Nordstrom, Staples Pivot to Agentic Search as Shoppers Seek AI Context

消费者寻求AI上下文支持，Nordstrom与Staples转向智能体搜索 零售商必须升级底层数据架构并优化模型延迟，以捕获对话式搜索流量，防止被外部大语言模型去中介化。 来自CI&T、Nordstrom、Mirakl和Staples的行业专家探讨了AI如何从商品发现到售后支持，从根本上重塑电商客户旅程。他们指出，搜索模式正显著转向包含多重约束条件的对话式查询，并讨论了生成式引擎优化（GEO）与智能体电商的兴起。专家组强调，借助AI取得成功需要干净且统一的数据底座，控制响应延迟以防购物车放弃，并在自主智能体交互与消费者信任之间取得平衡。 CI&T 调研数据显示，74% 的美国消费者已使用 AI 辅助购物，93% 的受访者认为 AI 帮助他们做出了更明智的购买决策，并直接带动了 55% 的购买量增长。 (0:45) 消费者的站内搜索正从简短关键词演变为包含预算、品牌和搭配等多重约束的长句，零售商必须通过扩展商品广度并补充丰富的 UGC 与问答数据，以适配生成式引擎优化（GEO）。 (0:58) 零售商正在部署基于 Transformer 的时序推荐模型以实时捕捉点击流意图，并力求在自有平台上成为“首选 Agent”（Agent of Choice），防止外部大模型中介化品牌与用户的关系。 (17:00) 底层数据现代化已成为零售 AI 落地的首要瓶颈，企业若不解决长期沉淀的孤岛式和重复数据问题，前端对话式 AI 极易发生价格推荐失准和偏离核心业务逻辑的事故。 (21:58) 在逐步探索 Agent 间自主交易（Agent-to-Agent）的过程中，高响应速度是避免转化流失的关键，零售商正借助会员忠诚度体系维持黏性，并通过成熟生态伙伴加速规模化部署。 (33:04)

As consumers increasingly adopt conversational tools for product discovery, retail enterprises must overhaul legacy data architectures and recommendation systems to support multi-constraint shopping journeys. Retail and technology leaders argue that long-term commercial survival depends on becoming the direct 'Agent of Choice' on proprietary platforms before external AI intermediaries capture and monetize the customer relationship. The Evolution from Keyword Search to Generative Discovery (0:45) Consumer shopping behavior is transitioning from transactional keyword queries toward educational, multi-constraint conversational interactions. According to a CI&T consumer study cited by Paula Macaggi and Casey Golden, 74% of US consumers have used AI for shopping assistance, 90% have used or are open to using it, and 93% report feeling they make smarter purchasing decisions, corresponding to a 55% increase in purchases. On-site retail search reflects this shift: Nordstrom's Kunal Sonalkar notes that user queries have expanded from two- to three-word product lookups into five- to six-word conversational prompts expressing complex, multi-item intent, such as assembling complete event outfits with specific aesthetic or budget constraints. Generative Engine Optimization and Catalog Requirements As product discovery shifts toward large language models (LLMs) and referral traffic from engines like ChatGPT and Perplexity, Generative Engine Optimization (GEO) is replacing standard search engine optimization (SEO). Mirakl's James Williams and Staples' Sidharth Thapar emphasize that because conversational prompts compress the discovery journey—leading shoppers to view fewer overall items before hitting a product detail page—retailers must maintain both expansive catalog breadth and rich, unstructured product data. Systems require user-generated content, customer reviews, and granular technical context (such as identifying that an HP '26X' toner cartridge yields 9,000 pages versus 4,000 for a standard unit) to prevent automated discovery agents from filtering out their inventory. On-Site Conversion and the 'Agent of Choice' Converting incoming AI-referred traffic requires replacing static batch systems with real-time transformer-based recommendation architectures. Sonalkar notes that tracking sequential user actions, such as real-time clickstreams and basket additions, allows retailers to infer immediate intent and address cold-start challenges. For catalogs exceeding 500 SKUs, conversational search provides direct intelligence on customer phrasing and context. Data Modernization and Autonomous Retail Realities (21:58) Panelists caution that front-end conversational interfaces fail without comprehensive data modernization. Golden states that enterprise investments are predominantly focused on consolidating decades of siloed, duplicate legacy data into unified systems. Williams illustrates the danger of deploying raw models without retail merchandising constraints: a kitchenware brand offloading search to a hyperscaler saw its AI recommend only 10% of true bestsellers and suggest $400 items for sub-$200 wedding gift queries. Autonomous retail demands verified pricing, live inventory, and validated substitution logic. Agent-to-Agent Commerce and Post-Purchase Trust (27:35) The panel identifies routine replenishment (such as grocery and CPG) and post-purchase customer support as immediate operational frontiers for AI agents. Thapar highlights implementations like Home Depot's conversational post-purchase support, which uses knowledge graphs to determine part compatibility and order statuses. In forward-looking agent-to-agent transactions, Williams points out that customer loyalty programs (such as Sephora's reward ecosystem) remain a powerful anchor even when products are widely accessible across other digital storefronts, while payment networks like Visa establish authorization protocols for machine-driven transactions. Because underlying LLM models are rapidly commoditizing, panelists conclude that retail differentiation relies on compute governance, data cleanliness, and latency reduction to avoid driving cart abandonment. An immediate operational test for the industry will occur in post-holiday periods, where return rates should logically decrease if AI-assisted shopping genuinely delivers more informed purchasing decisions.

在零售行业迈向智能体商务的过程中，生成式AI正促使消费者从传统搜索转向以意图为导向的对话式购物，这要求零售企业必须夯实底层商品数据基础并重构站内转化链路。来自CI&T、Nordstrom、Mirakl与Staples的专家指出，零售商需在应对外部大模型流量分流的同时，成为消费者首选的自有智能体，并通过低延迟架构与忠诚度机制保障商业闭环。 消费者行为演变与全渠道搜索重构 (0:45) CI&T的调研数据显示，74%的美国消费者已在购物流程中使用过AI辅助，90%的消费者已经使用或愿意使用该技术；93%的用户认为AI辅助使其购买决策更加明智，进而促成购买量提升55%。随着生成式工具的普及，用户的搜索习惯由过去简短的2至3个关键词，逐步转向包含5至6个甚至更多词汇的复杂自然语言提示词。例如在Nordstrom平台，消费者倾向于输入完整场景意图（如搭配整套婚礼着装），并在反馈结果中针对鞋履或配饰提出细粒度的局部修改需求。Staples的实践也显示，多重预算约束、品牌偏好与风格限定正迫使零售引擎处理深层次的语义逻辑。 生成引擎优化与商品数据基础设施 (0:08) 生成引擎优化（GEO）正在重塑流量漏斗顶端的商品曝光逻辑。Mirakl指出，消费者在跳转至商品详情页（PDP）前浏览的项目数量正在减少，这使得零售商的商品广度与非结构化富文本内容（如用户生成内容UGC、评价与问答数据）成为关键排位因素。例如在办公用品搜索中，系统必须理解“26X”墨盒型号代表9,000页的高打印量，而非仅输出基础标题。与此同时，CI&T强调绝大多数企业投资仍集中于核心数据架构的治理，零售商若未整合过去25年分散孤立的POS及商品主数据，直接调用顶层模型极易失误。Mirakl援引案例称，某厨具零售商直接接入超大规模语言模型后，仅有10%的推荐契合热销榜单，且在200美元以下的礼品搜索中错误给出了均价400美元的建议。 站内转化机制与自主智能体地位确立 为有效承接ChatGPT和Perplexity等外部平台引入的高意图推荐流量，零售商正在升级站内转化体系。Nordstrom采用基于Transformer架构的时序推荐模型替代隔夜批处理算法，通过实时捕捉点击流、加购及结账动作，即时区分冲动消费与长期调研行为。同时，CI&T建议拥有500个以上SKU的站点以自然语言交互替换僵硬的层级导航，从中沉淀消费场景语境。针对埃森哲提出的“首选智能体（Agent of Choice）”理论，Mirakl指出零售巨头必须在自有站点搭建对话式体验，避免客户关系被第三方大模型中介化，同时防范未来外部引流渠道高额商业变现带来的利润侵蚀。 智能体间交易、售后透明度与运营落地 (27:35) 在自动化交易方面，B2B采购与日常快消品补货成为智能体间（Agent-to-Agent）商务的早期试验场。尽管外部比价智能体活跃，但以Sephora为例，75%至80%的在售品牌虽可在其独立站直接购买，忠诚度积分机制仍促使买方智能体优先选择聚合平台；Visa等金融机构也在构建相应的授权与防欺诈协议。在售后端，类似Home Depot的知识图谱正在取代传统语音菜单以自动诊断零部件兼容性。CI&T提醒，品牌必须明确标示AI身份以维护用户信任，因为价格偏差或履约失误的责任将完全由品牌承担。Staples与Mirakl最后警示，模型本身已日趋同质化，推理延迟是导致弃购的关键因素，企业应依托成熟合作伙伴构建低延迟、重治理的轻量化系统，避免盲目全盘自研。 总体而言，生成式AI正将电商竞争核心从表层流量获取转向深层数据治理、履约确定性与忠诚度生态的建设。然而，跨平台实时库存同步、动态定价防抓取风控，以及智能体间交易结算的标准协议制定，仍是当前零售生态全面迈向自动化阶段悬而未决的技术与制度瓶颈。

## Background 本内容源自 Ai4 于 2026 年 8 月 22 日在 YouTube 平台发布的研讨会视频《从点击到转化：AI 如何重塑电商漏斗》。视频由《OFFBounds》主持人 Paula Macaggi 主持，邀请了来自 CI&T、Nordstrom、Mirakl 和 Staples 的四位零售及技术专家，共同讨论 AI 在获客、转化、留存以及智能体电商（Agentic Commerce）等环节的落地应用与未来趋势。 ## TL;DR 研讨会深入探讨了 AI 如何重塑消费者购物行为，指出消费者正从追求娱乐转向寻求教育与理性决策，长尾和场景化搜索已成为新常态。专家们剖析了企业从传统 SEO 向生成式引擎优化（GEO）演进的必要性，强调了高质产品数据与真实 UGC 对 LLM 推荐的重要性。同时，嘉宾们也指出了零售业面临的数据质量差、推荐脱节等挑战，并展望了未来 Agent-to-Agent 自动化交易与大规模个性化服务（Clienteling）的前景。 ## Atmosphere & style 这场讨论专业且兼具前瞻性，几位来自零售与技术前沿的专家对谈氛围热烈、言之有物。既有针对行业痛点（如零售数据如同“烂摊子”）的犀利洞察与务实吐槽，也有对智能体假日购物季和未来五年商业形态的积极展望，整体节奏明快且富有启发性。 ## Key takeaways - [04:38](https://www.youtube.com/watch?v=HjtvFQvCGnE&t=278s) CI&T 的报告显示，74% 的美国消费者已使用过 AI 辅助购物，90% 对此持开放或体验态度；93% 觉得自己的购买决策比两年前更明智，这直接推动增购了 55% 的商品。 - [04:38](https://www.youtube.com/watch?v=HjtvFQvCGnE&t=278s) 消费者的购物心理模式正在从单纯追求“娱乐”转向寻求“教育与信息”，AI 工具能够有效提升客户的长期价值（LTV），而不止于单纯的获客。 - [06:36](https://www.youtube.com/watch?v=HjtvFQvCGnE&t=396s) 搜索模式正在经历剧烈演变，用户搜索词变长，并从简单的“单品检索”转变为包含多重约束条件的“整套方案与综合场景搜索”。 - [10:01](https://www.youtube.com/watch?v=HjtvFQvCGnE&t=601s) SEO 正在向 GEO（生成式引擎优化）演变，为了让 LLM 在探索阶段精准匹配，零售商需要具备庞大的品类规模，并将真实的 UGC 与高分评价作为对话式属性融入产品数据源中。 - [15:22](https://www.youtube.com/watch?v=HjtvFQvCGnE&t=922s) 在即将到来的“智能体假日购物季”中，零售业绩将呈现极化分布（大赢或大输），而假期的退货率变化将是检验 AI 是否真正帮助用户做明智决策的关键验证指标。 - [17:38](https://www.youtube.com/watch?v=HjtvFQvCGnE&t=1058s) 站内推荐系统正从静态批处理转向基于 Transformer 的序列建模，以实时捕捉用户交互历史，精准分辨冲动消费与理性购物。 - [17:38](https://www.youtube.com/watch?v=HjtvFQvCGnE&t=1058s) 大型零售商面临着成为“首选智能体”（Agent of Choice）的战略窗口期，需抓住外部 LLM 导流的红利期，将客户转化为自身 Agent 的忠实用户。 - [23:12](https://www.youtube.com/watch?v=HjtvFQvCGnE&t=1392s) 零售业的数据目前普遍存在割裂、冗余及缺乏上下文的痛点；数据现代化是所有上层“华丽 AI 应用”落地生效的前提基础。 - [24:45](https://www.youtube.com/watch?v=HjtvFQvCGnE&t=1485s) “好的零售依然是好的零售”，AI Agent 的推荐如果忽视了商品企划（如畅销品匹配）与基础价格约束，就无法带来真正的转化提升。 - [27:35](https://www.youtube.com/watch?v=HjtvFQvCGnE&t=1655s) 售后体验正在通过语音 Agent 与智能客服重塑，建立完善的产品知识图谱是精准回答复杂商品细节与兼容性问答的技术核心。 - [30:12](https://www.youtube.com/watch?v=HjtvFQvCGnE&t=1812s) AI 将信任机制拉到了购买决策的前端，品牌方必须确保信息与价格的极高准确性，否则会直接损害消费者对品牌的信任。 - [32:58](https://www.youtube.com/watch?v=HjtvFQvCGnE&t=1978s) Agent-to-Agent（A2A）商务正在经历基础设施建设，高频标准化品类（如生鲜杂货）及 B2B/CPG 供应链将率先实现自动化交易。 - [37:21](https://www.youtube.com/watch?v=HjtvFQvCGnE&t=2241s) 零售商需要建立 Agent 订阅机制与数据防护护栏，防止外部 Agent 恶意抓取价格或诱导客户流失，同时确保全渠道价格一致性。 - [38:54](https://www.youtube.com/watch?v=HjtvFQvCGnE&t=2334s) AI 规模化落地中模型已成为通用商品，零售商应优先优化响应延迟、完善产品内容数据，并避免重复造轮子。 - [43:18](https://www.youtube.com/watch?v=HjtvFQvCGnE&t=2598s) 专家对未来 3 至 5 年智能体电商的前景看法不一：高频生鲜品类易实现 A2A 自动化交易，但全自动交易的全面普及仍面临挑战，同时专属客户服务（Clienteling）将迎来大规模个性化普及。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=HjtvFQvCGnE&t=4s) 研讨会开场与嘉宾背景介绍 小组讨论在展会第二天的尾声拉开帷幕，主持人与四位来自零售、电商平台及技术服务领域的专家共同探讨 AI 如何重塑电商漏斗的全流程。各位嘉宾分别介绍了各自的技术背景及当前企业最核心的 AI 策略与业务目标： * **Paula Macaggi**（主持人）：零售访谈节目《OFFBounds》的创始人兼主持人，每周与行业领袖深度对话，观察到过去一年“智能体 AI”（Agentic AI）迅速渗透行业，并深刻改变了消费者行为。 * **Casey Golden**（CI&T 北美零售与消费品主管）：介绍 CI&T 作为 AI 赋能与数字化转型合作伙伴，正帮助品牌将 AI 构想落地为具体解决方案。她强调，当前最核心的商业目标是实现 AI 商业模式的收益变现。由于传统的按工时与材料计费模式（Time and Materials）无法适配 AI 带来的超高效率（过去需要 6 个月完成的项目，现在仅需 6 周），公司正转向与结果挂钩的新型商业模式，确保高效交付的同时保证业务产出质量。 * **Kunal Sonalkar**（Nordstrom 数据科学家）：在 Nordstrom 工作超过五年，此前深耕搜索相关性与推荐系统，近期转向广告相关性领域，致力于为用户匹配最相关的广告内容。 * **James Williams**（Mirakl 代理电商主管）：Mirakl 是一家总部位于法国的技术公司，为全球绝大多数大型企业级零售商提供 Marketplace 平台技术支持。他提到坐在台上的 Nordstrom 正是 Mirakl 的核心客户之一。Mirakl 当前的战略重点是确保品牌商与零售商双端均做好迎接“代理电商”（Agentic Commerce）的准备，不仅要在当下的“探索阶段”（Discovery）奠定基础，更要为未来可能发生的“自动化交易”（Transactions）做好架构准备。 * **Sidharth Thapar**（Staples AI 经理兼高级应用科学家）：拥有超过十年的 AI 从业经验，专注于推荐、搜索及定价系统。他指出，目前企业级零售商最关注的核心诉求是：在跨平台交互中为客户提供一致的体验，同时建立对自营产品更深层次的理解，并将其无缝贯穿至各个业务系统之中。 --- ### [04:38](https://www.youtube.com/watch?v=HjtvFQvCGnE&t=278s) 消费者行为的深刻变革：从娱乐导向转向理性决策 主持人向现场观众发起举手互动，询问过去 30 天内是否曾使用 LLM 工具（如 ChatGPT、Claude、Perplexity 等）辅助购物，现场不少观众举手示意。随后，嘉宾们针对消费者行为的变化展开了深入探讨： 根据 CI&T 最新发布的报告，高达 **74% 的美国消费者已经使用过 AI 辅助购物**，而 **90% 的消费者对此持开放态度或已经体验过**。这一数据揭示了传统电商模式向现代化购物体验的巨幅跨越。 针对这份报告，Casey Golden 共享了几个震撼的洞察数据： * **93% 的消费者认为自己现在的购买决策比两年前更明智**。这种“决策明智感”直接驱动消费者增购了 **55%** 的商品。 * 与传统的社交媒体、Marketplace 等新渠道仅能带来“获客”（Customer Acquisition）不同，AI 工具正在显著提升客户的**长期价值（LTV）**。 * 消费者的心理模式正发生根本性转变：**从寻求娱乐（Entertainment）转向寻求教育与信息（Education）**。由于感到信息充分、决策精准，用户的购买频次和数量均有明显提升。 --- ### [06:36](https://www.youtube.com/watch?v=HjtvFQvCGnE&t=396s) 搜索模式演进：从单品检索到复杂综合场景 从零售商的实际数据来看，消费者的搜索行为正在发生质的飞跃： 1. **搜索词长度显著增长**：Nordstrom 的 Kunal Sonalkar 透露，过去用户的搜索极其简短，通常仅由 2 到 3 个词组成；而现在，5 到 6 个词的长尾搜索已成为常态。 2. **由“单品搜索”转变为“整套方案/场景搜索”**：随着对话式搜索的普及，用户不再单搜索某一件衣服，而是直接输入综合诉求（例如：“我想找一套参加婚礼当天的整体穿搭”）。 3. **高度个性化的实时微调**：在生成式交互中，用户可以针对推荐结果进行自由裁剪与替换——保留连衣裙但替换鞋子，或者保留整套穿搭但更换配饰。 针对这一趋势，Staples 的 Sidharth Thapar 补充道，现代搜索正变得极其复杂且充满限制条件。用户在寻找特定商品（如婚纱）时，往往会同时带入预算上限、特定品牌偏好等极其细化的约束条件。零售商不仅是在应对这种复杂性，更是在从底层基础架构上全面重构，以承接这种高度细分且具象化的需求。 --- ### [10:01](https://www.youtube.com/watch?v=HjtvFQvCGnE&t=601s) 探索阶段的变革：GEO 兴起与产品数据的深度重构 随着用户提问习惯的改变，传统搜索引擎优化（SEO）正在向生成式引擎优化（GEO）演变。在商品进入用户视线之前，AI 正在彻底改变产品探索（Product Discovery）的路径： Mirakl 的 James Williams 强调，在由 LLM 或 Instagram 等前端引发的探索体验中，用户在进入产品详情页（PDP）之前所能看到的商品数量急剧减少。这就意味着两件事情变得至关重要： 1. **极度严格的筛选条件与品类规模（Catalog Size）**：用户会带入更复杂的自然语言或严苛条件（如“能够保护踝关节的长跑鞋”）。零售商如果想在搜索层级变深时依然保持相关性，庞大的商品品类与供给规模是不可或缺的底座。 2. **对话式属性与真实 UGC 沉淀**：传统的产品数据整理（仅将描述中的属性转为键值对）已无法满足需求。现代 LLM 需要读取真实的“用户生成内容”（UGC）、高分评价（Reviews）以及问答（Q&A）。将这些“对话式属性”深度融合进产品数据源（Product Feeds）中，才能让 LLM 真正理解何为“好评如潮”或“护踝效果好”。 Staples 的 Sidharth Thapar 举例印证了这一点：以惠普 26X 墨盒为例，“26X”代表高容量（可打印 9,000 页，而普通版仅 4,000 页）。如果产品数据中没有将这种深层上下文（Context）提取并标明，AI 代理就无法在用户寻找“耐用/高页数墨盒”时精准匹配。 Nordstrom 的 Kunal Sonalkar 透露，自 2026 年 2 月起，Nordstrom 已经能从 ChatGPT 和 Perplexity 等平台看到明显的引流数据。LLM 将商品以精美的网格展现给用户并引导至官网。虽然目前大部分流量仍来自传统渠道（如 Google 和 Meta 广告），但他确信，随着 LLM 在购物场景下的进一步成熟（解决库存实时性、尺码准确度等问题后），流量格局将迎来彻底洗牌。 --- ### [15:22](https://www.youtube.com/watch?v=HjtvFQvCGnE&t=922s) 行业分化与迎接“智能体假日购物季” 针对零售商的应对现状，CI&T 的 Casey Golden 表达了非常犀利的观点： * **极少数零售商为即将来临的“智能体假日购物季”（Agentic Holiday Season）做好了准备**。 * 传统零售业绩看“红黑榜”（亏损或盈利），但在 AI 重塑的当下，业绩不会出现“平淡盘整”的状态，只有“大赢”或“大输”，这一切都会直接体现在 2027 年 1 月的 P&L 财务报表中。 * **退货率（Returns）将成为关键验证指标**：93% 的消费者认为自己做出了更明智的购买决策。每年 1 月都是零售业的退货高峰期，如果 AI 确实帮用户做出了更精准的匹配，今年假日季之后的退货率变化将直接检验这一假设。 ``` 传统探索模式 vs. 智能体电商探索模式 [ 传统模式 ] 用户 -> 简短关键词 -> 谷歌/站内搜索 -> 逐个点击筛选 -> PDP详情页 [ 智能体模式 ] 用户 -> 复杂场景/多约束条件 -> LLM/Agent (调用丰富UGC与品类) -> 组合推荐/微调 -> PDP详情页 ``` --- ### [17:38](https://www.youtube.com/watch?v=HjtvFQvCGnE&t=1058s) 站内转化与推荐系统：从静态批处理到 Transformer 序列建模 当用户通过外部 LLM 或搜索被引导至零售官网后，AI 正在深度重构站内的考虑（Consideration）与转化（Conversion）链路： 1. **推荐系统的实时进化**： Nordstrom 的 Kunal Sonalkar 介绍，传统的推荐系统多依赖于夜间运行的批处理任务（Nightly Batch Jobs），呈现的往往是静态推荐，且深受“冷启动”难题困扰。如今通过引入基于 Transformer 的模型，能够实时捕捉和序列化建模用户的交互行为（加购、点击、浏览历史），精准判断用户究竟是“冲动消费”还是“经过长期调研的理性购买”，进而输出极高上下文相关性的实时推荐。 2. **对话式商务（Conversational Commerce）与站内 AI 搜索**： Casey Golden 指出，对话式商务经历了数轮起伏，如今以“AI 购物助手/产品专家”的形式卷土重来。对于拥有超过 500 个 SKU 的零售商，直接引入 AI 搜索，让用户像对待 LLM 一样自然提问，具有极高的战略价值： * **提升用户体验**：用户无需在左侧分类栏中逐个勾选复杂的筛选按钮。 * **沉淀反向洞察**：零售商能直接获取用户描述产品的真实语言、场景与上下文（如：是用于生日派对还是职场通勤），从而反哺营销与选品。 3. **做“智能体的选择”（Choice of Agents）还是“首选智能体”（Agent of Choice）**： James Williams 引用了埃森哲的一个核心观点：零售商面临着一个选择与一个必然。对于像 Nordstrom 或 Staples 这样拥有强大品牌心智与自营品类的巨头，完全有实力成为用户的**“首选智能体”（Agent of Choice）**。零售商应当抓住目前的窗口期，将通过外部 LLM（如 ChatGPT）引流进来的客户转化为自身 Agent 的忠实用户，通过营销引导他们在自家平台上发起下一次购物。毕竟，目前 LLM 平台尚未开始对导流大规模收费，但这一红利期不会持续太久。 ### [21:58](https://www.youtube.com/watch?v=HjtvFQvCGnE&t=1318s) 超越聊天机器人的 AI 驱动转化优化 Sidharth Thapar 指出，像 Walmart 已经支持直接在 ChatGPT 上进行销售与下单，但对于典型零售场景而言，生成式 AI 的引入彻底改变了决策机制。过去是由消费者自行评估商品，而现在则是由 AI Agent 代表消费者进行评估与筛选。 这种转变对零售商的基础运营提出了更严苛的要求： * **精准定价**：价格数据必须更加准确。 * **替代方案与库存追踪**：对库存的监控必须极具前瞻性。一旦出现缺货，系统必须能实时推荐合规的替代产品。 * **决策影响力**：AI 交付给 Agent 的推荐结果会直接决定最终购买行为，因此后端数据的实时性与准确性至关重要。 --- ### [23:12](https://www.youtube.com/watch?v=HjtvFQvCGnE&t=1392s) 零售业的数据“烂摊子”与底层重构 Casey Golden 深入剖析了当前零售行业在落地 AI 时的现实困境。虽然零售商非常喜欢追逐各种“华丽”的新概念（如搜索优化、对话式商务等），但目前几乎所有品牌和零售商签约的核心项目都是**数据现代化**（Data Modernization）。 > “任何在零售业工作超过六个月的人都知道，我们的数据就是一个‘烂摊子’（dumpster fire），你根本无法在其之上构建智能系统。” 零售业数据的核心痛点包括： * **严重割裂与冗余**：数据孤岛现象严重，存在大量重复信息。 * **缺乏上下文**：历史数据缺乏上下文，甚至有长达50年的 POS 线下数据完全没有数字化上下文支持。 得益于新一代 AI 工具，曾经被认为需要 25 年才能解决的数据治理难题，现在正在以惊人的速度被破局： * 品牌能够将庞大的全球数据进行整合、统一与本地化，项目周期从几年缩短至不到一个季度（甚至 8 周以内）。 * 夯实数据基础是所有“华丽 AI 应用”能够真正发挥作用的前提条件。 --- ### [24:45](https://www.youtube.com/watch?v=HjtvFQvCGnE&t=1485s) 传统零售基本功在 AI 时代的适用性 针对聊天机器人及辅助搜索的落地实践，James Williams 分享了一个关于某知名厨房用品零售商的测试案例。该零售商将其网站的对话 Agent 完全外包给大型云服务厂商（Hyperscalers，如 Gemini、AWS 等）。在通过合成 Prompt（Synthetic Prompts）测试厨刀等大类目时，发现了明显的体验断层： ``` [ 用户的具体需求 ] --(如: 婚礼礼物 <$200) | v [ AI 推荐系统 (大模型Agent) ] | +---> 问题 1: 仅 10% 推荐商品为畅销品 (商品排布与商品企划脱节) | +---> 问题 2: 前 5 个推荐卡片均价 $400 (完全无视 <$200 价格约束) ``` 这一案例揭示了当前 AI 落地中的常见缺陷： * **畅销品匹配度低**：在推荐的产品卡片中，只有 10% 是该类目的畅销品（Bestseller）。这要么说明零售商的商品企划（Merchandising）团队一直选错了畅销品，要么说明 AI 根本没有理解商品企划的核心洞察。 * **忽视基础价格约束**：当用户明确提出“作为婚礼礼物，预算不超过 200 美元”时，系统推荐的前 5 款产品平均价格却高达 400 美元。 James Williams 强调，“好的零售依然是好的零售”（Good retail is still good retail）。正确的地点、正确的商品、价格、库存可用性及物流时效等传统零售要素依然没有改变。企业在应用 AI 时，必须优先做好数据丰富度（Product Enrichment）与用户体验实测（User Testing）。 --- ### [27:35](https://www.youtube.com/watch?v=HjtvFQvCGnE&t=1655s) 售后体验重塑与知识图谱的价值 话题延伸至购买后的环节（Post-Purchase），包括客户服务、忠诚度与个性化。Kunal Sonalkar 分享了语音 Agent 和自动化聊天机器人在客服领域的快速演进： * **传统与现代客服的对比**：以往的电话客服依赖于复杂的按键菜单（“按 1 选择这个，按 2 选择那个”），而现在像 Home Depot 等大厂已彻底转变了这一范式。用户体验如同与真人对话，Agent 能够精准回答复杂的商品细节与描述。 * **技术支撑**：这种智能交互的关键在于建立完善的**产品知识图谱**（Knowledge Graphs）。例如当顾客询问“我买的这个零部件能否适配我家里现有的某型号设备”时，知识图谱能够提供深度的兼容性判断并给出准确建议。 --- ### [30:12](https://www.youtube.com/watch?v=HjtvFQvCGnE&t=1812s) 消费信任移到决策前端与品牌责任 Casey Golden 提出，在 AI 介入购物全流程的背景下，构建消费者与零售商之间的**信任**（Trust）成为了贯穿全流程的主题： * **透明度与控制权**：消费者希望获得透明度，明确知道自己是在与 AI Agent 对话。如果用户以为在跟真人沟通，事后却发现是 AI，会产生被冒犯或被操纵的感觉（类似于看到由 AI 生成的宠物情感视频而流泪后的失落感）。 * **信任的时间节点前移**：过去，信任建立在售后阶段（产品质量过关、符合预期、按时送达）；而现在，AI 将信任机制直接拉到了**购买决策的前端**。 * **信息准确性**：消费者要求 Agent 提供极高的数据准确性。如果在 ChatGPT 或 Gemini 上看到某个价格，点击进入链接后却发现价格不一致，信任就会迅速瓦解。最终承担品牌承诺责任的不是 LLM 或 AI 技术本身，而是品牌方自身。 --- ### [32:58](https://www.youtube.com/watch?v=HjtvFQvCGnE&t=1978s) Agent 对 Agent (A2A) 商务的未来趋势与基础设施 随着技术演进，购物场景正逐步从“消费者使用 AI”向“Agent 代替消费者自主购物”（Agent-to-Agent Commerce）过渡。针对这一转变，专家们探讨了品牌如何应对： ``` [ 消费者 / 买方 Agent ] <---> [ 验证与支付层 (Visa/PSPs) ] <---> [ 商家 / 卖方 Agent ] | | (忠诚度/偏好记忆) (精确价格/库存/B2B预测) ``` 1. **实体与 Agent 购物并行**：James Williams 认为，涉及高情感体验的品类依然会保留实体店和网页 PDP（产品详情页），但高频标准化品类（如生鲜杂货 Grocery）会率先实现自动化。 