Retailers Deploy AI to Cut Credit Risk and Accelerate POS Operations

零售商应用AI降低信贷风险并加速POS运营 领先零售高管表示，通过AI实现信贷决策、门店入职及客户服务的自动化，能够提高转化率并减少运营摩擦。 Coppel、Ace Hardware和Vapi等企业的高管分享了人工智能在零售业的实际应用，涵盖从销售点（POS）实时信用评分模型到语音助手及员工内部支持等诸多方面。专家小组强调了人在推行AI过程中的核心作用，主张优先解决切实的运营痛点并实施严格治理，而非盲目追逐技术狂热。最后，专家们探讨了“自研还是购买”的策略，建议仅对能构成真正竞争壁垒的解决方案进行自主研发。 AREA15利用实时AI优化访客行程，将叙事元素融入实体体验，并促进其全新照片通行证的注册。 (19:39) Coppel在墨西哥同时作为零售商和金融机构运营，其80%的贷款是在门店内通过店内分期信用和“先买后付”（BNPL）直接发放，而非走传统银行渠道。 (19:39) Coppel在销售点采用实时预测算法，根据商品的耐用性（例如电熨斗与牛仔裤的对比）来评估违约风险。 (19:39) Coppel通过可联络性算法和OTP动态密码验证加强数据核验，以确保客户电话号码的安全可靠。 (19:39) 成功推行AI取决于人的因素：Ace Hardware将“Hey Arma”助手融入超过2700家门店的工作流程，在不增加员工负担的情况下加速入职培训。 (23:32)

Retail leaders are moving past artificial intelligence hype by anchoring deployments to concrete operational friction, real-time risk assessment, and human-centric workflows. Panelists emphasize that sustaining competitive advantage requires building proprietary core differentiators while maintaining rigorous governance over data and brand standards. Real-Time Underwriting in Retail Environments (19:39) Retailers extending point-of-sale financing operate under distinct operational risks, particularly when serving customers lacking traditional credit bureau records. Santiago Ontañon Alvarez reports that 80% of consumer financing at Mexican retailer Coppel is issued directly through in-store credit lines rather than bank loans. To underwrite these transactions, Coppel relies on multidisciplinary scoring models developed with psychologists, sociologists, and philosophers to predict repayment propensity. Product-Level Credit Risk Dynamics (19:39) Coppel's checkout scoring algorithms assess repayment risk based on the specific goods being financed rather than relying solely on basket value. Ontañon Alvarez explains that an identical purchase amount of 500 pesos ($25 USD) carries different risk profiles depending on whether the item is a durable household good, like an iron, or a non-durable item, like jeans. Because durable goods exhibit higher repayment rates, algorithms allow standard credit terms for the iron while requiring a down payment for apparel to offset risk. Integrating AI into Frontline Workflows (23:32) Deploying artificial intelligence in physical retail demands close alignment with frontline employee routines rather than imposing technical overhead. Melanie Hill details Ace Hardware's rollout of the "Hey Arma" assistant across more than 2,700 stores, designed to accelerate onboarding for new associates navigating complex inventories across product aisles. Rather than introducing administrative friction, the tool was tested directly with store associates to provide immediate product guidance during direct customer interactions. Automating Customer Support via Voice Agents (26:18) Customer service voice AI succeeds primarily when applied to focused, high-friction customer journeys with zero wait times. Nathalie Criou highlights that automated voice agents reduce drop-offs and human escalations when deployed for clear tasks such as payments, returns, and appointment scheduling. In the automotive market, Mexican platform Kavak implemented voice AI to handle repetitive inquiries and test-drive scheduling for car buyers, maintaining consistent support availability without expanding call center headcounts. Governance and the Build-Versus-Buy Framework Widespread access to AI tools allows non-technical employees to develop bespoke applications, but scaling across an enterprise introduces significant data unification and brand governance hurdles. Criou and Hill caution that companies must purchase commodity software for standard operational parity, while building and owning proprietary systems that define their core competitive advantage. To manage risk, retailers like Ace Hardware and Coppel enforce strict oversight via executive steering committees, departmental champions, and secure enterprise infrastructure. While AI lowers technical barriers to software development and operational problem-solving, retailers face ongoing challenges in enforcing unified data standards, security protocols, and brand values across widespread, decentralized implementations.

多位零售与技术专家指出，在人工智能重塑零售体验的过程中，企业制胜的核心不在于盲目追逐技术热点，而在于聚焦业务痛点与一线工作流，通过自主掌控核心差异化能力和严格的治理体系来实现可衡量的商业回报。 结合行为科学的结账即时信贷风控 (19:39) 墨西哥大型零售与金融机构 Coppel 展现了算法对传统零售金融的深度重构。Santiago Ontañon Alvarez 指出，虽然 Coppel 拥有 20 年的正式银行牌照，但其 80% 的消费信贷都是在门店内通过先买后付（BNPL）或循环信用等百货信用额度发放的。面对许多缺乏征信档案的客户，Coppel 组建了由心理学家、社会学家和哲学家构成的跨学科团队，构建自主信贷评分算法来预测还款行为。该机构在收银台运行即时机器学习评估已有超过 15 年历史。 该模型的核心特征在于根据商品属性动态调整风险。例如，当一名 60 岁女性同时购买价值同为 500 墨西哥比索（约 25 美元）的熨斗和牛仔裤时，系统评估耐用家居品（熨斗）的还款意愿显著高于常为家属代购的非耐用品（牛仔裤），从而仅对牛仔裤要求支付首付款以对冲违约风险。此外，Coppel 正利用现代云和 AI 基础设施来保障周日晚间等购物高峰期的即时计算速度，并结合一次性密码（OTP）对手机号进行即时验证。AREA15 的 Mark Stutzman 亦补充提到，他们通过即时 AI 构建连贯的叙事元素来优化实体照片通行证体验。 以人为本的一线赋能与语音交互体验 (23:32) Ace Hardware 与 Kavak 的实践表明，面向员工与客户的交互工具必须深度契合实际应用场景。Melanie Hill 介绍，Ace Hardware 在超过 2,700 家门店推广 AI 助手“Hey Arma”，其核心目标是解决被称为“红背心英雄”的新员工入职门槛过高的问题。由于店员需要在货架补货、过道导购与收银之间频繁切换，系统必须在油漆、电动工具等多品类中即时提供指导，且在全面推广前经一线员工实测，避免成为额外的行政负担。 在客户服务领域，Nathalie Criou 强调语音作为低摩擦沟通媒介的价值。通过消除长达 45 分钟的客服排队等待，AI 语音座席能够有效承接付款、退货或产品咨询等明确任务。例如墨西哥二手车交易平台 Kavak 引入语音 AI 协助处理试驾预约与购车咨询，面对因交易金额高而频繁来电确认的客户，全天候且极具耐心的语音交互显著提升了买家信任度，大幅减少了向人工客服升级转移的需求，甚至促使该企业围绕语音体验重新调整整体运营流程。 自下而上驱动业务成果与防炒作治理 (30:04) 针对 AI 技术的应用规划，与会嘉宾一致认为必须以实际业务成效优先于技术噱头。Ace Hardware 以“是否兑现了成为实用之地的品牌承诺”作为唯一的项目评估准则，在营销、采购、零售运营和国际业务等各部门设立“AI 冠军”，从基层挖掘实际操作中的阻力与痛点，再递交高管指导委员会评审立项，从而避免自上而下的教条化摊派。 在技术引入方面，Ontañon Alvarez 提出 Coppel 的三步走战略：购买底层平台、构建集成基础设施、自主拥有核心业务能力。例如其电商平台上线生成式搜索引擎后，转化率直接提升了 4 倍。为了防范概念炒作并把控风险，Coppel 建立了由首席运营官（COO）和战略团队领导的 AI 指导委员会，规定所有 AI 应用必须在严格签署保密协议（NDA）的安全企业级平台内运行。 购买通用标品与自研核心差异化能力 关于 AI 时代“自研还是外购”的权衡，Nathalie Criou 指出 AI 大幅降低了专业门槛，使非技术人员也能跨界完成任务，但这也带来了数据孤岛、业务规则冲突及品牌价值偏离等严峻的治理挑战。Melanie Hill 据此建议，企业应当“购买通用产品，自研核心差异”，即在仅需达到行业平价水平的常规能力上采购现成工具，而对形成核心护城河的关键点（如 Ace Hardware 的“Hey Ace”门店工具）必须坚持自主构建与迭代，切忌将核心战略思维和秘密武器外包给供应商。 在总结建议中，Mark Stutzman 列举其中学教师出身的儿子使用 AI 编写水池水质管理软件、以及自身数小时内自建工具替代企业级 SaaS 的实例，证明软件开发已被深度民主化；Santiago Ontañon Alvarez 则主张结合人类专业领域经验与 AI，重新审视并攻克过去被判定为无法解决的业务难题；Melanie Hill 亦提倡“胶带式创新”，通过观察员工与顾客的代偿性妥协动作来精准定位高价值痛点。 本场讨论形成的共识在于，AI 在零售领域的真正价值取决于能否解决一线工作流中的微观痛点并守住核心商业机密。随着开发门槛的全面降低，企业面临的最大挑战已从技术可得性转向组织治理——即如何在全员赋能的同时，确保跨工具、跨场景的数据标准与核心品牌价值不被稀释。

