StubHub Deploys Anthropic AI to Lift Customer Support Scores

StubHub部署Anthropic AI提升客户支持评分 在线票务与去中心化交易平台依托多模态AI和实时环境数据实现运营自动化，化解买卖双方的信任缺失，并防范代价高昂的交易失误。 Tom Crawford阐释了像StubHub这类高风险点对点交易平台所面临的独特运营挑战，公众信任度低和平台外数据碎片化对买家信心构成了严重威胁。他概述了AI在挂牌创建、动态发现、实时物流和自动化客户支持等全流程的介入，如何大幅减少交易失误并提升满意度。通过借助LLM和多模态视觉模型等工具，平台能够在数以百万计的实时交互中扩展即时决策能力，并建立长期的可靠性。 尽管全球去中心化在线交易平台年交易额达数万亿美元，但因履约失败和未审核卖家问题，该模式面临严重信任赤字，在Trustpilot上的平均评分仅为2星。 (0:00) 为应对分散供给与假票欺诈，StubHub在活动结束后才向卖家发放款项，以此消除伪造条形码的经济动机并缓解离线履约的数据盲区。 (3:30) 下游高达90%的买家问题源自上架环节的人工输入错误，平台正通过多模态AI图像识别自动提取属性并生成商品描述，大幅降低错误发生率。 (7:15) StubHub通过部署基于Anthropic Claude的大语言模型来处理自动化客服决策，推动客户满意度（CSAT）相比人工客服基准提升了30%。 (12:30)

Tom Crawford, Head of Product & Analytics for Supply Chain at StubHub, argues that decentralized marketplaces suffer from a systemic trust deficit that can be mitigated by deploying artificial intelligence across listings, dynamic pricing, and automated fulfillment resolution. Structural Vulnerabilities and the Marketplace Trust Deficit (0:00) Decentralized marketplaces facilitate massive trade volumes, accounting for nearly $4 trillion globally in 2024 and approximately $500 billion in the United States. However, these platforms face a pronounced trust deficit: review aggregator Trustpilot reports an average marketplace rating of only 2 stars, compared to 3.5 stars for general e-commerce and 4 stars across the broader web. In ticketing and peer-to-peer services, operational breakdowns carry disproportionate emotional and financial costs, as missing a live event cannot be rectified by a delayed replacement. Crawford identifies three core structural friction points inherent to the business model: opposing buyer-seller motives regarding speed and pricing, fragmented supply from unvetted sellers, and a lack of operational visibility. Because ticketing fulfillment frequently occurs off-platform through third-party apps such as Ticketmaster, StubHub cannot natively observe whether a digital transfer occurred. To manage fraud and unfulfilled deliveries, the platform is forced to rely on manual confirmations and withhold seller payouts until after an event concludes. Upstream Listing Accuracy and Semantic Discovery (7:15) According to Crawford, up to 90% of downstream buyer issues originate during initial listing creation because of seller input errors or misunderstanding platform rules. To resolve this, marketplaces are shifting toward multimodal AI models that automatically catalog items from a photograph. eBay, for example, demonstrated that 95% of sellers testing its visual listing tool accepted auto-generated titles and descriptions without making manual edits, saving listing time and removing descriptive errors before an item reaches the catalog. On the demand side, platforms are replacing rigid keyword searches with semantic clustering. E-commerce platforms like Etsy deploy machine learning models that group unstandardized handmade items into contextual 'taste buckets.' By interpreting initial touchpoints, search intent, and referral data, discovery engines can tailor search results to individual preferences without requiring rigid structured product data. Adaptive Operations and External Contextual Signals (10:00) Dynamic marketplace efficiency depends on ingesting live environmental indicators rather than relying solely on historical transaction trends. In freight logistics, Uber Freight uses AI-generated heatmaps incorporating live traffic, weather patterns, and regional demand to steer truck drivers toward high-volume corridors, reducing unladen 'empty miles.' At StubHub, external conditions directly inform pricing elasticity and demand prediction. Crawford notes that local weather forecasts serve as one of the strongest statistical predictors for event ticketing: impending rain sharply depresses demand and pricing for open-air Major League Baseball games while simultaneously increasing consumer interest in indoor theater performances. Automating Resolution, Authentication, and Support (12:30) To overcome the low customer sentiment typical of decentralized platforms, StubHub integrated an LLM powered by Anthropic's Claude to resolve customer support tickets. Despite initial internal apprehension that automated agents might alienate users, Crawford reports that customer satisfaction (CSAT) scores rose by 30% above the human agent baseline, driven by immediate, real-time dispute resolutions that bypassed manual queues. Automated visual validation also targets fraud detection before transactions occur. Computer vision systems parse uploaded ticket barcodes to detect structural discrepancies, mirroring authentication models deployed by luxury resale platforms like The RealReal to inspect and verify physical consignments at scale. Crawford concludes that successful marketplace operations depend on prioritizing long-term data logging over immediate storage costs and capturing compounding marginal gains—such as targeted automated fixes that improve only 1% of transactions. A persistent challenge remains resolving off-platform transfer blind spots when fulfillment depends entirely on external third-party software ecosystems.

