Mainframe Legacy Code Blocks Retail Artificial Intelligence Upgrades

大型机遗留代码阻碍零售人工智能升级 数十年历史的 COBOL 和 AS/400 系统阻碍了大型零售商扩展现代机器学习模型与自动化商业代理。 来自 Sound Ventures、7-Eleven、CI&T、Coupang 和 Klizer 的行业领袖齐聚一堂，共同探讨将生成式 AI 整合到零售与电商平台中的实际现状。专家组分析了将现代 AI 模型连接至传统基础设施面临的技术与运营障碍、数据质量和合理划定范围的重要性，以及证明供应商投资回报率（ROI）的策略。展望未来，嘉宾们强调了向 AI 原生敏捷性、多模态探索及代理商业（agentic commerce）转变的趋势，商家需针对 AI 代理和对话式购物来优化体验设计。 CI&T 与 7-Eleven 等企业分享了 AI 落地实践：CI&T 在向零售及 CPG 客户推广前已在内部人力、财务等部门构建专有 AI Agent，而 7-Eleven 则重点利用文档摘要优化内部工作流并以对话式 UI 改善客户旅程。 (2:21) 企业级 AI 落地的核心瓶颈在于传统老旧基础设施，运行数十年的 COBOL 与 AS/400 大型机缺乏 API 驱动的模块化架构，加之数据严重分散与低质，导致许多过早推出的 AI MVP 陷入停滞或失败。 (7:52) Coupang 指出早期软件架构设计常因未记录目标变量与时间戳而制约后续 ML 模型训练；同时，化解组织内部对 AI 的抵触需要将技术定位为消除繁琐重复劳动的协作工具。 (12:21) 电商体验正从传统关键词搜索转向多模态与对话式发现，未来不仅支持图像拆解推荐与宽泛意图对话，零售商也需逐步转向针对 AI Agent 的生成式引擎优化（GEO）与即时结账集成。 (24:18)

Industry leaders across retail, e-commerce, and digital consulting argue that realizing generative AI's full commercial potential requires moving past superficial front-end tools to resolve legacy infrastructure and data quality bottlenecks. While the technology enables conversational discovery and multimodal shopping, enterprise adoption remains gated by core system modernization, rigorous scoping, and workforce change management. Legacy Infrastructure and Siloed Data Bottlenecks (7:52) Marcelo Vessoni notes that enterprise-wide access to models like ChatGPT or Gemini is only a starting point; genuine transformation requires redesigning operational workflows. Corporate backbones, however, often rely on 30- to 40-year-old COBOL and AS/400 systems containing tens or hundreds of millions of lines of code, complicating rapid AI integration without deliberate modernization. Shahid Quraishi adds that monolithic legacy architectures lack API-first designs, making real-time scaling difficult. Furthermore, fragmented and uncleaned enterprise data routinely causes rushed Minimum Viable Products (MVPs) to fail. In contrast, panelists point out that young, AI-native startups founded within the last five years hold structural agility. Aaron Pallares compares legacy data cleanup to clearing an inherited phone, whereas startups begin with an open canvas. Supported by cloud-first, modular frameworks, newer players maintain clean, unified data catalogs from inception, facilitating smooth model scaling and direct improvements to operational key performance indicators like on-shelf availability. Scoping Deficits, Workforce Anxiety, and Contextual Design (12:21) Technical AI adoption frequently falters during software design. Adam Nguyen observes that product managers often design linear user journeys without logging outcome timestamps or target variables, leaving data teams without the labeled data required for supervised machine learning and reinforcement learning models. When vetting outside software to address internal pain points, Nguyen emphasizes that enterprises require proven customer references and real-world case studies before justifying third-party solutions over custom in-house builds. Adoption also encounters organizational friction from employee job security concerns. Nguyen explains that framing AI as an assistant that automates tedious tasks—such as eliminating daily hours spent manually entering data into spreadsheets—eases resistance and shifts cultural perspective. Complementing this, Pallares references Dr. Howard Moskowitz's realization that consumer markets require multiple tailored offerings rather than a single 'perfect Pepsi,' arguing that enterprise AI must reject one-size-fits-all deployments in favor of contextual models tailored to specific business strategies. The Evolution Toward Conversational and Agentic Commerce (24:18) The panel forecasts a substantial shift from keyword-based search queries toward assistant-driven conversational commerce. Quraishi anticipates that personal assistants will track household routines and preferences to curate purchase summaries automatically. Nguyen notes that search queries will shift toward broad situational constraints—such as asking an LLM for an entertainment setup that fits a specific room width—prompting systems to evaluate non-obvious alternatives like projectors rather than standard dimension-matched televisions. Vessoni points to integrations such as native checkout within chat interfaces between OpenAI and Shopify as evidence of an emerging agentic paradigm. In this environment, merchants must evolve their optimization strategies, utilizing Generative Engine Optimization (GEO) and frameworks like Model Context Protocol (MCP) servers to sell directly to autonomous AI agents rather than human shoppers navigating traditional product pages. Multimodal Database Demands and Experiential Retail Shifts (30:50) Supporting next-generation shopping behaviors requires substantial upgrades to enterprise data infrastructure. Nguyen explains that multimodal inputs combining imagery, voice, and text demand database overhauls to ingest, vectorize, and store rich media alongside standard text records. These capabilities allow for image decomposition—such as segmenting a photograph of a decorated tree to let a user add individual matching ornaments to a digital cart with one click—though they introduce high computational scaling costs for model retraining. Concurrently, physical brick-and-mortar storefronts must adapt as digital channels advance. Pallares argues that physical stores can no longer function solely as inventory distribution centers, warning that retailers must transform physical spaces into experiential destinations to counter foot traffic declines. Quraishi projects that immersive augmented and virtual reality will further disrupt physical footprints by allowing consumers to navigate and purchase inside virtual retail environments from home. While generative AI introduces powerful avenues for conversational discovery, agentic purchases, and multimodal engagement, enterprise realization remains tied to resolving legacy data debt, redesigning foundational infrastructure, and proving return on investment against custom builds. The discussion leaves open how traditional retailers with constrained margins will finance the costly database overhauls and physical real estate transitions required to keep pace with AI-native competitors.

在Ai4举办的研讨会上，投资机构与企业技术负责人指出，生成式AI在零售、消费品和电商领域的深化应用正面临数十年遗留系统与数据碎片化的严重阻碍，而行业未来的演进方向正快速转向对话式发现、面向AI智能体的销售优化以及多模态全渠道体验重构。 遗留架构与数据碎片化阻碍技术落地 (7:52) CI&T全球零售与消费品业务负责人Marcelo Vessoni认为，向员工开放ChatGPT或Gemini等工具仅是表层动作，实质性变革取决于核心业务流程的重构。然而，当前企业界大量核心业务仍运行在已有三四十年历史的COBOL和IBM AS/400大型机系统上，代码量高达数千万至数亿行。在缺乏渐进式现代化升级策略的情况下，直接接入现代AI平台面临巨大的技术阻力。 7-Eleven研发工程经理Shahid Quraishi指出，单体架构欠缺以API为优先的模块化设计，限制了机器学习模型的实时弹性扩容。同时，传统企业普遍存在严重的数据孤岛与数据质量问题。若企业未先完成底层数据的清洗与结构化治理，就盲目仓促推出AI最小可行产品（MVP），往往会直接导致项目搁浅或失败。 组织抵触情绪、规划缺陷与场景化适配 (12:21) Coupang零售分析负责人Adam Nguyen分析了AI推进中的技术规划缺陷与人员心理障碍。在技术层面，产品经理在早期系统架构设计时常忽略数据标注规范，仅关注端到端功能实现，而未能同步记录结果时间戳或目标变量，导致后期工程团队因缺失带标签数据而无法训练监督学习模型。在人员组织层面，业务员工普遍存在被算法淘汰的焦虑。Nguyen通过将AI重新定义为剔除繁琐手工表格录入的辅助工具，有效化解了运营团队的抵触情绪。 Klizer合伙人Aaron Pallares借用霍华德·莫斯科维茨（Howard Moskowitz）博士为百事公司细分口味的案例强调，企业部署AI不存在一劳永逸的普适方案。成功的关键在于提出正确的业务上下文问题，针对特定用户群体与商业场景定制模型，而非追求单一的通用解法。 原生数字敏捷性与外部采买的ROI验证 (19:41) 在对比近五年成立的AI原生品牌与传统零售企业时，Pallares与Quraishi认为，新兴品牌无需像接收二手设备那样耗费精力去清理历史冗余与脏数据，而是能够直接从零构建云原生、模块化的数据目录，其干净的数据基础能直接赋能货架可用率等核心零售指标。传统企业受制于固化思维和分散的数据源，在模型扩展上面临先天劣势。 针对企业自建与采购第三方AI方案的抉择，Nguyen表示，大型电商平台在评估外部供应商时，高度依赖已获验证的客户案例与契合自身复杂数据环境的真实落地证据。供应商若无法在前期清晰证明投资回报率（ROI），很难推动企业在内部自研力量之外促成采购决策。 对话式发现、智能体交易与多模态重塑 (24:18) 专家预测，传统基于网页关键词的搜索正在向对话式和意图驱动体验迁移。Quraishi与Nguyen指出，未来的个人AI助理能够结合家庭偏好自主生成购物推荐，搜索查询也将更加宽泛与情境化（例如用户要求寻找适配特定窗宽的观影方案，AI将主动推荐投影仪等替代商品）。Vessoni补充道，随着即时结账功能嵌入对话界面，商家必须开始推行生成引擎优化（GEO），从直接面向人类买家转变为面向AI智能体销售。 为了支撑下一代多模态消费行为，Nguyen与Quraishi指出，电商底层数据库需经历重大改造，以在大规模计算成本约束下摄取、向量化并存储高维图像与音频数据，支持自动图像拆解与单品直接加购。与此同时，Pallares与Quraishi表示实体店必须摆脱单纯的库存流转功能，转向依靠沉浸式体验和AR/VR远程漫游维系消费者触达。 生成式AI虽为零售与电商开拓了从供应链自动化到多模态交互的巨大空间，但企业若要实现规模化价值，必须首先承担解决数十年遗留主机技术债、重构底层数据工程与梳理组织流程的高昂转换成本。