2. **忠诚度机制依然有效**：以 Sephora 为例，其销售的品牌中有 75%–80% 属于 Shopify 品牌（消费者本可以直接从品牌官网购买，甚至更便宜），但用户依然选择 Sephora，核心在于忠诚度计划（Loyalty）。用户的 Agent 同样会记录用户的忠诚度偏好，并在决策中保留这些激励机制。 3. **基础设施建设**： * **支付与防欺诈**：如 Visa 正在针对授权 Agent（Authorized Agents）开发基础设施，以确保 Agent 间交易的安全与可信。 * **人机回环（Human in the Loop）**：在全面自动化之前，许多场景仍需要人类进行二次确认（例如点击 iPhone 完成最终 Checkout），如何优化这一环节而不阻碍转化率是关键。 4. **率先落地的场景**：Casey Golden 指出，Agent 到 Agent 的模式目前已经在 **B2B 和 CPG（快速消费品）供应链**（如经销商订单自动化、需求预测）中进入实际生产环境，随后将扩展至生鲜杂货，最后才是服装或房产等高决策成本品类。 --- ### [37:21](https://www.youtube.com/watch?v=HjtvFQvCGnE&t=2241s) 零售商应构建的 AI 核心能力与合规防线 针对 Agent 互动的普及，零售商需要在网站与后端搭建全新的能力： * **订阅与需求预测**：Sidharth Thapar 建议建立 Agent 对 Agent 的订阅机制，通过获取周期性需求更好地规划供应链与需求预测。 * **AI 护栏与数据防抓取**：必须建立严密的安全护栏，防止外部 Agent 恶意抓取价格数据或通过极微小的价格差（如便宜 2 美分）诱导用户流失。 * **跨渠道一致性**：无论是线上官网、线下门店还是第三方平台，必须保证价格与内容的一致性。如果 Agent 或消费者发现同一商品在不同渠道的价格不一，信任将不复存在。 --- ### [38:54](https://www.youtube.com/watch?v=HjtvFQvCGnE&t=2334s) AI 规模化落地（Operationalizing at Scale）的建议 讨论最后，专家们就如何将 AI 从实验室试点（Pilots）扩展到企业级规模给出了具体建议： * **模型已成为通用商品（Models as Commodities）**：Kunal Sonalkar 指出，现在不需要盲目追求更大或更复杂的模型。零售商的重点应放在**产品内容的准备度**、系统的高可用性以及借鉴成熟的软件工程规范（如治理、LLM 护栏和算力管理）。 * **控制延迟（Latency）**：对于搜索、推荐和聊天机器人，响应延迟至关重要。过高的延迟会导致极高的放弃率（Abandonment）。切忌在生产环境中使用过于复杂而牺牲了响应速度的大模型。 * **善用生态，避免重复造轮子**：James Williams 建议零售商不要试图完全自建（Build It Yourself）所有 Niche 工具。市场上有很多专注于特定领域的优秀供应商，寻求合作能大幅降低落地风险。 * **领导力与战略清晰**：Casey Golden 总结道，零售商在扩大 AI 规模时，首先要明确自己是否在正确的方向上进行扩展，领导层需要具备超越传统财务 P&L 的零售前瞻视野。 ### [42:43](https://www.youtube.com/watch?v=HjtvFQvCGnE&t=2563s) 守护品牌灵魂与数据基石下的商业梦想 讲者强调，品牌在拥抱新兴科技以实现规模化增长的同时，核心在于“保护品牌在消费者心中的灵魂与完整性”（"protect the soul of the integrity of what that brand means to their consumers"）。技术是放大器，但前提是必须正确地传达品牌价值，并且拥有坚实的数据基础（data foundation）。 在当今的技术环境下，“没有什么是不可能的，一切都变得可以负担且切实可行”。只要品牌清晰知道自己想要实现什么，就能将最狂野的购物遐想变为现实。 - **品牌核心**：坚守品牌灵魂与消费者认同，避免技术规模化稀释品牌价值。 - **技术可行性**：强大的数据基础让过去高昂的 AI 方案变得触手可及。 --- ### [43:18](https://www.youtube.com/watch?v=HjtvFQvCGnE&t=2598s) 展望五年未来：智能体电商与个性化服务 在论坛接近尾声、仅剩两分钟时，主持人 Paula Macaggi 向嘉宾提出了一个快速闪电问答：假设五年后大家再次相聚在 Ai4 论坛，智能体电商（Agentic Commerce）会变成什么样？ 嘉宾们给出了截然不同但极具启发性的“热辣观点”（hot take）： 1. **Sidharth Thapar (Staples Inc.)** 认为，在日常高频的重复性消费品类（如杂货、生鲜 groceries）中，完全可以实现“智能体对智能体”（agent to agent）的自动对接与交易；但在探索与发现（discovery）环节，他依然希望保留人类自主购物和逛街的乐趣。 2. **James Williams (Mirakl)** 提出了相当尖锐的见解：五年之后，大家可能依然在讨论“探索（discovery）”阶段的 AI 应用，而未能真正普及全自动的“交易（transactions）”。 3. **Casey Golden (CI&T)** 则展望了更具深度的服务模式，她预测在 3 年内，规模化的个人定制体验（即客户关系维护/Clienteling）将彻底普及。她还风趣地吐槽道，Clienteling 这种商业概念自 1909 年起就已经在实际运营中存在，也许 3 年后《韦氏词典》（Merriam-Webster）终于能把它正式收录进去。 ``` +-----------------------------------------------------------------+ | 未来五年智能体电商 (Agentic Commerce) 展望 | +-----------------------------------------------------------------+ | +------------------------+------------------------+ | | | v v v 【Sidharth Thapar】 【James Williams】 【Casey Golden】 - 高频复购(如生鲜): - 行业仍将聚焦于: - 3年内普及大规模 Agent-to-Agent 自动下单 商品探索 (Discovery) 个性化服务 (Clienteling) - 探索探索(Discovery): - 自动交易(Transactions)- 呼吁词典正式收录 保留人类逛街乐趣 仍需更长时间突破 百年商业术语 ``` - **高频重复消费**：生鲜杂货类等标准品易于被 AI Agent 替代决策与自动下单。 - **探索与交易之争**：智能体在帮助用户筛选探索上进展迅速，但完全替代决策的自动交易仍面临信任与体验门槛。 - **Clienteling 规模化**：借助 AI 技术，原本仅限于高端零售的专属客户服务（Clienteling）将实现大规模自动化与个性化呈现。 --- ### [44:48](https://www.youtube.com/watch?v=HjtvFQvCGnE&t=2688s) 会议结语与后续议程 主持人 Paula Macaggi 对各位专家嘉宾的分享表示感谢，并特别推荐了嘉宾 Casey Golden 的优秀 Podcast 节目。由于时间限制，现场无法安排 Q&A 环节，但 Paula 提示观众，接下来即将进行关于 Mars（玛氏）在 CPG（消费品）领域智能体电商应用的相关对话，并鼓励大家在接下来的大会期间与各位 panel 嘉宾交流互动。 ## Notable quotes > [04:38](https://www.youtube.com/watch?v=HjtvFQvCGnE&t=278s) 93% 的消费者认为自己现在的购买决策比两年前更明智。 — Casey Golden > [04:38](https://www.youtube.com/watch?v=HjtvFQvCGnE&t=278s) 从寻求娱乐（Entertainment）转向寻求教育与信息（Education）。 — Casey Golden > [23:12](https://www.youtube.com/watch?v=HjtvFQvCGnE&t=1392s) 任何在零售业工作超过六个月的人都知道，我们的数据就是一个‘烂摊子’（dumpster fire），你根本无法在其之上构建智能系统。 — Casey Golden > [24:45](https://www.youtube.com/watch?v=HjtvFQvCGnE&t=1485s) 好的零售依然是好的零售（Good retail is still good retail）。 — James Williams > [42:43](https://www.youtube.com/watch?v=HjtvFQvCGnE&t=2563s) 保护品牌在消费者心中的灵魂与完整性（protect the soul of the integrity of what that brand means to their consumers）。 — Casey Golden > [42:43](https://www.youtube.com/watch?v=HjtvFQvCGnE&t=2563s) 没有什么是不可能的，一切都变得可以负担且切实可行。 — Casey Golden > [38:54](https://www.youtube.com/watch?v=HjtvFQvCGnE&t=2334s) 模型已成为通用商品（Models as Commodities）。 — Kunal Sonalkar > [43:18](https://www.youtube.com/watch?v=HjtvFQvCGnE&t=2598s) 五年之后，大家可能依然在讨论‘探索（discovery）’阶段的 AI 应用，而未能真正普及全自动的‘交易（transactions）’。 — James Williams ## Mentioned Paula Macaggi, Casey Golden, Kunal Sonalkar, James Williams, Sidharth Thapar, Ai4, OFFBounds, CI&T, Nordstrom, Mirakl, Staples Inc., ChatGPT, Claude, Perplexity, Instagram, Google, Meta, Walmart, Home Depot, Sephora, Shopify, Visa, Gemini, AWS, Mars, Merriam-Webster ## Video description Casey Golden - North American Leader, Retail & Consumer Goods, CI&T Paula Macaggi - Founder and Host, OFFBounds James Williams - Head of Agentic Commerce, Mirakl Kunal Sonalkar - Data Scientist, Nordstrom Sidharth Thapar - Manager AI/ Senior Applied Scientist, Staples Inc. In the competitive world of eCommerce, every click counts—and AI is quickly becoming the difference-maker at every stage of the funnel. This panel brings together leaders from top digital-native brands and marketplaces to discuss how AI is being used to optimize acquisition, conversion, and retention. Topics will include dynamic product recommendations, personalized promotions, predictive merchandising, AI-generated creative for A/B testing, and real-time customer segmentation. Panelists will also explore how LLMs and chat agents are being integrated into the buyer journey—from guided shopping to post-purchase support—and what it takes to operationalize these tools at scale. If you're building or selling in a digital storefront, this session will provide actionable insights on how to drive growth and efficiency with AI across the full customer lifecycle. [Source](https://www.youtube.com/watch?v=HjtvFQvCGnE)

[4:38] Casey Golden: 93% 的消费者认为自己现在的购买决策比两年前更明智。 [4:38] Casey Golden: 从寻求娱乐（Entertainment）转向寻求教育与信息（Education）。 [23:12] Casey Golden: 任何在零售业工作超过六个月的人都知道，我们的数据就是一个‘烂摊子’（dumpster fire），你根本无法在其之上构建智能系统。 [24:45] James Williams: 好的零售依然是好的零售（Good retail is still good retail）。 [42:43] Casey Golden: 保护品牌在消费者心中的灵魂与完整性（protect the soul of the integrity of what that brand means to their consumers）。 [42:43] Casey Golden: 没有什么是不可能的，一切都变得可以负担且切实可行。 [38:54] Kunal Sonalkar: 模型已成为通用商品（Models as Commodities）。 [43:18] James Williams: 五年之后，大家可能依然在讨论‘探索（discovery）’阶段的 AI 应用，而未能真正普及全自动的‘交易（transactions）’。

[0:04] Casey Golden: Traditional time-and-materials pricing breaks down when tasks that used to take six months now take six weeks, forcing a pivot toward outcome-based commercial agreements so speed does not compromise quality. [0:45] Casey Golden: 93% of consumers report feeling that they make smarter purchasing decisions today compared to two years ago. This sense of confidence directly drives behavior, resulting in customers making 55% more purchases. [0:58] Kunal Sonalkar: Instead of searching for an individual item, users increasingly input complex, multi-item intent—such as requesting an entire outfit for a wedding. [0:08] James Williams: Because consumers enter LLMs with precise, hyper-specific constraints... retailers require expansive catalogs to remain relevant as query precision rises. [0:14] Casey Golden: Pointing to the post-holiday period, she predicts that return rates in January will serve as a definitive litmus test for AI efficiency [0:17] James Williams: Capturing conversational intent directly on their own platforms prevents external LLMs from disintermediating the customer relationship [21:58] Sidharth Thapar: In this new paradigm, AI agents—rather than end consumers—frequently evaluate products and compare options. [21:58] Casey Golden: Retail data has historically been a siloed, duplicate-filled "dumpster fire" lacking context across decades of point-of-sale systems. [21:58] James Williams: On basic price-constrained queries (e.g., looking for a wedding gift under $200), the agent suggested products averaging $400. [27:35] Paula Macaggi: Consumers demand absolute transparency and control; feeling tricked into believing an AI is a human creates friction and emotional pushback. [27:35] Casey Golden: Explicitly identifying AI interactions (e.g., "Hey Casey, I'm an AI agent, how can I help you?") preserves long-term brand equity and customer confidence. [33:04] James Williams: even though 75% to 80% of products sold at Sephora are available directly from Shopify brand storefronts, consumers—and by extension, their AI shopping agents—will prioritize buying through Sephora to aggregate loyalty rewards. [38:55] Sidharth Thapar: latency remains a major driver of cart abandonment. Deploying overly complex models that slow down site search or conversational responses directly damages conversion rates.