StubHub供应链产品与分析负责人Tom Crawford指出，去中心化交易平台虽然承载着数万亿美元的交易规模，却普遍面临严重的信任赤字与履约断层；通过在商品发布、供需发现、动态定价及售后风控中深度引入多模态与生成式AI，平台能够从源头消除人为失误并实现高吞吐量的实时决策。 撮合平台的信任赤字与链外履约困境 (0:00) 根据Crawford引用的行业数据，2024年全球平台型贸易额已接近4万亿美元，其中美国市场平台交易额约5000亿美元，占其电商总额的三分之一，亚马逊第三方卖家的销售占比也达到约三分之二。然而，评价聚合网站Trustpilot的数据显示，去中心化交易平台的平均评分仅有2星，显著落后于常规电商平台的3.5星和全网平均的4星。在现场票务等高风险场景中，错失关键赛事的损失具有不可替代性，任何履约失误都会对用户信任造成持久破坏。 去中心化平台通常受制于流动性冲突、高度分散的供给与链外数据缺失三大瓶颈。买卖双方诉求天然对立，买方追求低价即时履约，卖方则追求收益最大化与快速提现。更严峻的是，多达90%的下游买家纠纷源于最初发布环节的人为填报错误或对规则的误解；加之大量门票流转发生在场馆第三方应用（如Ticketmaster）中，平台难以直接监控交付全貌，目前只能通过活动结束后才放款给卖家的防御性手段来遏制假票欺诈。 多模态录入与语义聚类重塑供需匹配 (7:15) 为从源头阻断发布端失误，二手交易平台正逐步转向多模态AI驱动的自动化上架流程。eBay与Mercari等平台允许卖家仅凭一张实物照片，便由AI自动提取特征属性、比对历史销售数据库并自动生成优化后的标题与描述。eBay的数据表明，高达95%的测试用户未对AI生成的描述做任何修改便直接采纳，不仅为单件商品节省数分钟操作时间，还显著降低了因参数填报不当带来的后续纠纷。 在买家发现侧，传统的关键词匹配已被深度语义搜索所取代。例如，手工艺品交易平台Etsy利用AI将数以百万计的非标独特商品归入算法驱动的“品味桶”（taste buckets）。系统通过解析买家的初始接触点、搜索关键词及外部来源，能够跳过机械的属性筛选，根据个人偏好动态聚合出视觉与情境高度相关的商品集合，大幅提升了长尾货品的匹配效率。 实时环境信号与外部变量驱动动态优化 (10:00) 动态交易系统的高度复杂性要求算法能够实时接入环境数据。以Uber Freight为例，其调度引擎结合实时交通、局部气象与即时需求等多维指标生成动态热力图，有效引导卡车司机前往高需求区域，从而大幅降低无载空驶里程并优化动态定价与运力激励。 在票务预测模型中，非传统外部变量展现出极强的统计相关性。Crawford分享指出，天气数据是StubHub门票价格弹性最显著的预测指标之一。例如，降雨预报会大幅抑制美国职业棒球大联盟（MLB）等室外赛事的门票需求，但同时会显著推高室内剧场演出的关注度与订票意愿。将此类实时环境数据嵌入定价引擎，能使平台在不可控的外部冲击下维持供需平衡。 大语言模型客服与机器视觉构建风控闭环 (12:30) 在售后与真伪核验环节，自动化决策正替代传统的排队审核机制。StubHub接入由Anthropic的Claude大模型驱动的客服系统后，不仅打破了自动化可能损伤客户体验的内部顾虑，反而将客户满意度（CSAT）基准提升了30%，并通过全自动即时裁决缩短了纠纷处理流程。 与此同时，机器视觉模型被广泛应用于前端防伪与资质审查。StubHub利用自动条码扫描工具比对票据参数，在售前拦截伪造或不合规条码；奢侈品寄售平台The RealReal同样借助计算机视觉验证模型对实体商品进行规模化真伪鉴定。这种在订单上架与流转节点的即时机器审核，从根本上降低了人工抽检的漏网风险。 冷启动个性化与边际累积的AI落地法则 (14:45) 针对在复杂供应链中规模化落地AI，Crawford提出了多项运营准则。首先是在跨国交易中部署实时多语言翻译，并在会话中同步进行违规与欺诈扫描；其次是针对新用户的冷启动问题，放弃依赖长期交易历史画像的传统方式，转而基于进站即刻的上下文信号进行即时意向推断。 此外，平台必须将全量系统交互与事件日志的记录放在首位，因为用于模型训练的长远数据资产价值远高于服务器存储成本。Crawford强调，AI开发成本的下降使工程团队能够专注于解决长尾边缘案例，即便某项改进仅能优化1%的用户体验，在数百万级交易的复合放大下，也能产生重塑整个平台竞争力的边际效益。 Crawford认为，去中心化交易平台的未来不在于简单的人力替代，而在于利用AI将脆弱的低信任撮合改造为高确定性的履约闭环；但对于如何彻底打破跨场馆第三方应用的“链外数据盲区”，平台目前仍高度依赖事后结算约束与第三方接口协同，数据完全可见性仍是悬而未决的行业难题。

## Background 本视频由 Ai4 于 2025 年 8 月 28 日发布在 YouTube 平台。内容为 StubHub 供应链产品与分析负责人 Tom Crawford 的主题演讲，探讨了在 StubHub、Uber、Amazon 和 Etsy 等去中心化双边交易平台中，如何运用 AI 解决复杂的供需交互难题，以提升系统可靠性、消除摩擦并规模化建立买卖双方的信任。 ## TL;DR 演讲深入剖析了双边交易市场在流动性匹配、非标分散供给以及跨平台数据断裂三大结构性挑战，并指出多数履约问题根源在于最初的上架环节。通过剖析 eBay 智能图像上架、Etsy“品味桶”语义搜索和 Uber 动态调度等案例，展示了 AI 在全链路的重塑作用。StubHub 自身通过引入天气等多维数据辅助定价，并部署基于 Claude 的大语言模型将客户满意度（CSAT）提升了 30%。最后，演讲提炼出构建 AI 驱动型组织的核心能力，倡导全量数据打点、降低创新试错门槛并持续累积微小边际收益。 ## Atmosphere & style 演讲风格务实透彻且富有同理心，开场借由 Oasis 演唱会与球迷错失季后赛的生动比喻，迅速拉近了冰冷算法与人类强烈情感体验之间的距离。整体叙事条理分明、案例详实密集，既有跨行业头部企业的宏观视角，又有 StubHub 内部落地的微观数据指标。现场氛围既充满前瞻探索的启发性，又洋溢着以技术守护用户美好瞬间的诚恳与热忱。 ## Key takeaways - [00:00](https://www.youtube.com/watch?v=3wFCLj7C6R8&t=0s) 双边交易平台（Marketplace）面临流动性匹配、高度分散的非标供给以及断裂不完整的数据链条三大核心结构性挑战。 - [00:00](https://www.youtube.com/watch?v=3wFCLj7C6R8&t=0s) 门票与现场体验承载极高情感价值，履约一旦失败会造成难以逆转的用户伤害，然而双边市场在 Trustpilot 上的公众信任度平均仅有 2 星。 - [05:00](https://www.youtube.com/watch?v=3wFCLj7C6R8&t=300s) 交易链路中 90% 最终导致买家端出问题的故障，其根源都出在卖家最开始挂单（Listing）时的信息录入错误或规则理解偏差。 - [05:00](https://www.youtube.com/watch?v=3wFCLj7C6R8&t=300s) eBay 利用 AI 图像识别自动生成商品标题与描述，95% 尝试该功能的卖家完全保留生成文案一字未改，大幅降低上架摩擦并消除了跨语言障碍。 - [05:00](https://www.youtube.com/watch?v=3wFCLj7C6R8&t=300s) Etsy 通过 AI 将非标长尾商品归类至“品味桶（Taste Buckets）”，结合买家实时搜索行为，实现了兼顾美感与精准度的语义级搜索体验。 - [10:00](https://www.youtube.com/watch?v=3wFCLj7C6R8&t=600s) 跨领域环境数据价值巨大，StubHub 发现门票价格最佳预测指标之一是天气，雨天会导致室外赛事门票暴跌，同时推高室内百老汇剧院需求。 - [10:00](https://www.youtube.com/watch?v=3wFCLj7C6R8&t=600s) StubHub 部署基于 Claude 的 LLM 替代并辅助人工客服，不仅实现了快速自动化判决，还使 CSAT（客户满意度评分）整体提升了 30%。 - [15:00](https://www.youtube.com/watch?v=3wFCLj7C6R8&t=900s) AI 赋能双边市场的四大核心维度包括：秒级实时自主决策、打破语言壁垒、基于行为线索的冷启动即时个性化，以及小步快跑的敏捷 A/B 测试。 - [15:00](https://www.youtube.com/watch?v=3wFCLj7C6R8&t=900s) 企业必须坚持全量数据埋点与记录，因数据资产在未来的潜在商业价值远远超过当下的数据存储成本。 - [15:00](https://www.youtube.com/watch?v=3wFCLj7C6R8&t=900s) AI 时代创新门槛降至历史最低，团队应勇于承担试错风险，并拥抱微小收益（Marginal Gains）的持续复合累积以构建平台护城河。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=3wFCLj7C6R8&t=0s) 现场活动的极高情感价值与双边市场的核心挑战 Tom Crawford 首先向全场介绍了自己在 StubHub 担任供应链产品与分析负责人的工作职责。StubHub 作为全球领先的现场活动门票交易平台，覆盖了从拉斯维加斯 Sphere 的后街男孩（Backstreet Boys）演唱会到布宜诺斯艾利斯足球赛的各类门票交易。Tom 的团队核心任务就是确保整个履约链路万无一失：“如果你在 StubHub 买了一张门票，我们必须确保你能顺利入场，拿到的票准确无误、真实有效，并享受一段美妙的体验。”这份工作固然极具成就感，但难度也非同寻常。 ```mermaid flowchart TD accTitle: 双边交易平台三大核心挑战 subgraph S1["流动性与匹配难题 (Liquidity)"] A1["买卖家利益冲突: 卖家求高价速销 / 买家求低价优质"] A2["非标商品单一性: 售出即下架，无法像标品仓库随时补货"] end subgraph S2["分散且非标的供给 (Fragmented Supply)"] B1["履约意愿与时效参差不齐 (秒发票 vs 临开场前遗忘)"] B2["欺诈与假票风险 (平台采用活动结束后打款机制)"] end subgraph S3["断裂与不完整的数据 (Incomplete Data)"] C1["跨平台流转黑盒 (如 Ticketmaster 转赠不可见)"] C2["端到端供应链可视性缺失"] end ``` 这种挑战首先源于双边交易平台（Marketplace）本身的复杂网络结构——平台上连接着全球数以百万计的买家与卖家，任何一个履约节点断裂都会导致交易失败。更为关键的是，现场演出和体育赛事是一种具有极高情感附加值的商品。