Home Depot Builds In-House AI Layer to Stop Vendor Lock-In

家得宝构建自研AI层以防供应商锁定 随着基础AI模型日趋商品化，专有运营数据与解耦架构成为企业唯一持久的竞争护城河。 Ningyu Chen概述了在《财富》前25强零售企业中规模化应用AI的严谨运营模式，强调智能化转型必须首先掌握实时数据管道与运营数字化。在“自研谋求差异化，外购实现对等化”的原则指导下，他详细阐述了在批准规模化部署前，涵盖业务投资回报率（ROI）、技术可行性及一线用户采纳度的严格三阶段关卡评估流程。Chen列举了具体的企业落地应用，包括星巴克的Barista Assist以及家得宝（The Home Depot）的Magic Apron和Pro Blueprint Takeoff工具，阐明统一的核心平台如何带来复合的“滚雪球效应”。最后，他主张采用与模型解耦的抽象层来防止供应商锁定，将专有数据采集视作唯一持久的护城河，并指出汽车穿梭餐厅（drive-thru）中的对话式语音AI是一个尚未攻克的百亿美元级前沿领域。 企业 AI 的成熟需遵循严格演进路径：企业必须先搭建实时运营数据管线并实现数字化触点，才能成功规模化落地智能与个性化的 AI 应用。 (0:04) 在 AI 时代，速度已取代企业规模成为核心竞争要素；企业应恪守“差异化自研，同质化外购”的原则，将前沿基础模型、专有领域数据与初创公司生态合作深度整合。 (3:15) 任何 AI 概念从试点走向规模化推广，都必须经历验证业务 ROI、技术可行性以及一线用户接受度的三阶段门禁评估框架。 (6:07) 基于统一的企业 AI 平台底座开发可产生滚雪球效应，使 Magic Apron 购物助手、专业图纸算量等多个业务场景得以迅速复用共享基础设施。 (9:16) 随着基础模型和基础设施逐渐同质化，沉淀并掌控专有领域数据成为企业唯一的长期竞争护城河；同时架构上应设立抽象层以实现模型解耦与快速切换。 (12:42)

Ningyu Chen, Senior Vice President of Technology at The Home Depot, argues that enterprise AI adoption requires sequential operational prerequisites, a strict division between proprietary and commodity software, and disciplined stage-gate validation across distinct organizational productivity tiers. Prerequisite Architectures and Platform Foundations (0:04) Chen frames enterprise technology evolution across three sequential eras: real-time operational data pipelines (exemplified by tracking flights, bags, and crews at Delta Air Lines), digitization and omnichannel infrastructure (such as store and supply chain digitization at Macy's and Inspire Brands), and the intelligent enterprise driven by contextual AI at scale (at Starbucks and The Home Depot). He contends that organizations cannot bypass earlier phases, asserting that intelligent AI deployment at 'Point C' requires completing real-time data integration at 'Point A' and operational digitization at 'Point B.' Build versus Buy and Vendor Abstraction (3:15) In the AI era, Chen identifies deployment speed as the decisive operational factor, whereas earlier transformation waves allowed enterprises months or years to evaluate implementations. To guide resource allocation, he adheres to the principle of building for differentiation and buying for parity. For The Home Depot, frontline store associate tools and customer shopping journeys represent core competitive advantages that cannot be outsourced, while standard administrative functions like HR and finance rely on commercial off-the-shelf software. To balance differentiation with platform scale, Chen combines foundation model partnerships—collaborating with Anthropic, OpenAI, Google, and Microsoft—with selective investments or acquisitions of early-stage startups to secure proprietary intellectual property. Architecturally, he mandates placing an abstraction layer in front of all large language models and commercial APIs. This decoupling insulates the enterprise from vendor lock-in and allows technical teams to substitute underlying models within days or weeks if vendor pricing, stability, or capabilities fluctuate. Stage-Gate Governance and Frontline Deployment (6:07) Before committing engineering resources, Chen enforces a three-stage evaluation framework: validating the business problem with quantifiable return on investment (ROI), assessing technical maturity and architectural feasibility, and securing business sponsors alongside verified frontline user demand. Initiatives progress through a gated delivery funnel moving from proof-of-concept to controlled pilots, limited scale, and finally enterprise rollout. Chen illustrates this disciplined rollout with Starbucks' Barista Assist, an OpenAI-powered generative tool built with Microsoft to reduce recipe memorization and operational onboarding strain amid industry-wide turnover. Following limited testing and a 300-store pilot, the system scaled across 11,000 domestic retail locations. At The Home Depot, the company developed 'Magic Apron,' an AI shopping assistant operating online and via an in-store mode to help DIY shoppers navigate an inventory of 3 million SKUs. Chen describes a 'snowball effect' where this foundational platform subsequently enabled additional use cases, including Blueprint Takeoff—which automates material estimates from construction blueprints for professional contractors—and post-purchase logistics support in call centers. Data Harvesting and Three Productivity Tiers (12:42) Chen maintains that foundational models, orchestration tools, and compute hosting will inevitably become commoditized, leaving an organization's proprietary domain data as its sole durable competitive advantage. Consequently, he advises enterprises to harvest their own operational data and maintain strict ownership terms when negotiating vendor contracts. To manage organizational execution, Chen categorizes AI use cases into three distinct governance tiers. Personal Productivity focuses on broad workforce literacy via commercial tools like Microsoft Copilot and Google Gemini Pro without rigid ROI requirements. Team Productivity targets departmental efficiencies using low-code or no-code tools built by citizen developers. Enterprise Productivity targets core operations, supply chains, and customer journeys using custom internal platforms; these initiatives demand measurable financial returns and rigorous security oversight across all deployed agents. Chen identifies automated voice ordering for quick-service restaurant drive-thrus as an unsolved application dating back to his work in 2019. Despite representing what he terms a 'billion-dollar opportunity' to mitigate labor shortages and automate order upselling, reliable enterprise deployment remains difficult due to persistent challenges with background noise, varied accents, and complex conversational patterns.