[0:04] Hey everyone, thank you so much for joining us on  the end of the second day. Uh it's a pleasure to [0:09] talk about a topic that has been so discussed  in my life. Uh if you don't know me, I'm Paulo [0:15] Makaji. I am the host and founder of Offbounds. I  interview retail leaders every week and from one [0:22] year to now uh we are hearing a lot about agent  AI. This is taking all our conversations and [0:29] I have here with me guests who can support both  two technologies that support the retailers and [0:36] two guys that are living this living and breathing  the changes in consumer behavior. And today we're [0:42] going to talk about the changes from the consumer  side and how retailers are tackling them. Uh I [0:49] would love to you guys to introduce yourselves.  uh if you can say in two sentences who you are, [0:56] what you do and what are the re the AI challenges  they are tackling most now. Casey, do you want to [1:03] start? Sure. Um I'm Casey Golden, a North America  leader for retail and CPG with Cint. uh we are a [1:13] AI enablement and digital transformation partner  helping brands uh with their AI ambitions and [1:21] turn them into delivered solutions. Um the most  important business objective right now is really [1:30] materializing the gains from AI commercial models.  Um traditional pricing and doesn't really support [1:39] speed and now we can do things in 6 weeks that  used to take 6 months. So time and materials is [1:45] really not working that much anymore. Um, so  we've got a lot of new commercial models and [1:51] it really helps align the client with um them  getting the speed and us getting a commercial [2:02] model that's tied to outcome. So just because it's  fast doesn't mean that it has to be crap because [2:08] we're tying ourselves to the outcome. Yeah, that's  why your perspective is so important here. Canal, [2:13] you work in Nordstream. uh tell us what are who  you are, what exactly you do there and what are [2:19] the AI objectives you have now. Yeah. Um hi  everyone uh I'm Kunal. I'm a data scientist [2:25] at Nordstrom. U I've been with them for more than  5 years now. Uh I started out in the search space, [2:33] search relevance recommendation systems and very  recently I have moved to the ads space. So ads [2:39] relevance um showing the most relevant ads to  customers. Uh, so I'm working in that space. [2:45] Happy, happy to be here. James, you can want to  go next. Yeah, perfect. Lovely to meet everybody. [2:51] James Williams. I work for Miracle. Miracle.  We're a French tech company and we power the [2:57] marketplaces of most big enterprise retailers uh  globally. And if you looked at the program, you'd [3:04] see Dimmitri who was meant to be here from Leroy  Milan. And and I thought when he wasn't able to [3:09] make it uh I would be let down. But fortunately,  Nordstrom, who's another huge Miracle customer, [3:14] is joining us here on stage. So, uh, great  to have uh to have Nordstrom as a substitute, [3:19] but in no way a substitute. Um, in terms of  what we're working on at Miracle, you know, [3:25] we we sit at a really interesting place of, you  know, between brands and retailers, you know, [3:32] and powering that marketplace ecosystem. And so  today the focus is ensuring that both sides of [3:38] the equation are ready for uh gentic commerce.  You know in today's world that means discovery [3:46] uh but but it also means laying the foundations  ready for uh if if and when transactions maybe [3:52] take place at some point. Thanks. I can go  next. Hi everybody. I'm Sadhart Tapper. I [3:59] work as a manager applied science u applied AI  at staples in Boston and um I have been in the [4:07] AI space for 10 plus years primarily working  on recommendation systems, search systems, [4:12] pricing systems and uh one thing that I would  say that enterprises are currently focusing on [4:20] u is giving their customers the same experience  across different platforms that you interact with [4:25] them on and having a better product understanding  having better understanding of your own products [4:32] and then exposing that through all these systems.  Thank you so much. And before we go into the [4:38] topics that we're going to discuss, I want to ask  you, can you please raise your hand if you used [4:44] any type of loom could be chatbt cla perplexity  to shop something the next in the last 30 days. [4:55] Very good. So, Cint uh just released a report,  Casey, that was very interesting. It researched [5:03] several consumers and it found out that 74%  of US consumers have already used AI to help [5:10] them shop. And 90% have either used it or are  open to doing so. That is a big shift already [5:18] in how we used to do commerce to how you're  doing now. Uh what was the what surprised [5:25] you the most about this report? I think what I  found the most you guys can download the report [5:32] um it's on our site or come up to me after and  scan my QR code because it's really insightful. [5:38] Um so what I found most interesting is that 93%  of consumers felt that they're making smarter [5:47] purchasing decisions today than 2 years ago. And  that is the primary driver to those customers [5:56] making 55% more purchases which in other channels  when we have new channels, new discovery channels, [6:06] new sales channels, marketplaces, it's always  driving uh customer acquisition and this is [6:13] actually driving LTV. And so I think moving from  entertainment to education is a pretty compelling [6:23] shift for the consumer wanting more information  rather than entertainment. So because they feel [6:31] they're more informed about their decision,  they make they purchase more and buy. Yeah. [6:36] They're buying more. Yeah. Now from the retailer  perspective, have you seen any shift happening [6:42] with the consumers? Yeah, definitely. I think  um one of the major shifts has been um the [6:48] way people search on the website. Um we used to  have like really short short-term search search [6:55] queries and I think they have been expanding.  Um we used to have two three words like very [7:01] succinct queries but now we are more like five to  six word queries. Uh and with the conversational [7:07] search coming into picture I think a lot of people  don't want to search just one product at a time. [7:14] uh we see a lot of people end up searching for  outfits in general. So a search query could be [7:19] hey I'm looking for the entire wedding day outfit  or I'm I want to attend a wedding please suggest [7:26] me an outfit and then it can go to like a super  custom level as well. Uh if uh if the search [7:32] system suggests you the shoes which you don't  like, you can just replace the shoes, keep the [7:37] dress. Uh if you don't like the jewelry, replace  the jewelry, keep everything else. So that is that [7:43] has been like a very big shift at least that's  what we have seen uh where the search used to [7:48] be focused on like a single product but now it has  become more aggregated more broader yeah very good [7:56] James you from the marketplace perspective how has  this shift happen in the consumer behavior I think [8:07] right now not a huge change like ultimately  our retailers are seeing different, you know, [8:13] levels of traffic. The Google direct Google search  is going down. LLM's attribution is going up, [8:20] but but ultimately it hasn't significantly  changed the way our businesses run today and [8:26] and the impact on marketplaces. Doesn't mean it  won't in the future, but today I would say it [8:30] hasn't changed. That's probably also because  we're in categories discretionary spending. [8:36] discretionary spending, you know, more time  for consideration, etc. versus, you know, [8:42] other fullergogentic experiences that maybe have  less consideration. Yeah. From uh the report we I [8:49] was reading that consumers are using to research  to make more informed decisions and considering [8:55] prices. Uh do you think uh sweatart that consumers  are going from a search perspective to more [9:05] uh assisted p perspective when they're shopping?  