Tom 幽默地反问观众是否还记得自己第一次吃过的餐厅、第一次打的 Uber 或第一次在亚马逊买的东西；但如果问到第一次去看的音乐会，几乎每个人都记忆犹新（例如 Tom 自己在米尔顿凯恩斯看 Oasis 绿洲乐队的经历）。如果某个球迷等待了一代人的时间终于迎来主队打进季后赛，却因为 StubHub 的平台失误没能进场，这种创伤绝不是第二天重新补下一单就能抚平的，用户很可能会记恨一辈子。 在庞大的商业体量背后，双边市场整体面临着严重的信任赤字与运营瓶颈： - **市场规模与渗透率**：2024 年全球双边市场电商交易总额接近 4 万亿美元（其中由中国头部电商平台占据很大比重）；在全美也有约 5000 亿美元规模，占全美电商交易总量的三分之一。哪怕是亚马逊，据报道也有约三分之二（2/3）的销售来自第三方卖家。 - **公众信任度低迷**：在评论聚合网站 Trustpilot 上，双边市场平台的平均评分只有 2 星（满分 5 星），而普通独立电商平均得分为 3.5 星，全行业整体基准为 4 星。这凸显出平台在规模化扩张与保持可靠性之间的巨大矛盾。 - **三大核心结构性挑战**： 1. **流动性匹配（Liquidity）**：卖家希望“回款越快越好、价格越高越好”，买家希望“价格越低越好、速度越快越好”。演出门票不同于拥有成千上万标准库存的图书，每一张门票挂单售出后就立即消失，具有高度实效性与非标性。 2. **分散且异构的供给（Fragmented Supply）**：全球各类卖家服务意识参差不齐，有的卖家秒发门票，有的卖家拖延到最后一刻甚至遗忘；此外还存在欺诈风险（StubHub 采取活动结束后才向卖家打款的策略，以从经济动机上杜绝假票）。 --- ### [05:00](https://www.youtube.com/watch?v=3wFCLj7C6R8&t=300s) 供给端痛点与 AI 在上架与搜索上的破局实践 双边市场面临的第三大核心挑战是“不完整的数据链条（Incomplete Data）”。在传统自营业务中，企业对供应链拥有完整的端到端掌控力，清楚知道商品何时打包、何时出库、物流状态以及客户反馈。但在双边交易市场中，信息极度分散。以门票为例，大量电子票是通过 Ticketmaster 等第三方 App 的“转赠给好友”功能流转的，平台完全依赖卖家主动上报是否完成转赠，在系统层面上存在严重的跨平台数据黑盒，极难提前预判并消除履约摩擦。 针对这些痛点，AI 技术正展现出巨大的改造潜力。Tom 指出，在整个交易链路中，**90% 最终导致买家端出问题的故障，其根源都出在卖家最开始上架挂单（Listing）的初始环节**——要么是卖家填错了信息，要么是平台未能有效传达政策与激励规则。eBay 和 Mercari 在这一环节的 AI 应用堪称典范： ```mermaid flowchart LR accTitle: AI 赋能的端到端交易优化流程 subgraph Seller["卖家端：极简上架"] S_Photo["上传一张实物照片"] --> S_AI["AI 图像识别与历史数据比对"] S_AI --> S_Listing["自动生成高质量标题/规格/描述"] end subgraph Platform["平台端：跨语言与意图理解"] S_Listing --> P_Match["多语言自动转换与意图匹配"] end subgraph Buyer["买家端：语义与品味搜索"] B_Query["模糊搜索/上游行为偏好"] --> B_Bucket["Etsy 品味桶 (Taste Buckets) 分组"] P_Match --> B_Bucket B_Bucket --> B_Result["精准高转化呈现"] end ``` - **eBay 与 Mercari 的图片智能上架**：卖家只需拍摄并上传一张商品照片（如一张黑胶唱片），AI 模型就能瞬间解析物品特征，自动匹配平台历史销售数据库中的同类产品，生成高质量的标题、属性和商品描述。 - **显著提升转化率与准确率**：AI 生成的描述大幅缩短了多语言买家与卖家之间的沟通障碍。eBay 官方数据显示，**95% 尝试该功能的卖家完全保留了 AI 自动生成的文案，一字未改**，既为卖家节省了大量繁琐录入时间，也从源头上消除了后续交易纠纷。 - **买家端搜索与 Etsy 的“品味桶（Taste Buckets）”机制**：过去在电商平台搜索“复古乐队 T 恤（vintage band t-shirt）”，系统只能机械地匹配包含该字样的商品。手工艺品交易平台 Etsy 面对海量非标长尾商品，利用 AI 双向重塑搜索体验：一端从买家的搜索词、流量来源及漏斗顶层的行为特征构建实时用户画像；另一端利用算法将非标商品归类至不同的“品味桶（Taste Buckets）”，让搜索结果既精准契合个人审美，又条理清晰便于浏览。 --- ### [10:00](https://www.youtube.com/watch?v=3wFCLj7C6R8&t=600s) 物流匹配、动态定价与基于 Claude 的信任与服务重塑 在供应链匹配与物流调度方面，Uber 及 Uber Freight 的实践展示了 AI 如何在复杂网络中引导供需平衡。Uber Freight 为卡车司机提供了基于 AI 与实时优化的“热力图（Heatmap）”，核心目标是减少卡车的空驶里程（Empty Miles）。系统不仅计算即时供需关系，还将天气变化、交通拥堵等全维度外部数据输入模型，动态调整对司机的路线引导、激励措施和运价策略，从而在三方博弈中实现整体利益最大化。 Tom 分享了一个来自 StubHub 自身的惊人发现：**平台门票价格的最佳预测指标之一居然是“天气数据”**。例如，一旦遇到下雨天气，MLB（美国职业棒球大联盟）室外赛事的门票价格会显著跳水，而室内百老汇剧院的门票需求和价格则会逆势上涨。这种反直觉的跨领域数据应用，证明了将更多维度的环境数据注入 AI 决策模型的巨大价值。 ```mermaid flowchart TD accTitle: AI 在 StubHub 履约与客服中的应用体系 subgraph Verification["入库防欺诈校验"] V1["门票图像/条形码智能扫描"] V2["自动比对挂单信息与票面真实数据"] V3["高价值交易风险阻断 (参考 The RealReal 奢侈品鉴定)"] end subgraph Service["客服与自动化争议解决"] S1["部署 Claude LLM 替代/辅助人工客服"] S2["实时自动化判决 (Auto-Resolution)"] S3["CSAT 客户满意度提升 30%"] end Verification --> Service ``` 在“信任与满意度（Trust & Delight）”板块，StubHub 在客户支持与防欺诈方面取得了突破性进展： - **引入 Claude 模型重塑客户服务**：StubHub 部署了基于 Claude 的 LLM 处理客户咨询。管理层原本仅期望借此降低运营成本和人工坐席负荷，但实际效果出人意料——**由 Claude 驱动的客服系统使平台的 CSAT（客户满意度评分）相比过去的人工客服提升了整整 30%**。 - **毫秒级自动判决（Auto-resolution）**：双边平台面临的核心难题在于超大规模并发下的快速决策，AI 的实时自动仲裁与处置能力完美填补了这一诉求。 - **图像识别与真实性核验**：StubHub 开发了门票智能扫描系统，自动对比卖家上传门票的条形码、座位号与上架填写信息之间的细微差异；同类技术也在二手奢侈品交易平台 The RealReal 中广泛用于高客单价商品的真伪智能鉴定与风险防控。 --- ### [15:00](https://www.youtube.com/watch?v=3wFCLj7C6R8&t=900s) 规模化落地经验与构建 AI 驱动型组织的四大法则 总结各家平台在 AI 转型中的成功路径，Tom 提炼出四大关键能力： | 关键能力维度 | 传统模式瓶颈 | AI 赋能后的突破模式 | |:--- |:--- |:--- | | **实时决策能力** | 依赖人工审批（Human in the loop），响应迟缓且难以承载流量洪峰 | 90% 的场景由 AI 在安全边界内自主秒级决策，效率与质量双赢 | | **语言与全球化** | 跨国交易存在语言壁垒，海外扩张需本地团队重度介入 | 跨语种双向实时自动翻译，支持无障碍沟通并实时过滤欺诈与违规行为 | | **即时个性化** | 依赖老用户多次复购沉淀的静态历史画像 | 用户初次进站即可通过行为线索实现“冷启动”高精度个性化推荐 | | **敏捷 A/B 测试** | 试错成本高、上线周期长、风险难以快速度量 | 小步快跑快速迭代，量化负面影响的同时最大化创新探索空间 | 为了在企业内部真正落地上述能力，Tom 强调组织必须培养深厚的“AI 原生文化”，并遵循以下三条行动准则： 1. **坚持全量数据记录（Quality & Logging of Data）**：“如果你的系统正在发生某项业务流程而你却没有将其打点记录下来，那么这些数据资产在未来的潜在美元价值，将远远超过当前支付的存储成本。”只要产生行为就全量记录，存储成本极低，后续若确实无用随时可以删除。 2. **降低创新试错门槛、勇于承担风险**：在 AI 辅助研发的时代，打造一款新产品的机会成本已经降到了历史最低点。相同的研发团队规模现在能交付成倍的产出，团队应该大胆尝试更多新想法、承担更多合理的试错风险。 3. **拥抱微小收益的累积（Embrace Marginal Gains）**：产品演进不再局限于耗时数月才能上线的大型轰动性功能。借助 AI 的高效率，团队应该去打磨那些哪怕只影响 1% 用户的细微痛点。这些边际收益随着时间推移不断叠加复合，最终将铸就平台强大的护城河。 Tom 在结尾总结道，双边市场虽然极其复杂难做，但 AI 正在为这个古老的商业形态注入全新生命力。技术的终极价值，正是让更多人能够顺利走进现场，在季后赛的看台上或在梦中乐队的演唱会里，享受那些一生难忘的美好瞬间。 ## Notable quotes > [00:00](https://www.youtube.com/watch?v=3wFCLj7C6R8&t=0s) 如果你在 StubHub 买了一张门票，我们必须确保你能顺利入场，拿到的票准确无误、真实有效，并享受一段美妙的体验。 — Tom Crawford > [05:00](https://www.youtube.com/watch?v=3wFCLj7C6R8&t=300s) 90% 最终导致买家端出问题的故障，其根源都出在卖家最开始上架挂单（Listing）的初始环节 — Tom Crawford > [05:00](https://www.youtube.com/watch?v=3wFCLj7C6R8&t=300s) 95% 尝试该功能的卖家完全保留了 AI 自动生成的文案，一字未改 — Tom Crawford > [10:00](https://www.youtube.com/watch?v=3wFCLj7C6R8&t=600s) 平台门票价格的最佳预测指标之一居然是“天气数据” — Tom Crawford > [10:00](https://www.youtube.com/watch?v=3wFCLj7C6R8&t=600s) 由 Claude 驱动的客服系统使平台的 CSAT（客户满意度评分）相比过去的人工客服提升了整整 30% — Tom Crawford > [15:00](https://www.youtube.com/watch?v=3wFCLj7C6R8&t=900s) 如果你的系统正在发生某项业务流程而你却没有将其打点记录下来，那么这些数据资产在未来的潜在美元价值，将远远超过当前支付的存储成本。 — Tom Crawford > [15:00](https://www.youtube.com/watch?v=3wFCLj7C6R8&t=900s) 在 AI 辅助研发的时代，打造一款新产品的机会成本已经降到了历史最低点。 — Tom Crawford > [15:00](https://www.youtube.com/watch?v=3wFCLj7C6R8&t=900s) 产品演进不再局限于耗时数月才能上线的大型轰动性功能。借助 AI 的高效率，团队应该去打磨那些哪怕只影响 1% 用户的细微痛点。 — Tom Crawford ## Mentioned Tom Crawford, StubHub, Ai4, Sphere, Backstreet Boys, Oasis, Uber, Amazon, Etsy, Trustpilot, Ticketmaster, eBay, Mercari, Uber Freight, MLB, Claude, The RealReal ## Video description Tom Crawford, Head of Product & Analytics - Supply Chain, Stubhub Abstracts How will AI transform the user experience in the most complex form of ecommerce - decentralized marketplaces? In this talk, I’ll explore the unique challenges of platforms like StubHub, Uber, Amazon, and Etsy, where millions of transactions happen in real time between unconnected buyers and sellers. These environments demand instant trust, frictionless interactions, and hyper-personalized experiences - an ideal playground for AI. I’ll share how we’re designing smarter systems to improve reliability, reduce friction, and build confidence at scale. Whether you’re working in B2C, B2B, C2C, or hybrid ecommerce, the strategies and frameworks I’ll discuss are highly applicable. Join this session to understand how AI is reshaping digital marketplaces. [Source](https://www.youtube.com/watch?v=3wFCLj7C6R8)

[0:00] Tom Crawford: 如果你在 StubHub 买了一张门票，我们必须确保你能顺利入场，拿到的票准确无误、真实有效，并享受一段美妙的体验。 [5:00] Tom Crawford: 90% 最终导致买家端出问题的故障，其根源都出在卖家最开始上架挂单（Listing）的初始环节 [5:00] Tom Crawford: 95% 尝试该功能的卖家完全保留了 AI 自动生成的文案，一字未改 [10:00] Tom Crawford: 平台门票价格的最佳预测指标之一居然是“天气数据” [10:00] Tom Crawford: 由 Claude 驱动的客服系统使平台的 CSAT（客户满意度评分）相比过去的人工客服提升了整整 30% [15:00] Tom Crawford: 如果你的系统正在发生某项业务流程而你却没有将其打点记录下来，那么这些数据资产在未来的潜在美元价值，将远远超过当前支付的存储成本。 [15:00] Tom Crawford: 在 AI 辅助研发的时代，打造一款新产品的机会成本已经降到了历史最低点。 [15:00] Tom Crawford: 产品演进不再局限于耗时数月才能上线的大型轰动性功能。借助 AI 的高效率，团队应该去打磨那些哪怕只影响 1% 用户的细微痛点。

[0:00] Tom Crawford: How will AI transform the user experience in the most complex form of ecommerce - decentralized marketplaces? [0:00] Tom Crawford: These environments demand instant trust, frictionless interactions, and hyper-personalized experiences - an ideal playground for AI. [0:00] Tom Crawford: Worldwide marketplace trade reached nearly $4 trillion in 2024, heavily influenced by large platforms in China. [0:00] Tom Crawford: In the United States, marketplace GMV stands at roughly $500 billion, representing about one-third of total domestic e-commerce. [0:00] Tom Crawford: Third-party sellers account for approximately two-thirds of all items sold on Amazon. [0:00] Tom Crawford: Online review aggregator Trustpilot shows an average rating of only 2 stars for marketplaces, compared to 3.5 stars for general e-commerce platforms and 4 stars across all web categories. [3:30] Tom Crawford: Sellers aim to maximize payout speed and amount, whereas buyers demand top quality, fast delivery, and low prices. [3:30] Tom Crawford: To curb fraud, StubHub holds seller payouts until after the event takes place, eliminating financial incentives for uploading counterfeit barcodes. [7:15] Tom Crawford: Up to 90% of downstream buyer issues originate at the initial listing stage due to seller input errors or misunderstood policies. [7:15] Tom Crawford: eBay reported that 95% of users who tested this feature adopted the AI-generated description without making any edits [10:00] Tom Crawford: weather data serves as one of the strongest statistical predictors for ticket pricing elasticity [12:30] Tom Crawford: Customer Satisfaction (CSAT) scores increased by 30% over human agent baselines. [14:45] Tom Crawford: The monetary value of long-term data for model training far outweighs server storage costs. [14:45] Tom Crawford: Small improvements that impact only 1% of users compound across millions of transactions to elevate the entire platform experience.