## Background 本视频由 Ai4 频道于 2025 年 8 月 27 日发布在 YouTube 平台上。根据页面简介，本次圆桌论坛邀请了来自 Sound Ventures、7-Eleven、CI&T、Coupang 以及 Klizer 的多位行业专家与投资人，共同探讨生成式 AI 如何通过内容创作、产品设计、营销自动化、供应链优化及客户互动重塑零售、CPG 和电子商务行业，帮助企业降低成本并加速产品上市。 ## TL;DR 与会嘉宾深入探讨了生成式 AI 在零售和电商领域内部流程重构与外部客户体验提升上的真实落地场景。讨论重点剖析了阻碍 AI 规模化应用的四大核心瓶颈：陈旧的遗留系统（如 COBOL 和 AS/400）、严重的数据碎片化与训练变量缺失、组织内部员工对被替代的恐惧心理，以及代际认知偏差。嘉宾们强调，企业必须重构数据架构并保持“人的温度”，以应对从传统视觉界面向语音交互和智能助手演进的陡峭变革曲线。 ## Atmosphere & style 整场研讨会气氛专业而务实，充满了行业一线实践者的高碰撞洞察。主持人的敏锐提问引导着嘉宾们避开空洞的行业热词，直面技术落地中的真实痛点与组织阵痛。现场讨论既有对未来技术重构零售交互的展望，也有对旧架构解耦与员工心理疏导等现实问题的深思，呈现出一种理性、前瞻且紧扣商业本质的探讨氛围。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=Yv_ACluUNi0&t=4s) 本场研讨会由 Sound Ventures 的 Lexi Burby 主持，邀请了来自 CI&T、7-Eleven、Coupang 及 Klizer 的专家，共同探讨生成式 AI 在零售、CPG 和电商领域的最新应用与影响。 - [00:27](https://www.youtube.com/watch?v=Yv_ACluUNi0&t=27s) 咨询机构 CI&T 采取“吃自己狗粮”策略，先在内部重塑营销、财务、HR 及法务流程，再将总结的经验赋能客户端的供应链和商业平台。 - [00:27](https://www.youtube.com/watch?v=Yv_ACluUNi0&t=27s) 零售与电商企业应用 AI 的通用模式主要分为内部工作流优化（如文档自动摘要）和外部客户体验提升（如对话式互动界面）。 - [00:27](https://www.youtube.com/watch?v=Yv_ACluUNi0&t=27s) Coupang 在 AI 研发初期聚焦于零售供应链优化，通过数据收集、模型训练及收集反馈三大环节，同时服务消费者与平台供应商。 - [00:27](https://www.youtube.com/watch?v=Yv_ACluUNi0&t=27s) Klizer 提出应对 AI 冲击的三步走策略：梳理与规范数据、安抚与赋能员工、回归教育与业务基本原则。 - [00:52](https://www.youtube.com/watch?v=Yv_ACluUNi0&t=52s) 运行在 COBOL 或 IBM AS/400 等遗留系统上的单体架构缺乏模块化与 API 接口，是阻碍大型零售企业对接 LLM 和 AI Agent 的核心技术瓶颈。 - [00:52](https://www.youtube.com/watch?v=Yv_ACluUNi0&t=52s) 数据高度碎片化及软件早期设计缺乏“输出结果变量”标注，导致许多企业在训练监督学习模型时因缺乏关键数据而导致 MVP 尝试失败。 - [00:52](https://www.youtube.com/watch?v=Yv_ACluUNi0&t=52s) 推行 AI 最大的组织阻力来自于员工对岗位被取代的恐惧，企业需将沟通重点放在消除机械性繁琐劳动、实现“效率提升”的共识上。 - [00:52](https://www.youtube.com/watch?v=Yv_ACluUNi0&t=52s) 向员工发放 ChatGPT 或 Gemini 账号仅停留在浅层体验，真正的 AI 变革在于结合业务哲学对核心流程进行深度重构。 - [34:09](https://www.youtube.com/watch?v=Yv_ACluUNi0&t=2049s) 随着语音搜索与智能助手的发展，传统网页和手机界面的购物模式正加速向智能体与语音交互过渡，其颠覆速度将超过当年的移动互联网革命。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=Yv_ACluUNi0&t=4s) 研讨会开场与嘉宾背景介绍 研讨会由 Sound Ventures 的投资人 Lexi Burby 主持。作为一家由 Ashton Kutcher 和 Guy Oseary 领导的早期风险投资机构，Sound Ventures 在各阶段均有投资，而 Lexi Burby 主要负责生成式 AI 应用领域的投资。本场圆桌论坛旨在深入探讨零售、电子商务及消费品（CPG）领域的最新趋势，以及 AI 正在对这些行业产生的深远影响。 随后，四位行业专家依次介绍了各自的背景与业务领域： - **Marcelo Vessoni**：来自数字技术咨询公司 CI&T（CINT），担任零售与 CPG 业务线全球高级副总裁，长期帮助全球顶级零售与 CPG 品牌构建数字平台并应对技术革新。 - **Shahid Quraishi**：来自 7-Eleven 的研发部门（R&D），担任工程经理，专注于跨业务线的数字创新，探索新的商业机会并将其落地推进。 - **Adam Nguyen**：来自韩国最大的电子商务平台酷澎（Coupang），担任零售分析总监。他生长于美国加州，目前常驻首尔。 - **Aaron Pallares**：来自 Klizer，担任合作伙伴渠道销售经理，致力于通过整合 WMS（仓库管理系统）、PIM（产品信息管理系统）、CMS（内容管理系统）等系统，帮助制造商和分销商实现电商体验的数字化转型。 --- ### [00:27](https://www.youtube.com/watch?v=Yv_ACluUNi0&t=27s) 企业在内部与外部业务中应用 AI 的基线现状 在谈到当前企业如何将 AI 融入内部运营和面向客户的外部业务时，各位嘉宾给出了不同的侧重点与实践路径： - **CI&T 的“吃自己狗粮”策略**：Marcelo 表示，作为一家帮助客户打造“AI 优先”企业的咨询机构，CI&T 自身首先必须在内部全面实践。在过去两年里，CI&T 研发了一系列内部 AI Agent 和平台，重塑了市场营销、财务、人力资源、法务等部门的流程。只有将内部试错总结出的经验进行提炼，才能更好地赋能客户端的供应链、商品规划（Merchandising）及 B2B/B2C 商业平台。 - **行业通用模式：内部工作流与外部体验**：Shahid 观察到，行业内部主要将 AI 用于工作流优化（例如文档自动摘要）和传统流程升级；在外部，则侧重于提升客户体验，如推出以对话/聊天（Chat-based）为核心的互动界面。 - **供应链与供应商端的赋能**：Adam 提到，Coupang 在加入 AI 研发初期，将核心精力放在了零售供应链的优化上。这不仅服务于最终消费者，更服务于平台上的供应商（Vendors）。整个研发路径包含三大核心环节：定义数据需求、利用数据训练机器学习模型、以及将模型上线并收集供应商反馈。 - **应对 AI 冲击的三大内部重组抓手**：Aaron 强调，AI 浪潮到来时企业经历了一场“神圣时刻（Come to Jesus moment）”。为了应对客户和团队的焦虑，Klizer 采取了三步走策略： 1. **整理数据**：将分散的数据梳理清晰，规范内部工作流。 2. **安抚与赋能员工**：正视员工对于“AI 是否会取代我”的恐惧，倡导诚信与透明，引导团队拥抱 AI。 3. **回归教育与基本原则**：将学习和实施方法带入业务中，使员工不仅能在工作中使用，还能应用到日常生活中。 --- ### [00:52](https://www.youtube.com/watch?v=Yv_ACluUNi0&t=52s) 阻碍企业全面落地与规模化应用 AI 的核心瓶颈 针对主持人 Lexi 关于“阻碍企业从 AI 试点跨越到长期大规模落地”的提问，嘉宾们从技术遗留问题、数据质量以及组织心理等维度进行了深层次剖析： #### 1. 遗留系统（Legacy Systems）与架构瓶颈 企业无法快速推进 AI 的最大阻碍在于底层系统过于陈旧。全球大量大型零售企业的核心业务依然运行在拥有 30 至 40 年历史的遗留平台（如 COBOL 或 IBM AS/400）上。这些单体架构（Monolithic Architecture）缺乏模块化和 API 驱动的特性，导致企业无法顺畅对接最新的大语言模型（LLM）或运行新的 AI 代理。 #### 2. 数据碎片化与设计阶段的数据缺失 - **数据碎片化**：Shahid 指出，绝大多数企业面临数据分散、格式不统一的问题。很多企业为了追求速度盲目推行 MVP（最小可行性产品），最终却因为缺乏高质量、清洗过的结构化数据而以失败告终。 - **缺乏“结果变量”的早期设计**：Adam 补充道，许多产品经理在初期设计软件系统时，只关注业务流从 A 点到 B 点的流转，却忽略了对“最终输出结果”（如商品是否成功上架、最终交付的时间戳）的记录与标注。由于设计早期没有分析师或工程师参与，导致后期训练监督学习（Supervised Learning）模型时缺乏关键的客观变量（Objective Variables）。 ``` [遗留架构与数据瓶颈对 AI 模型训练的影响] +--------------------+ +--------------------+ | 遗留单体系统 | --> | 数据高度碎片化 | | (COBOL / AS/400) | | (无统一清洗/规范) | +--------------------+ +--------------------+ | v +--------------------+ | 缺少关键训练变量 | | (无最终 outcome) | +--------------------+ | v +--------------------+ | AI 模型无法落地/ | | MVP 尝试以失败告终| +--------------------+ ``` #### 3. 员工对于岗位被替代的恐惧 在组织层面，对 AI 取代工作的焦虑构成了巨大的推行阻力。Adam 举例称，在与运营负责人沟通 AI 路线图时，对方常直接质问“这是否会取代我的团队”。解决这一问题的关键在于沟通艺术——通过引导对方意识到 AI 是为了消除每日手动填写 Excel 等繁琐的机械性劳动，从而将“替代焦虑”转化为“效率提升的共识”。 #### 4. 视角转变与代际差异 Aaron 引用了 Malcolm Gladwell 在 TED 演讲中关于 Howard Moskowitz 博士（为百事可乐和 Prego 意大利面酱进行市场研究）的经典案例：通过将数据分类并构建不同的偏好模型，最终开创了“特浓（Extra Chunky）”等多样化产品线。AI 的应用也是如此，需要企业根据自身的业务目标提出正确的问题。此外，企业内部管理层与新一代员工之间存在代际认知偏差，在推进技术落地前，必须先在业务哲学和问题定义上达成一致。 - 许多企业以为向员工发放 ChatGPT 或 Gemini 账号就是“AI 革命”，但这仅停留在浅层体验，真正的变革在于核心业务流程的重构。 - 解决遗留系统的技术债务不是一蹴而就的，需要制定长期的渐进式替换策略。 - 推动机器学习（ML）与大模型上线时，必须保留“人的温度（Human Touch）”，通过前期沟通化解员工疑虑，确保系统获得高采用率。 <ElicitationsGroup message="想要进一步探索本场研讨会讨论的内容，您可以选择："> <Elicitation label="了解如何解决零售企业数据碎片化与 AI 训练数据的提取策略" query="请详细分析在零售和电商领域，企业应如何重构数据架构以解决数据碎片化问题，从而更好地训练 ML/LLM 模型？"/> <Elicitation label="深入研究遗留系统（AS/400/COBOL）向现代 AI 架构演进的重构方案" query="企业如何将运行在 COBOL 或 AS/400 等遗留系统上的核心业务解耦，并过渡到 API 驱动的现代 AI 架构？"/> </ElicitationsGroup> I'm having a hard time fulfilling your request. Can I help you with something else instead? ### [34:09](https://www.youtube.com/watch?v=Yv_ACluUNi0&t=2049s) 交互范式变革与 AI 治理 随着 AR/VR 技术的演进与语音搜索的普及，未来的零售交互体验将发生根本性重塑。Aaron Pallares 现场向观众抛出一个引人深思的互动问题：“有多少人记得苹果早期刚推出 App Store 时，把 iPad 玩耍的孩子突然背着家长搞出 50 美元扣费账单的经历？”他指出，这种场景即将以语音和智能搜索的形式重化：当孩子随口说出一句“我想喝巧克力牛奶”，系统便自动下单并送货上门，导致账单悄然增加。面对这种技术演进，不仅家庭内部需要引入约束机制，企业层面更必须构建严密的 AI 治理体系。 Marcelo Vessoni 进一步深化了这一观点，直言“消费者就是真正的老板”。消费者行为的改变将对品牌方施加巨大的生存压力，不跟上节奏的企业必将被淘汰。他将当前的变革与 2010 年至 2015 年间的移动互联网革命进行了类比：当时许多公司对手机端这种新型数字界面持观望态度，最终智能手机彻底颠覆了视觉体验。而当下的 AI 变革比起当年的移动化进程，适应曲线将更加陡峭、颠覆速度也会更快。传统网页和手机界面的购物模式，正不可逆转地向智能体、语音交互和数字助手过渡。 - 数据佐证：近期的行业调查显示，未来将有 **30% ## Notable quotes > [00:27](https://www.youtube.com/watch?v=Yv_ACluUNi0&t=27s) 吃自己狗粮 — Marcelo Vessoni > [00:27](https://www.youtube.com/watch?v=Yv_ACluUNi0&t=27s) Come to Jesus moment — Aaron Pallares > [00:27](https://www.youtube.com/watch?v=Yv_ACluUNi0&t=27s) AI 是否会取代我 — Aaron Pallares > [00:52](https://www.youtube.com/watch?v=Yv_ACluUNi0&t=52s) 这是否会取代我的团队 — Adam Nguyen > [34:09](https://www.youtube.com/watch?v=Yv_ACluUNi0&t=2049s) 有多少人记得苹果早期刚推出 App Store 时，把 iPad 玩耍的孩子突然背着家长搞出 50 美元扣费账单的经历？ — Aaron Pallares > [34:09](https://www.youtube.com/watch?v=Yv_ACluUNi0&t=2049s) 我想喝巧克力牛奶 — Aaron Pallares > [34:09](https://www.youtube.com/watch?v=Yv_ACluUNi0&t=2049s) 消费者就是真正的老板 — Marcelo Vessoni ## Mentioned Lexi Burbey, Shahid Quraishi, Marcelo Vessoni, Adam Nguyen, Aaron Pallares, Sound Ventures, Ashton Kutcher, Guy Oseary, CI&T, 7-Eleven, Coupang, Klizer, Howard Moskowitz, Malcolm Gladwell, WMS, PIM, CMS, COBOL, IBM AS/400, LLM, ChatGPT, Gemini, App Store, iPad ## Video description Lexi Burbey, Investor, Sound Ventures Shahid Quraishi, Engineering Manager, R&D, 7-Eleven Marcelo Vessoni, SVP, Head of Retail & CPG, CI&T Adam Nguyen, Director, Analytics, Retail Product, Coupang Aaron Pallares, Partner Channel Sales Manager, Klizer Abstracts This session will explore how generative AI is transforming the retail, CPG, and eCommerce sectors by enabling new levels of innovation and operational efficiency. Panelists will discuss how companies are using generative AI for content creation, product design, and marketing automation, as well as optimizing supply chains and customer interactions. Learn how generative AI is reshaping product offerings and customer experiences while reducing costs and speeding up time-to-market. Gain insights into real-world applications of generative AI and how to integrate these technologies into your business to stay ahead of the competition. [Source](https://www.youtube.com/watch?v=Yv_ACluUNi0)

[0:27] Marcelo Vessoni: 吃自己狗粮 [0:27] Aaron Pallares: Come to Jesus moment [0:27] Aaron Pallares: AI 是否会取代我 [0:52] Adam Nguyen: 这是否会取代我的团队 [34:09] Aaron Pallares: 有多少人记得苹果早期刚推出 App Store 时，把 iPad 玩耍的孩子突然背着家长搞出 50 美元扣费账单的经历？ [34:09] Aaron Pallares: 我想喝巧克力牛奶 [34:09] Marcelo Vessoni: 消费者就是真正的老板

[12:21] Adam Nguyen: Do you like manually working with these Excel files inputting data on a daily basis for a couple hours every day? [12:21] Adam Nguyen: Well, we're going to help you get rid of that. [19:41] Aaron Pallares: It's like getting my secondhand iPhone from my sister... she's got all her pictures in there, I have to delete all that stuff and reset it. I don't have to deal with that when it's a fresh start. [28:33] Aaron Pallares: Can you imagine what AI would have looked like with Toys "R" Us? I mean, we'd definitely have to work two or three jobs just to keep up with that because of the experiences that our own kids are doing. [24:18] Adam Nguyen: find a setup to watch movies in 1080p that fits my window width [34:09] Aaron Pallares: How many of you remember when Apple came up and they were like doing all these apps... and then unfortunately we gave our kid the iPad or whatever and then we're like,