The Home Depot技术高级副总裁Ningyu Chen指出，企业要将人工智能从孤立的实验性试点推向全企业规模化应用，必须建立在成熟的数据与数字化底座之上，并通过“自研求差异、采购求均势”的原则及严格的阶段门控机制进行落地推进。他强调，基础模型终将商品化，唯有深植于核心业务场景的专有数据与统一平台能力，才能形成持续复利的竞争壁垒。 企业转型的三大阶段与底层演进顺序 (0:04) Chen结合其在达美航空（Delta Air Lines）、梅西百货（Macy's）、Inspire Brands、星巴克（Starbucks）及家得宝（The Home Depot）的任职经历，将企业技术演进归纳为三个不可逾越的阶段：首先是依托航班、行李和机组等实时运营数据建立的“实时企业”（Point A）；其次是围绕门店和供应链推进全渠道与业务线上化的“数字化转型”（Point B）；最后才是基于上下文理解、用户意图与个性化交付的“智能企业”（Point C）。 根据Chen的观点，企业无法跳过A点和B点直接实现C点，若缺乏扎实的实时数据管道与数字化业务触点，AI的规模化部署便无从谈起。无论处于哪一波转型浪潮，企业若想实现可持续的成效，必须依托一套兼顾准确性、响应速度、扩展弹性、系统安全性与成本效益的企业级技术底座。 差异化自研、生态合作与模型抽象解耦 (3:15) Chen指出，与以往允许企业耗费数月甚至数年进行论证的数字化转型相比，AI时代的核心变量是“执行速度”，任何规模的企业都面临极高的时效要求。在架构决策上，Chen践行“自研求差异，采购保均势”（Build for differentiation, buy for parity）的原则。家得宝作为财富25强零售商，其核心客户旅程与一线员工工具被视为不可外包的战略护城河，且商业现成软件往往无法承载其业务负荷与复杂度；人力资源与财务等通用后台功能则选择直接采购。 面对快速更迭的大模型技术，Chen主张采取双轨合作策略：一方面与Anthropic、OpenAI、Google及微软等前沿模型厂商直接合作，在基础模型中注入家得宝专有的零售经验与交易数据；另一方面通过对初创企业联合投资或收购锁定核心知识产权。为杜绝供应商锁定风险，Chen建议在所有大语言模型与第三方API前部署抽象层，确保企业在供应商面临宕机、涨价或能力滞后时，能在一天至一周内无缝替换底层模型。 三阶段门控机制与三层生产力治理模型 (6:07) 为杜绝脱离实际业务的无效技术试点，Chen确立了严格的三阶段门控评估框架（Stage Gate Framework）：第一阶段评估业务背景与投资回报率，明确具体业务痛点并量化财务收益；第二阶段进行技术可行性审查，评估模型成熟度及架构支撑部门级或全企业级负荷的能力；第三阶段核验组织就绪度与变革管理，确保获得业务部门主管的鼎力支持，并验证一线员工或消费者的实际采纳意愿。只有顺序通过这三道门槛，方案才会进入概念验证（POC）、受控试点、有限扩展乃至全企业推广的部署流程。 在日常治理方面，Chen将组织内的AI应用划分为三个层级：第一层是利用微软Copilot或谷歌Gemini等商用助手提升个人AI素养，该层级不设严苛的ROI考核；第二层是赋能业务人员利用低代码/无代码工具实现部门流程自动化（如账目结算、采购审批），考核部门任务周转速度；第三层则是依托自研平台驱动收入增长与核心业务（如供应链与核心客户体验），必须承担明确的财务回报与业务价值。Chen强调，中小企业必须设法沉淀自有业务数据并在对外采购中牢牢锁定数据所有权，因为数据是抵御模型商品化的唯一持久屏障。 统一底座驱动的“滚雪球效应”与落地实践 (9:16) 在星巴克任职期间，Chen团队联合微软基于OpenAI模型打造了“咖啡师助手”（Barista Assist），以缓解快餐零售业普遍存在的高流失率与培训压力。该助手帮助一线员工快速掌握复杂的饮品配方、定制参数及设备排障流程，在历经严密受控的阶段测试后，项目从300家门店试点最终扩展至全美1.1万家零售门店。 加入家得宝后，面对非专业家装消费者难以检索300万种在线SKU的痛点，团队构建了覆盖网页端、移动端及实体店“店内模式”的AI导购助手“Magic Apron”。Chen将这种扩张模式总结为“滚雪球效应”：依托单一标准化的企业AI底座，持续衍生出多样化的业务能力。例如针对专业建筑承包商推出蓝图解析工具（Blueprint Takeoff），自动提取民宅图纸中的材料用量清单、核算价格并匹配实时库存；同时将对话智能引入客服中心，协助处理售后配送追踪、数量变更和工地缺料等复杂履约问题。 尽管生成式AI在多项业务中展现出扩展潜力，Chen指出餐饮快餐行业的语音自动点餐（如车载外卖通道）仍属极难攻克的复杂领域。自2019年开始探索以来，该场景因受到极端背景噪音、各地口音及多样化对话习惯的干扰，至今仍未获得彻底解决；但Chen认为这一难题背后蕴藏着数十亿美元的市场空间，是未来亟待突破的关键前沿。

## Background 该内容由 Ai4 频道于 2026 年 8 月 21 日发布在 YouTube 平台，记录了家得宝（The Home Depot）技术高级副总裁 Ningyu Chen 与《Inc.》杂志资深撰稿人 Jennifer Conrad 展开的炉边对话。简介表明，多数 AI 试点受阻并非因为技术缺陷，而是卡在概念验证到企业级采纳之间，本次分享聚焦于推动 AI 走出试点、迈向全企业规模化落地的核心框架、自研与采购策略及通用实战教训。 ## TL;DR Ningyu Chen 结合达美航空、星巴克与家得宝的变革历程，指出 AI 转型必须以坚实的实时数据底座与业务数字化为基石，并确立了“为差异化自研，为通用对齐采购”的决策原则。他提出了从业务 ROI、技术成熟度到一线采纳意愿的三阶段评估门禁，并以星巴克助手落地 11,000 家门店及家得宝“Magic Apron”和“蓝图算量”为例，展示了分级放量与平台滚雪球效应。此外，他建议中小企业以沉淀专有数据构筑护城河，倡导建立涵盖个人、团队与企业级的三层分级治理模型，并强调通过构建模型抽象层以应对底层大模型快速迭代的不确定性。 ## Atmosphere & style 整场对话务实、严谨且充满一线实战的硬核底色，摒弃了泛泛的行业炒作，直击大企业技术落地最棘手的架构与变革管理难点。Ningyu Chen 的叙述条理清晰、坦诚从容，既有对技术更迭速度的敬畏与敏锐，又有将复杂系统推进至万家门店的大将风范。主持人的提问紧扣企业核心痛点，嘉宾的回答环环相扣且案例生动，让整场交流如同一次顶级技术领袖的闭门战略复盘，极具启发性与实操指导价值。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=yy14kGsCBd0&t=4s) 企业技术演进历经三个阶段：以实时数据互通为核心的“实时企业时代”、以全渠道协同为核心的“数字化转型时代”，以及以理解业务上下文和用户意图为核心的“智能化企业时代”；AI 转型无法跳过前两个阶段直接实现。 - [00:04](https://www.youtube.com/watch?v=yy14kGsCBd0&t=4s) 底层企业级技术平台必须兼顾准确性、响应速度、扩展规模、安全性与成本效益五大核心指标，否则任何试点都无法沉淀为可持续的商业价值。 - [03:15](https://www.youtube.com/watch?v=yy14kGsCBd0&t=195s) 在生成式 AI 时代，“执行速度”已取代单纯的企业体量成为决定胜负的最关键变量。 - [03:15](https://www.youtube.com/watch?v=yy14kGsCBd0&t=195s) 技术选型应遵循“为差异化而自研，为通用对齐而采购”（Build for differentiation, buy for parity）的心智模型，核心客户体验与一线员工系统坚定自研，通用大宗场景优先采购成熟方案。 - [06:07](https://www.youtube.com/watch?v=yy14kGsCBd0&t=367s) AI 试点走向全企业规模化需通过三道门禁评估：明确的业务上下文与 ROI、成熟且可扩展的技术可行性、一线用户的采纳意愿与业务骨干的变革管理推动。 - [06:07](https://www.youtube.com/watch?v=yy14kGsCBd0&t=367s) 星巴克通过阶梯式推进路径（单店试点、小范围微调模型、分批放量至 11,000 家门店），成功将 OpenAI 赋能的咖啡师辅助系统全量落地，显著降低一线员工的认知负荷。 - [09:16](https://www.youtube.com/watch?v=yy14kGsCBd0&t=556s) 家得宝推出面向消费者的“Magic Apron”购物助手，依托统一的企业级底层平台，有效解决了 300 万 SKU 复杂家装商品的发现与选购痛点，并实现了线上商城与线下实体门店模式的无缝融合。 - [10:49](https://www.youtube.com/watch?v=yy14kGsCBd0&t=649s) 通过平台的“滚雪球效应”，家得宝将 AI 从 DIY 散客拓展至专业承包商（Pro），推出自动解析施工图纸的“蓝图算量提取”（Blueprint Takeoff）功能，并打通呼叫中心构建全流程售后生态。 - [12:42](https://www.youtube.com/watch?v=yy14kGsCBd0&t=762s) 基础大模型和编排框架将高度商品化，中小企业与初创公司的核心护城河在于自身专有数据资产与业务认知，在外部合作中必须牢牢掌握数据所有权。 - [13:51](https://www.youtube.com/watch?v=yy14kGsCBd0&t=831s) 家得宝推行 AI 三层心智模型：个人生产力（提升全员 AI 素养，不设严苛 ROI）、团队生产力（通过低代码平台实现业务自动化）、企业级生产力与增长（驱动核心供应链与营收的高稳健系统），底层依托统一的安全治理框架进行监控管理。 - [16:05](https://www.youtube.com/watch?v=yy14kGsCBd0&t=965s) 在架构设计上应搭建“模型抽象层”，解耦业务逻辑与底层开源/闭源模型接口，以对抗技术迭代不确定性，确保系统在供应商故障或落后时具备快速平滑切换的能力。 - [17:21](https://www.youtube.com/watch?v=yy14kGsCBd0&t=1041s) 复杂真实环境下的全双工“语音对话式 AI”因方言、环境噪音与长尾交互差异至今极难完美落地，但一旦被攻克将成为颠覆线下服务与点餐业态的十亿美元级商业机遇。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=yy14kGsCBd0&t=4s) 数字化转型的三个时代与演进框架 Jennifer Conrad 开场介绍了对话嘉宾——家得宝（The Home Depot）技术高级副总裁 Ningyu Chen（陈宁宇），并引出本次炉边对话的核心议题：企业如何规模化落地 AI 试点项目，以及在何时应当推进、何时不宜盲目扩张。Ningyu Chen 回顾了自己先后在达美航空（Delta Air Lines）、梅西百货（Macy's）、Inspire Brands、星巴克（Starbucks）及家得宝等标志性企业推动技术变革的历程，将自己的职业生涯总结为“通过技术变革提升客户体验、一线员工体验以及整体业务运营效率”。 Ningyu Chen 指出，企业技术演进历经了三个截然不同但环环相扣的时代： 1. **实时企业时代（以达美航空为代表）**：核心在于大规模构建业务领域的实时数据底座，使航班、旅客、行李、机组人员等数据能够全量实时互通，这也构成了现代出行体验的底层支柱。 2. **数字化转型时代（以梅西百货与餐饮集团 Inspire Brands 为代表）**：以实时数据为基础，推动商品运营、供应链和门店全方位数字化，为消费者构建全渠道（Omnichannel）体验。 3. **智能化企业时代（以星巴克与家得宝为代表）**：适逢 ChatGPT 横空出世，核心转向理解业务上下文（Context）与用户意图（Intent），为顾客和一线员工提供高度个性化的交互体验。 基于这三个时代的沉淀，Ningyu Chen 提炼出两条关键方法论：其一，“若不经历 A 点与 B 点，你绝不可能直接跳到 C 点”，即 AI 转型不能空中楼阁，必须以大规模实时数据（A）和完善的业务数字化（B）为坚实基石；其二，必须始终构建一个具备“高准确性、高速度、大规模、高安全性和高成本效益”特征的企业级技术平台，缺乏底层成熟平台的支撑，任何转型试点都无法形成可持续的企业级价值。 ```mermaid flowchart LR accTitle: 数字化与智能化转型的演进路径 subgraph S1["阶段 A：实时数据底座"] A["大规模实时数据<br/>(航班/旅客/库存)"] end subgraph S2["阶段 B：全渠道数字化"] B["业务流程数字化<br/>(供应链/商品/门店)"] end subgraph S3["阶段 C：智能化转型"] C["AI 与意图理解<br/>(个性化体验/决策赋能)"] end A --> B --> C ``` - **实时数据为先**：缺乏底层实时信息同步，后续的数字化与 AI 应用将失去数据准确性支持。 - **平台核心指标**：企业级平台必须同时满足准确度（Accuracy）、响应速度（Speed）、扩展规模（Scale）、安全性（Security）与成本效益（Cost Effectiveness）。 --- ### [03:15](https://www.youtube.com/watch?v=yy14kGsCBd0&t=195s) AI 时代的新变量与“自研 vs 采购”策略 谈及近年来行业环境的剧烈变化，Ningyu Chen 强调最关键的变量就是“速度”（Speed）。在过去的数据转型或数字化转型时代，企业往往有数月甚至数年的时间去规划、论证和逐步推进，大型企业凭借体量优势通常能占据主导地位；但在生成式 AI 时代，速度已成为决定胜负的首要因素，无论企业规模大小，只要具备正确的构想并以极致的速度执行，就能脱颖而出。 针对企业在 AI 布局中面临的“自研还是采购”（Build vs. Buy）抉择，Ningyu Chen 分享了一个清晰的心智模型：“为差异化而自研，为通用对齐而采购”（Build for differentiation, buy for parity）。对于家得宝而言，客户体验以及门店一线员工体验是核心竞争力所在，这类直接决定企业命运的场景“绝不会外包给第三方”，加之财富 25 强企业的超大规模与业务复杂性，市面上极少有成熟供应商能直接承载，因此必须坚定自研；而对于 HR、财务等通用大宗业务功能，则优先采用成熟商用方案。 此外，Ningyu Chen 提出了 AI 时代至关重要的第三种路径——战略合作（Partnership）： - **与前沿基础大模型厂商深度合作**：家得宝与 OpenAI、Anthropic、Google 和 Microsoft 等顶尖机构紧密协作，采用其最先进的工具与模型作为底座层，再注入家得宝独特的行业专业知识、私有数据和历史交易沉淀，构建专属领域模型。 - **与高潜力初创团队共建共赢**：密切关注具有突破性构想的初创公司，在早期阶段通过联合投资或技术收购的方式将其纳入生态，将关键知识产权（IP）内化为企业自研能力的一部分。 --- ### [06:07](https://www.youtube.com/watch?v=yy14kGsCBd0&t=367s) 从试点走向规模化：评估模型与星巴克案例 将 AI 项目从概念验证（PoC）推向全企业级规模化，需要经历严谨的阶段性门禁（Stage-gate）评估。Ningyu Chen 阐述了落地前必须回答的三个核心前置问题： 1. **业务场景与 ROI 明确度**：是否具备清晰的业务上下文、可量化的投入产出比（ROI）以及对成功标准的明确定义。 2. **技术成熟度与可行性**：底层技术是否足够成熟，能够稳健支撑部门级或企业级规模的运行要求。 3. **一线采纳意愿与变革管理**：最终用户（无论是消费者还是基层员工）是否具备积极的使用意愿，是否有业务侧的核心业务骨干（Business Champion）全力推动组织变革管理。 Ningyu Chen 结合其在星巴克的实战经历阐述了这一框架的应用。