[0:00] Hi there, how is everybody? I am Tom and I work for StubHub, which hopefully most of you have heard of, but we're basically a marketplace for trading tickets to live events. So whether it's the Backstreet Boys of the Sphere or a soccer game in Buenos Aires, we've got you covered around the world. You can buy and sell really any ticket to live experience that we can offer. And specifically my role is to manage the technology and the teams that make sure that if you buy a ticket to a live event on StubHub, you get into the event. We get you the ticket, it's the right ticket, it works, you have an amazing experience. which is an amazingly rewarding thing for me. I love live events and I hope a bunch of you too, too, as well. But it's also a pretty tough gig, right? Firstly, we're a marketplace and marketplaces rely on having sellers around the world send tickets to the buyers. There's millions of different users you're trying to connect to any one time to get the good from A to B, which means things can go wrong. And secondly, live events are an extremely emotional thing to be selling, right? I don't know how many of you remember the first restaurant you ate at, or the first Uber ride you took, or the first time you bought something on Amazon. What about the first time you went to a concert? Probably quite a lot of you can remember the first time you went to a concert, I certainly do, Oasis in Milton Keynes, it was amazing. And you don't forget things. And if things go wrong, if we are going to go and see your team in the playoffs the first time in a generation and StubHub messes up and so you can't go, that's not like ordering something again the next day that you can get along with and you'll be fine with it in a week or two. You'll probably remember that for a very long time, maybe the rest of your life. So the stakes are very high, and the challenges are pretty huge as well, operating any marketplace. The good news is that AI is starting to show real promise in solving many of the problems that lead to these bad experiences. So today I'm going to talk to you about some of those challenges. What are the challenges that marketplaces like StarPub and others face? What are the opportunities that AI can start to solve within that space and across the supply chain? We're going to look at a few case studies, people who are doing a really good job already of using that technology for a great buyer and seller and platform experience. And finally, I'll try and tie that together into some lessons that hopefully you can take even if you don't work for a marketplace like me. you First of all, what is a marketplace? So, you know, marketplaces have been going since the beginning of time, since the beginning of human civilization. People have been buying goods from individual people in halls and marketplaces across the world. But increasingly, it's all about e-commerce. Nearly $4 trillion of goods and services were traded on marketplaces worldwide in 2024. Now, that figure is pretty dominated by some really large marketplaces in China, but in the US as well, you have about half a trillion, which is about a third of all e-commerce happening across the country. It's a big business, also growing, the shares growing. And even someone like Amazon now has reportedly something like two-thirds of the stuff you buy on Amazon is from third-party sellers. Whew. *sigh* But you can see some of the challenges that marketplaces face in the way they're perceived. This is just one data point, but I think it's pretty interesting. If you look at Trustpilot, which is like an online reviews aggregator, you'll see the marketplaces have an average score of two stars only, which is really poor. And like who would buy from a website that has two stars on the website? Compared to like three and a half stars for average e-commerce scores and four stars in general. So this just shows some of the perception and the issues you get because of how marketplaces are difficult to operate and how there's a real opportunity if you can operate a marketplace that is both big at scale and is also trustworthy and reliable every time. *sniff* So what are those challenges? There's a lot, but I'd try and bucket them into kind of three broad areas. Firstly, liquidity. So matching up of buyers and sellers. As I said before, you might have millions of different buyers and sellers around the world who are all trying to sell their goods or buy those goods. Matching them up is challenging. They also have very different interests in terms of what they're trying to get out of that transaction. The seller wants to get paid as much as possible, as quickly as possible. The buyer wants the best quality of good, as quick as possible, and at the best possible price. Matching them up at scale instantaneously is really tricky and needs constant validation. It's not like you're selling a book where you have a thousand in the storehouse and you're shipping them out one at a time. There's millions of different individual listings that once they're sold are gone. And so you have a real problem with real-time availability that can also impact on your pricing if you're doing it dynamically and intelligently. Secondly, fragmented supply. So it goes without saying from what I've said before, right, if you have all these different suppliers all around the world potentially