[0:04] So today we have a pretty exciting panel [0:06] for you um chatting through really a lot [0:08] of the trends that we're seeing across [0:10] retail, e-commerce, and consumer and how [0:12] AI is impacting each of those individual [0:14] categories. I'll introduce you or I'll [0:17] actually let the the team introduce you [0:19] that's on stage today. But just for a [0:21] brief introduction on myself, I'm Lexi [0:23] Burby. I'm an investor at Sound [0:24] Ventures, which is an early stage VC [0:26] fund uh led by Ashton Kutcher and Gio [0:29] Siri. So, we invest across stages, but I [0:32] lead our AI application investing. I'll [0:35] go down the line and just have each of [0:36] you introduce yourself, a little bit [0:38] more about you, your role, and then the [0:40] organization that you're with. [0:42] >> I'll go first. [0:43] >> Perfect. [0:44] >> All right. So, um good morning [0:45] everybody. Thanks for the opportunity. [0:47] It's going to be awesome. I'm Marcelo. I [0:49] I lead the retail and CPG practice for [0:53] CINT on a digital perspective and uh [0:57] CINT is a global digital specialist [1:01] a a partner that works with some of the [1:04] most iconic brands in the world, retail, [1:07] CPG and other verticals building their [1:09] digital platforms and helping them [1:11] navigate all the complexities of the [1:14] digital world and even more now with AI [1:17] and everything. That's it. [1:20] >> Hey. Hi. Uh my name is Shahed. Uh I work [1:23] for 7-Eleven at R&D department. Um my [1:27] main focus is on uh digital innovation [1:30] uh going across 7-Eleven and so mainly [1:33] we'll try to find any new business [1:36] opportunity and try to take that to uh [1:38] to the finish line. So that's where I [1:40] mainly work. [1:43] >> Hi everybody. It's a pleasure to be [1:45] here. Uh my name is Adam. I lead retail [1:47] analytics for Kong uh which is the [1:50] largest e-commerce platform amongst [1:52] other things in South Korea. Uh I'm [1:55] currently based in Soul uh but I grew up [1:56] in the US and from California. [2:01] >> Hello, good morning. My name is Aaron [2:02] Polaris. I am the partner channel sales [2:05] manager at Kaiser. And what we do is we [2:07] help simplify manufacturers and [2:09] distributors their e-commerce experience [2:12] digitally transforming that their WMS, [2:15] the PIMS, the CMS and everything else [2:18] that comes with an acronym. [2:21] >> Thank you all. So I think you know one [2:24] of the things that would be helpful to [2:25] kick us off is for everyone to go [2:27] through and give us a baseline. How is [2:29] your organization incorporating AI today [2:32] both internally and what you see uh [2:34] externally facing customers? Well, maybe [2:37] Marcel we'll start with you. [2:40] >> Okay. Thank you. Good question. [2:43] um as a as a company that we are in this [2:46] journey of uh helping our clients become [2:50] can I say AI first companies right [2:53] whatever that means and all the [2:55] complexities and opportunities we have [2:57] to do the same inside right we have to [3:00] eat our own dog food and uh we are super [3:04] proud technology one aspect but the more [3:09] complicated aspect is the adoption of [3:12] any sort of AI platforms and tools [3:15] inside an organization that will [3:16] actually you know be uh change the way [3:21] we do business change the way we run [3:23] some internal processes and I think [3:25] that's that's the most important thing [3:27] that we've been trying to do inside CNT [3:29] to kind of learn with our own journey [3:31] and then apply the learnings towards our [3:35] clients right so in the last two years [3:38] been building our own set of agents and [3:41] platforms arms that are changing the way [3:43] we work internally in marketing, [3:45] finance, HR, [3:48] legal, all sorts of business process and [3:51] then learn from that and then apply the [3:53] same learnings to our clients right and [3:55] of course uh when it comes to retail CPG [3:59] clients it's way beyond that I'm talking [4:02] about um back back offs business cases [4:06] but when it comes to those clients it [4:08] has to be related to the main business [4:10] process the supply chains, the [4:12] merchandising, the B2B, the B2C [4:14] platforms. So, it's been quite a journey [4:17] to learn you know using AI in those [4:21] contexts. [4:23] >> Um, I will speak more on like what's the [4:26] current industry trend that is going on. [4:28] uh like what I've seen internally it's [4:31] mainly um focusing on improving the [4:34] workflows like uh document summarization [4:38] and then improving some of the internal [4:40] process that's a more legacy uh way of [4:43] doing it and externally I've seen mainly [4:47] focusing on customer experience for [4:49] example how we can have more uh chat [4:52] based uh oriented conversation more [4:55] conversationdriven [4:57] uh output [4:58] and how we can improve the current [5:00] customer experience on using some of the [5:03] uh internal flows. So yeah, that's what [5:06] I have seen across. [5:08] Um so on our side when I started our [5:12] focus was to improve the supply chain uh [5:14] part of the retail business and what [5:16] that means is that our customers are not [5:18] just the customers sitting at home [5:19] purchasing on the e-commerce platform uh [5:22] but also the vendors that sell products [5:25] to these customers. How do we make their [5:27] lives more efficient? How do we make [5:28] things easier for them and how can we [5:31] leverage data? A lot of the work when I [5:33] first joined was to scope what that work [5:36] uh what that data needs and what those [5:38] requirements from them could be and [5:41] figure out ways with the product team to [5:43] leverage AI. Um, a lot of these things [5:47] start off with defining data [5:49] requirements, [5:51] figuring out how they can be used to [5:53] power and train ML models, and then [5:56] lastly, working with our vendors, [5:58] getting feedback, uh, if we do launch [6:00] something, um, and seeing how the market [6:03] adopts it. uh if you don't have these [6:05] three components, it makes it very hard [6:07] to to leverage what we have to what's [6:11] going on nowadays in in sort of the Bay [6:13] Area and tech space using LLMs and and [6:16] all that. Um [6:20] I like what my colleagues said. I think [6:22] uh we all had that uh come to Jesus [6:24] moment when AI hit the market because [6:26] you know here we have a tool that we [6:28] have to be on the forefront of and be [6:30] able to explain it to our customers [6:31] while simultaneously building our [6:33] companies and at the same time they have [6:35] access to that resource right so we're [6:37] having to explain well how do you use it [6:39] better you know what do you do it so you [6:41] know what we did is you know we [6:43] basically took the bull by the horns and [6:44] we said look this is how we're going to [6:45] do it we started first with our data [6:47] right we started looking at our data and [6:48] said look it's all over the place let's [6:50] just you know like as I mentioned get [6:52] our workflow straight right internally [6:53] again paralleling all of the things that [6:55] are coming down the pike and the growth [6:57] plan that we had the other thing was the [6:59] people right we had to get our team in [7:01] in in line to be able to understand that [7:03] hey look AI is here there's nothing we [7:05] could do about it you know uh we got to [7:07] confront it we're going to do it with [7:08] you know integrity and honesty um we [7:11] know what we know we don't know what we [7:12] don't know and we're going to do this [7:14] and we're going to be able to explain it [7:15] and try to calm the nerves of your team [7:17] that obviously come up to you and say [7:19] hey am I going to