## Background 该内容由 Ai4 于 2026 年 8 月 21 日发布在 YouTube 平台上。视频记录了一场由 varejocast 播客主持人 Caio Camargo 主持的行业圆桌论坛，汇聚了来自 Coppel、Vapi、AREA15 和 Ace Hardware 的高管，共同探讨零售与 CPG 领军企业如何利用 AI 升级营销物流、重构线上线下客户旅程并沉淀核心运营能力。 ## TL;DR 论坛嘉宾深入探讨了零售行业 AI 从早期的“尝鲜好奇”迈向“业务深度融合与价值变现”的深水区实践。Ace Hardware 分享了覆盖数千家门店、赋能前线店员的语音外脑“Hey ARMA”；墨西哥商业巨头 Coppel 介绍了日活破万的 AI 综合平台“Coppel Genius”与独特的实时信贷风控逻辑；沉浸式娱乐地标 AREA15 则拆解了 AI 驱动的动态客流票务调度系统以及自研替代 SaaS 的实操路径。结合 Vapi 在语音 AI 领域的突破，嘉宾们一致强调应摆脱“为了用技术而用技术”的噱头，确立以真实业务痛点驱动、标品直接采购、核心护城河坚决自研与分层解耦的务实战略。 ## Atmosphere & style 整场对话务实、坦诚且极具幽默感，嘉宾们既有对“行业谈 AI 宛如玩喝酒游戏”等盲目跟风现象的犀利吐槽，也毫无保留地复盘了底层数据混乱及耗时一年却以失败告终的惨痛教训。主持人的敏锐穿引与各位老牌零售、技术掌门人的真实碰撞相得益彰，没有枯燥的技术布道，取而代之的是围绕一线员工动线、真实商业 ROI 与企业生存护城河的深度思辨。现场节奏明快而充满行业实战体感，为听众呈现了一堂极具现实参考意义的企业级 AI 落地大师课。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=ZFH9QDPzDkA&t=4s) 零售业对待 AI 已从最初的“好奇心试水期”迈入深水区，核心转向如何支撑企业最核心的护城河、实现可衡量的 ROI 与盈利能力。 - [01:51](https://www.youtube.com/watch?v=ZFH9QDPzDkA&t=111s) Ace Hardware 推出耳佩式语音助手 Hey ARMA，以赋能而非替代一线员工为目标，将万级 SKU 知识库装入店员耳旁，在 2,700 多家门店规模化复刻标准化贴心服务。 - [03:18](https://www.youtube.com/watch?v=ZFH9QDPzDkA&t=198s) Coppel 融合百货零售、银行与养老金闭环，深耕墨西哥下沉市场，依托跨消费与金融维度的立体数据流构筑坚实的数据与算法壁垒。 - [05:35](https://www.youtube.com/watch?v=ZFH9QDPzDkA&t=335s) AREA15 将客流预测模型（误差≤7%）与自研 AI 票务中枢深度融合，实现收益管理优化、场馆容量动态均衡与餐饮等二次消费的高效转化。 - [07:47](https://www.youtube.com/watch?v=ZFH9QDPzDkA&t=467s) 语音 AI 正在重塑客户交互体验，从早期追求缩短通话时长的成本节约工具，跃升为能够提供深度情感连接与一对一专属陪伴的品牌资产。 - [10:22](https://www.youtube.com/watch?v=ZFH9QDPzDkA&t=622s) 零售企业在推进 AI 时必须直面底层数据泥潭与老旧系统的客观现实，AI 正从浮于表面的附加功能深入到业务运营与组织治理的骨干体系中。 - [12:38](https://www.youtube.com/watch?v=ZFH9QDPzDkA&t=758s) Coppel 搭建统一中台平台 Coppel Genius 解耦底层复杂系统，实现超万名跨 IT 与业务线员工的日常高频采纳，同时利用 AI 反哺老旧架构的现代化重构。 - [15:07](https://www.youtube.com/watch?v=ZFH9QDPzDkA&t=907s) AREA15 实行“两周冲刺 AI 实验室”机制并采用提示词驱动编程（Vibe Coding），快速自研替代外部昂贵商业 SaaS 软件以显著削减采购授权开支。 - [19:39](https://www.youtube.com/watch?v=ZFH9QDPzDkA&t=1179s) Coppel 将信贷风控与催收置于核心，结合心理学与耐用品/快消品的实时还款倾向模型，并在结账时借助动态 OTP 验证确保联系方式真实可靠。 - [23:32](https://www.youtube.com/watch?v=ZFH9QDPzDkA&t=1412s) AI 落地必须坚持“先爱上问题再爱上技术”，深入货架通道开展面对面测试以顺应一线员工的真实动线，避免沦为自上而下摊派的考核负担。 - [26:18](https://www.youtube.com/watch?v=ZFH9QDPzDkA&t=1578s) 在汽车交易等长决策链条场景中，语音 AI 凭借无等待、高耐心与低摩擦交互，能大幅提升转化意向且争议转人工处理的比例极低。 - [30:04](https://www.youtube.com/watch?v=ZFH9QDPzDkA&t=1804s) 拒绝“AI 喝酒游戏”式的盲目跟风炒作，通过在各业务线设立 AI 拥护者（AI Champion）推动自下而上的“AI 民主化”，以是否兑现核心品牌承诺作为唯一检验准绳。 - [34:55](https://www.youtube.com/watch?v=ZFH9QDPzDkA&t=2095s) 盲目使用大模型攻坚复杂长尾问题（如 80% 准确率的租约提取）往往远不如标准化流程高效，必须时刻保持“人在回路”并警惕技术自嗨陷阱。 - [39:09](https://www.youtube.com/watch?v=ZFH9QDPzDkA&t=2349s) 确立清晰的“买还是造”战略边界：通用标准化能力直接采购以快速达到行业基准，体现自身差异化竞争力和独家洞见的核心能力坚决自研自持。 - [43:56](https://www.youtube.com/watch?v=ZFH9QDPzDkA&t=2636s) 采用“分层解耦”的技术架构将顶层业务逻辑与底层可插拔的模型供应商隔离，从而在快速吸收前沿大模型红利的同时有效抵御模型频繁更迭的重构风险。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=ZFH9QDPzDkA&t=4s) 开场与嘉宾亮相：零售 AI 迈入新阶段 主持人 Caio Camargo 以一段轻松幽默的玩笑开场，戏称台上有巴西人登台后，明年大会现场肯定会被巴西同行“完全占领”。随后他言归正传，指出过去几年零售业对待 AI 的态度主要是出于纯粹的“好奇心”，大家都在探寻它到底能做什么。如今，这一探索摸索的试水期已经彻底过去，行业必须聚焦于如何利用 AI 打造出更加个性化、响应敏捷且具盈利能力的真实体验，并直面第一阶段成效显现之后的深层变革。 紧接着，来自法国的 Nathalie Criou 接过话茬打趣互动，并介绍了自己所在的语音 AI 平台公司 Vapi。Vapi 致力于帮助各类企业构建、测试并在生产环境中持续优化语音 AI 智能体（Voice AI Agents）。她透露，Vapi 的客户中有大量零售领域的企业，沉淀了许多一线实操落地与客户互动的实践洞见。 随后，Melanie Hill 介绍了北美五金连锁巨头 Ace Hardware（快客硬件）。作为一家拥有约 5,000 家门店的百年五金合作社，Ace Hardware 的业务还拓展至拉丁美洲、中东及亚洲。Melanie 笑称大家虽然都熟知其“贴心之地”（The Helpful Place）的品牌定位，但她今天不会强迫现场听众合唱那首著名的品牌广告歌，而是希望探讨如何借助 AI 真正践行这一服务承诺。 - **主持人**：Caio Camargo（零售播客 varejocast 主持人） - **嘉宾 Nathalie Criou**：Vapi 产品副总裁，主营语音 AI 智能体开发、评测与部署优化平台 - **嘉宾 Melanie Hill**：Ace Hardware 高级概念部主管（创新与战略），负责洞察与前沿业务 --- ### [01:51](https://www.youtube.com/watch?v=ZFH9QDPzDkA&t=111s) Ace Hardware 的“Hey ARMA”：为门店店员打造数字外脑 在主持人的引导下，Melanie 详细分享了 Ace Hardware 刚在春季上线的一项核心 AI 项目——“Hey ARMA”。ARMA 是“Ace Retail Mobile Assistant”（Ace 零售移动助手）的缩写。该项目的初衷非常明确：在全美 5,000 家门店内，以完全一致的高水准向顾客提供专业服务，而实现路径并不是去替代店员，而是全面为前线店员赋能。 Melanie 坦言，新加入 Ace Hardware 的员工通常必须亲自下门店轮岗，自己当年在门店工作时就表现平平，因为店内通常陈列着数以万计的 SKU，普通人根本不可能把每件商品的特性与用途都背下来。如今通过 Hey ARMA，店员只需佩戴耳机，就可以通过语音随时与这位 AI 助手对话。 无论是新员工还是资深老手，当面对越来越有备而来的专业顾客时，店员只需发问，Hey ARMA 就会直接在耳机中提供精准答案——从专业产品推荐、家居改造项目施工建议，到两款竞品的深度参数对比。这一工具把庞大的商品知识库直接装进了店员耳旁，不仅消除了店员面对专业顾客时的知识盲区，更让顾客进店时能持续获得惊喜与超预期的服务。 - **项目名称**：Hey ARMA（Ace Retail Mobile Assistant） - **硬件形态**：店员佩戴式语音智能耳机 - **核心功能**：实时检索万级 SKU 信息、匹配施工/改造建议、横向对比竞品参数 - **战略定位**：赋能而非替代前线员工，确保“贴心互助”体验在 5,000 家门店无缝复刻 --- ### [03:18](https://www.youtube.com/watch?v=ZFH9QDPzDkA&t=198s) Coppel 庞大商业生态与下沉市场的 AI 差异化 墨西哥商业巨头 Coppel 的 CIO Santiago Ontanon Alvarez 介绍了这家拥有 80 多年历史的私人企业集团。Coppel 绝不仅仅是一家零售商：它在全墨设有约 2,000 家门店，同时下辖一家银行以及名为 Afore 的全额退休养老金管理机构。更关键的是，这三家实体服务的是同一批客户——以墨西哥处于 C 级及以下的低收入群体为主。这种横跨零售、金融贷款和养老金管理的立体闭环，成为 Coppel 挖掘数据与训练 AI 的核心护城河。早在 ChatGPT 面世之前，Coppel 就已经深耕 AI 与机器学习多年。 Santiago 描绘了 Coppel 极其独特的全国网络：其业务遍及墨西哥 550 多个城市，而很多竞品在全国的总门店数甚至都达不到 550 家。