Definitely I completely agree with that. I think [9:11] as Kunal said the searches are getting more and  more complex. You have more nuance search that [9:16] a pro that a customer might want and with these  conversational agents allowing you to shop they [9:22] might come in with a lot of constraints too.  So taking the same example, let's say if you're [9:26] searching for a particular uh wedding dress,  you might come in with a constraint of a budget [9:31] and let's say a few brands that you only want to  buy from. And I think that makes search even more [9:37] complex. So retailers are not just struggling,  but they're like preparing themselves for all of [9:42] these different complexities. Yeah. And especially  because the numbers are still small of consumers [9:47] doing this way of shifting. because the shift  is not big yet like we're preparing for when [9:53] it's mass when become a massification of ways of  shopping right um but let's talk about the the [10:01] journey from finding the product I think from the  last 10 12 maybe 20 years we were really focused [10:09] on found find something uh on Google on search or  uh Tik Tok or Instagram and now we actually asking [10:20] our own and that became like the whole the death  of SEO and the beginning of the GEO. Um James, [10:29] I want to start with you. Um how AI is changing  product discovery before the customer get to the [10:36] website. Um we tried this year to everybody. I'm  sure a big chunk of people go to NRF and we tried [10:46] to put on our banner for miracle that catalog size  matters. Um they didn't let us do it but uh the [10:53] point the point still stands that ultimately in an  in an LLA like discovery that starts on an LLM or [11:01] even on other surfaces like an Instagram etc. you  see less products before you end up on a PDP. And [11:11] if you see less products, then two things then  matter more than anything, I I believe, right? [11:17] So the first piece is that a consumer is coming  in with much more strict criteria. So you could [11:25] either say they're making longer searches because  they're using really woolly language. I I want a [11:32] a running shoe that makes my ankles feel good. But  realistically, what's happening is they're adding [11:37] criteria. They're adding a running shoe for long  distance running with a high arch, high foot arch, [11:45] whatever that might be. And your ability to as  a retailer stay relevant as the level of detail [11:53] goes up becomes really important. And that's the  first pie. That's why catalog size does matter. [11:58] The second piece um is that you product data has  always been important and you look at the best [12:05] retailers in the world and this hasn't changed.  Product data is always important but but it [12:10] becomes even more important and I think this is  not product data for product data's sake. This [12:16] is not you know taking an attribute that was in  the description and and putting it in a key value [12:21] pair. The semantic layers that exist in Google and  JB can solve that. This is about taking genuine [12:28] user generated content and enriching product feeds  so that when someone writes the words, you know, [12:37] highly reviewed, then there's a, you know, there's  a way that the LLM can see what the reviews have [12:43] said and and see what the Q&A that's generated.  And so that's why I see the value of the [12:48] conversation attribute. So, um, yeah, two things.  because I think catalog size matters and you know [12:53] product data all the way from like the foundation  to conversational attributes also matters. Yeah, [12:59] sweetheart I see that you are agreeing a lot.  How has AI changed the way u retailers approach [13:10] product content and merchandising and search? I  think it's u completely completely overlapping [13:18] with what just James said that they want to  focus on how they can enrich their content and [13:24] it matters a lot now given the same example of  trying to find something that is good for your [13:29] ankles. U I I think you can only do that if that's  that information about that product is available [13:35] within your content. And taking an example from  yesterday's uh talk that I gave uh there was an [13:42] HP cartooner cartridge that a customer wants to  search and they give a model number like 26x. Now [13:48] 26x actually means a high yield product. So it can  print 9,000 sheets versus let's say printing just [13:56] 4,000 sheets uh which could be another model. So  that context needs to be captured within the the [14:02] within your product content. You cannot just  have a good title or good description. Your [14:07] specifications, your price points, all of those  matter even more now so that they can influence [14:12] this these decisions. Yeah. How can how was that  for you? Uh I think it's it's been a very similar [14:18] experience uh for us as well. um at least like  I have started seeing actual data since February [14:28] uh to this month where we have started getting  traffic from chat GPT and perplexity where we [14:33] actually send out our product catalog and people  do visit from those sites. Uh one aspect which [14:39] really work works out in that uh in the favor of  those LLMs is you get to see a good assortment [14:47] of products in a very nicely laid out grid and  then you can just redirect to our website. Um [14:53] right now the majority of the traffic still  comes from Google and meta ads but I I won't [14:59] be surprised if that number sort of shifts as  the LLMs become a little more sophisticated [15:04] from the shopping standpoint. uh there are like  certain things which need to be figured out like [15:10] uh out of stock availability the correct sizing  that sort of thing uh in a normal conversational [15:15] flow uh but I think we'll we'll get there like I  I don't even doubt that. Yeah. And uh we obviously [15:22] talking about severs is not something that simple  to as just to put a description in one product, [15:29] right? Uh Casey, you work with retailers from  different sizes of different levels of maturity, [15:37] from infrastructure to how they serve their  customers. Uh who what the retailers that are [15:43] winning, the ones that are falling behind are  doing now? Yeah. I mean, I think what really [15:49] makes a successful retailer stand out is action  and upskilling. Very few retailers are prepared [16:02] for an agentic holiday season and we will see  these numbers come January in red and green. [16:09] Um I'm not projecting any KPIs in black. I don't  think we're running anything flat through holiday. [16:19] You're either winning or you're losing and it's  going to be on the P&L. Um, right now there is [16:30] you have consumers thinking that they're making  smarter purchases. It's it's in the data. It's [16:36] in your conversations at the dinner table when  you're having lunch with your friends. These are [16:42] conversations are happening. You're you're proud  of your purchase. You feel like so much easier. [16:49] I'm so I I'm not going to return this item. So,  everybody who works in retail knows what January [16:58] looks like on returns. And I think that this  is going to be very interesting to see how many [17:04] how returns in the holiday season afterwards  affect how those 93% of customers think that [17:13] there's making smarter decisions. Thank you.  Let's see. Let's let's talk again in January. [17:23] So okay like we see that consumers are  finding products and finding brands and [17:30] finding your website before in different ways is  starting to get there right let's say they found [17:38] uh your website or the product from your website  from an LM and goes to your website how is AI [17:47] changing the consideration and conversion once  they are in the website. Kunal, I would love to [17:53] hear from you. Yeah, absolutely. Um, so if once  the customer sort of gets on the website, I think [18:00] there are quite a few experiences which you can  trigger to make the decision more informed. Uh, [18:06] recommendations is a really upand cominging  space where um, LLMs and agentic AI has uh, [18:13] affected uh, a lot. Um like in in our case the  recommendation systems were like nightly batch [18:20] jobs which we used to run and most of them were  like static recommendations and we also used to [18:26] have problems on the cold start problems. Uh but  with some of the transformer based models you can [18:32] actually model sequences on how the customers have  interacted with the products which products have [18:38] they clicked on added to bag and checked out. Um  and you can actually develop really good contexts [18:44] on what exactly are they looking for? What um was  it like a impulse purchase or was it they have [18:51] been researching for a long time. Uh so you can  model like some of those approaches really well [18:57] uh with some of these transformer based models.  