selling goods and potentially selling tickets, it's hard to be sure they all have equal quality. both in terms of the actual physical good or service, or in terms of the, you know, the look and feel of what they're selling, but also in terms of the way they treat the buyer on the other end of the transaction. So I work in tickets, and that's a big variance in terms of [5:00] Someone might sell a ticket and immediately send it to the buyer and that buyer gets the ticket straight away and they have a great experience. Versus others might be lazy, they might forget to send it. For various reasons, they might send it at the last moment. Those two experiences are really, really different. And it's often very hard to know upfront which seller is going to give the former experience and which is going to give the latter experience. There's also a high fraud risk, right? You know, if people are selling goods online, there's an opportunity for people to think they can make a quick buck. On StarPub we only pay sellers after the event to avoid the risk of people uploading fake tickets because there's no incentive to do that financially. But it really does exist in a bunch of marketplaces. Bye. And finally, and I think arguably the most interesting one is the problem of incomplete data. So in a traditional business, you have pretty good visibility over your whole supply chain. You know where the good is. You know if it's been sent out. You know if the customer's got an issue with it and they reported it to you. With marketplaces, you often don't know that, right? You're relying on information and data from disaggregated people. On the ticketing side, for example, a lot of tickets these days are sent via Ticketmaster's app, where you transfer to a friend. We rely on sellers to tell us when that happens. We don't know. That's off our platform if they click that button, which makes it very hard for us to understand the dynamics on the buyer and seller side and make sure there's no friction in the experience that could stop the buyer going to the event. Thank you. Now, I'm hoping, given this is an AI conference, and because of some of the problems I've been lining up there, there's been light bulbs going off around the room in terms of the ways that AI can improve some of those problems. So I'm going to start to go through some of the things that are going to transform any marketplace. And I'm going to try and pull from a broad range of marketplace as well as my own experience at Stop Hub and try and roughly group them into a few themes I think would be helpful to take away. *sniff* Firstly, on the liquidity problem. The liquidity problem is obviously two-sided. You have the buy side of the equation, you have the sell side of the equation. Supply and demand, classic problem. And really, like both of them in terms of experience, start right from the moment the seller first lands on your site or onto your app and starts filling information in about the listing that they're going to create and try to sell to the buyers. 90% of the mistakes that are made that cause issues on the buyer side happen right up at that front point. either because the seller makes a mistake in listing information or because you'd fail to communicate policies or incentives to them that would mean they gave a good experience to the buyer at the end of the day. And two companies I think are doing really well with this and using AI, pretty interestingly, are eBay and Macari. So both of these companies try and take away the complexity of filling all that information in and allow you to upload a photo of the good or thing you want to sell. And it interprets what it is. It matches it to similar stuff that's been sold on the platform before. and useful information that can be used to buy it. So that one reduces mistakes, but it also improves the conversion on the buyer side, because it makes it much easier to match buyers with what they want to buy. If there's different languages on the seller and the buyer side, it can automatically resolve that in the way you couldn't really do before. And the descriptions are generally way higher quality than what you get by filling it in on your own, which means that you can improve conversion on the buy side when they go through the checkout process. And eBay, for example, quoted online as saying that 95% of the people who tried that feature kept the description exactly as it was from AI. So it saved like many minutes for the seller and reducing loads of mistakes down the funnel. On the buy side, there's also a ton of innovation happening. I was thinking back about one of my first experiences buying something on e-commerce, and it was buying a vintage band t-shirt. If you search a vintage band t-shirt 