在许多过去没有像样零售渠道、缺乏从中国、越南、美国或本土大批量组织货源能力的偏远小镇，一家 Coppel 门店的设立直接将具有消费力的平价商品与金融信贷引入当地，以一己之力推动了墨西哥偏远小村落向城镇化的蜕变。他们不仅是售卖商品，更是在通过提供必需的生活物资与金融流动性兑现“让生活更美好”的品牌口号。 - **企业架构**：综合零售（约 2,000 家门店）+ 零售银行 + 养老金管理机构（Afore） - **客群定位**：墨西哥 C 级及以下低收入大众市场 - **渠道纵深**：覆盖全国 550+ 座城市，深入偏远缺乏零售基础设施的腹地 - **数据壁垒**：打通同单一客户在零售消费、银行信贷及养老理财维度的全周期数据流 --- ### [05:35](https://www.youtube.com/watch?v=ZFH9QDPzDkA&t=335s) AREA15 的沉浸式娱乐科技与 AI 实时票务中枢 主持人向全场发问是否有人没去过拉斯维加斯的著名沉浸式娱乐综合体 AREA15，并极力向现场观众“安利”当晚务必亲自前往体验。兼任 Fischer Brothers（拥有百年历史的家族地产集团，也是 AREA15 的母公司）全球首席技术官以及 AREA15 首席运营官与首席技术官的 Mark Stutzman 登台分享。过去七年里，他一直主导这家沉浸式体验地标的数字化底座建设。 Mark 指出，AREA15 运用机器学习和 AI 已经有相当长的历史。从开业之初起，他们就构建了高精度的客流预测算法模型，将每日天气条件、整座拉斯维加斯城内举办的各类会展活动以及历史客流数据进行多源交叉分析。该算法能够将当日访客预测精度控制在 7% 的误差范围以内，进而直接驱动每小时轮班排班系统，优化一线灵活用工成本。 如今，AI 已经全面渗透进 AREA15 的每一个业务纵向。作为一项年售出数百万张门票的娱乐圣地，联票捆绑销售（Bundling）已占据绝大多数份额。为此，他们正在自主研发新一代票务平台，整个底层完全由 AI 驱动——包括端到端的收益管理优化（Yield Management）、动态定价、各沉浸式场馆的容量实时规划，以及通过算法引导游客在合适的时间前往等候时间最短的项目，以实现全场客流在空间和时间上的动态均衡调度。 - **客流预测系统**：结合天气、全市大型会展活动与历史客流量，预测精度达 93%（误差 ≤ 7%） - **核心收益业务**：年销售数百万张门票，以高复杂度的组合套票（Bundling）为主 - **自研 AI 票务平台**：实时统筹收益管理、客流动态分流排队、场馆运力承载规划 --- ### [07:47](https://www.youtube.com/watch?v=ZFH9QDPzDkA&t=467s) 语音 AI 体验跃迁：从客服降本到情绪连接 Caio Camargo 将对话引向当下行业的最核心命题：零售业是否已真正跨越单纯追求“产出初步成效”的阶段？现在究竟处在哪个发展水位？ Vapi 的 Nathalie 从跨行业客户的数据反馈中给出了一个反直觉的洞察。长期以来，电话呼叫中心是消费者最反感、最痛恨的沟通渠道，尽管人类天生偏好口头交流，但受制于企业对人工坐席的高昂成本压力与服务时长的严苛管控，客服电话往往变成了机械式应付和漫长等待的代名词。然而，语音 AI 的成熟正在推动该渠道迎来“新生”。 Nathalie 强调，企业起初部署语音 AI 纯粹是为了填补人手短缺，但现在客户体验正从“自动化补充”升级为“令人愉悦的沉浸交流”，客户甚至主动拉长通话时间，敞开心扉向 AI 倾吐心声。这带来了一个令人啼笑皆非的现象：很多呼叫中心客户开始抱怨“现有的 KPI 考核指标全都失灵了”。传统的联络中心 KPI 旨在尽可能缩短每次通话时长以压低服务成本，但现在的真实情况却是，大量用户只是非常享受与语音 AI 智能体的流畅交谈，甚至在电话里主动聊起家里的宠物狗、儿子刚办完毕业派对周末要回家等完全与业务无关的私人家常琐事。这种前所未有的一对一专属陪伴感，让语音 AI 正在从纯粹的成本中心转变为充满潜力的品牌情感资产与新增收入来源。 --- ### [10:22](https://www.youtube.com/watch?v=ZFH9QDPzDkA&t=622s) 从“万能瑞士军刀”回归业务核心：组织深水区的博弈 Melanie Hill 回顾了 AI 在零售领域的演进历程：早期它被包装为无所不能的“瑞士军刀”，声称能搞定全业务链条；随后进入具体职能赋能期，被应用到劳动力排班、销量预测和前端消费体验。而当下的零售业竞争极其白热化，Ace Hardware 及许多头部同行的视角已经彻底转变——从“AI 能做什么”聚焦到“AI 如何支撑我最核心的护城河”。 Melanie 直言，这一转型也让企业内部迎来了真正的阵痛与摩擦。AI 不再只是贴在现有系统外层的“创可贴”或附加功能，而是要真正融入日常业务运营骨干。这时，极其现实的组织治理问题便全面爆发：优先级的权衡、底层数据架构的归属，以及由谁来主导治理——到底该归 IT 部门、市场营销部门，还是归每个独立的业务线？如何将零散的构想转化为具备可衡量 ROI 的概念验证项目？ Melanie 还打趣地向全场发问：“现场有谁的底层数据系统是完美无瑕、完全为 AI 准备好的？请举手。”全场仅有一人举手，引得台上会心大笑。Melanie 坦言，像 Ace Hardware 这样拥有百年历史的企业，底层遗留系统错综复杂。但大家已经从早期的天真好奇，真正迈入了“直面现实挑战、深耕业务融合”的深水区。面对 Caio 追问未来是否会演变为“利用 AI 彻底颠覆旧业务模式”，Melanie 坚决赞同，并补充强调核心在于如何利用 AI 成倍扩充企业日常的运营容量，实现跨维度的效率提升。 --- ### [12:38](https://www.youtube.com/watch?v=ZFH9QDPzDkA&t=758s) Coppel Genius：用中台打通底层混乱，释放万人级 AI 生产力 面对同样的数据系统混乱难题，Santiago 分享了 Coppel 的应对之策。由于底层系统年久失修且架构复杂，Coppel 选择直面现实，在遗留系统之上快速开发了一层专有连接平台，将 AI 工具与底层尚未现代化的“数据泥潭”解耦对接。 Coppel 集团拥有 125,000 名员工，其中约 20,000 人为白领知识工作者。一年前，Santiago 团队做出了一个极其果断的战略决策：要求所有知识工作者停止自行在外部付费购买第三方商业 AI 软件，统一接入由公司自建的综合平台——**Coppel Genius**。该平台整合了原本单人订阅成本高达每月 800 美元的各类尖端 AI 工具矩阵，并免费开放给全员。 这一举措引发了内部的“爆发式采纳”。如今，Coppel Genius 拥有超过 10,000 名日活跃用户（DAU），规模相当于本次大会的参会总人数。 Santiago 分析了这一庞大活跃群体的构成分布，发现了一个极具价值的黄金平衡点： ``` Coppel Genius 日活用户构成 (~10,000 DAU) +-------------------------------------------------------------+ | 30% 方案/软件工程与 IT | 30% 纯一线业务部门 | 30% 业务与 IT 融合跨界团队 | +-------------------------------------------------------------+ ``` 这一分布意味着 AI 不再只是工程师的极客玩具，而是真正渗透进各条业务线来解决现实的经营问题。更为巧妙的是，原本旨在充当过渡缓冲层、屏蔽底层陈旧系统的这层 AI 平台，如今反而成为推动系统现代化的加速器——负责升级供应链和电商平台底层架构的技术团队，正大量借助 AI 代码生成与自动化工具，大幅缩短老旧系统的重构周期。 --- ### [15:07](https://www.youtube.com/watch?v=ZFH9QDPzDkA&t=907s) AREA15 的务实三步走战略与“两周冲刺”AI 实验室 Mark Stutzman 深入拆解了 AREA15 与 Fischer Brothers 在落地 AI 时的实用主义进阶路线图。他首先以宏观视角打了个形象的比方：即便是目前最顶尖的前沿基座大模型，其成熟度也仅仅相当于“蹒跚学步的幼儿”，企业界现在就必须警惕并提前考虑当模型进入“叛逆青少年期”时所面临的治理与控制风险。回到落地执行层面，他们制定了清晰严整的三步走方案： 1. **第一步：基模型标准化与分层进阶培训** 全公司统一采用 Anthropic 的 Claude 模型作为主力工具。针对管理层及以上骨干，每隔两三个月组织一次体系化培训。Mark 强调，培训的重点绝非教大家如何打字聊天，而是深入讲解协同办公机制（Co-work）、构建自主 Agent（智能体）以及自动化日常琐碎流程的方法论。 2. **第二步：用 Vibe Coding 彻底重构 SaaS 技术栈，削减百万授权费** 梳理集团所有子公司采购的数百万美元 SaaS 软件授权。技术团队通过标准化技术栈，采用提示词驱动编程（Vibe Coding）从零自研替代方案。在统一技术栈的规范下，任何开发者输入指定提示模板即可快速编写系统。目前已成功替换掉约 8 个外部商业 SaaS 平台，并在接下来的几个月内计划再干掉 20 个同类商业软件，大幅节约采购开支。 3. **第三步：设立硬指标的 AI 实验室（AI Lab）** 当各业务单元对 AI 越来越熟练并开始主动提出复杂的流程优化需求时，由于业务体量庞大（超出各部门负责人个人摸索的范畴），AI 实验室正式介入。实验室设立了一条铁律：**所有承接的项目必须在两周冲刺周期（Two-week Sprint）内交付**。一旦评估耗时超过两周，便剔除出 AI 实验室，交由集团重大项目委员会统筹排期。 此外，他们自主研发了一整套内部配套工具链，包括支持员工提交业务痛点的需求提报工具、全公司共享的精选提示词库（Prompt Library），以及一款定制的 Chrome 浏览器插件。该插件能在员工与 Claude 对话时实时介入，结合全公司验证有效的历史最佳实践，辅助员工自动优化与保存高阶提示词。Mark 总结道，集团大举推进 AI 的终极目的绝非借机裁员，而是要把核心员工从低附加值的繁杂事务中解救出来，让他们把精力全部扑在具有全局价值的商业战略推演上。 --- ### [18:28](https://www.youtube.com/watch?v=ZFH9QDPzDkA&t=1108s) 动态客流与全场域行程编排 在回应 AI 如何直观重塑 AREA15 现场游客体验时，Mark 详细拆解了他们脱胎于票务中枢的核心应用案例。在周末和节假日，沉浸式园区的客流峰值逻辑与 Uber、Lyft 的高峰打车场景如出一辙。早期他们采购并深度定制了一套第三方虚拟排队系统，而现在他们通过前面提到的“先授权体验、再自研替代”策略，完全自主构建了一套属于 AREA15 的专有虚拟排队中台。 这套系统远远超越了“扫码取号并在排到时发送短信通知”的初级排队功能，它实质上成为了游客置身于整个 AREA15 园区期间，连接游客与场馆数据中枢的实时交互界面。 基于 AI 算法，该系统在游客入园的几小时内持续进行全域行程的动态编排（Dynamic Itineraries）： - **实时穿插分流**：根据全场各沉浸式展馆当前的实际排队热力图，在游客等待 A 项目的空档期，精准推荐前往当前几乎零等候时间的 B 项目，平抑各展馆的瞬时拥堵； - **二次消费转化**：在算法预测到游客即将结束游玩且处于就餐低谷与高峰转换期时，AI 会在绝佳时机主动向其推送自营特色餐厅的空位订座提醒，将游客的现场动线、游玩节拍与餐饮零售收益进行深度绑定。 ### [19:39](https://www.youtube.com/watch?v=ZFH9QDPzDkA&t=1179s) 实体娱乐体验与Coppel的实时风控革命 Mark Stutzman 谈到，AREA15 正借助其实时 AI 技术为访客打造无缝体验，例如刚刚上线的“照片通行证”（Photo Pass）系统，实时 AI 不仅能在现场为宾客创造绝佳体验，还能围绕整个游览过程编织出沉浸式故事。随后，主持人 Caio Camargo 将话题抛给 Coppel 的首席信息官 Santiago Ontañon Alvarez，询问当信用评分从后台走到前台、直接面对顾客的实时交互时，业务究竟发生了怎样的质变。 Santiago 澄清并重塑了 Coppel 的定位：尽管 Coppel 拥有银行牌照已超过 20 年，但在建行之前他们就已从事信贷业务，目前店内高达 80% 的放贷并不是通过银行完成，而是依靠百货商店自身的信用额度（类似“先买后付”BNPL 或循环信用）。