快餐与饮品行业（QSR）面临普遍的咖啡师离职率挑战，新员工入职时往往难以在短时间内熟记繁复的操作规范、海量的个性化饮品配方以及各类复杂设备的故障排查步骤，承受着巨大的认知负荷（Cognitive Load）。星巴克当时的业务目标非常明确：通过智能化工具简化操作流程、加速新员工上手，从而确保门店交付标准一致且高品质的体验。 该项目携手微软，依托 OpenAI 底层大模型搭建知识辅助系统。团队从极小范围的功能和受限门店切入，根据咖啡师的真实反馈不断微调模型、校准回答准确率，随后稳步推进推广路径：从最初的单个试验门店，逐步扩展至 300 家店、数千家店，并最终于去年 11 月全量推向全美 11,000 家星巴克门店，实现了跨越式规模化。 ```mermaid flowchart TD accTitle: AI 项目从构想到规模化推广的三阶段评估门禁 subgraph G1["阶段一：业务契合度"] Q1{"业务场景与 ROI<br/>是否明确？"} end subgraph G2["阶段二：技术可行性"] Q2{"技术成熟度能否<br/>支撑目标规模？"} end subgraph G3["阶段三：组织与采纳"] Q3{"用户是否有采纳意愿<br/>且有变革推动者？"} end subgraph G4["执行阶段：分级规模化推广"] P1["PoC 验证"] --> P2["受限试点 (小范围门店)"] P2 --> P3["持续调优模型与准确率"] P3 --> P4["全量企业级部署 (11,000+ 门店)"] end Q1 -- 是 --> Q2 Q2 -- 是 --> Q3 Q3 -- 是 --> P1 ``` - **认知负荷减负**：通过 AI 助手将配方查询、设备排障等流程极简化，让咖啡师能将精力重新聚焦于顾客连接。 - **阶梯式推进路径**：严格遵循“概念验证 $\rightarrow$ 局部试点 $\rightarrow$ 准确率微调 $\rightarrow$ 分批放量（单店 $\rightarrow$ 300店 $\rightarrow$ 最终万店全量）”的节奏。 --- ### [09:16](https://www.youtube.com/watch?v=yy14kGsCBd0&t=556s) 家得宝“Magic Apron”购物助手的落地实践 Ningyu Chen 随后分享了家得宝在面向消费者场景中的代表作——AI 购物助手“Magic Apron”（魔法围裙）。家装零售领域具有极高的专业壁垒，不同于沃尔玛或塔吉特等综合商超，家得宝门店内陈列着海量复杂的建材与五金工具。除了经验丰富的专业承包商（Pros）或资深 DIY 爱好者，普通消费者甚至 Ningyu Chen 本人走进门店都会感到无所适从。此外，家得宝线上电商平台拥有超过 300 万个 SKU，商品发现与选购决策的门槛极高。 为了解决这一痛点，家得宝启动了“Magic Apron”项目，旨在帮助顾客大幅简化商品发现流程，建立选品信心并顺利完成采购。在启动初期，团队就清醒地认识到该场景未来将承载全企业级的影响力，因此从一开始就着手打造统一的底层支撑平台，系统性地夯实了准确性、运行性能、成本控制和安全合规等关键能力。 在完成试点与局部灰度部署后，“Magic Apron”已全面上线至家得宝官网（.com）及官方移动端 App；不仅如此，该能力还被集成到移动 App 的“店内模式”（In-Store Mode）中，让顾客在实体门店穿梭选购时，也能实时获得智能导购与专业指引。 - **300 万 SKU 发现痛点**：家装商品专业度高、分类复杂，传统搜索难以满足非专业用户的模糊意图表达。 - **全渠道融合协同**：从线上电商平台延伸至线下实体门店场景，通过统一的底层技术平台实现线上线下无缝体验衔接。 ### [10:49](https://www.youtube.com/watch?v=yy14kGsCBd0&t=649s) 平台的“滚雪球效应”：从专业客户蓝图算量到全流程呼叫中心 在企业级 AI 转型中，仅完成单一场景的验证绝非终点。要实现持久且具变革性的成功，必须规划好“第一步与第二步”。在底层核心平台搭建就绪后，企业应当不断向其叠加新的业务用例，形成 Ningyu Chen 所称的“滚雪球效应”（Snowball Effect）——在取得第一个全垒打后，紧接着打出第二个、第三个全垒打。基于此，家得宝（The Home Depot）将 AI 能力从普通 DIY 散客进一步拓展至复杂的专业客户（Pro）。与 DIY 客户“只为修补特定物件、寻找单个零件”的简单诉求不同，专业承包商（Pro）面临的是盖房子、旧房整建或开发整栋公寓楼等庞大工程，需求极为繁复。 针对专业客户的痛点，家得宝推出了名为“蓝图算量提取”（Blueprint Takeoff）的 AI 功能。该系统能够直接读取住宅建筑蓝图，并自动将其解析转化为包含材料品类、采购数量、价格预估与实时库存的主清单（Master Sheet），极大简化了工程采购的前期规划。不仅如此，团队还将该能力进一步向售后环节延伸。家得宝的客户旅程并不会在付款那一刻终止，后续涉及大量送货至工地、临时修改订购数量、查漏补缺增购材料等售后服务。通过将该平台与呼叫中心系统全面打通，家得宝构建了一套基于统一底座生长出来的完整客户体验生态系统。 - **DIY 与 Pro 场景差异**：DIY 客户重在单品检索与快速维修；Pro 客户涉及整屋建造或公寓楼开发，需要多品类、大批量的复杂物料协同。 - **蓝图算量提取（Blueprint Takeoff）**：自动化解析施工图纸，输出涵盖物料清单、数量计算、价格核算与库存状态的项目采购总表。 - **售后生态联动**：将蓝图与采购数据同步给呼叫中心，高效支撑工地配送变更、增补物料等售后高频变动需求。 --- ### [12:42](https://www.youtube.com/watch?v=yy14kGsCBd0&t=762s) 中小企业的破局之道：以自有专有数据构筑核心护城河 针对缺乏家得宝海量历史客户数据沉淀的中小企业或初创团队，Ningyu Chen 给出了明确建议：最核心的原则在于必须找到采集与沉淀“自身专有数据”（Proprietary Data）的途径。在 AI 技术演进的终局中，基础大语言模型、底层基础设施甚至编排工具（Orchestration）都会高度商品化，未来获取成本将趋近于极低乃至免费，真正能拉开企业间差距的唯有专属于自身的业务知识与私有数据资产。 在实际落地中，中小企业即便选择与初创公司（Startups）或第三方供应商合作开发用例，也必须在合作架构与合同条款中明确数据主权，确保业务沉淀下来的数据完全归企业自身所有，而不是拱手让给合作厂商。只有持续积累这些专有业务认知，企业才能建立起坚实的技术与商业壁垒。 - **核心差异化资产**：模型算力与编排框架终将同质化，企业的私有知识库与专有数据才是不可替代的核心壁垒。 - **数据所有权控制**：采用外部 Vendor 或 Startup 工具时，架构设计必须保证自身对生成和处理数据的绝对所有权。 --- ### [13:51](https://www.youtube.com/watch?v=yy14kGsCBd0&t=831s) 家得宝的 AI 三层心智模型与统一安全治理 面对个人、部门与企业级应用不同维度的落地需求，家得宝在内部推行了一套清晰的“三层心智模型”（Three Buckets），对不同维度的用例采取分级建设与评估标准。第一层是**个人生产力（Personal Productivity）**，旨在提升全员的“AI 素养”（AI Literacy），通过推广 Microsoft Copilot 或 Google Gemini 等通用工具，让员工在日常办公中习惯与 AI 协作；这一层不强制考核严苛的投资回报率（ROI），重心在于拉高组织整体技能基线。第二层是**团队生产力（Team Productivity）**，聚焦于部门内部或跨团队的明确业务目标自动化（如财务月结、采购流程等），主要依靠低代码/无代码（Low-Code/No-Code）平台赋能业务部门的“公民开发者”（Citizen Developers）自主搭建流程。 