get replaced you know [7:20] that's a lot that's going on, you know. [7:22] And then lastly, it's, you know, just, [7:24] hey, you know what? Let's go back to our [7:27] education, right? Let's go back to our [7:28] our principles of how we learn and how [7:30] we're going to implement that because [7:32] not only are we, you know, teaching them [7:34] how to do it in the business, but [7:35] they're going to take it again and do it [7:37] in their homes. They're going to take it [7:38] and do it like that. So, it was all that [7:41] mixed in with sales and marketing and, [7:43] uh, you know, everything that happens in [7:45] e-commerce. So it's a challenge that [7:47] we've, you know, embraced and we hope [7:50] that you guys too. [7:52] >> So when we meet with early stage [7:54] companies, one of the things that we [7:55] always see is there's, you know, a [7:56] couple stages of grief with AI when it [7:59] comes to enterprises that they're trying [8:00] to sell into. The first is fear, the [8:03] second is curiosity, the third is [8:05] experimentation, and then the fourth is [8:07] kind of a long-term contract or [8:10] something that ends up being a longer [8:11] term engagement or usage. What are the [8:13] things that prevent you from or prevent [8:16] organizations broadly if it's not yours [8:18] from actually adopting AI today? And [8:21] what are the biggest pieces that you [8:22] think uh push us farther and farther [8:24] from being at the enterprise AI adoption [8:27] goals that you know I think each of us [8:29] from our most ambitious selves hopes [8:31] for. [8:34] >> I'll go first. All right. So some of the [8:36] things that we see that are [8:39] um well one thing a quick step back one [8:42] thing is to say that our organization is [8:45] is going towards an AI revolution by [8:50] uh the leadership deploying chat DBT or [8:53] Gemini or any other model to the people [8:55] inside organization [8:57] that's okay I think everybody should be [9:00] trying to do that and doing that and [9:02] running experiments the other aspect is [9:05] actually change the way some of the [9:07] business process that are key to the [9:10] operation of those companies will [9:12] actually be transformed. Right? So [9:14] that's a second stage of this revolution [9:16] inside companies in adoption. Um, one [9:19] one thing that we see that is preventing [9:22] some of those companies to moving into [9:25] actually [9:27] activating value from from those pilots [9:30] and PC's and running experiments is [9:34] there's a gap between what AI and the [9:37] platforms and the dreams are allowing [9:40] them to build versus the reality in [9:43] terms of the whole ecosystems that those [9:45] companies operate on. Right? As you all [9:49] know, the world runs on legacy [9:51] platforms. Platforms that have have been [9:54] running for 25, 30, 40 years. [9:58] Do you know the amount of a Cobalt AS400 [10:01] code out there is huge? We're talking [10:03] about millions and millions and hundreds [10:05] of millions of code that have been [10:07] running for 40, 30 years, right? So it's [10:11] not easy to kind of deploy a new way of [10:15] working, the new way of deploying [10:17] technology, new technologies when the [10:20] pretty much everything is running on [10:22] legacy platforms. So we need some [10:25] strategy to little by little change the [10:28] way those underlying platforms are [10:30] operating and that's a journey right. So [10:33] people companies are kind of realizing [10:35] that nowadays and say yeah it's very [10:38] very hard to move [10:40] um at speed when you have those [10:44] ecosystems but they have to start [10:46] somehow somewhere right so it needs to [10:50] have an strategy for that so I think [10:53] that's one thing that is preventing [10:54] things from moving even faster nowadays [10:56] but I'll compliment later I will let my [10:59] my colleagues here to compliment [11:01] >> um so I do agree with what Marcelo just [11:03] mentioned right it's mainly our legacy [11:05] system that we're using which is not [11:08] modular and we are not having a focus of [11:10] having an APIdriven first architecture u [11:13] that's where we actually have all those [11:16] issue where we can't scale AI at the [11:18] same pace for example if you want to run [11:20] some new model uh that is something that [11:22] we can't do it with our monolithic [11:25] legacy uh architecture that we have it [11:28] one thing I would like to add is legacy [11:31] system is one of reason but I based on [11:34] my 15 plus years of experience working [11:36] in I have seen that data play a huge [11:39] role in that too um most of the system [11:42] that we are seeing right now have issue [11:44] with the fragmented data u the data is [11:46] pretty scattered across and we don't [11:49] have one well structured clean data [11:51] where we can uh fit this to the system [11:54] and try to get the more meaningful [11:56] output um I've seen some uh companies [12:00] where they try to uh rush um like going [12:05] with their MVP model and it usually fail [12:08] and one of the reason primary is due to [12:11] uh data not being uh fragmented prop [12:14] clean properly and yeah that's lead to [12:17] the failure for some of the quick MVP [12:22] >> uh I'll try to tackle this uh question [12:24] into two two parts uh first on the [12:27] technical side from what I've seen uh [12:29] and then secondly more from maybe the [12:31] human or political side. Um, similar to [12:34] what my colleagues have said, legacy [12:36] systems tends to be a bit of a problem, [12:38] but also for relatively new companies, [12:41] um, issues with maybe knowing how to [12:43] scope data systems to train AI models, [12:47] uh, is a problem as well. um younger [12:50] companies might not have the experience [12:52] or know what the endgame looks like if [12:54] you're trying to build a data set to [12:56] train um a machine learning model, [12:58] especially if they use supervised [13:00] learning on top of reinforcement [13:01] learning. Um [13:04] sometimes when someone is scoping out [13:06] like say a product manager scoping out a [13:08] software system to track um maybe a tool [13:11] to support vendors, right? [13:14] uh they might think, "Oh, we should [13:16] implement the time stamp to know when [13:18] it's in um maybe in the the beginning [13:20] part of the the uh uh sign on funnel for [13:24] items." Um but oftent times the focus is [13:27] always uh how do we get from point A to [13:29] point B to point C and they forget to [13:33] also include the outcome uh of the the [13:36] software system. So has the product been [13:38] released to the customer? What does that [13:40] time stamp look like? [13:41] um do we track it in binary or in some [13:44] continuous uh variable. Uh these types [13:47] of discussions need to happen in the [13:49] early stages during the design phase. Uh [13:52] for many companies it's very [13:54] productdriven and not enough analysts or [13:56] engineers are in those types of [13:58] meetings. Uh that creates gaps in what [14:01] you can actually use to train an ML ML [14:04] model. Uh, one of the big takeaways I [14:06] found from starting my job was to always [14:09] be in the room when you're in the design [14:11] phase at the very beginning so that you [14:13] have a a seat at the table. Uh, if you [14:16] don't do that, you might realize, oh [14:17] crap, I have to launch this LLM and I [14:20] don't have objective variables or uh end [14:22] goal variables to to train it on. Uh, [14:25] and then that can cause a whole set of [14:26] problems. Um, and for the second one, [14:29] maybe these days there's a bit more [14:31] acceptance of it, but about a year and a [14:33] half ago, there were a lot of fears [14:35] about AI taking over jobs, especially [14:36] for operations folks and and for [14:38] vendors. I've had several conversations. [14:40] I think Erin, you mentioned fears about [14:42] this too of your experience uh where I'm [14:45] sitting with uh an operations lead uh to [14:49] talk about a new ML roadmap item on our [14:51] team and they ask straight to me like, [14:54] is this going to eventually replace my [14:56] team? And although I don't know what the [14:59] future holds