他语出惊人地指出，外界可能以为 Coppel 是个零售商、靠卖货或放贷为核心，但其实公司“真正的业务是催收”。在墨西哥，大量消费者在征信局（Credit Bureau）根本没有任何记录，Coppel 必须靠自己的算法自主完成信用评分，甚至专门聘请了心理学家、社会学家和哲学家来共同研究到底谁会按时还款。 针对实时交易，Santiago 分享了一个运行了 15 年的经典案例：一位 60 岁左右的女性顾客来到柜台，手里拿着两件价格相似（约 500 比索，折合 25 美元）的商品——一台电熨斗和一条牛仔裤。算法会立即识别出耐用品与非耐用品的还款意愿差异：牛仔裤极大概率是买给她女儿的，而电熨斗是自己带回家用的耐用品，因此顾客为电熨斗还款的意愿明显高得多。系统在结账瞬间计算出还款倾向，若分值较低则现场要求支付首付款。这种实时风控虽然精准，但以前在周日傍晚客流高峰时会导致系统严重卡顿，如今 Coppel 正利用 AI 将核心系统迁移至云端提速，并结合“可触达性算法”（Contactability Algorithms）进行手机号验证——不同于美国只要报手机号就给折扣的做法，墨西哥门店必须当场发送 OTP 一次性动态验证码，确保收集到的还款联络方式绝对真实。 - Coppel 80% 的借贷发生在线下百货门店体系内，而非通过其旗下银行。 - 团队吸纳了哲学家、心理学家和社会学家，跨学科构建底层信用还款模型。 - 耐用品（如 25 美元的电熨斗）与消费品（同价位的牛仔裤）具有完全不同的实时还款倾向模型。 - 为避免无效数据，在墨西哥门店登记电话号码必须当场完成 OTP 短信动态密码核验。 --- ### [23:32](https://www.youtube.com/watch?v=ZFH9QDPzDkA&t=1412s) Ace Hardware的Hey Arma：让AI融入门店伙伴的真实动线 主持人随后问及 Ace Hardware 针对门店员工推出的数字化助理应用“Hey Arma”，并提到它已进入 2,300 家门店。Melanie Hill 纠正并更新了这一数字，透露该系统目前已覆盖超过 2,700 家门店，并深入剖析了规模化落地过程中的两大核心收获。 第一个核心启示在于“AI 的人性化一面”（The human side of AI）。Melanie 直言，行业峰会上人们往往沉迷于令人惊叹的技术参数，却容易忽略终端实际使用者的感受。在零售行业一线工作极其繁杂繁重，员工需要一边理货、一边穿梭在货架通道回答顾客提问，还得随时跑去收银台结账。如果把 AI 做成又一个“公司高层强行摊派的任务”、多加一项考核指标，员工只会产生排斥心理。因此团队耗费大量时间让一线员工在实际货架通道中进行面对面（belly-to-belly）测试，确保它能自然嵌入员工极其紧凑的工作动线，而不是停留在实验室或高管汇报 PPT 里的花哨摆设。 第二个核心启示是“始于业务痛点而非技术本身”。Hey Arma 最初并不是由技术实验室发起的自嗨项目，而是来自门店系统团队的真实诉求：如何缩短新员工的入职培训周期？Ace Hardware 将其一线伙伴亲切地称作“红马甲英雄”（red-vested heroes），但新人刚入职就必须迅速通晓从户外烤炉、油漆调配到第 45 号货架上某颗小零件的所有细节，这带来了极大的挫败感。正是为了化解这个具体的零售痛点，团队才引入 AI 作为赋能工具，践行了“先爱上问题，再爱上技术”的研发哲学。 --- ### [26:18](https://www.youtube.com/watch?v=ZFH9QDPzDkA&t=1578s) 语音AI的破局之道：以Kavak汽车交易为例 Vapi 的产品副总裁 Nathalie Criou 针对“消费者何时愿意接纳 AI 语音代理而非人工坐席”做出了深度剖析。她指出，语音是人类最原始、摩擦力极低的交互方式，哪怕不识字的人也能无障碍表达并获得丰富反馈。过去消费者排斥自动语音是因为预期体验极差，而现代语音 AI 的首要破局点就是直接砍掉等待时间——比起在电话线上苦等人工坐席 45 分钟，能够随时秒接电话的 AI 代理显然极具吸引力。同时，成功的语音应用往往聚焦于边界清晰的目标（如退货处理、催缴付款或特定答疑），让用户预期明确，也方便企业定向调优。 针对行业原本担心的“AI 解决不了问题导致投诉激增并升级给人工坐席”（Escalation）的隐忧，Nathalie 表示市场数据呈现出截然相反的结果。只要 AI 语音代理打磨精细，用户不仅不会急于转人工，反而非常渴望与 AI 交流。她以墨西哥二手车电商平台独角兽 Kavak 为例：在二手车买卖场景中，客户线索一旦从聚合平台发出，谁先打通电话响应，谁就最有机会拿下订单。然而买车是昂贵且复杂的决策，消费者需要经过多次沟通来建立信任，Kavak 根本没有足够的真人坐席来支撑这种高频来回的咨询。 上线了专门预订试驾的语音 AI 之后，情况发生了意想不到的变化。消费者不仅乐于接听，甚至频繁回拨电话提问：“我刚在网站看到这款车，它的情况如何？”在用户看来，AI 展现出了真人难以比拟的超级耐性，甚至有一种“渴望为用户解决问题”的真诚感，客户能完全按照自己的节奏获取信息。这项尝试取得了巨大成功，Kavak 甚至顺势围绕以语音为核心的客户体验重构了整个公司的业务流，而服务争议升级至人工处理的比例变得“极其罕见”。 --- ### [30:04](https://www.youtube.com/watch?v=ZFH9QDPzDkA&t=1804s) 商业成果大于技术噱头：摆脱AI“喝酒游戏”式的盲目跟风 当主持人提问“更先进的技术还是更好的商业成果更重要”时，Nathalie 斩钉截铁地回答“绝对是商业成果”，因为企业存在的本职是盈利与节约成本，AI 归根结底只是服务于业务与客户体验的工具。Melanie 完全赞同这一观点，并透露 Ace Hardware 内部评判一个 AI 项目从概念验证（POC）推向规模化量产，有且仅有一个准绳：它是否践行了 Ace 成为“最热心助人的地方”（the helpful place）这一品牌承诺？如果答案是否定的，项目就会被立即叫停。 谈到如何避免被新技术热潮绑架，Melanie 幽默地打了个比方：“如今 AI 简直就像是一场喝酒游戏，如果每提到一次 AI 就喝一杯，全场早就彻底烂醉了。”现在企业里经常出现高管冷不丁拍拍下属肩膀问“你们部门打算怎么用 AI”，让人措手不及。Ace Hardware 的破局之策是组织内的“AI 民主化”——公司在执行指导委员会的统筹下，在国际业务、商品企划、零售运营、市场营销等每一个业务线内设立“AI 拥护者”（AI Champion）。通过自下而上的方式让真正了解一线痛点的人提出业务卡点，再由技术治理团队提供支持，使 AI 变成实实在在的日常工作方式，而非自上而下为了赶时髦打勾应付的任务。 Santiago 也分享了 Coppel 的技术平衡哲学。在科技落地方面，Coppel 奉行三步原则：“采购顶尖的基础技术，构建自研平台，并彻底自持转化商业成果的核心能力”。例如，Coppel 在其电商平台上线了生成式搜索引擎，直接带来了惊人的 4 倍转化提升。在治理层面，公司由 COO 与战略部门领衔 AI 委员会，技术部门则把好安全和保密协议（NDA）红线，绝不放任业务团队无节制地野蛮试验。Santiago 坦言在行业峰会上每个人都会被无数个“要不要喝杯 15 分钟咖啡聊聊项目”的推销请求淹没，零售企业必须保持清醒，不能盲从炒作。 --- ### [34:55](https://www.youtube.com/watch?v=ZFH9QDPzDkA&t=2095s) 买、建还是自持：30年CTO反思AI试点失败的教训与平台化思考 拥有超过 30 年技术老兵背景的 Mark Stutzman 针对“买（Buy）、建（Build）还是自持（Own）”分享了一段发人深省的惨痛教训。当他接管 Fisher 与 AREA15 的技术事务时，迎面撞上了一个典型的“彻底失败的 AI 试点项目”：团队怀揣巨大野心，试图用大模型去解析抽提长达上百页、极其繁复的商业租约合同，企图打造一个无所不知的“Fisher 大脑”。这个项目耗费了整整一年时间，当团队兴奋地向 Mark 汇报“AI 已经能搞定 80%”并打算继续攻关剩余的 20% 时，Mark 当场泼了冷水。 ``` 传统人工/标准化流程 (99.9% 准确度) [结构化数据录入] ──> [即时投入消费与流转] ──> 高效沉淀 失败的早期 AI 尝试 (80% 准确度) [长达数百页租约] ──> [AI抽取 80%] ──> [人工核对排查剩余 20%] ──> 极度耗时低效 (犹如大海捞针) ``` Mark 敏锐地指出，在严肃商业合同场景下，剩下的 20% 错误犹如大海捞针，人工逐条核验这 20% 所耗费的时间，远超过用成熟系统进行人工录入的时间；过去二十年间使用的成熟录入工具准确率高达 99.9%，而企业的终极诉求只是把数据规范地放进系统以便消费使用。于是他果断叫停了该项目，改用标准化的简易模板，让租约在签署时就直接规范录入。Mark 感慨道，AI 毫无疑问是他这辈子见过最强大、最深刻的技术变革，但它目前仍处于幼年期，必须有人类在回路（Human in the loop）扮演“保姆”的角色，千万不能被闪亮的新玩具晃瞎了双眼。 Nathalie 从技术平台（以 Vapi 语音平台为例）的角度接续了讨论，总结了企业选择接入平台来构建 AI 体验的深层逻辑： - **隐藏底层复杂度**：平台把复杂混乱的技术细节封装起来，抽象为业务人员能够理解和调用的交互界面，防止企业陷入“拿着锤子找钉子”的技术自嗨陷阱。 - **划定核心自持资产**：企业必须分清什么是真正属于自己的护城河——凡是关乎极高商业价值的核心成果或极其重要的客户关系纽带，企业哪怕使用外部工具加速，也必须牢牢自持并掌控其主导权。 ### [39:09](https://www.youtube.com/watch?v=ZFH9QDPzDkA&t=2349s) AI 技术的平民化、治理挑战与“买还是造”的战略抉择 Nathalie Criou 深入剖析了 AI 技术的本质转变及其在组织落地中面临的最大阻碍。在过去，企业若想使用高精尖技术，必须依赖工程师或特定领域的专业专家；但如今，AI 大幅降低了准入门槛，让任何职能部门的员工都能释放创造力，去完成以往从未受过培训的任务。然而，这种“人人都可使用 AI”的现状也引出了各大客户最关切的核心问题——企业治理（Governance）。企业必须在数据、业务上下文以及行为准则上建立统一的规范与约束。以 Ace Hardware 为例，其品牌灵魂在于“对顾客极致友好与乐于助人（Helpful）”，因此如何通过技术与工具将这种企业价值观固化并全面执行，成为 AI 在整个组织内规模化落地的必要前提。 ``` 【 自研 vs 采购决策矩阵 】 ┌───────────────────────────┬───────────────────────────┐ │ 采购 (BUY) │ 自研 (BUILD) │ ├───────────────────────────┼───────────────────────────┤ │ • 标准化大宗商品能力 (Commodity) │ • 核心业务差异化 (Differentiation) │ │ • 仅为达到行业基准水平 │ • 自有业务洞察与商业护城河 │ │ • 通用基础设施与现成工具 │ • 如 Ace Hardware 的 Hey Arma │ └───────────────────────────┴───────────────────────────┘ ``` 随后，主持人将“是买、是造还是掌控（Own）”的问题抛给了 Ace Hardware 的 Melanie Hill。Melanie 给出了非常明确的战略边界：对于标准化的大宗商品型能力（Commodities）选择直接采购，以此迅速追平行业基准；而对于能体现自身独特竞争优势的核心业务，则必须坚持自研。Melanie 提到，Ace Hardware 之所以自主构建店内核心服务工具“Hey Arma”，正是因为他们必须牢牢掌控店内服务的交付方式、掌握数据洞见并拥有自主迭代优化的能力。在 AI 仍处于“婴儿期/蹒跚学步期”的当下，企业绝不能把战略认知拱手让人，必须明确界定哪些地方适合接入 AI、哪些地方坚决不碰，“守护好企业独有的秘密配方（Secret Sauce）与护城河”。 - **平民化赋能**：AI 彻底打破了专业工程壁垒，各业务部门均能跨界进行自主创新。 - **治理与价值观对齐**：必须统一数据与上下文规范，确保 AI 表现严格契合企业品牌定位（如 Ace 的“Helpful”准则）。 - **采购与自研分流**：标品直接买入保基准，核心护城河与差异化能力必须自主研发与掌控。 --- ### [41:52](https://www.youtube.com/watch?v=ZFH9QDPzDkA&t=2512s) 终极圆桌建言：零售商下一步该如何落地 AI 在讨论进入尾声的压轴环节，主持人要求每位嘉宾用最简炼的语言向现场零售商给出下一步的实操建议。AREA15 的 Mark Stutzman 率先打破“一句话”的限制，分享了一个生动的真实案例：他的儿子作为一名教龄仅两年的英语老师，在毫无软件工程背景的情况下，竟然利用 AI 独立构建了一套“泳池化学品管理应用”。Mark 直言，AI 已经彻底实现了软件开发与知识获取的平民化。他强烈呼吁大家回家后亲自动手实践，不要只停留在“用 AI 查旅游攻略”的闲聊阶段，而要尝试把一个想法在几小时内做成真实的软件或业务模型。Mark 甚至透露，自己在早晨一边看新闻喝咖啡的短短两小时内，就已经用 AI 自行写出的工具替换了多个昂贵的 SaaS 订阅软件。 Santiago Ontanon Alvarez 则从问题求解的角度给出建议：零售商应当回过头去审视那些过去被认为“根本无法解决”的顽疾，尝试引入 AI 寻找全新的突破切入点。尽管单靠 AI 无法百分之百解决所有难题，但人类智能与人工智能的结合能够为复杂业务挑战带来前所未有的解决视角。 - **动手构建业务**：跳出浅层对话玩法，利用低门槛的开发能力验证商业创意。 - **降本增效实证**：利用业余碎片时间即可通过 AI 快速自研替代现有商用 SaaS 产品。 - **重访不可能的问题**：用“人类智能 + AI”的组合拳攻克过去零售业务中的死结。 --- ### [43:56](https://www.youtube.com/watch?v=ZFH9QDPzDkA&t=2636s) 业务洞察导向、分层解耦架构与 Vapi 平台体验 Melanie Hill 对 Santiago 的观点表示高度赞同，并强调零售、CPG 和电商企业必须始终坚持以“业务洞察和痛点”为核心驱动。她引用了自己在创新领域常用的概念——“胶带式创新（Duct Tape Innovation）”，即在观察到员工或顾客为了应对某种不便而产生代偿性行为（Compensating Behaviors）时，精准地介入产品或服务来弥补这一摩擦点。企业应当将 AI 精准部署在这些能够产生差异化竞争力的关键节点上。 ``` 【 分层解耦技术架构示意图 】 ┌───────────────────────────────────────────────────────┐ │ 顶层：业务逻辑与品牌体验 │ │ (保持稳定，不随底层模型变动而重构) │ └──────────────────────────┬────────────────────────────┘ │ ┌──────────────────────────┴────────────────────────────┐ │ 中层：灵活接入与隔离层 │ │ (统一上下文、治理规则、业务流编排与 API) │ └──────────────────────────┬────────────────────────────┘ │ ┌──────────────────────────┴────────────────────────────┐ │ 底层：可插拔的最佳模型/供应商 │ │ (每隔几周自由切换最前沿的 SOTA 大语言模型) │ └───────────────────────────────────────────────────────┘ ``` 最后发言的 Nathalie Criou 总结了两大核心策略： 1. **立即行动并精准切入**：AI 应用已不再是可选项，企业应选定一个范围极其明确、边界清晰且投资回报率（ROI）容易衡量的细分场景切入。 2. **构建分层架构以隔离技术演进风险**：AI 底层模型的迭代速度史无前例，企业必须采用“分层解耦”的架构设计，确保上层业务逻辑独立稳定，底层的模型与算法组件能够随行业前沿每隔数周无缝热插拔替换，避免“底层一出新模型，业务就得推倒重来”。 在结语中，Nathalie 介绍了 Vapi（vapi.ai）平台。该语音 AI 平台无需深厚开发背景即可上手，既能支持极小规模的原型验证，也能轻松承载大型企业每天数以千万计的高并发通话调度，并鼓励大家亲自体验其震撼的交互表现。 - **胶带式创新**：识别现实场景中的代偿摩擦点，以 AI 作为精准的效能放大器。 - **架构分层隔离**：以模块化解耦对抗高频的模型更迭周期，守护顶层业务系统的稳定性。 - **Vapi 赋能语音**：从单次调用到千万级全渠道并发，提供低门槛且高弹性的语音 AI 解决方案。 ## Notable quotes > [07:47](https://www.youtube.com/watch?v=ZFH9QDPzDkA&t=467s) 现有的 KPI 考核指标全都失灵了。 — Nathalie Criou > [10:22](https://www.youtube.com/watch?v=ZFH9QDPzDkA&t=622s) 现场有谁的底层数据系统是完美无瑕、完全为 AI 准备好的？请举手。 — Melanie Hill > [15:07](https://www.youtube.com/watch?v=ZFH9QDPzDkA&t=907s) 所有承接的项目必须在两周冲刺周期（Two-week Sprint）内交付。 — Mark Stutzman > [19:39](https://www.youtube.com/watch?v=ZFH9QDPzDkA&t=1179s) 外界可能以为 Coppel 是个零售商、靠卖货或放贷为核心，但其实公司真正的业务是催收。 — Santiago Ontanon Alvarez > [23:32](https://www.youtube.com/watch?v=ZFH9QDPzDkA&t=1412s) 先爱上问题，再爱上技术。 — Melanie Hill > [30:04](https://www.youtube.com/watch?v=ZFH9QDPzDkA&t=1804s) 如今 AI 简直就像是一场喝酒游戏，如果每提到一次 AI 就喝一杯，全场早就彻底烂醉了。 — Melanie Hill > [30:04](https://www.youtube.com/watch?v=ZFH9QDPzDkA&t=1804s) 采购顶尖的基础技术，构建自研平台，并彻底自持转化商业成果的核心能力。 — Santiago Ontanon Alvarez > [34:55](https://www.youtube.com/watch?v=ZFH9QDPzDkA&t=2095s) AI 毫无疑问是我这辈子见过最强大、最深刻的技术变革，但它目前仍处于幼年期，必须有人类在回路（Human in the loop）扮演保姆的角色。 — Mark Stutzman > [39:09](https://www.youtube.com/watch?v=ZFH9QDPzDkA&t=2349s) 守护好企业独有的秘密配方（Secret Sauce）与护城河。 — Melanie Hill > [43:56](https://www.youtube.com/watch?v=ZFH9QDPzDkA&t=2636s) 构建分层架构以隔离技术演进风险，确保上层业务逻辑独立稳定，底层的模型与算法组件能够随行业前沿每隔数周无缝热插拔替换。 — Nathalie Criou ## Mentioned Caio Camargo, varejocast, Santiago Ontanon Alvarez, Nathalie Criou, Mark Stutzman, Melanie Hill, Vapi, Ace Hardware, Hey ARMA, Coppel, Afore, Coppel Genius, AREA15, Fischer Brothers, Anthropic Claude, Kavak ## Video description Caio Camargo - Host, varejocast Santiago Ontanon Alvarez - CIO, Coppel Nathalie Criou - VP, Product, Vapi Mark Stutzman - Chief Operating & Technology, AREA15 Melanie Hill - Head of Advanced Concepts (Innovation & Strategy), Ace Hardware Corporation Retailers are using AI to deliver more personalized, responsive, and profitable experiences. This panel features leaders from retail and CPG companies discussing how they’re using AI to enhance marketing, streamline logistics, and reinvent the customer journey—both online and in-store. Learn which tools are driving results and how teams are building AI into core retail operations. We’ll also examine innovations across omnichannel retail, from checkout-free physical stores to AI-powered customer service agents. Join this panel to see how leading technologists and retailers are shaping the future of the industry. [Source](https://www.youtube.com/watch?v=ZFH9QDPzDkA)

[7:47] Nathalie Criou: 现有的 KPI 考核指标全都失灵了。 [10:22] Melanie Hill: 现场有谁的底层数据系统是完美无瑕、完全为 AI 准备好的？请举手。 [15:07] Mark Stutzman: 所有承接的项目必须在两周冲刺周期（Two-week Sprint）内交付。 [19:39] Santiago Ontanon Alvarez: 外界可能以为 Coppel 是个零售商、靠卖货或放贷为核心，但其实公司真正的业务是催收。 [23:32] Melanie Hill: 先爱上问题，再爱上技术。 [30:04] Melanie Hill: 如今 AI 简直就像是一场喝酒游戏，如果每提到一次 AI 就喝一杯，全场早就彻底烂醉了。 [30:04] Santiago Ontanon Alvarez: 采购顶尖的基础技术，构建自研平台，并彻底自持转化商业成果的核心能力。 [34:55] Mark Stutzman: AI 毫无疑问是我这辈子见过最强大、最深刻的技术变革，但它目前仍处于幼年期，必须有人类在回路（Human in the loop）扮演保姆的角色。 [39:09] Melanie Hill: 守护好企业独有的秘密配方（Secret Sauce）与护城河。 [43:56] Nathalie Criou: 构建分层架构以隔离技术演进风险，确保上层业务逻辑独立稳定，底层的模型与算法组件能够随行业前沿每隔数周无缝热插拔替换。

[26:18] Nathalie Criou: The consumers were saying, 'No, I have the answers that I want when I want it.' They found the AI to be much more patient with them. They found the AI to want to please them. [30:04] Melanie Hill: Are we delivering on our promise to be the helpful place? [39:09] Melanie Hill: Build what is your point of differentiation. Own the insight of whether you use AI or not. Do not let someone else have that mindset for your company... own your secret sauce and your moat.