第三层则是**企业级生产力与增长（Enterprise Productivity & Growth）**，直接作用于供应链核心调度、财务主系统或驱动营收增长的客户体验平台。此类系统必须构建在极其稳健、经过严密架构设计的企业级底座之上，严密平衡成本效益、系统性能、高并发扩展性与安全性。贯穿这三个层级的核心基石则是**统一的 AI 治理（Governance）**，确保管理层能够全面监控和调度所有已部署的个人 Agent、团队 Agent 及企业级 Agent，使所有智能化动作均在安全可控的环境中运行。 ```mermaid flowchart TD accTitle: 家得宝 AI 三层应用分级与底层治理架构 accDescr: 展示个人生产力、团队生产力、企业级系统三层划分及其共同依赖的安全治理底座。 subgraph Layers["AI 应用落地三层模型"] L1["【个人生产力】\n提升 AI 素养 (Copilot / Gemini)\n不设严苛 ROI 考核"] L2["【团队生产力】\n业务流程自动化 (财务结账/采购)\n低代码平台 + 公民开发者"] L3["【企业级生产力与增长】\n核心供应链 / 财务 / 客户体验\n高可用、低成本、高扩展架构"] end subgraph Gov["统一底座"] G["【底层统一治理 (Governance)】\n监控与管理所有个人、团队与企业级 Agent | 确保数据与合规安全"] end L1 --> G L2 --> G L3 --> G ``` --- ### [16:05](https://www.youtube.com/watch?v=yy14kGsCBd0&t=965s) 开源闭源之争：构建模型抽象层以对抗技术不确定性 在探讨开源权重模型（Open-weight）与闭源商业工具的选择时，Ningyu Chen 强调团队必须始终“从业务问题出发逆向推导”，而非从技术选型出发。根据具体业务场景的技术可行性、成本开销及性能指标，综合决定采用开源还是闭源方案。 在架构设计上，他强烈建议企业在所有模型和供应商接口前端搭建一层**模型抽象层（Abstraction Layer）**。底层 AI 技术更迭极其迅速，“今天最好的模型明天可能就落后了”。通过构建抽象层解耦业务逻辑与底层模型，企业能够拥有极强的议价权与容灾弹性；一旦某个供应商面临倒闭风险、服务中断或出现重大技术瓶颈，工程团队可以在一天或一周内无缝平滑切换到底层替代平台，而无需推倒重构上层业务代码。 - **逆向推导原则**：从业务痛点定义技术可行性，综合考量 Cost、Performance 及安全指标后再做模型选型。 - **构建模型抽象层**：屏蔽底层开源与闭源模型的接口差异，确保系统具备在数天乃至一天内快速切换模型供应商的敏捷迁移能力。 --- ### [17:21](https://www.youtube.com/watch?v=yy14kGsCBd0&t=1041s) 难以攻克的领域：语音对话 AI 的挑战与十亿美元机遇 当被问及是否有任何 AI 项目在当前看来“永远无法走出试点阶段”时，Ningyu Chen 笑称在当前的技术爆发期“永远不要轻言不可能”（"Never say never"）。但他坦言，自己长期深入思考且当前极难彻底攻克的方向是**语音对话式 AI（Voice Conversational AI）**。 早在 2019 年，他便在快餐连锁（QSR）行业探索利用语音 AI 实现得来速点餐自动化、削减门店人力成本并进行智能交叉销售。然而即使在生成式 AI 大幅演进的当下，语音对话场景依然极难做到尽善尽美——真实世界中的环境噪音、各地口音与方言、千奇百怪的说话方式以及人机交互时的微妙心理差异，给系统带来了极高的复杂度与长尾异常。Ningyu Chen 直言，尽管门槛极高，但“如果有人能真正攻克这一难题，那将是一个十亿美元级别的巨大商业机遇”。对话最后，主持人 Jennifer Conrad 感谢 Ningyu Chen 的实战经验分享，对谈在现场观众的掌声中圆满结束。 - **历史探索与痛点**：2019 年在快餐行业（QSR）尝试语音点餐自动化与交叉销售，长尾变体、噪声和交互差异导致落地极难。 - **百亿商业前景**：真实复杂环境下的全双工语音交互一旦被完美攻克，将彻底重塑点餐与线下服务业态，蕴含极高的商业价值。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=yy14kGsCBd0&t=4s) 通过技术变革提升客户体验、一线员工体验以及整体业务运营效率 — Ningyu Chen > [00:04](https://www.youtube.com/watch?v=yy14kGsCBd0&t=4s) 若不经历 A 点与 B 点，你绝不可能直接跳到 C 点 — Ningyu Chen > [03:15](https://www.youtube.com/watch?v=yy14kGsCBd0&t=195s) 为差异化而自研，为通用对齐而采购 — Ningyu Chen > [10:49](https://www.youtube.com/watch?v=yy14kGsCBd0&t=649s) 在取得第一个全垒打后，紧接着打出第二个、第三个全垒打 — Ningyu Chen > [12:42](https://www.youtube.com/watch?v=yy14kGsCBd0&t=762s) 基础大语言模型、底层基础设施甚至编排工具都会高度商品化，未来获取成本将趋近于极低乃至免费，真正能拉开企业间差距的唯有专属于自身的业务知识与私有数据资产 — Ningyu Chen > [16:05](https://www.youtube.com/watch?v=yy14kGsCBd0&t=965s) 从业务问题出发逆向推导，而非从技术选型出发 — Ningyu Chen > [16:05](https://www.youtube.com/watch?v=yy14kGsCBd0&t=965s) 今天最好的模型明天可能就落后了 — Ningyu Chen > [17:21](https://www.youtube.com/watch?v=yy14kGsCBd0&t=1041s) Never say never — Ningyu Chen > [17:21](https://www.youtube.com/watch?v=yy14kGsCBd0&t=1041s) 如果有人能真正攻克这一难题，那将是一个十亿美元级别的巨大商业机遇 — Ningyu Chen ## Mentioned The Home Depot, Ningyu Chen, Jennifer Conrad, Inc. Magazine, Ai4, Delta Air Lines, Macy's, Inspire Brands, Starbucks, ChatGPT, OpenAI, Anthropic, Google, Microsoft, Magic Apron, Walmart, Target, Blueprint Takeoff, Microsoft Copilot, Google Gemini ## Video description Ningyu Chen - SVP Technology, The Home Depot Jennifer Conrad - Senior Writer, Inc. Magazine Most AI pilots don't fail because of the technology. They stall somewhere between proof of concept and enterprise adoption. Join me for an honest conversation about what it really takes to move AI from promising pilots to enterprise-wide value. I'll share lessons from my own journey, what worked, what didn't, and the framework I use to identify, develop, and