because these are [15:00] directives that might come from our uh [15:02] seauite uh I don't think ML right now is [15:06] anywhere near uh the ability to replace [15:08] humans. So my question back to him was [15:12] do you like manually working with these [15:14] Excel files inputting data on a daily [15:16] basis for a couple hours every day? And [15:18] he said no. [15:22] So my answer to him was well we're going [15:24] to help you get rid of that. And then [15:25] that that helped with the adoption [15:27] discussions uh that that I had. If it [15:30] weren't for these, you know, types of [15:32] meetings though and clearing up this [15:35] misconception, you may run into friction [15:38] when you're trying to launch an ML model [15:40] at scale. And so it's it's always [15:42] important to have that human touch as [15:44] well uh just to to ease any concerns. [15:47] But yeah, th those are just two examples [15:49] of what I've seen. [15:51] >> It's hard to follow up with my [15:52] colleagues here, but uh I'll give it a [15:54] shot. So for us was the perspective, [15:56] right? Um the greatest way to explain it [15:59] is um what Malcolm Gladwell said in his [16:02] YouTube uh uh TED talk about Dr. [16:05] Moscowitz. Dr. Moscowitz was hired uh by [16:08] Pepsi to be able to see how uh he could [16:12] get the perfect Pepsi. This is back when [16:14] aspartane was trying to be added to [16:16] Pepsi. And he talks about how like you [16:19] know when this guy is a psycho physicist [16:23] and psychoanalysist and stuff like that. [16:24] So he's well degree but he was asking [16:27] the wrong questions and he basically [16:29] lost the account right and it bothered [16:31] him so much and then all of a sudden he [16:33] gets you know contacted by you know the [16:35] makers of PGO and stuff and and then all [16:37] of a sudden he's like you know what I'm [16:39] going to do is I'm going to go ahead and [16:40] basically set up these datas and these [16:42] individuals in different type of [16:43] clusters and then from then I'm going to [16:44] make you know some some [16:48] data models out of the information that [16:50] I got and then all of a sudden that's [16:51] how we got like you know extra chunky [16:53] and and you uh traditional and all these [16:55] flavors kind of deal. So AI is the same [16:57] way, right? It's being exposed to the [16:59] world and now it's going to tailor to [17:02] your liking according to your business, [17:04] according to your vision, according to [17:05] your models. So in our field, we have to [17:07] figure out what that is, right? Because [17:08] you know, not only are you exposed to it [17:11] as business owners, but you're also [17:12] exposed, you know, your kids are exposed [17:14] to it, right? And your kids are like the [17:15] people that are yapping in your ear and [17:16] say, "Hey, dad, you need to do this." [17:18] Whether it be, you know, your [17:19] 10-year-old or it be somebody who's [17:20] working in your field and stuff. And so [17:22] there's generational gap there as well. [17:24] So we embrace it through a perspective [17:26] of really good answering questions and [17:28] stuff like that before we even get into [17:30] the tech. And again, I agree 100% that [17:33] the the a lot of the tech is really [17:35] legacy. So we're having to figure that [17:36] out and stuff like that. So we go [17:38] through a process of the bots, we go [17:39] through the process of automation, we go [17:40] through the process of integration and [17:42] all that. So now how do we take this [17:44] thing called AI and how do we make it [17:45] take it to the next level? That's [17:47] depending on each and each and [17:49] everybody's definition of that. So, [17:52] >> and Adam, I just wanted to touch on, you [17:55] know, something that you mentioned about [17:56] how your organization [17:59] gets ready for AI politically, and I [18:01] think each of you mentioned this in your [18:02] own ways. What would it take for a brand [18:05] that is trying to sell or a a company [18:07] rather that is trying to sell into [18:10] Coupang as a third party vendor? What [18:13] would it take for them to prove ROI to [18:16] essentially uh you know get buy in [18:19] internally and push it over the finish [18:21] line if you are going with the buy [18:23] versus build decision? [18:26] >> That's a great question, Lexi. [18:29] I would say when I speak to vendors on [18:32] potential software uh systems or tools, [18:35] I look for happy customers and example [18:39] cases of it being used because for me I [18:42] know that there are very unique problems [18:44] for a company like ours and not one [18:48] glove fits all um will you know help [18:51] with the issues maybe or complex data uh [18:56] challenges that we we deal with on a [18:57] day-to-day basis. [18:59] So if I see an example that matches my [19:02] own experience without, you know, being [19:04] prompted or or me having to prompt them [19:06] on what I'm looking for, uh I would be [19:09] more inclined to to uh see the tool as a [19:12] potential solution to issues that we we [19:15] are constantly working on. Uh and then [19:17] of course there's always another um [19:19] aspect to it where uh is it something [19:21] that can add enough value versus um us [19:23] building it inhouse and then all these [19:25] other things start to come into play. [19:27] But if you don't have that first step, [19:29] it makes it very difficult to to just be [19:31] in that same conversation, especially if [19:33] I need to sell to you know my team or [19:36] colleagues as well that this is the [19:37] right approach. [19:41] And one thing I think each of you, this [19:43] is a challenge on perspective here, but [19:46] one thing each of you may have an [19:47] interesting um thought on is where AI [19:51] native brands have advantages over more [19:54] legacy or industry players who have been [19:57] around for in some cases hundreds of [20:00] years. Where do you think brands that [20:02] have started in the last five will have [20:04] advantage? And I'll start with you [20:05] Aaron. [20:06] >> Thank you. So there's that old saying [20:09] that says you can't pour, you know, new [20:10] wine into old wine skins and stuff like [20:12] that. It basically talks about a [20:14] mindset, you know. Um I am 100% in awe [20:19] of the people that have come before me [20:20] and I respect them and um enjoy working [20:23] with many of them. As a matter of fact, [20:24] many of them are are my elders and stuff [20:26] like that. Um I've had the privilege of [20:28] being able to be in an organization that [20:30] is fluid with the change, but you have [20:32] organizations that are not. and uh you [20:34] know this is how we've done it for the [20:36] longest and and this is how we remain to [20:38] do it and you know so you're having to [20:40] battle that uh and it's and it's a [20:43] challenge because again you're painting [20:44] the picture in their mind that it's not [20:46] going to necessarily be anything that's [20:48] going to hurt them but actually benefit [20:50] them right so you have to storytell with [20:52] them and say look this is where you at [20:53] this is where you could be so now with [20:55] the people that are starting off fresh [20:58] they don't have to deal with old data [20:59] they don't have to deal with with with [21:00] that stuff they can start fresh they can [21:02] start new they can start you know basic [21:03] basically just like an open box like you [21:05] know it's like you know me getting my [21:07] secondhand iPhone for my sister I mean [21:09] she's you know she's got all her [21:10] pictures in there I don't want to see [21:11] them and stuff I have to delete all that [21:13] stuff and maybe reset it and she's got [21:14] the scratches right I don't have to deal [21:15] with that when it's a fresh start when [21:17] it's a fresh start it's like okay what [21:19] can we do what have we we have what can [21:21] we architect that you know that we [21:23] haven't architect before you know so [21:25] then the sky's is the limit there and [21:27] again no disrespect to the legacy and [21:29] stuff like that because they are the [21:31] foundation that we stand on and stuff we [21:33] actually like and