[0:04] I promise in next year we're going to have a lot  of Brazilians here. Once a Brazilians on the stage [0:08] next year, we dominate to the whole event. Thank  you very much for joining us in this panel. I [0:14] briefly introduce our our panelists over here. I  have a few words for you guys to understand this [0:19] panel. What's going on working on um for the past  few years retail approach and AI with curiosity. [0:26] What can it do? That phase is behind us. Today  we need to understand how AI creates personalized [0:32] responsive and profitable experiences. Then  what comes after the first results with me [0:37] joining this panel I want you my my panelists  here to present theirelves with a little [0:42] uh talk. Uh please quickly introduce yourself your  company and how your company's approaching smart [0:48] retail with in AI today. Please Natalie start with  you. Very good. Hi everyone. Everyone from France [0:54] since you're doing the Brazil thing I do this.  Yes. Good. All right. Uh, so my name is Natali [1:01] Cuo. I am on the product team at a company called  Vappy. And Vapy is a platform to help others build [1:09] voice AI agents. Uh, you can build them, you  can test them, you can then optimize them uh, [1:14] as they run in production. And a lot of our  customers are real customers. So we have a [1:20] few insights that we can share today. Okay.  Melanie. Hi, I'm Melanie Hill. Um, I lead the [1:26] consumer insights and innovation division at Ace  Hardware. Um, ACE is a hardware cooperative with [1:32] about 5,000 stores in the US and also stores in  Latin America and the Middle East and Asia. Um, [1:40] we're known as the helpful place. I'm not going to  make anyone in the audience sing the jingle. Um, [1:46] but we are looking at approaching AI to really  deliver on that promise [clears throat] and we'll [1:51] talk a little bit about that today. Yeah. Well, I  wanted to understand to audience to understand too [1:55] about the name of the project some projects  the ACE hardware is doing right now. Yeah, [2:00] one of the projects that we just released  this past spring with AI is called Hey Arma [2:05] and ARMA [clears throat] is an acronym for ACE  retail mobile assistant. And the idea here is how [2:11] do we deliver helpful in a way that's consistent  across our 5,000 stores? Um, and how do we use AI [2:18] to empower the associate? Right? We don't want to  replace the associates. People come to our stores [2:23] because they know they're going to get help. It's  fully for the associate. It's fully direction to [2:28] the associate. Correct. And so it's it's within  the earpiece of the associate. Um they can ask [2:33] questions whether they are new to ACE. So you know  as we started ACE you have to work at a store. So [2:39] I worked at a store. I was a horrible associate.  Um but you realize that you have to know [2:45] everything about the tens of thousands of SKs that  are in our stores. And so this um AI assistant, [2:52] this digital assistant talks within the earpiece.  It can give that associate the knowledge they need [2:57] to provide that helpful answer, whether it's a  product recommendation or a project advice or [3:03] even being able to compare two products against  each other. It really helps and empowers that [3:08] associate to surprise and delight the consumer.  Because sometimes uh the the customer comes [3:13] to the store even more prepared than just the  associate. Correct. Correct. [laughter] Okay. Hey, [3:18] Santiago, tell us about Copel. Uh,  thank you, Kaio. Hello, everyone. Um, [3:23] Santiago Antion from Copel. Copel is a private  company in Mexico. U, it's a group of companies [3:30] actually. We we are a retailer. Uh, we have a  bank and we have a pension fund administration [3:37] um company which is called AFO in Mexico. And and  actually that's a key differentiator for us in our [3:43] usage of AI because we serve the same customer  with these three companies. We're not only their [3:48] retailer. Uh we have about 2,000 stores all  all around Mexico. Uh we are their bank and [3:56] sometimes we even manage their their pension.  So that that that is a very compelling way of [4:03] bringing [snorts] and integrating all of that data  of our same customer which is typically C see C [4:09] below. Uh so we cater to the lower income market  in Mexico and uh and data has always been our our [4:17] biggest differentiator and we have been using AI  uh for many many years before Chat GPT even came [4:25] into into the arena. Uh an 80year-old company very  successful and as I said private okay so copel [4:33] works like an ecosystem. Yes, exactly. It's a it's  an ecosystem where we would like our customers [4:39] from go to go from the the retailer to the to  the bank and to the pension. But yeah, it's a [4:45] uh everything the customer needs copel can offer.  Yes, we actually our our tagline is we would like [4:52] to make your life better. uh not only by providing  uh I mean if you if you historically see how Copel [4:59] uh grew we have stores in more than 550 cities in  Mexico some of our competitors don't even have 550 [5:08] stores uh so we are in very uh all AC all across  Mexico in places where there used to be no stores [5:16] and they have there were no purchasing power in  terms of bringing goods uh that were coming from [5:21] either China, Vietnam, the US or even Mexico that  had this uh purchasing power. So a couple store [5:29] basically changed uh a small town into a city in  Mexico. I understand this. Mark, tell us about [5:35] the area 15. But just just wait a little bit.  Uh does everyone does anyone here never heard [5:42] about area 15 before? Raise the hand. Let me see.  Okay, there's a tip for you. Go to Era 15 tonight. [5:50] go over there and see what they have because it's  awesome. But please now present Arrow 15. It it [5:56] sounds like I owe him a referral fee later in  the day. But yeah, hello everybody. I'm We were [6:01] there yesterday and it was awesome. Good. Uh I am  I'm Mark Stzman. I'm the global chief technology [6:07] officer at Fischer Brothers, which is a 100 plus  year old family-owned real estate business, and [6:14] the owner of Area 15, where I've been the chief  technology officer for the last seven years. and [6:19] the chief operating officer as well for the last  two AI as you can imagine for us um we we've been [6:28] using it for many years from a machine learning  perspective when we first opened we're able to do [6:33] um predictive visitor counts so we look at we look  at kind of things like weather and events that are [6:38] happening in the city and historic visitor counts  and we can generally within 7% accuracy predict [6:44] how many visitors we're going to have in our  current day which then lets you use AI to actually [6:48] help build schedules for hourly staff and whether  too you said the weather you cross information [6:54] about the weather to understand have a people or  not. Yeah. Yeah. We feed it all into the algorithm [6:58] and and and that's like kind of the basic stuff  and then there's you know honest honestly we we [7:03] are using AI in literally every vertical of the  business today and I think we can talk a lot about [7:09] that moving forward but I can give you an example  of a fairly large initiative is as you can imagine [7:14] we sell a lot of tickets at area 15 millions of  tickets and so we're we're actually building our [7:20] own ticketing platform because bundling has  become extremely important to us and has kind [7:25] of gobbled up the vast majority majority of  our ticketing sales, but the entire platform [7:30] will basically be in many ways powered by AI uh  including revenue optimization, yield management, [7:38] capacity planning for our attractions, routing of  guests to the right attraction at the right time [7:42] for this for the shortest wait. It will all be  powered in real time by AI. So, we're we're doing [7:47] lots of cool stuff. Okay, moving on. When it comes  to AI, retail has moved beyond curiosity. It's not [7:53] a what is AI what's working for um I also believe  that's my belief my opinion it has moved beyond [8:00] simple delivering results we're not arguing here  or discussing this event just about the results [8:06] they can bring for you guys what stage are we in  now it's results stage or another stage starting [8:13] with you Natalie I can give a a perspective  um we work across a lot of different clients [8:20] in fact a lot of different industries and there's  a lot in common across Ross all these industries. [8:25] uh and voice AI is of course a means of relating  to customers and interestingly this was a channel [8:36] that was hated by customers even though humans  as humans we prefer to talk because of limitation [8:45] in cost human power availability all of this  it became very difficult to man uh and with [8:52] voice AI agent what we're seeing is a rebirth of  that channel where people start to enjoy those [9:00] experiences. So it has gone from I just want to  augment what I have because I don't have enough [9:07] humans to serve my customers to this experience  is delightful and I'm going to have more of it [9:15] because I could now develop deeper relationship  to customers and now it is at the point where [9:20] customers are pulling and they want to spend more  time with AI which we found open more information [9:29] when you talk with AI. So now you for example an  example that I can give which is actually we get [9:37] complaints now from customers like my KPIs are  wrong because if I am say a contact center I am [9:44] measuring on how short I can keep a call but the  desired outcome from a lot of their customers is [9:50] I'm just enjoying this conversation so much and so  they can they have their customers saying oh let [9:57] me tell you about my dog and my son is coming this  weekend, you know, after his graduation party has [10:04] nothing to do with why they're calling, but they  just want to talk to someone and they enjoying [10:07] that conversation so much. So, there is an  amazing potential for revenue and branding um in [10:15] um AI because you can get this personalized touch  one-on-one and about Melon when you talk about a [10:22] hardware do you think is in the results stage or  in another stage right now for you? Yeah, I would [10:28] say, you know, in my time in retail, I've seen AI  go from it can do everything, right? It was like [10:35] the the Swiss Army knife. It can do everything.  Please do everything. The second stage, you know, [10:40] I would come [laughter] to conferences like this  and it was like, hey, it can help your workforce [10:44] management. It can take help forecasting. It  can help with consumer experiences. And where [10:50] I think um a lot of the industry is at and Ace  Hardware as well is moving from what it can do [10:57] to how does it support what you do best. Retail  is a very competitive environment. Um you really [11:04] have to lean in to what is your differentiation  factor. And so now we're approaching AI from that [11:10] lens of how does it support my like calling  card to the consumer? How does it support [11:16] how I need to deliver business? Um, and I think  that's where the tension is showing up, frankly, [11:22] is we've moved from, okay, I'm going to bolt  it on to existing experiences to it becoming [11:28] integrated with your day-to-day. And that's where  conversations on prioritization and governance, [11:34] you know, like who owns this? Is it IT? Is it  marketing? Is it every department? Um, you know, [11:41] how do we translate use cases into provable pilots  with ROI? Um, but yeah, that's how I see the [11:48] continuum going. And I would say ACE is definitely  in the middle of it. We've learned a lot. We've [11:52] made, you know, some successes, some mistakes as  as everyone has. We're a hundred-year-old company, [11:58] so you can imagine what our data systems look  like. Um, raise your hand as your if your data [12:03] is completely amazing and ready for AI. Um, but  um, just one, right? We got one. I need we need [12:11] to talk. Um but um yeah I think that curiosity  has now moved into oh man it's getting real and [12:17] how do we really integrate it into our day-to-day  you said to the to do the best uh don't you think [12:23] is the next level going to be is not what I can  do best with an AI but what can I do totally [12:28] different than I used to do correct as well as  increase the capacity how can I increase what I'm [12:33] doing dayto-day and actually do more and provide  it more efficiently and more productively okay [12:38] Santiago for you what's the jar copel right Well,  we we moved uh aggressively from uh curiosity from [12:46] exploring curiosity to explosive adoption. Uh we  we launched a year ago a a platform because we we [12:56] have the same problem as Mel. Uh our our systems  are a mess uh in terms of data, in terms of a [13:03] lot of uh structure. We are working on modernizing  those. We started playing with AI at the beginning [13:10] and then everybody said okay I need this. So we  developed a platform to connect the AI with the [13:18] mess the underlying mess and then we did something  that I think was very smart which was we we said [13:24] to everyone couple has 125,000 employees of which  20,000 are knowledge workers. Uh so we said to all [13:32] the knowledge workers stop using uh commercial  uh platforms or commercial tools. We're going to [13:39] give you access to them. Uh and actually we are  g giving them through coppel genius which is the [13:45] name of the platform access to tools that it would  cost them $800 a month to have them on their own [13:52] and basically exploded. Today we have more than  10,000 active users per day uh daily active users [14:01] which is about the same number of people that are  in the conference uh these days. So that that is [14:07] incredible. I I mean and the the amounts of things  and the the kind of things that people are doing [14:12] are incredible. Obviously a large group of those  is my team uh which is doing solution engineering, [14:18] software engineering. uh that's about a third of  the people doing uh some work there 30% 30% but [14:25] this is really interesting because I was looking  at the at at the numbers so 30% is it 30% is the [14:32] business and 30% is like uh like integration  of IT to the business so that means that we [14:37] are really using it to solve the problems of the  business with IT and the operating model and and [14:44] and interestingly enough I mean we build this  this layer to to help us uh deal with uh things [14:50] are not modernized yet. Well, it's helping us  modernize quickly because the teams that are [14:55] modernizing or implementing the new platforms  on supply chain or e-commerce or things like [15:00] that are actually using AI to go faster faster  and better. [laughter] Yes, I I I would hope so. [15:07] Definitely. Mark, [clears throat] for you guys  in 15, how is this is the stage of a result? In [15:12] what stage are you? Yeah, I think I mean look I  think let's start at the higher level of what's [15:17] the maturity life cycle of AI in general right I  think I think even the frontier models are still [15:23] toddlers and I'm sure every one of us is concerned  about when these model become teenagers right and [15:32] and governance and control of these models and  all that good stuff but let's put that aside [15:36] that's for a different panel um for us on the area  15 and Fiser side we've kind of taken a fairly [15:44] structive and structured and practical approach  to this. So the first was to standardize on [15:50] an LLM and so we've rolled out claude across  the company uh both area 15 and Fiser and our [15:57] other subsidiaries we have trained the teams and  we're actually bringing in structured training as [16:02] well that will be ongoing so sessions for kind of  management and above uh every couple of months and [16:08] that's really to teach them not just to use chat  but how do they use co-work how do they actually [16:13] use agents and automate their own processes  and things like that So that was really step [16:18] one. Step two was we looked at all of the software  that we're licensing across the companies and we [16:24] said okay this is a few million dollars in in SAS  licensing costs and I prioritized the software and [16:31] we started replacing it with vibe coded stuff that  we've basically created a standardized tech stack [16:37] uh and so any of the developers can now work on  these things and like basically the first prompt [16:41] is this is how we're building it and this is the  tech stack. Um, and so we've replaced, you know, [16:47] probably eight different platforms today. We've  got another 20 that are actually going to be [16:52] done in the next couple of months. So that was  really phase two. And phase three for us is okay, [17:00] we've now built this AI lab. The AI lab's  job is once the business units become kind [17:06] of more familiar and comfortable with their  own AI tools and starting to play with agents, [17:11] they get better at understanding what processes  could be optimized with AI. But they're bigger [17:18] than a bread box. It's not something that our  SVP of finance could handle on his own. But we [17:22] do have one rule with the AI lab. Every project we  work on has to be completed within a twoe sprint. [17:29] If it goes beyond two weeks, it's not a project  for the AI lab. It gets prioritized with all [17:34] other major projects across the organization.  