deploy AI use cases at scale. We'll explore the critical success factors that separate isolated wins from sustainable transformation. Every AI journey is different, but the lessons are often universal. I hope you'll leave with new ideas, fresh perspectives, and a clearer path to impact. [Source](https://www.youtube.com/watch?v=yy14kGsCBd0)

[0:04] Ningyu Chen: 通过技术变革提升客户体验、一线员工体验以及整体业务运营效率 [0:04] Ningyu Chen: 若不经历 A 点与 B 点，你绝不可能直接跳到 C 点 [3:15] Ningyu Chen: 为差异化而自研，为通用对齐而采购 [10:49] Ningyu Chen: 在取得第一个全垒打后，紧接着打出第二个、第三个全垒打 [12:42] Ningyu Chen: 基础大语言模型、底层基础设施甚至编排工具都会高度商品化，未来获取成本将趋近于极低乃至免费，真正能拉开企业间差距的唯有专属于自身的业务知识与私有数据资产 [16:05] Ningyu Chen: 从业务问题出发逆向推导，而非从技术选型出发 [16:05] Ningyu Chen: 今天最好的模型明天可能就落后了 [17:21] Ningyu Chen: Never say never [17:21] Ningyu Chen: 如果有人能真正攻克这一难题，那将是一个十亿美元级别的巨大商业机遇

Ningyu Chen: Most AI pilots don't fail because of the technology. They stall somewhere between proof of concept and enterprise adoption. Ningyu Chen: Every AI journey is different, but the lessons are often universal. [0:04] Ningyu Chen: driving technology change to elevate the customer experience and the frontline employee experience, and also the overall business operations. [0:04] Ningyu Chen: You cannot get to Point C without going through Point A and Point B. [3:15] Ningyu Chen: Build for differentiation, buy for parity. [9:16] Ningyu Chen: snowball effect [12:42] Ningyu Chen: You must find a way to harvest your own data. [16:07] Ningyu Chen: never say never [17:49] Ningyu Chen: a billion-dollar opportunity

[0:04] I'm Jennifer Conrad. I'm a senior writer with  Inc. Magazine. I have Ning Yu Chen with me who [0:09] is the senior vice president for technology at  the Home Depot. And as you heard, we're going [0:15] to talk about how to scale AI pilots and when  to do it, when maybe it's not a great idea. [0:22] So, um, to start, I would love to hear if you  could talk us through your background because I [0:26] know prior to Home Depot, you had very senior  roles at Starbucks, Inspire Brands, Macy's, [0:31] Delta. So, talk us through that and how that  helps you develop a model for when to scale AI, [0:37] when to use AI. Great. So, I would say I I've  been very fortunate to be part of the business [0:43] transformation of these many iconic brands. Um  if I summarize my career for those brands I would [0:49] say one sentence is about driving technology  change u to elevate the customer experience and [0:56] the frontline you know employee experience  and also the overall business operations [1:01] um there's actually distinctive three kind of era  of transformation so remember early days in delta [1:08] um that kind of transformation is about building  out real time enterprise so what that means is by [1:15] is to creating this business domain data being  real time at scale. So think about flight, the [1:20] passenger, the bag, the crew be able to have that  in real time at scale that become the foundation [1:27] for all the elevated experience you experience  today in your travel. Um and we move on to the [1:32] um you know retail world uh Macy and later QS  industry uh inspired brands is about digital [1:39] transformation but it was built on the real time  information because without that you cannot do [1:45] really good digital transformation. So there we  are digitizing the merchandising digitizing supply [1:50] chain digitizing the store to create omni channel  experience for our customer. Now fast forward I [1:56] move on to the uh Starbucks and later uh Home  Depot. That was