10 years ago, the only thing you'd see on most e-commerce platforms would be literally just things called vintage band t-shirt, Rolling Stones t-shirt, something else like that. I don't know if you had the same experience, but it was pretty bad. The searching experience was pretty awful. And AI and machine learning and other similar technologies are really unlocking some changes in that. I think Etsy is a really good example of this because the goods they sell are so diverse. They're like kind of a marketplace for crafts and individual kind of traders. Um, And which means they have a wide variety of things they're trying to sell, and that must be really hard to match buyers with the things they might be interested in. And what they're doing is using AI on both sides of the search equation. Firstly, getting a profile of who the buyer is, what they're looking for, from all the data points you can pull from what they're searching for, where they've come from, anything else you can gather right at the top of the funnel. And then they're bracketing different goods into what they call taste buckets. So things that are kind of broadly similar based on past experience with similar users based on what the buyer is trying to buy. And it's both showing the most relevant stuff, but also grouping it into things that are useful and easy to navigate. This is just an example of the eBay one, where you can see that the experience of selling a vinyl record has gone from multiple information, filling in loads of data, to just upload a Pick something that's kind of similar to what's been seen before. And then with a few clicks of a button, you are through and your item is listed and ready with a way better quality than [10:00] här beforehand. Tack. The second broad area I think there's kind of quite a lot of literature on but worth calling out is around kind of logistics of either kind of subtly adjusting supply and demand or finding a way to get things from A to B. Det är no surprise att se Uber listed här. Jag har ju en Uber Freight, det är en som jag finder att det är intressant här. Det är den här att de subtlyna införsning och demaner och matcher det tillsammans. Det har en cool feature där det är en heatmap som de ger till truck drivers. Det är att de är att de är trynna att göra en empty miles där någon är med nogen eller servis och ingenting i van. Och de har en heatmap att ta till dem där man ska gå till i varje de mest lika tillgör. Det är det som är leder av AI och real-time optimisation av de kinder som de är för. Det är för två reasons. One is att du är att subtlyna inflyt perception och insentiv för individuels. Så det är både bra produkt design och AI. And secondly, the broad range of data that they are putting into this system is super interesting. So it's not just where demand and supply is, it's like, what's the weather in this location? What's traffic look like? Really putting anything we can in to make this model more adaptive and more responsive. An interesting anecdote from StubHub even is we found that one of the best predictors for ticket price is actually the weather information. Because people don't want to go to see an MLB game when it's raining, they might go to the theatre. And that was kind of a surprising thing for us because we realised similar data that other companies are using, you can input into your systems. And you might be really surprised about the applications and the value you see from them. It's definitely worth trying. Tack. Det har det här deep red model, som jag tror är en bra namn, som du kan lära ut en ton av online. Det är som att använda AI i en super intelligent way till real-time optimise where du sender drivers till, de incentives du ger dem, de pricing du ger dem till optimise och att göra all tre providers i den relationen har en bra outcome. Tack. Finally, I'm calling trust and delight. I've waited till this slide to start talking too much about StartHub, because this is an area that I particularly care about in StartHub. I think it's so important that when you're engaging with live events, you have an amazing experience from the moment you buy a ticket. And one of the ways we're starting to see some real improvements is on the customer service side. And there's two applications I talk about. One is kind of obvious. We're using a clawed LLM model to actually answer customers, which at first we were worried would have some downsides. Vi trodde vi att vi hade haft en kost vi hade eller kontakt vi hade reducer. Vi hade inte väntat det för att förändra customer experience men det var. Vi hade precis som har hänt. Vi hade CSAT scores, som Perception av Ability och Customer Service 30% höger än vi hade för human agent. The other thing is auto-resolution. A