But really the the lab is now working with the [17:40] different business units to say, okay, what are  the use cases here? We've developed an internal [17:44] tool where everybody can submit their own use  cases. We've developed an internal prompt library [17:49] so we can share prompts across the organization.  We've developed a Chrome plugin for Claude so that [17:55] basically when you're act interacting with Claude  you can save your prompt or you can optimize your [18:00] prompt using best practices for prompt writing but  also historic prompts that we know are effective [18:05] within the organization. So we're really I mean we  are all in in this stuff and our goal is not our [18:13] goal is not to remove staff. Our [snorts] goal  is to free up staff to focus on the strategic [18:22] stuff the big picture stuff. Yeah. Okay. Keep on  you you uh with all this process what is changing [18:29] the guest experience at the area 15. Well  there's a lot of things in the works. I mean, [18:33] I can give you an interesting example that'll  hang off of our ticketing platform, but we have [18:39] peak periods. We're like we're like Lyft or Uber,  right? And so, we've used kind of a third-party [18:45] virtual queuing platform that we've enhanced the  heck out of to make it work for us. We've now [18:51] built our own version of that. Like everything  else, we license and then we build and then we [18:56] save a bunch of money, but we make it the way we  need it at Area 15. And we're kind of unique for [19:01] a bunch of reasons. again a separate panel, but  basically that virtual queuing becomes not only [19:07] the way to basically sign into an attraction and  get a notification when it's time to go to the [19:14] attraction. It really becomes kind of your access  to that guest the entire time they're at the area [19:21] 15 district. So, it allows us to use AI to build  dynamic itineraries for them. It allows us to use [19:28] AI to recommend which attraction to go to next  because they won't have a wait. It allows us to [19:33] use AI to recommend that they make a reservation  at one of our owned restaurants. It recommends [19:39] that they go sign up for the photo pass experience  that we're launching. So, it's it's really it's [19:45] really using AI in real time to create the best  guest experience possible. And in our world, [19:51] there's opportunities to create story around that  as well to make it more immersive for the guest. [19:57] Okay. Santiago, you say Copel was also a bank  for the customers. What changes when credit score [20:03] moves from the back office to the live customer  interest direction? Well, that that is that is [20:10] very interesting and let me reframe it a little  bit. Uh Copel does have a bank but we have we have [20:16] had a bank for 20 years but years before we had  a bank we were giving credit actually today 80% [20:23] of the money the money we lend is not through  the bank is through the store it's through a [20:28] department store credit line it's like buy now pay  later or even like a revolving credit we we have [20:34] several types of credit but interestingly enough  we have always used some sort of machine learning [20:42] or AI to enhance the the experience and also let  let me put it this way. So you would think that [20:51] Copel is a retailer and there are businesses  to sell goods maybe or our business is to loan [20:56] money maybe. a real business is to collect in a  in a country where where where you are lending [21:04] uh money to people that are not on the on the  credit bureau that you're basically doing the [21:09] credit scoring yourself with algorithms. I  mean we hire psychologists, philosophers, [21:16] uh sociologist uh all types of people that help  us understand who's really going to be paying [21:23] and not. So it it has always been part of our  experience in the store. that you go to the teller [21:29] and uh and and have two pieces of things. So, for  example, I always give this example. You have an [21:36] iron which cost I don't know 500 pesos um $25 and  you have some jeans and they cost about the same [21:44] thing. So, the propensity to pay for the couple  customer is different for a durable good than for [21:50] a non-durable good. So, in real time we we have we  have been doing this for 15 years. In real time, [21:57] we calculate the propensity to pay and if the  propensity is low, we ask for a down payment. So, [22:04] for example, the genes, if I mean, I'm not trying  to profile anybody, but that we actually profile [22:09] them in real time. 60 year old woman comes to the  teller with those two goods. Uh, most likely the [22:16] jeans are for her for her daughter and most  likely she's more likely to pay for the iron, [22:20] which is a durable good that she's going to use at  home for her. So we we we have all that data and [22:27] we have been using that data to make decisions in  real time. Now obviously that has created a burden [22:33] on the customer experience because sometimes the  systems especially on a Sunday on on a Sunday [22:39] evening where everybody's at the store it becomes  slow. So we we are now using AI not only to do it [22:47] faster but also AI to move faster to the cloud to  modernize our decision making systems to enhance [22:55] our we we have these contactability algorithms  where we figure out okay this customer do I have [23:01] their phone number or not oh no maybe I haven't  I haven't reached them for so long so yeah give [23:07] me your phone number and by the way don't it's  not like in the US I mean in the US you go to a [23:11] store and they said if you give me your number.  I will give you 15% and you say whatever number [23:16] and they give you no no in Mexico we we say okay  I'm going to send you an OTP a onetime pin and you [23:22] have to tell give it to me otherwise you're not  giving me your number you're not giving me your [23:26] number so these are the types of things that we  are using AI to enhance the customer experience [23:32] but also enhance our possibility to collect okay  Melanie you you told us about hey Arma what did [23:39] they learn from putting he ara in more and 2,300  stores. Is 2,300 the number? Well, now we're a [23:46] little over 2,700. 27 just 40 100 other stores,  but [laughter] right nothing. Um but um I think [23:56] we we learned twofold. One is um we learned about  the human side of AI and I think we forget about [24:03] that as we're in conferences like that. We're  talking about what the technology can do. the [24:07] technology is so amazing, but we for forget about  the human component of the people that end up [24:14] um adapting and using the AI. And so one of the  key learnings that we learned when we launched [24:19] this mobile assistant, this digital assistant was  it had to fit within the work stream of a standard [24:25] associate. We if you've ever been and you've ever  worked retail, it's a lot. You're you're stocking, [24:32] you're answering someone in the aisle, you got  to move up to the register to check someone out, [24:37] someone's asking you a question, right?  You're doing a million things at once. And so, [24:41] as we thought about a digital assistant, it had to  fit that busy work stream. It couldn't be, oh man, [24:47] it's just another thing that corporate wants  me to do [laughter] to save time and service [24:52] the consumer. No matter the contra, it's the  same thing. like [laughter] definitely like, [24:56] oh, it's another metric that I'm going to have to  how many times what they invented now AI tool. We [25:01] really thought about the human side of it and  we perfected it with humans. Um, we also tested [25:06] it with associates before it went out. So, we  really thought about and took the time to say, [25:12] hey, how is this going to work in our stores? Just  not in a lab or on a PowerPoint or in an executive [25:18] presentation. like when it's bellyto belly in the  aisle is this helping that associates to give the [25:23] better answer and to service the consumer better.  So that's that's one. The other is always lead [25:30] with a consumer insight. AI needs to be grounded  in that consumer insight or that business problem. [25:36] And um hey Arma actually started out of our store  systems team. They said hey like how do I reduce [25:42] the onboarding time for a new associate? You know  we've got someone who just joined the ACE family. [25:48] Um, we call them redvested heroes. Um, they just  joined the Ace family and now they've got to learn [25:54] everything from grills to paint to, you know, the  widget on aisle number 45. Um, how do we make that [26:01] better for them? And that's how Hey Arma actually  started was that tension point, that consumer [26:07] insight, that retail insight, and then we said,  you know what, AI is something that can help with [26:12] that. First being love for the problem, then you  be in love with the technology. Huh. Natalie, talk [26:18] about voice AI for sure. What makes a customer  accept an AI voice agent instead of asking for [26:25] a human? It's interesting. Um, great question. So,  number one, I will go back to my first statement. [26:35] We like to speak to other people. It is a very  low friction way for humans to engage. You don't [26:42] even need to know how to read uh any anyone has  a language. it's immediate and we can get a very [26:48] rich response. So the key challenge particularly  also for retailer was to overcome the expectation [26:57] that most of us had which was it was going to be a  terrible experience and so the way uh the industry [27:04] handled this was to say the number thing that  we're going to cut is wait time with AI because [27:10] with AI anyone can pick up the phone at any time.  So that was one thing that would make someone [27:15] want to speak to an AI because the alternative  is I have to wait in line for 45 minutes. The [27:21] second thing is they typically pick use cases  or objectives where there was a very clear [27:30] outcome or objective definition and a very clear  measurement of that outcome. And it could be I'm [27:36] here to sell you something or I'm here to answer  your questions or I'm here to take your payment [27:41] or I'm here to whatever it is hand out a return  and um the expectation was then set very clearly [27:49] with the consumer themselves and it was very easy  then for the company to also tweak their agents to [27:56] get the business result they wanted and we thought  ultimately it was going to be leading to a lot of [28:03] escalation. What we saw in the market was exactly  the opposite. Provided that the agent is wellmade [28:11] and it does take either expertise that you have  to invest yourself or it takes you using a tool [28:18] that gives you that expertise that includes that  expertise but people just want to talk. And so one [28:26] example that I love to talk about is a company  Mexican company. It's called Kavak. They have [28:32] a call marketplace and they started to use voice  AI agents to first help book uh test drives and it [28:42] was really important for them because the as soon  as someone wants to have some request for test [28:47] drive they go through some aggregator and then the  first people that respond to them is more likely [28:51] to get the business and then they realize that  it's a very difficult decision for the customers. [28:57] It's a lot of money. The customers need more and  more calls to increase the level of confidence in [29:05] buying the car. And so they would want to engage  with them and Kai didn't have enough people to [29:10] put in front of them in order to answer all these  questions. But when there was this voice AI agent, [29:16] they just keep calling back and again and oh,  I just saw this, what about this, what about [29:20] that? And so the consumers were saying, "No, I  I have the answers that I want when I want it." [29:27] They found the AI to be much more patient with  them. They found the AI to want to please them. [29:35] Um, and so they wanted to spend more and more  time and they've reorganized their entire company [29:42] around that customer experience. So now everyone  of their function touches in this case it's voice [29:49] because voice is important to their business but  could be true for any any AI obviously depending [29:54] on the Yeah. no matter what and then so they  organized everything so that then it's a complete [29:58] pull from the customer uh and escalations is  extremely extremely rare and just said about [30:04] any AI what matters more better technology or  better business result business results hands down [30:14] uh at the end of the day all these businesses are  here right either to make money save money because [30:18] we're all here to make a profit AI is a technology  in the service of a business in the service of a [30:25] customer experience. So always start with that and  then expand. Okay. B for you. 100% I I agree with [30:33] you. Business result. Um at ACE we look at AI and  we decide which use cases actually go to pilot and [30:41] then go on to scale based on one r rubric. Are we  delivering on our promise to be the helpful place? [30:47] If we can't say yes to that um we're not doing  it. Um, and so it really is about the business, [30:53] not necessarily technology for technology's sake.  