the right time of the the Chai [2:02] Gibb come out and uh we basically into the era of  intelligent uh enterprise uh transformation and [2:10] in this case really about context about intent  and how you personalize experience for these [2:16] uh customers and associates. Um I would say the  framework I kind of generate over the years are [2:22] really two two kind of key points. One is you  cannot get to point C without going through point [2:28] A and point B. What we mean by that is that if you  say hey I want to do a transformation today. You [2:33] need a foundation. The foundation is a real-time  information data at scale then you go digitization [2:39] on top of it then you can do AI transformation  in a right way. Right? So that's number one. [2:44] Number two is what has been constant throughout  this era of transformation is you have to think [2:49] about a platform an enterprise platform that is  really had a characteristic of accuracy speed [2:57] um scale and security and cost effectiveness  because without that enterprise platform in place [3:04] build out mature enough to be able to empower  all the use cases your transformation not going [3:09] to be sustainable and cannot last long. So you  talked about the things that kind of stayed the [3:15] same. So you're looking at performance, accuracy,  whether things can scale, how much they cost. What [3:20] are some things that have changed in recent year  years? And specifically when you're looking at AI, [3:24] how are you maybe looking at things differently  now as far as build versus buy and things [3:29] like that? Yeah. So that you have two question  there. So let me start with question number one, [3:32] what changed? I would say the biggest change  variable is speed. If you think about in the [3:38] in the era earlier the the the data transformation  and digital transformation you have years months [3:44] to kind of adjust contemplate and go after  sometimes size matters if you're a big company [3:50] you do have an advantage but I think in the era of  AI I think speed is a dominant factor of success [3:57] it doesn't matter you're small you're big if you  have the right idea you with the right speed you [4:01] will come out and that's one question so the other  question is about build versus buy right So, so [4:08] the the way to think about this is actually a very  interesting question. So, again, I have a mental [4:11] model very simple uh kind of philosophy. I think  you many of you probably heard about this. It's [4:16] a it's a beautiful differentiation and buy for  parody. So, what did it mean in the AI world? So, [4:21] if you think about what's most important  for your business uh in Home Depot case, [4:26] it will be the customer experience, the frontline  associate experience in the store. And we would [4:31] never outsource that piece to others because we  want to control our destiny, right? So that's [4:35] piece you want to buy, you want to build and also  on top of that Home Depot has a large scale like [4:40] if it's a fortune 25 company scale complexity  means that very few vendor can kind of handle [4:47] the kind of uh complexity and scale. So that's on  that front. Now on the other hand if you look at [4:51] HR look at finance yeah that's a commodity you can  you can find something that is scalable use it and [4:59] um now there's third piece to it which is kind  of what lean towards called AI age difference the [5:05] partnership going to be important the partnership  has