But delegated authority [35:42] gets really tricky here. I mean, maybe [35:43] if I can ask a different question [35:45] because this is probably the more [35:46] important one. Um, you saw that Amazon [35:49] recently blocked Muse and there are many [35:52] reasons for this. It's mainly that [35:54] Amazon is an advertising company and [35:56] they want you and your eyeballs going to [35:58] their website. They don't want your [35:59] agent because the agent, you can't [36:00] monetize an agent's eyeballs because the [36:02] agent doesn't have eyeballs. But I kind [36:04] of think of it as you can either be [36:06] antagonistic to agentic commerce or you [36:09] can be highly receptive to agentic [36:11] commerce. And I don't actually think I [36:13] think it's an oversimplification to say [36:15] anybody who's antagonistic is is a [36:17] lighteyed and they're gone. It's like [36:18] well Amazon has a reason like a real [36:20] financial reason to be antagonistic [36:23] because advertising is is so much more [36:25] lucrative for them than just selling [36:27] things. Like the highest value skew is [36:29] when they they get your eyeballs and [36:31] they sell the eyeballs. But I guess how [36:33] do you think about this this kind of [36:35] playing out? Um because it feels like [36:38] being enabling technology for merchants [36:40] to be receptive to AI commerce is highly [36:44] highly highly valuable. And but by the [36:47] way at the same time helping the [36:50] antagonistics ones filter out the agent [36:53] commerce is also highly valuable because [36:54] again sometimes they have a reason to do [36:56] that. But if I can make it so that you [36:59] know chat GPT services like yeah you [37:01] should buy here and I make that as [37:02] simple as possible. It just it goes [37:04] without saying like I'm probably going [37:05] to get a lot more transactions than [37:07] somebody that inserts 15 captures and [37:09] says in order to proceed with this and [37:11] then Muse is going to say hey interrupt [37:14] your your nap you got to go do these 15 [37:16] things. No no no you're going to go with [37:17] the one that is receptive to AI [37:19] commerce. Yeah, I mean I think this is [37:21] one of the most fun um uh sort of fronts [37:26] to to watch in in Agent and Commerce and [37:29] I think we're going to get multiple [37:31] different versions of this. We have [37:32] Amazon which I think has many reasons to [37:34] to to block agents as you said they're [37:36] effectively you know in the advertising [37:38] business. You have Walmart which is very [37:40] sort of similar right now hasn't you [37:43] know hasn't taken that step and in fact [37:45] you might argue that if you are one of [37:47] the largest retailers on earth with the [37:49] best last mile delivery that this is all [37:51] good for you and then you have the [37:54] series of merchants that are sort of [37:55] under the Shopify umbrella which are a [37:57] lot of small businesses that really this [37:59] is probably a new like vital channel so [38:02] maybe there's like fear on the Amazon [38:04] side maybe there's greed on the Walmart [38:05] side and maybe there's like you know no [38:08] choice on uh on the SMB side and each of [38:11] these actors or each of these merchants [38:13] are going to experiment with with how [38:16] they deal with agents. I think this [38:18] reminds me a little bit I know we sort [38:20] of have a different view on this of how [38:22] open banking you know was rolled out [38:24] which is do I have the right as an [38:26] Amazon customer to delegate my login to [38:30] an agent and does Amazon have the right [38:32] to say no? This is what happened with [38:34] Plaid and you know Chase and and the [38:36] banks and in the end what tended to [38:38] happen was there was this like long [38:40] negotiated settlement where the likes of [38:42] Chase said hey you can access you you [38:44] you do have the right to this [38:46] information but only on my APIs and my [38:49] APIs limit what information I give you. [38:51] They tokenize the bank accounts [38:52] >> and I might charge a lot [38:53] >> and I might charge you some money for [38:54] it. So actually I think the interesting [38:56] question is I think we may have plenty [38:59] of opportunities or or plenty of [39:01] merchants who just welcome agents. We [39:04] may also have merchants who have their [39:05] own agents. Uh in fact this is what [39:07] happens if I go to a Best Buy and I want [39:10] to buy a flat screen TV. I assume [39:12] there'll be a salesperson there with an [39:14] agenda and he or she will say let me [39:17] show you some flat screen TVs. Now their [39:18] job is to help me but their job is also [39:21] to get a commission. Right? So we might [39:22] have in in the case of Chase and open [39:25] banking it was much more a case of like [39:27] security and control. In the case of [39:29] Amazon it might be I am an opinionated [39:32] API. I have an opinion on what you [39:34] should buy Alex based on many many [39:36] different sort of you know data points [39:38] about you. And then my agent may have an [39:41] opinion about what I want. And so that's [39:43] going to be a fascinating battle or or [39:46] communication. And in the end, I think [39:49] consumers probably win, but we don't [39:52] know what that's going to look like [39:53] going forward, [39:54] >> right? Well, maybe on that, like what do [39:56] you think is overhyped versus underhyped [39:59] in terms of everything that's happening? [40:00] I mean, in general, my rule of thumb is [40:02] that Silicon Valley overhypes things, [40:04] but then the rest of the world [40:05] underhypes things. So, you know, [40:07] everything here, every conversation is [40:09] AI. Um, and it feels like you go to [40:11] Iowa, nobody's talking about AI, but [40:13] then you show it to them like, "Holy [40:15] how is this real?" because they [40:16] tried it three years ago and uh like you [40:19] I just bought a new Tesla and I remember [40:21] I had I had an old Model 3. I think this [40:23] is a good metaphor for what we're [40:24] talking about. Like I had an old Model 3 [40:26] in 2017 with self-driving didn't work at [40:29] all. It was a joke. Eventually got, you [40:31] know, got a different car and now it's [40:33] like it actually works, but I was [40:35] skeptical because I was used to the 2017 [40:38] thing. Yeah. Right. And I feel like a [40:39] lot of that is playing out with AI right [40:41] now where if if you had tried chat GPT [40:43] with all the hallucinations and [40:44] everything in 2022. You're like, "Oh, [40:45] it's a cute toy for like writing poems. [40:48] Um, but or doing my homework in English [40:50] class or something. This is not going to [40:52] actually buy things for me. This is not [40:54] going to write code. It does all these [40:55] things right now." So, I guess what what [40:57] do you think is underhyped versus [40:58] overhyped in terms of everything that's [41:00] going on? [41:01] >> Yeah. So, maybe maybe I'll start with [41:02] with with overhyped. Like I don't think [41:04] we're going to get a bank run on savings [41:08] deposits where agents are going to sort [41:10] of maximize yield and move my money [41:13] around between different accounts. I [41:15] mean a perfect example I am a very happy [41:18] you know Mercury personal account uh [41:21] customer. I'm also a Chase customer and [41:25] you know ideally I should put all my [41:27] savings in mercury where I get a very [41:28] good savings rate and then at the last [41:31] moment when I have to make bill sort of [41:32] pay bills I'll move I'll I'll move over [41:34] the the the cash to you know over to um [41:37] over to my existing bank Chase but their [41:40] bill pay just works better and it's it [41:42] it has more integrations and I have more [41:44] kind of setup. So I I think there's [41:46] enough friction in in the world where we [41:50] we won't see some of the crazy [41:51] volatility in in financial services. Um [41:55] I think maybe what's underhyped [41:58] there is a lot of demand that is [42:01] unmonetized. Um and this is both at the [42:05] sort of part of the world where you [42:08] could expect more GDP. An example of [42:10] this is we actually have a group cycling [42:13] WhatsApp group and we have a person on [42:14] that group who every I don't seems like [42:18] every six months when you know our [42:19] friend Seth comes up with a new [42:21] recommendation for what tires to get [42:23] because he looks at rolling resistance [42:25] and every time he does this I'm like oh [42:26] I should remember to buy those tires cuz [42:28] they're safer but I only buy it you know [42:31] every sort of you know two years. If my [42:34] agent was in that group chat, I would be [42:36] buying tires every time Seth makes a new [42:38] recommendation. You're dramatically [42:40] undermonetized on your on your bike [42:42] equipment, Alex. If the agent could find [42:45] things that could help you with cycling [42:47] and just bought them for you, there's [42:50] more GDP coming out of Alex on cycling, [42:52] >> right? [42:53] >> That