It's about how can we serve our customers and make [30:59] sure when they come to our stores or they shop  our our um online platform, they're having the [31:04] best and most helpful experience. Let me add this  to this question. How you run off the hype of a [31:10] new technology, for example? Say that again. How  you run off a new the hype of a new technology, [31:16] for example. Yeah. I think for for in just for  results. Yeah. I think for us the way that we're [31:22] approaching, you know, I I I guess I can say  this like AI almost feels like a drinking game [31:29] sometimes and I feel like if we were to say AI is  a drink, we'd all be drunk because it comes up in [31:34] every single conversation. There definitely  is a lot of hype behind it, right? And we [31:38] um if you've worked for a company, you probably  have an executive that's tapping you on the [31:42] shoulder and saying, "Hey, how are you using AI  in your division?" And you're like, "I don't know, [31:47] but I'll get back to you on that." the way that  you translate it so it doesn't just become hype, [31:51] it becomes what you do and how you service that.  It's really to think about how do how have you [31:56] democratized AI across our organization. One  of the things that a um ACE did at the very [32:02] beginning um I serve on our executive steering  committee um for the corporation and one of the [32:08] things that we did in the very beginning is we  designated AI champions in every single division [32:13] across the company. Everything from international  to merchandising to retail ops to marketing, they [32:19] have an AI champion. And the idea here was let's  not make it something that you need to do and it's [32:25] hype and you feel like okay, I've checked that box  and move on to something else. Let's democratize [32:30] it and have it led from the bottom up versus the  top down. let those people in those departments [32:36] figure out where the friction is or where there is  a better consumer experience that can be supported [32:42] by AI and bubble that up to a steering committee,  a governance arm, an IT group that can then work [32:48] on it. And that's how we made it more of this is  how you're doing work versus this is what I need [32:54] you to do. And so it's less hype and more how do  we deliver that experience. Santiago, how in Copel [32:59] do you deal with this with new technologies and  the hype of new technology be and um in the other [33:05] hand in the other side you have the the looking  for results first. Exactly. Well, that that is [33:11] a very interesting balance because I I mean the  obvious answer is is results. But let let me give [33:16] you how how we approach this this balance. Uh and  it has to do a little bit with what people talk [33:22] about the build versus buy versus own. We we like  to think in copel that uh in terms of technology [33:30] we want we want to buy the best basic technology  build the platform and own the capability of doing [33:40] the business results and what type of business  results those that are very dramatic in terms [33:45] of for example we we just introduced a generative  search engine for our e-commerce well it's giving [33:52] us 4x results and and we're still shy about about  that but it's that's the type of results that we [33:59] want but how do we do that also al also with  governance so we have we we have a a steering [34:05] committee on AI that is led by by our COO and also  by the strategy area but also on technology we say [34:12] okay this is this is how we're going to build it  and we're not just going to unleash you to to do [34:19] everything that you you want to do in your way do  whatever you want but let's do it in a control way [34:25] with uh platforms that have NDA that that have  the security and all of those things. So we we [34:31] try to balance both both of those. And how how  do we uh manage the hype? I don't know. It's it's [34:37] difficult. I mean coming to to a conference  like this, I don't know if if my colleagues [34:41] feel the same, but you you get uh swamp pressured  pressure. You get swamped with all these requests [34:47] for a 15-minute uh coffee or whatever. I mean uh  we don't have that time. Sorry. Don't worry, Mark. [34:55] There I got to use the same question that the that  Santiago defense a little bit. Uh what kind of [35:01] technology you should buy, you should build, and  you should own. Well, I I want to go back to the [35:07] original question. I I for me, tech is always  just a tool. Whatever the technology is, it's [35:13] just a tool to get to better business outcomes  or a better guest experience, which gets you to [35:19] better business outcomes. I I I the whole time  everyone was thinking speaking, I was thinking [35:25] When I was asked to kind of oversee tech across  Fiser and Area 15, I had walked into what I'm sure [35:32] all of you have seen, which is what I called  like a miserably failed AI pilot. And it was [35:40] an overeager use of this technology to solve this  really big problem. In this case, it was how do we [35:47] abstract our leases and basically create this kind  of Fisher brain? And if any of you have looked at [35:52] these, you know, multiundpage commercial leases,  they're out of control. And they had spent a year [35:58] trying to get the AI to do this. And I before I  even saw the results, I said, "It we'll get you [36:03] 80% of the way there." And they said, "Yeah, and  then we'll just, you know, we'll we'll solve the [36:08] other 20%." But I said, "Yeah, but that other  20% is like a needle in a hay stack. So, you're [36:14] going to spend more time than you would have spent  manually abstracting a lease with a system and a [36:19] tool that you've been using for 20 years that does  it at 99.9% accuracy when what you really want is [36:25] to get that data into a system so you can actually  consume it and use it and be more efficient. And [36:32] so, we scrapped that and we built that system and  we used kind of a standardized process. And by the [36:39] way, we also created a very simple template  for one of Lisa's sign for someone just to [36:43] enter the data cleanly into the system where you  can consume it in a million ways. So like look, [36:51] I'm I'm like a 30 plus year CTO. I've been  building this [ __ ] forever. It is so easy to [36:57] get absorbed by the shiny new thing. AI is a once  in a-lifetime technology. It is more significant [37:05] than anything any of us will and have ever seen.  The power of it is extraordinary. The ability to [37:12] change our lives is extraordinary. It is still  in its infancy. Technology is the technology, [37:19] right? It needs a babysitter and it still needs  a human in the loop. Can I add one thing because [37:25] something that we think so vap is is a a platform  and so people use platforms to build uh their eye [37:32] experiences and there are um three reasons that we  hearing customers move to platforms regardless of [37:41] what they are happens to be a voice AI platform  they exist for any kind of AI um technology but [37:47] one of them is they actually to not get there  they want the technology to be almost transparent [37:55] to them or is it complexity to be transparent to  them. So they do state their purposes say I want [38:01] this outcome or I need you to do this for me and  platforms have an advantage in that they hand they [38:07] take away the complexity of the technology they  present it as a level of abstraction that you as [38:12] a business person or any kind of function you  are you can actually interact with it and that [38:17] helps with the problem of um finding a technology  in search of a problem as opposed to you have a [38:24] problem and you can use that technology to solve  it. So that's number one. Sometimes people find [38:28] first the technology that let me see there's a  problem to resolve with this problem. Let's find [38:32] a problem now. But it gets you to at least all  right after that you can always decide like do [38:37] I want to invest more myself in there or not. So a  lot of people use it for that. The second thing is [38:45] um ownership versus writing like what is it  again that you want to own or what is it that [38:50] you want to you're okay with someone else doing?  And anything that is about an extremely important [38:57] business icon to you or if it is a customer  relationship something like this that is a [39:02] real true asset to your business you want to own  this even if you use tools to achieve that you [39:09] want to be in control of just that bit and so it's  another attraction platform or if you decide to [39:15] build it yourself where you how you want to decide  what to own and not own. Uh AI is also an amazing [39:22] technology in that anyone can use it. In the past,  if you wanted to have access to great technology, [39:30] you needed to be a specialist. You needed to be an  engineers. You need to know about communication, [39:36] whatever it was. Definitely. Um and with AI now,  a lot of people can do a lot of things that they [39:42] were not taught to do, they were not born to do.  And so it also opens up a million opportunities [39:49] for um creativity for innovation. It lowers the  barrier to access uh to all of the things. So [39:57] that's also one of the magical things about AI.  Any function can now pretty much do anything [40:02] which poses the last problem which is the one that  we're hearing the most from our customers in terms [40:06] of what is preventing AI to be deployed across  organizations and that that of governance. Um so [40:13] unifying data, unifying context, unifying rules,  whatever you know obviously ACE for example is [40:20] like very strong like are we helpful to customers.  So making sure that this is respected across the [40:24] board and then forced in the technology or in  the tools that you're using has become very [40:29] important as a condition for AI to then be adopted  by organizations at large. Okay. ML passing to you [40:37] the first of the same question what's to buy  was what's to build and what's to own. Yeah, [40:42] I would say um buy the commodity. Um there are  a lot of tools that you can buy that you're just [40:49] trying to be at par. You're trying to build in  some new capabilities. This is not going to be a [40:54] point of differentiation. That's where I think you  build you build your points of differentiation. [41:00] We built Hey Arma. It's literally how we deliver  our services and store. We want to own that and [41:05] we want to build that and we want to be able to  iterate that and learn from that. So I would say [41:10] build what is your point of differentiation. Own  the insight of whether you do and use A or or not. [41:17] Do not let someone else have that that kind of  mindset for you or own that mindset for your [41:24] company. Especially in this world of we're trying  to figure out governance or Marky said, you know, [41:29] AI is is is in its teenage or baby stage. Um,  toddler stage, [laughter] even younger. Um, [41:36] but you really need to own where you're going  to put AI and where it's going to be integrated [41:41] into your company and where it should not live.  and those decisions and under what Natalie said [41:46] it's basically own own your secret sauce and your  moat yeah definitely definitely definitely Mark [41:52] starting with you right now final round for us  the unfortunately the time is almost up in one [41:58] sentence what should every retailer do next with  an AI in one sentence what will area 15 do next [42:05] with AI is that what you said you know any kind  of retailer here what they do do next with AI what [42:11] what what they need to look for what should they  do next with AI. Yeah. Yeah. Oh, [laughter] no. [42:17] Hard question. Listen, one sentence. Come on.  Oh, Ku's Jew. I don't know. [laughter] Look, my [42:23] my son is an English teacher and he's a two-year  English teacher and all of a sudden he's building, [42:31] you know, pool chemical management applications.  Honestly, go home and play with AI. Forget about [42:39] its massive impact to the world. just do something  and and realize that it has and this was said [42:46] earlier it has democratized software development.  It has democratized the access to knowledge and [42:53] data beyond go beyond chatting with it. Don't use  it to plan your next vacation itinerary. Think of [43:00] an idea. Build the idea. Prove to yourself that  you can actually build a business and software in [43:06] an evening or over coffee while you're sitting  there watching the news in the morning. I'm [43:11] telling you it's fact. I have replaced multiple  SAS applications that I have built over coffee in [43:16] two hours. Jason Santiago. Well, my my my advice  or suggestion would be go back and look at one [43:27] problem or a set of problems that you thought  were impossible to solve and try solving them [43:34] with AI or at least try to approach them with AI.  You probably won't solve them if you fully with [43:39] AI. I mean you will it will be a combination of  artificial intelligence and your intelligence but [43:45] uh but yeah I think I think AI gives you a  different angle to problems that you at least I [43:52] thought were impossible to solve. Mel yeah I want  to double down on what Santiago say I would say [43:58] for all retail CPG e-commerce really lead with  that insight and [clears throat] that business [44:03] problem. there's one thing you can do, do that.  It'll show you where AI needs to show up. Um, [44:10] you'll see those point of friction. Um, I've  done innovation in most of my career and we [44:15] call it duct tape innovation. You see someone  doing compensating behaviors and then you [44:21] fill it with a product or a service. Do the same  thing for your your retail environment, your CPG [44:27] uh company. think about where there are points  of differentiation and you level that insight of [44:33] like this is where AI can show up and help us be  more competitive in the market. Natalie, you too. [44:39] And finally, where people can know a lot a little  bit more about here on the expo. All right, I'll [44:45] I'll answer the first question. So, um I would say  do two things. Number one, use AI. I don't think [44:51] it's optional anymore and identify something very  bounded something narrow scope with a very clear [44:58] RARI that you can measure. Number two, insulate  yourself from technology changes. The rate of [45:05] innovation in the AI world is unprecedented.  So build your tech in layers so that you can [45:13] change models underneath. you can change the  best of the breed as it evolves every couple [45:19] weeks without having to change anything on top.  So don't change your business just because a new [45:25] model has come out. So it'll be my two things. Um  where find VP only here on the expo. So yeah well [45:33] uh if you want to any question about VP obviously  we have vap.ai though um I would just recommend [45:38] that you play with it. VP is a platform you don't  have to be a developer. No problem is too small. [45:45] You can do it for very large scale businesses.  You can handle millions and millions of calls a [45:49] day or you can make one call. Um, you're going to  love it for sure and you just Yeah, just play with [45:55] it. It will wow you. It wow me. Okay, time's  up. Thank you everyone. Thank you everybody.