Uh so yeah I that that would be my perspective. [19:03] Yeah. Yeah. uh Casey, do you see any use cases of  probably like assisted shopping like AI chat bots [19:13] uh on your client's websites? Yeah, I mean  conversational commerce has been in and then [19:21] it was out and then it was in and now  it's AI assisted but it's essentially [19:27] um very similar experience, right? Shop with the  stylist, shop with the product expert. Now it's [19:34] AI. Um, this is becoming very much in again. Uh,  I think the most interesting thing that I'm seeing [19:42] if you have if you have over 500 SKUs, adding the  AI search and letting customers search your site [19:55] the way that they talk to an LLM is incredibly  insightful, not just for the customer to find [20:04] what they're looking for rather than having  to click all these little cute little buttons, [20:09] on the taxonomy on the left, but it's very  telling to learn how your customer thinks [20:16] about your product directly. Um, what they're  using, the words they're using, the context [20:22] they're framing it in. Are they talking about a  birthday party? Are they talking about work? Um, [20:28] I think there's a lot of insights to learn from  that. So, I want to see more of that. Me, too. [20:34] See, Har, do you want a compliment? You go ahead.  Sorry, James. Do you want a compliment? I I I love [20:39] uh sorry for interrupting and if I make a crap  point you just say I should have gone first. Um [20:46] the I I I like that I like that framework and I  heard a really good framework from Accenture. Um [20:51] which is you know retailers they face one choice  and one inevitability right the first one is [20:58] being the choice of agents and the second one is  being the agent of choice and and certainly for [21:05] the Nordstrom of the world for the Staples of the  world like the these have own categories and they [21:10] have the right to be the agent of choice. What  do you do with that customer differently when [21:16] their source was chat GPT or Gemini is really  important because if you feel really good about [21:23] your own agent you should be emailing your your  marketing to them should be next time you have [21:29] a product search start on my agent because you  then you know today chat GPT don't charge for [21:35] the referral but you know that isn't going to  last for very long and so there's a real place [21:41] where you can cement yourself becoming the agent  of choice and I think actually it's still open [21:46] uh you know the doors still open for retailers  to walk into that position. Yeah. I mean, [21:50] I spend eight hours a day talking to my agents.  It' be very cool to not have to leave it. [21:58] Sidar, uh, from your perspective, beyond chat bots  and assisted assisted purchasing or conversational [22:05] commerce, uh, what are other ways to better the  conversion rate in the website that is AIdriven? [22:15] I think I I have different perspectives there.  Like I noticed Walmart now allows direct selling [22:21] on chat GBD. You can place an order on chat GBD  but assuming like a typical retail where you're [22:28] receiving these questions from an conversational  AI. I think the difference is that now you're [22:34] influencing a decision. You're you're basically  it's not the customer who's evaluating it's the [22:41] agent who's evaluating. So your prices need  to be more accurate. You probably need to [22:46] have alternative products in place. Let's say you  might you might have you need to have you need to [22:51] be more proactive about tracking your inventory.  If you're in if you're out of inventory, then you [22:56] need to have substitutes in place. So all of that  matters even more now given that what you're going [23:01] to deliver is in directly going to influence the  decision. Absolutely. So retail loves the sparkly, [23:12] you know, little item, but when it comes to  like what are retailers and brands buying [23:18] right now? There's a couple search, there's  a little bit of the conversational commerce, [23:24] there's a little bit of the the sparkly items.  Almost every single brand that we're talking to [23:29] and contracts that are getting signed is  data modernization because anybody who's [23:34] worked in retail for more than six months know  that it's just in our data is a dumpster fire [23:43] and um you just can't build off you can't build  intelligence off of it. It's it's siloed. It's [23:49] duplicative. Our historical data has no context.  There's there's 50 years of no digital data and [23:56] just POSOS data. Um we are working on projects  that are taking these 25 year ambitions of what [24:04] good data looks like um and consolidating it  and unifying it in like record time like less [24:11] than a quarter. and it's a completely different  environment and being able to um localize it, [24:18] add context um and not be so siloed that like  we're aggregating, consolidating global data [24:27] in like 8 weeks. Like that's really what the  brands and and retailers are doing right now is [24:33] is fixing that foundation that we thought we'd  never get fixed. Exactly. Otherwise, the shiny [24:38] things don't work. Yeah. So data is the foundation  and I think we keep hearing that uh all the time [24:45] in our conferences as well. Um James, I'm curious  for um your clients or things that you're seeing [24:52] in the market. Did you have any surprise when  implementing some of these technologies despite [24:58] not having the data when implementing uh chat bots  or uh acided search or any of these tools to serve [25:07] this new customer? um you know we because of the  position I tell a story and then a lot of the [25:15] chat bots today on retailers sites have been uh  offloaded to Gemini AWS etc and we the those guys [25:25] are incredible at what they do I'll give you a  story and uh is a kitchen retailer who did exactly [25:31] this they offloaded it to one of the hyperscalers  and then you we ran some synthetic these are syn [25:39] like I agree synthetic prompts are not a great  way but they directionally right and we were doing [25:46] kitchen knives so we basically replicated all the  top level categories on their agent can it solve [25:52] the problem and uh we even set things like pricing  parameters and the first thing is only 10% of the [26:02] product so they're usually five product cards only  10% of the products uh were ever bestsellers. So, [26:09] one of two things is wrong, right? If you're  only seeing 10% of the products that would be [26:14] the best seller on the product card, like either  you guys, the merchandising team doesn't know what [26:19] they're doing and they've been picking the wrong  bestsellers for all this time or the AI is not [26:24] doing isn't taking the merchandise insight that  it should be doing and presenting the best. That's [26:29] the first thing. The second one, its ability to  answer a simple price query like I don't want [26:34] it's a wedding gift. I don't want to spend more  than $200. the average price $400 for the for [26:39] the first five products. Like you those little  things that you don't think matter, but they do [26:45] they do matter. And so like the retail basics in  this world haven't changed, right? Right place, [26:52] right product, you know, right price, right  availability, right inventory, shipping dates, [26:56] whatever that might be. And so I think there  that's that's one of the things that we keep [26:59] going harking on is like good retail is still good  retail. That hasn't changed. And how do you make [27:05] sure the experience you're creating continue  to that be? And