and and and enjoy [21:35] those conversations because we want to [21:36] get him out of those legacies into the [21:38] new and it's cool to see, you know, the [21:40] one gentleman that stands up and stuff [21:42] like that who would go in with his [21:44] Johnson Murphy's and stuff and then all [21:45] of a sudden at the next conference you [21:46] see him with like Nike shoes and you're [21:48] like no way this guy has jumped on the [21:50] bandwagon to be able to say hey the you [21:52] the new is in kind of deal. So that's [21:54] awesome and stuff like that. So yeah. [21:57] >> Yeah. So just to add to what uh Aaron [22:00] just mentioned right I have seen the [22:02] same thing with the new companies if [22:04] they are following the modular [22:06] architecture it's very easy for them to [22:08] do something new and they are mainly [22:11] cloud first environment so they don't [22:13] have to worry about what their legacy [22:15] system look like and one more thing they [22:18] have advantage is from day one they can [22:21] have a very clean well structured data [22:24] to work with we have seen like legacy [22:26] system are pretty tight to um data from [22:29] different sources. So they are it's [22:31] become little bit hard for them to [22:34] structure the data and then do it. But [22:36] that's one of the advantage uh they have [22:38] it like if you start new uh we can start [22:40] with a very flexible and clean data on [22:43] that [22:44] >> just to complement um you're right I [22:47] agree with that it's uh companies that [22:50] were born in the digital transformation [22:53] era has they have those advantages [22:56] and now we are entering into this AI [22:58] transformation era and one of the [23:01] biggest advantages is the access to data [23:05] a more universal unified catalog of way [23:09] of transacting that data in the supply [23:11] chain, the value chain. Um, and you are [23:14] able to provide a better experiences, [23:16] better personalization experience to [23:18] your customers. You can have access to [23:21] realtime data coming from your [23:24] customers. You can optimize your [23:26] process. you can you have better ways to [23:29] reduce for example [23:31] onshelf availability which is a very [23:34] important metric that perhaps some of [23:36] those big brands they are still [23:38] struggling to fix for real right so [23:41] things related to those metrics and KPIs [23:45] that actually are related to the whole [23:47] value chain of the products and the data [23:52] I think that's the key advantage [23:54] >> and uh one more thing I've seen that [23:55] with the legacy system if you wanted to [23:57] scale your AI model u it's very it's [24:00] very difficult your u legacy [24:03] architecture will not be able to handle [24:05] or do that but if you have a complete u [24:08] more like more like a modular [24:12] architecture it's become very easy to [24:14] scale your AI models okay [24:18] >> and in that vein how do you think that [24:20] AI will continue to transform the way [24:22] that consumers shop both on their [24:25] preference es on discovery on the actual [24:29] experience of buying. What are some of [24:31] the things that you've been seeing or [24:33] hope to see over the next couple of [24:34] years? [24:36] >> Um I I think that most um the way we [24:40] search will be completely different. It [24:42] will be less browser search. It will be [24:44] more like conversation having with your [24:46] assistant and for example it will try to [24:49] understand what you're trying to do like [24:51] based on your search habit. uh it can [24:54] find out hey for example if you're [24:56] trying to buy something for your family [24:58] members it can easily identify what your [25:02] like your family is pretty much into [25:04] outdoor activities or some other uh [25:06] activities uh it can identify that and [25:09] then curate and summarize some of the um [25:13] like what you can buy and then it will [25:15] give you suggestion so I feel it will be [25:18] completely different we will be using [25:19] less browser for and it will be more [25:22] like uh assistant driven. [25:26] >> Just to add to that, I I think search [25:28] will be a bit more broad uh moving [25:31] forward. Um you know, nowadays you have [25:33] to search for like a specific TV uh with [25:36] dimensions and and all those things, but [25:40] with the power of AI, you might be able [25:42] to make something a bit more broad. um [25:46] like I want to be able to see a movie um [25:49] you know in 1080p [25:53] um and specifically that captures um [25:57] this width of um my window or something [26:00] like that and they might not an LLM [26:02] might recommend you not just TVs but a [26:04] projector or some other um venue to to [26:07] watch the the movie on that might be in [26:10] existence of the e-commerce platform [26:12] things that you probably wouldn't have [26:13] thought about before. Um, so more [26:15] discovery work will will come up because [26:17] of it and you might be surprisingly [26:19] delighted. [26:22] I think the the future of shopping is [26:24] being designed right now as we speak and [26:27] in 5 years will be very different. [26:29] There's a potential to be very [26:31] different. [26:32] It's going beyond the search which is [26:35] extremely important. Those are the [26:37] initial experiments of the way you [26:40] transform the way you shop starts with [26:43] the discovery and the search. Um you may [26:48] you may have seen that um OpenAI and and [26:52] and Shopify for example are kind of [26:55] working together to deploy checkout [26:58] experiences inside the chat interface. [27:01] Right? So that's the beginning of a [27:04] potential revolution and things will [27:06] accelerate even faster moving forward. [27:08] Right? The way you need to optimize your [27:12] content, your products, the description [27:15] of your products or services on how to [27:18] sell to AI, right? You're not selling to [27:21] people anymore. You're selling to AI, [27:23] which will be an agent, will be one of [27:25] those models, the the core models out [27:28] there. the the concept of a GEO, right, [27:31] is going to be extremely important. I [27:33] think you can talk about it as well, [27:35] right? This the beginning of a a a huge [27:38] revolution and the technology to enable [27:42] that revolution [27:44] is taking shape as we speak, right? One [27:46] of the potential candidates that I I [27:49] think people on the technology side here [27:52] could take a look at is the concept of [27:55] those MCP servers, right? uh which is a [27:59] way for a agent to understand an [28:02] underlying platform to pull data and [28:05] then actually activate an action using [28:08] this kind of a middleware right so more [28:11] and more companies I see that I'm having [28:13] conversations with our clients and [28:16] everybody is talking about running [28:18] pilots exposing some of the underlying [28:21] services via right now MCP servers it [28:25] might change in the future might become [28:27] another standard from Google, for [28:29] example. But that's the future. It's [28:31] inevitable. [28:33] >> Most definitely. I would agree with [28:34] that. Uh, you know, the one thing that [28:37] just kind of popped out when this [28:38] question was asked was actually [28:40] buildings, right? When I think about, [28:42] you know, everything that's coming down [28:43] the pike and especially like e-commerce [28:45] and stuff, I think about what's [28:46] happening with the retails about the [28:48] buildings that they have, right? So, [28:49] you're not getting the traffic that you [28:50] used to get because everything is like, [28:52] "Hey, just send it to me, right?" Uh, [28:53] Amazon did us disservice back when it [28:56] started doing its thing because of the [28:57] fact that it kind of set the tone where [28:58] everybody was like, I want an Amazon. I [29:00] want an Amazon. I want an Amazon. And [29:02] uh, we were like, well, you're not an [29:04] Amazon, so therefore you don't have an [29:05] Amazon budget, so therefore we have to [29:07] kind of like figure out what you can do [29:08] at your level kind of deal. And we [29:10] helped a lot of, you know, individuals [29:12] grow into that model kind of per se. But [29:14] the biggest thing is to see, okay, now [29:16] we are going into the world of retail [29:19] and we're going to go into the [29:20] e-commerce and now it's gotten basically [29:22] steroids by AI. How is that going to [29:25] work out with your buildings? What [29:26] experiences are you going to do and [29:27] continue to do? Are you just going to [29:28] turn out the situation for like you're [29:31] not going to have a building and [29:32] everything's going to be just [29:33] distribution? So, it's it's not like a [29:36] onesizefitit all because again, AI is [29:38] not. Um, it's the really trying to [29:40] figure out what that model long term is [29:42] going to be because you can have a great [29:44] store and then, you know, nobody shows [29:45] up to your store and want to go and walk [29:47] inside kind of deal if that's your [29:48] model. Or you can be like, you know [29:50] what, we're just going to have a [29:51] different thing. The one experience that [29:52] blows my mind is Toys R Us. Man, I love [29:54] Toys R Us, man. I used to go there. I [29:55] used to actually work at Toys R Us. It's [29:57] gone. I like I go on in Christmas and [29:59] I'm like, where do I go? I want to go on [30:00] live. No, but I want to go touch it. I [30:02] wouldn't take my kid to go touch it and [30:03] do have that feeling, you know. Uh uh uh [30:06] Gary Vee, uh Vaynerchuk talked about how [30:08] they could have done a whole experience [30:10] better that if they would have just kind [30:11] of like tweet and been along for the [30:13] ride about what's going on. Can you [30:15] imagine? I mean, what AI would have [30:16] looked like with Toys R Us. I mean, we'd [30:18] definitely have to work two or three [30:19] jobs just to keep up with that because [30:21] of the experiences that our own kids are [30:23] doing. And believe me, they're coming [30:24] for your kids, believe it or not. So, [30:26] they're going to come into education. [30:27] They're going to come into different [30:28] things. So, it's all intertwined in that [30:30] aspect. So, if we think it just is going [30:32] to do it in the retail part, well, you [30:34] are the retail. You are the person that [30:36] they're looking for and they're going to [30:37] be looking into where you shop, how you [30:39] shop, when you shop, you know, what [30:40] color, what, you know, what time frame, [30:42] what mood, everything. All of that is [30:44] data. It has it now. Now, it's just a [30:46] matter of tailoring it to an experience. [30:48] So, [30:50] >> it's it's funny that you mention, you [30:52] know, the way that these companies are [30:54] influencing our kids. And I think it's [30:56] something that's really relevant today [30:57] as we think about, you know, the next [30:59] generation and how their preferences [31:00] change and how that relates to [31:02] consumers. When we think about the next [31:05] generation, what we're seeing is there's [31:06] a lot of companies that are actually [31:08] changing the way that those consumers [31:10] shop or the way that they really [31:12] discover anything, not even just [31:14] shopping, not a an intent to purchase, [31:16] but the way that they discover. And [31:18] that's I think the the magical moment [31:20] that used to be you know going to Macy's [31:23] for back to school shopping is now a you [31:26] know a moment in time where you're [31:29] really excited about seeing a technology [31:32] in a new interface or an interface that [31:34] you're familiar with. So what we're [31:35] seeing is voice we're seeing uh [31:38] applications within video. We're seeing [31:40] discovery and chat that is intentbased [31:43] instead of searchbased. And you're [31:45] seeing really companies that are [31:46] tackling this next generation and [31:49] meeting them where they are. And that's [31:51] something that I think has uh you know [31:53] really been something that's been unique [31:55] to the AI transformation relative to [31:57] other previous technology that we have [31:59] seen uh at the innovation layer. Some of [32:02] the things that we we kind of spoke [32:04] about as a group beforehand were also [32:06] how these behavior changes then [32:08] influence infrastructure changes. So [32:10] when you think about building your [32:12] companies from here and continuing to [32:15] iterate on the back-end technology that [32:17] powers some of these AI applications, [32:19] what are the biggest changes that you [32:22] think you'll be able to see in the next [32:24] couple of years that really are as a [32:26] direct result of the behavior changes of [32:28] your consumers or your clients [32:30] consumers? [32:34] >> Anyone wants to go first? [32:36] >> Um I mean I I can go first a little. Um [32:40] I think multimodal input might be a big [32:42] thing over the next year or so. Uh [32:44] multimodal will just means that instead [32:46] of just text you can in include things [32:48] like pictures for example or uh voice um [32:51] on what you want to come out of your [32:53] shopping experience. Um, imagine a world [32:56] where maybe you you take a picture of [32:58] your Christmas tree and then you upload [33:00] that to a search platform on an [33:01] e-commerce site and then that search [33:03] platform will take apart each part of [33:05] that image, get you the ornaments, where [33:07] to get the tree, give you a dozen [33:10] options for each of these things and [33:11] lets you pick actually it'll probably [33:14] even recommend what which ones look the [33:16] best for you and then you just add to [33:18] cart and and purchase. Uh I think to do [33:21] that type of streamlined experience, [33:23] teams that work on this work, uh the [33:26] stuff will have to update their [33:28] databases to store image data, to store [33:31] um sound data, not just text data. Uh [33:35] and that will also be a reinforcement [33:38] for what the machine learning models [33:39] will get to be trained on as well. It's [33:42] kind of like a loop, right? Um but all [33:45] of this stuff, it's going to be [33:46] expensive. if it's going to going to [33:48] require a lot of scale. Companies are [33:50] going to have to figure all that out [33:51] over the next year. [33:53] >> Yeah. Um just to add to that, it will be [33:56] more like conversational and rather than [34:00] actually having a textual based uh [34:02] interaction. So for example, you can [34:04] have a very visual representation of how [34:07] you are going to shop in a store. So you [34:09] will have a AR VR uh experience where [34:12] you can sit at your home and [34:15] just shop at uh any stores. So it will [34:17] be the way we are going to interact is [34:19] going to completely change. [34:24] >> I wanted to jump in with a question to [34:25] the audience. How many of you remember [34:27] when uh Apple came up and they were like [34:29] doing all these apps and like you know [34:31] people were downloading the apps and [34:32] then unfortunately we gave our kid you [34:34] know the iPad or whatever and then we're [34:35] like where did this $50 charge come from [34:38] from you know like how many people had [34:39] that experience? Well, imagine that [34:41] experience in voice and search, right? [34:43] You're going to have it. And the kid and [34:44] and they know referring to the people [34:47] and and and the the people that are are [34:49] the voice, right? You know, you guys are [34:51] all the voice. You know, you you you [34:52] speak and and you're heard in your [34:55] community and everything like that to [34:56] the people that excuse me are creating [34:59] this voice, they're going to be the ones [35:00] to be able to dictate, right? And again, [35:02] as always, there's always going to be [35:04] some that going to rise up to the top [35:05] and in that, you know, companies are [35:07] going to merge, etc., etc., etc. I say [35:09] that because you know we talked about [35:10] the kids and stuff like that. So don't [35:12] be surprised later on if all of a sudden [35:13] you get an now a 99 thing like that just [35:15] because of the fact that your kids said [35:17] hey I want chocolate milk and then all [35:18] of a sudden you have chocolate milk in [35:19] front. So those experiences are going to [35:21] be changing and stuff. So it's something [35:23] to embrace. Yes, there's got to be some [35:25] you know some some accountability in [35:27] that aspect you know not only at the [35:29] home right but uh but also with the [35:31] governance of AI and stuff which is a [35:33] huge topic as well. Um, but to embrace [35:36] the the the the natural progression of [35:38] what this is going to look like and you [35:40] guys get to decide what that looks like [35:41] for your clients. So [35:43] >> to to rephrase it, the customer or the [35:46] consumer is the big boss, right? [35:48] >> Yeah. [35:49] >> The consumer behavior will change and [35:52] that will put pressure on brands to [35:54] adapt very fast. If the brands are not [35:57] adapting, [35:59] they will die for sure. That's that's a [36:02] certain fact. Just like for example [36:05] maybe in 20 2010 15 years ago everybody [36:10] talking about the mobile revolution but [36:12] just a few companies were kind of [36:14] embracing still a new digital interface [36:17] which is your phone right everything was [36:20] web based [36:22] and then of course because of iPhone and [36:24] and Android that was the beginning of a [36:27] revolution on the way people behaved [36:30] when it comes to all sorts of visual [36:33] experiences. [36:34] And some companies took some time to [36:37] adapt to that. I think we're going [36:39] through the same cycle again, but now [36:42] it's going to be faster than that. [36:44] Right? I think the adaptation curve [36:47] might be even shorter than the previous [36:49] transformation, the previous disruption [36:52] that we all face it. Right? the new [36:54] generation, our kids, the young people, [36:58] ourselves will change the way we do lots [37:01] of things starting with shopping and [37:03] that'll put pressure on brands to adapt [37:05] very fast. It's clear the shopping [37:08] experience will transition from the [37:11] traditional web and cell phone interface [37:13] just to something else. Whether or not [37:16] you want to call the agents or voice or [37:19] chat bots or digital assistants [37:22] up to us to define and then brand it but [37:25] that's clear. So um just to add one more [37:28] thing uh there was a recent survey um [37:31] out recently where they were saying that [37:33] around 30 to 40% of the way the people [37:37] are going to shop is going to change [37:39] where it will be uh assistant based and [37:41] more personalized and around 40 to 50% [37:45] less traffic coming through uh browser [37:47] search. So yeah, [37:51] >> our last couple of minutes here, maybe [37:52] we'll go through a lightning round of [37:54] questions and so try to answer in one [37:58] sentence or less, ideally one word. Uh [38:01] we'll start with one AI term people [38:04] pretend to understand but don't. [38:08] >> Um [38:11] >> I'll start llm. Yes. Um yeah, I I was [38:16] going to add to that like more towards [38:18] like hallucination like trying to think [38:20] that okay u whatever ML is outputting we [38:24] assume that it's right but it it's not [38:27] it's all depend on what's uh your [38:29] underlying data is and it's more you [38:32] need to have a human uh in the loop in [38:34] middle to make sure that uh most of the [38:36] outputs are correct. [38:38] >> It's real intelligence or not. [38:41] >> Yes, exactly. That's a good one. [38:43] >> I'll say hallucinations. [38:47] >> And what's a job in retail or e-commerce [38:49] that doesn't exist today but will exist [38:52] in three years thanks to AI? [38:56] maybe um [38:59] AI agent orchestrator, someone that [39:02] drives [39:04] the operation of um multi- aent platform [39:08] that is responsible for tapping into [39:12] several business process in the whole [39:14] value chain. Maybe that's a new job that [39:17] will kind of revise and review the [39:19] decisions that those future agents will [39:22] do on behalf of the current platforms. [39:24] Yeah, I think it will be like a workflow [39:27] moderator. So, basically trying to [39:29] understand the tone of your uh company [39:32] products and trying to set that tone. [39:35] So, you'll have couple of moderators who [39:37] are going to uh set that tone for your [39:39] product. [39:41] I would say AI auditor um teams that [39:44] help make sure that the AI you're [39:46] building um is doing actually the right [39:49] thing and they must be a separate team [39:51] from the teams that are actually [39:52] building the AI otherwise there's a [39:55] conflict of interest. We we actually [39:57] have sort of needs that now uh we're [39:59] just handling it more informally. I [40:01] definitely love that everybody, if you [40:03] listen to the undertone, is basically [40:05] almost talking about checks and [40:06] balances, right? You know, we're [40:07] embracing it, but we also need to be [40:09] able to make sure that we have kind of [40:11] like a middle ground where everybody, [40:12] you know, is playing fair because just [40:14] as much as the good guys are using it, [40:15] the bad guys are also using it for other [40:17] stuff they're not supposed to be doing [40:18] it kind of deal. So, at Kaiser, what [40:20] we're trying to do is to make sure that, [40:22] you know, our customers are taken care [40:23] of in the aspect of what their vision [40:26] and model is. And you know if they don't [40:28] align and stuff like that or they need [40:29] another perspective we you know [40:31] encourage them to say hey look at it [40:33] from this perspective right you know um [40:35] because we would rather be in a project [40:37] that we're saying like it's going down [40:38] the right path and just because of the [40:40] simple reason that you know of a feel or [40:43] because of a client that we're going to [40:44] go and and make something that we're [40:45] just not proud of and we're just not [40:47] going to go ahead and and support [40:48] morally. So yeah. [40:53] And then AI will never replace blank in [40:57] retail and e-commerce. [41:00] >> My favorite restaurant, it will never [41:02] replace my favorite restaurant. It might [41:04] have replaced the experience, but not my [41:05] favorite restaurant. I'll still be able [41:06] to go there and stuff. So, [41:08] >> um I I think that it will be human u [41:12] emotions. Uh that will never be [41:15] replaced. Um [41:18] >> yeah, it won't replace my shopping [41:20] preferences. [41:21] >> Yeah. [41:23] Well, they might [41:26] I've seen what under [41:28] >> I don't know. [41:32] I think the brand, the core of the brand [41:34] itself. I think the the taste, the [41:38] feelings, the heart of the brand. I [41:40] think that's it. [41:42] >> And then last one, uh, in 10 years, what [41:45] percentage of shopping will be done [41:46] without ever seeing a product page? [41:49] >> Yeah. Um yeah I I just mentioned about [41:52] the recent survey around 40 40 to 50% of [41:54] the stock will completely change and it [41:57] will be more uh like personalized [41:59] assistant driven. [42:02] >> U maybe I'm I'm less uh bullish on it. [42:04] I'd say maybe like uh still maybe up to [42:07] 20% won't be seen. [42:10] >> I think what they'll be able to see is [42:12] more they're like did I really purchase [42:14] that? Who purchased this? Because at [42:15] that point they're like they have it in [42:16] front of them and if anything returns [42:18] are going to spike to the roof and be [42:19] like I didn't order this. Did you order [42:21] this? [42:23] >> I'm with you. I think all of us will [42:25] have some kind of a [42:27] agent that knows everything about you. [42:30] Your preferences, your taste, your past [42:32] behavior and all of a sudden you are [42:35] hyperargeted for everything. So I think [42:37] it's going to be close to 100%. [42:41] Well, that that is all the time that we [42:43] have, but in summary, I think, you know, [42:45] enterprise brands like those that are up [42:47] here and the customers that they work [42:49] with are embracing AI one step at a time [42:52] and are thinking about the implications [42:54] of what that means for consumers [42:56] broadly. You can find each of us after, [42:58] but really appreciate the time. Thank [43:00] you all. Thank you.