lot of the problems that marketplaces face is about making decisions quickly in real time at massive scale. That's something that AI is absolutely perfect for. Tack till elever och personal vid Säkerhetssäkerheten. Second is in terms of fraud and verification. I could use Slapapp again here. We do have scanning technology to look at the tickets that get uploaded to our system, look for differences in barcodes, differences in what the seller's provided versus what's actually on the ticket. But one example I find really interesting is the RealReal, which is high-quality luxury goods sold on a marketplace format, which is obviously a real risk for fraud. You have big problems with Tack till elever och personal vid Säkerhetssäkerheten. A bunch of companies are doing it well. I want to try and pull out a few themes in terms of what is making these companies do well and what can you take forward from it. Tack. [15:00] Firstly, as I said, For me, so much of the value of AI is about that real-time decision-making ability. When there's things happening at scale and you've got humans in the loop making decisions, you should always be asking, are there ways we can get AI to do this in a safe way and a quality way? Because nine times out of 10, that's going to result in faster decisions and better decisions for you. Secondly, language and globalization. As a very like multinational company with Startup, this is something that we see really truly every day. Which is that there is no real reason we shouldn't be selling in every country in the world right now, because we can automatically translate everything. And that's even things you can do where on a marketplace you can connect buyers and sellers together to have conversations, to talk to each other, to engage with each other in both of their languages separately. And while constantly verifying there's no like fraud going on or no like bad behavior from the seller or buyer in a way that you never could have done before. *sigh* Thank you. Third, personalization. This is kind of like the Etsy example. But I think previously, like, in order to really understand your customer, you'd need them to come back and buy a few times to know what they liked, what they're looking for, what their preferences are. That's no longer true. You can learn a ton of stuff about people you can use as soon as they land on your site to give them a way more personalized and better experience. I'm not. And finally, we're at a tech conference, so I don't think I need to extol the virtues of A/B testing too much, but it is just such a massive unlock because it allows you to move quickly. You both know the impact of what you're doing, and you can also take more risks because you know what the downside will be quickly, and you can measure things, and you can iterate, and you can innovate. But to do all these things, you need a culture of AI, right? I'm hoping that everyone in this room is starting to build these in their organizations. But really a few things that I think have really set apart these companies and that I think would want to build in any company. First is just data, quality of data, right? If a process is happening in your system and you're not logging it, you're not recording it, I think the value of that in terms of dollars is going to way exceed the cost in terms of storage space. Just log it and see if it works and you can always delete it later if you need to. Second, innovate and take risks. The opportunity cost of building a new product has never been lower. You can do so much more now with the same number of engineers. So that means you should be trying more. The output should be more. You should be doing more things and taking more risks. And finally, I put this, I think it's an interesting one, like embracing marginal gains. Like one of the outputs of that fact that you can do more is you no longer need to build like the big showstoppers and that's all you can do. You can get huge value from the small things that maybe only 1% of customers see and are affected by, but those add up over time. Because you can do more, you can get those marginal gains, and you can continually drive a much better platform. So, that's that. Hopefully some interesting stuff you can take from this. In summary, I think marketplaces are hard, basically. They're tough. But AI is finding really great new ways to unlock that. And I'd love to talk with any of you over the course of the day about other ideas you have to solve to solve some of these problems, because they're super fascinating. And as I say, the end result for me at least is more people getting to go to live events, seeing the band with their dreams or seeing their team in the playoffs and things like that. So, thank you.