so that that's why we always [27:09] what we're starting with today is always like  that data foundation and then making sure that [27:14] you are really user testing what your customers  experience. Yeah. And that comes we were talking [27:21] about it yesterday right how it's important the  product enrichment and having the right content [27:26] for the product so the chatbot and the agents can  match. Um and that's why your work is so complex. [27:35] But let's let's move on a little bit from um  the consideration to purchase and let's think [27:41] about other ways that is shifting on the the post  purchase right like either loyalty personalization [27:50] and even customer service is not here uh but  he shared uh he's the CIO of Leroy Merlin if [27:57] I'm pronouncing it correctly uh and he shared that  they have voice customer service agents that are [28:06] not people but they talk to people like they as  if they were and I didn't test it myself and I'm [28:14] a bit skeptical but could now are you're seeing  some changes in the relationship with customers [28:22] after the purchase. Yeah, absolutely. So, uh I  think voice agents is definitely a big piece on [28:28] the customer service side. I have not personally  worked on that area. Uh but I know for a fact like [28:35] we have been testing out um auto chat bots after  the post purchase. Um there was like a really good [28:42] talk over here uh from Home Depot today uh on how  they have sort of routed this aspect completely [28:50] where um starting out they had press one for this  press two for that press three and so on so forth [28:56] and now completely that uh whole paradigm has  shifted where you feel like you're talking to [29:02] a person and it will answer literally all of  your queries even if it is associated with a [29:08] product detail product description So there is  definitely like a big shift happening uh post [29:14] purchase as well. Yeah. Where is my where is my  order? Has this beened or things like that too, [29:20] right? Exact. Yeah. I was can you imagine a  world where we no longer have to hear uh the [29:26] menu options have recently changed. I I think one  interesting part u again this is from the Home [29:33] Depot talk which I attended. uh the customer was  asking about compatibility like hey I'm ordering [29:38] this product and will it be compatible with  certain product which I'm trying to fix through [29:44] u and the chatbot was able to answer like okay yes  the compatibility will be fine and you should go [29:49] ahead and order I think that that's a really  big step on just asking uh from post-purchase [29:54] standpoint and I think just just one point I  wanted to add like that is where having knowledge [30:00] graphs is important for your product so that you  have like you have that deep understanding of if [30:05] what is compatible with bot. Yeah. Yeah. And  Casey, I'm curious um because all that like [30:12] I think if I'm talking to a chatbot and if I I'm  talking to a agent that talks to me on the phone, [30:20] I think like at first I have like I don't know  I don't trust it. And I think trust between [30:27] the customer and the retailer is very important  especially as we navigate into the agentic world [30:33] right. Um what are what can companies do to build  this trust between retailers and customers? Trust [30:41] was a reoccurring theme throughout our entire  report. Um consumers want transparency. They [30:50] want control. um they want to know that they're  talking to AI before figuring it out and feeling [31:02] it almost feels kind of weird there.  There's weird feelings about it when [31:06] you think you're talking to a human and then  you realize like oh I'm not. It's even it's [31:13] even happening. I'm sure everybody here  you kind of feel like taken advantage of [31:20] when you start crying over some like pet Tik  Tok that was totally II because it's like how [31:26] dare you play with my emotions. It was fake.  This should be transparent like, "Hey Casey, [31:31] this is an AI agent. How can I help you?" Yeah.  Like I'm yours. You know, they can make it more [31:35] personal, but um this trust used to be in the post  purchase. it as long as you had a good product [31:43] um that it was communicated properly, they got  what they expected, it was delivered timely, [31:49] you were able to earn that trust. Um but now  AI is pulling that trust into the beginning [31:56] of the purchasing decision and they want  accuracy, the agent accuracy. When will I [32:04] receive it? Is this the right price? I think  I've clicked on a lot of things that had a [32:10] price on ChachiPT or Gemini and then when I go  to the link it's completely different price. [32:18] um not where really sure where that disconnect  happens but um they want the accuracy and I think [32:26] that's really where um enriching the catalog data  ensuring that this product descriptions and are [32:37] correct and being able to deliver on that brand  promise because it's not going to be the LLM or [32:45] the AI assisted technology ology that's going  to be held to the brand promise. It's going to [32:50] be the brand and they're going to have to have  accountability. Yeah. And I think it gives me [32:58] another type of curiosity here. We are here so  far we are talking about the consumers that use [33:04] AI when they're shopping. But we know that agent  commerce will soon become agent to agent that my [33:13] agent and this is the topic of the next talk. Um  the agents will shop on our behalf and they will [33:19] shop autonomously. Um and I think that trust  you usually gain with the relationship right. [33:28] So if I am not having the relationship with the  brand, if my agent is shopping with your agent, [33:34] who is going to get that trust and who is going to  get that relationship between these two parties, [33:40] right, James, as we're getting into this  agent world where agents shop on our behalf, [33:47] what do you think it needs to happen from the  retailer side before it becomes mainstream? [33:57] I think there's certain product categories  where agentto agent shopping will will never [34:02] take place like agent assisted shopping but I feel  really confident there's still going to be shops [34:09] there's still going to be physical shops there's  still going to be a you know a PDP that you go to [34:14] um and there obviously categories where  there will be agent to agent shopping but [34:20] but maybe more broadly what does it take to be  ready for agent to agent shopping. Uh I think [34:25] like loyalty still matters in that world and  I'll give you an example why. So uh Sephora, [34:34] great retailer uh in the US. I think 75 or 80%  of its brands that they sell today are Shopify [34:42] brands. And so you technically could probably get  them from the brand directly, maybe even cheaper. [34:49] Why do people go to Sephora? Everybody answers  loyalty and your agent is also going to know [34:57] that you're a loyal Sephora shopper and the agent  has an incentive to to keep in your good books, [35:03] right? Like if you we've got to be polite to  the agents and hopefully they're polite to us [35:07] and follow our means. And so those those retail  fundamentals I think still become really really [35:12] important. I think then like that's that's its  relationship, your agent's relationship with you [35:18] and why that's still important. on the loyalty  aspect. I think then there's the standard things [35:22] you think about. So product data continues to  be even more important if you're not looking at [35:27] it being like that's mispriced. That's the first  thing. The second piece is um the PSPs and like [35:34] the financial uh element to this and the fraud  prevention and all of that. Visa recently I think [35:42] announced they're doing a a big piece of work  around you know authorized agents. So I think the [35:47] infrastructure is being built around making sure  that they're trusted transactions and so retailers [35:52] will allow them to take place. Um and then I  then I think the there still like the completely [36:00] autonomous piece might be grocery, right? Like  yes, I've bought milk and pear avocados every [36:07] Friday for the last 6 weeks. Like if it showed up  at 7:00, I don't need to authorize that as long [36:11] as it's within the realms of, you know, normal  spending. Um, I still think they'll there'll be [36:16] an agent touch or human touch like human in the  loop. And so how do how do you make sure that [36:21] that human in the loop doesn't prevent conversion  because that would be the other risk is that you [36:26] know it effectively prevents conversion because  the human never gets I don't know if you like the [36:31] human never gets back to say yes I authorize that  you click the iPhone to to complete the checkout. [36:36] So we're doing um some agentto agent work right  now and where we're seeing it is actually on the [36:43] B2B side at CPG. So demand forecasting placing  orders from distributors um and it's agent to [36:51] agent on the B2B side. Um I think this is where  the training ground is. Uh the next one to go like [36:59] you said James grocery another consumer product.  