is underhyped. Um, at the same [42:56] time, agents are going to save us money. [42:58] So like you know there's probably many [43:00] places where I'm either I I have some [43:03] subscription and it's just been a hassle [43:05] to cancel or I did some trial and I [43:08] forgot about it. So I think agents will [43:10] simultaneously save us money maybe use [43:12] that that savings to to help us buy you [43:16] know buy sort of you know new new items. [43:19] So I I think we are have yet to [43:21] experience the always on nature of these [43:24] agents and as far as I can tell human [43:28] desires and needs are infinite. So [43:30] there's always something else that I [43:33] will want and so I I I think that is [43:34] that's [43:35] >> maybe on this kind of antagonistic [43:37] versus um consiliatory or whatever the [43:40] opposite of antagonistic would be here. [43:42] I mean, a lot of what people are very [43:44] excited about with the agents is it gets [43:46] something for you that is intentionally [43:48] hard and is not really subject to the [43:50] laws of economics. The laws of economics [43:52] are like, you know, the cure for high [43:54] prices is high prices. [43:55] >> If it's really hard to get a restaurant [43:57] reservation because you have to call at [43:59] 7 a.m., well, now everybody has a tool [44:02] that calls at 7 a.m. And therefore, uh, [44:05] the first five people that have the [44:07] tool, now they have an amazing tool. the [44:10] last five people like no no no no like [44:12] it's it's you know it's so crowded [44:13] nobody goes here anymore kind of thing. [44:15] So um how do you think the world changes [44:19] because it feels like right now this is [44:21] the weird this is the weird position [44:22] that we're in is that a lot a lot of the [44:24] early wins if you will the consumer [44:26] successes are around like oh it saved me [44:28] money it did this because of the [44:30] asymmetry. Yeah, [44:31] >> right. Like a very small number of [44:33] people have the tool. The merchant or [44:35] the counterparty has no idea what's [44:37] going on. I can't wait. You know how [44:39] like as a consumer I've been trained to [44:41] say, "No, no, I want to talk to an agent [44:42] or press zero." Or it's like you can [44:44] look up how do you get out of the stupid [44:45] phone tree. I can't wait for the other [44:48] side to say, "No, I have to talk to [44:49] Terry. I ain't talking to Muse." Yeah. [44:52] Right. Like wouldn't that be funny? You [44:53] call Comcast, you're like, "I hate you. [44:54] You suck. Your service is terrible. You [44:56] went down 10 times. This is I do this [44:57] all the time." And then but I of course [45:00] have uh I actually do this with codeex [45:01] because it turned out Comcast is [45:03] blocking [45:03] >> musees. [45:04] >> They know the IP addresses or something [45:07] so it won't let you log in. So I just [45:09] have chat GPT you know Codex Astra like [45:11] login on my computer does everything but [45:14] I'm waiting for the day when Comcast [45:16] says we know this ain't you right? Like [45:18] put Alex on the line because I'm not [45:20] giving something to his agent. And [45:22] that's going to because right now again [45:23] that's what I mean by the asymmetry. [45:24] It's like the small number of insiders [45:27] have this tool. But you know what? It's [45:29] not going to you can't scale that [45:31] infinitely because the world is going to [45:33] adapt. [45:34] >> Although I do have news for Comcast. The [45:36] GPT Astra and Alex are about the same uh [45:39] sort of foe. So So be careful if you if [45:42] you get them on the phone. [45:43] >> Um yeah, I mean I think you're pointing [45:45] out something that that we have yet to [45:47] appreciate, which is [45:49] >> what happens when you actually read the [45:51] terms and conditions and go to the edge [45:54] of what is allowed and disallowed. So, [45:57] you know, an example, the other day I [45:59] was looking at flights. I wanted to buy [46:02] a flight that was on the edge of the [46:05] price that I could pay that would get me [46:07] an upgrade. And that's a kind of a hard [46:09] thing to do. Sometimes I overshoot it. [46:11] My agent did that perfectly. Um or um [46:15] you know, sort of any any case where [46:17] it's very difficult to uncover the terms [46:19] and conditions. Like a good example [46:22] might be what happens to rewards? Like [46:25] rewards on cards are only there because [46:28] I get you this very sort of big headline [46:30] reward on dining, but then you also [46:33] spend it on fuel. And if agents are [46:36] intermediating all of our transactions [46:38] and I'm only getting points on the card [46:41] that gets me I'm only using the card [46:43] where I get points, the whole points [46:45] reward ecosystem falls apart. many many [46:48] of these things could start any anything [46:50] with a with a sort of breakage model in [46:52] it is going to start to fall apart and [46:55] then we will have to have new terms and [46:56] conditions. So maybe that's better, [46:59] maybe that's worse. I don't know, but [47:01] it'll certainly be be different. [47:03] >> And maybe what what are you kind of with [47:05] all of these themes like what where do [47:07] you think VGS is is going to go? Because [47:10] I again I find it so interesting because [47:11] it's the world's most boring business [47:13] around it's this giant vault that stores [47:17] all of the payment credentials for like [47:19] most like everybody who's listening to [47:21] this podcast I almost guarantee you have [47:23] a credential that's stored at VGS [47:25] because you interact with the merchant [47:27] that uses VGS for storage of [47:29] credentials. So that's a it's a really [47:31] formidable position to be in and it's [47:34] been growing remarkably just because [47:36] like yeah you need to every merchant [47:38] needs to store the credentials of the [47:40] consumers that they work with but like [47:42] where do you think is the most exciting [47:44] of everything that we've talked about [47:46] for VGS's future? [47:47] >> Yeah. So uh I think there's sort of [47:50] three eras of VGS that are kind of [47:53] overlapping and compounding and you [47:55] talked about the first era which is the [47:56] vault. This is 10 years as sort of [47:58] becoming the most trusted store for [48:02] credentials across the internet. Very [48:05] hard business to to to execute on, but [48:07] that's you know that's the sort of [48:09] privileged position that the company's [48:10] in. The second era has been what you [48:13] talked about sort of you know [48:15] information going into and out of that [48:17] vault. So that's a third of those [48:18] credentials have to be updated. That's [48:20] become very interesting and complicated [48:24] because [48:25] each network has their own standard for [48:28] how you update their credentials and [48:30] each network has created multiple [48:32] network tokens that represent the vault. [48:34] If you were just storing the information [48:36] be one thing, but you have to store it [48:39] and uh get a uh a series of tokens from [48:43] the networks um so that so that the data [48:47] is effectively you know refreshed [48:48] >> right. The third era is what we call the [48:51] network era which is this this role of [48:54] uh brokering the credentials among [48:57] multiple parties in a way that makes [49:00] sense according to the rules and that [49:02] adds value to to all to all of the [49:04] ecosystem players and we're sort of [49:06] starting to do that now with Aentic [49:08] commerce. Um there are other cases where [49:10] VGS is actually working with the banks [49:12] on the tokenization front. So if you're [49:14] if you're tokenizing the cards and [49:17] you're storing the cards, you're on both [49:19] sides of this of this ecosystem, which [49:22] allows you to do some of the information [49:25] swaps that that you you know that you [49:26] that you sort of talked about earlier. [49:28] So and and and all all of these combine [49:31] all of these combine will continue to be [49:32] a vault. We'll continue to add services [49:35] and as we add more services on more [49:37] merchants and and more issuers we [49:40] continue to get this network effect [49:41] where you can broker credentials you [49:43] know among the parties and I think [49:45] agentic commerce is just fuel into that [49:48] ecosystem. [49:49] >> Well Terry this is uh this has been fun [49:52] down memory lane. Um one of these days [49:55] we should do did we ever I think we [49:56] actually did a history of trial pay one [49:58] a long time ago. [49:59] >> Yeah I think we should we should redo [50:00] it. It was it was one of the first [50:02] podcasts that we did, the the studio [50:03] wasn't, you know, wasn't as nice. But I [50:06] think what's consistent between that [50:08] first podcast and this one is you should [50:10] be buying more flowers for for your [50:11] wife. [50:12] >> Yes. I think that I think that's a great [50:13] idea. All right. Well, thank you very [50:14] much. [50:15] >> Thank you for having me.