It's it's a natural fit and I think retail will be [37:07] the consumer shopping for clothing or a new house.  I mean I think we're farthest away from there but [37:15] the training grounds are already in production  like in environment. Yeah. And sweetheart, what [37:21] do you think are the capabilities that retailers  need to build in the website for these agents that [37:28] are going to show? I think uh one of the points  may be agent to agent subscriptions because we [37:36] already do that right we place orders with two  months 6 months time period repeated orders how [37:42] about the demand forecasting can be planned better  by knowing the demand that's coming through the [37:47] subscriptions but you're not directly integrated  with them I think it's also about like you could [37:53] uh you probably need to put a lot of AI guardrails  in place and like these agents might be trying to [37:59] scrape your information or that Shopify example  that you just gave of maybe you're trying to find [38:03] a product on Sephora and then because it's like  2 cents lower on Shopify you recommend that to [38:08] the customer that loyalty needs to be in place  that trust has to be built but from the retailer [38:14] perspective I think accurate pricing is one aspect  but while making sure the margin is uptight you [38:21] don't you don't want to be just going down just  because you have that competitive intelligence you [38:25] want to make sure it's margin positive deal for  you but have being accurate and then giving that [38:31] same experience not just from pricing perspective  even from content perspective across all channels [38:36] that you sell in. I think that really builds that  experience because if I see a product price as [38:43] $10 versus a $12 on a re on on your stores versus  your website uh agents and humans are not going [38:50] to trust you. So I think that that matters. Yeah.  And I I'm going I'm going to stay with you really [38:55] quick. Let's talk about scaling AI, right?  Like there there are a lot of these pilots, [39:02] but operationalizing it uh to the level of so  many products, so um many customers, it's quite [39:12] challenging. Um what is the biggest lesson you've  learned about operationalizing AI and what is one [39:20] advice you give to the retailers in the room?  I think what I've realized is few years ago we [39:27] used to try to make a better model or use a bigger  model for everything. I think that problem is more [39:33] a commodity now. Models are commodities. The pro  the product content or your readiness with having [39:40] all the capabilities to work at scale is what  retailers should focus on. And when I say that, [39:46] I mean things like again AI guardrails, LLM  guardrails or let's say your your infrastructure, [39:53] your compute needs to have governance around that  system. Like all of these best practices already [39:59] exist in software engineering space, but you need  to make sure those are incorporated in your agent [40:04] to agent systems as well. Yeah. James, would you  like to answer the questions as well? Do you want [40:12] me to repeat? Sure. So at scale yeah um building  the these ideas of AI is easy but scaling to the [40:21] level of enterprise with so many products and so  many customers and maybe so many businesses it's [40:27] a bit challenging uh what would be your advice  to the group here of retailers I think to be [40:34] honest with you I think there's I saw a bunch of  sessions on like why pilots fail why all I feel [40:40] like coming to this conference is probably like  half the battle and just go and look at where you [40:45] should be focusing your time. There's been so many  great sessions on good examples of why they were [40:49] successful. I thought Home Depot was a really  great example of that. Like I think half Yeah, [40:53] half the battle is understanding that just because  it's a good idea and I can do it one off doesn't [41:01] mean I should do it. And it certainly doesn't mean  if I'm I should build it myself. Like I, you know, [41:07] that's the other thing I'd say is that if you  go into the hall today or the expo hall, I mean, [41:13] there's there's thousands of vendors in there  that are are trying to solve very niche products [41:20] and they're all looking for that their enterprise  brand, their first enterprise brand or customer [41:25] to work with. Go and find the right one and work  with them and and if let it become their problem, [41:30] I guess is maybe what I'm uh trying to say. Yeah,  I have one point to add on the operational part. [41:36] Uh I feel with the whole search recommendations  and like chat bots latency is a very big aspect [41:43] of how customers search. So if it's taking a lot  of time to load your results, there will be a [41:49] lot of abandonment which people will do. So if a  retailer is trying to build out their own models, [41:55] uh my one advice or like two cents would be uh  don't use a super complex model even though the [42:01] results might seem very relevant but if it's  hurting you on the latency you won't be able [42:05] to productionalize it and u if you can't take it  to production then what good is it going to do [42:11] for you? So yeah how about you casing? You need to  know if you're scaling the right thing. Um, every [42:18] brand on the planet was founded by a visionary  and over the last couple decades, uh, leadership [42:27] is they are skilled financial operators, but they  have little to no experience in retail outside of [42:36] the P&L or the boardroom. Right now, every single  brand needs leaders that have a vision for the [42:43] brand to protect the soul of the integrity  of what that brand means to their consumers, [42:49] be able to communicate it and embrace technology  to scale the right thing cuz and they need to [42:59] have the data foundation. Everything is nothing.  No, like everything is affordably possible now. [43:08] like it can be done. So like you can bring your  wildest shopping dreams to reality if you know [43:18] what they are. We have two minutes left. So I  I like to ask something really quick. Let's say [43:27] we meet again five years from now here at  AI4 to have exactly the same conversation. [43:34] What do you think commerce agent tech commerce  will look like in five years? But it has to be [43:41] super short because we just have less than two  minutes. Sweetheart, I'll go next. I'm thinking [43:52] okay. Um short answer would be um for the repeated  categories like groceries, I can totally match an [43:58] agent to agent. Um but I really wouldn't want to  take the discovery side to an agent. Like I like [44:04] shopping. I like exploring. So I would like to  keep that. James, this is a bit of a hot take, [44:13] but I think we might still be talking about  discovery, not transactions. in 5 years. Okay, [44:19] Casey. Um, in 3 years I believe we will be able  to have individualized experiences scaled which [44:31] is called clientelling and I think in 3 years  maybe Miriam Webster will add it to the dictionary [44:38] after it being operational since 1909. All right,  let's see in five years then. Thanks everyone. Uh [44:48] Casey has a great podcast too. You guys should  follow her and follow what she has to say. Um [44:55] my guess we don't have time to Q&A today. We have  an agent commerce conversation coming right after [45:02] from Mars. So the CPG side is going to be very  interesting. And if you want to talk to any of [45:08] them, they're going to be around the conference  I hope today and tomorrow. Thank you so much.