multiple multiple factors so sometimes you [5:10] build yourself it's not fast enough so you want to  partner with somebody build together or build upon [5:15] uh their foundation so in in the Home Depot case  we partner with all the frontier here um you know [5:21] companies so anthropic openai Google and Microsoft  using their latest tools and also the model as our [5:28] foundation layer right so that's that's one part  and we we supply our expertise knowledge data [5:34] um and all the kind of transaction in the past to  help create specific around home depot model now [5:40] that's one side of the other side of partnership  is really with startups these days I think startup [5:45] a lot of great ideas you have the scenario  of starter one and maybe come with a million [5:50] billion dollar company right you want to identify  this kind of company early on either co-invest or [5:55] you is a is a you know buying out that particular  company so you can have the IP as part of that [6:01] um you know build of your own kind of piece going  forward um and I'd love to sort of step back [6:07] here and see could you tell us h about a recent  project that you've done and how you've taken it [6:12] from pilot to scaling it across the organization  yeah I can give you two how about a bonus So one [6:19] example is from my era of uh at the Starbucks. So  it's about uh frontline associate experience. The [6:26] example will be with Home Depot is more customer  experience. Now before I kind of talk about [6:30] example, let me first give you some kind of mental  model in terms of how we think about how you go [6:35] from you know idea to pilot to scale right. So um  think about more like stage process. So the number [6:41] one step is hey do we have the right use case to  start with? What I mean by that is do we have the [6:47] right business context? Do we have the clear ROI  associated with it? Do we know what the business [6:52] outcome look like? What success look like? So  that need to be done first. Now next step will [6:58] be okay is a technical feasible. Is a technology  mature enough to support a use case. The use case [7:03] could be a very departmental use case or could  be very kind of enterprise scale use case. Right? [7:08] So depend on that you have to think about the  maturity of the technology for that scenario. [7:13] If you pass that stage, the stage number three  is do we have enthusiastic adoption at the you [7:19] know audience level whether it's a customer  or whether associate employee uh do we have [7:23] a business champion uh to support that change  management because anything we introduce new [7:28] into the space require change management so if  you got that three stage done then you can say [7:33] okay let's start start with you know PC pilot  and then limited scale and then you know scale [7:40] uh you know companywide so the two example kind  of walk you through it. The number one is around [7:45] um a Starbucks uh space. So the problem they are  dealing with is really around store operations. [7:51] So if you think about what really makes Starbucks  experience shine is the store experience and if [7:58] you think about who's there is the barista and in  the QR industry there's a lot of turnovers in in [8:04] the barista. No Starbucks has the best turnover  among all the in terms of the loyalty but still [8:09] you have turnovers right. So what it means is new  bris come in they do not understand the standards [8:15] procedures uh we have a lot of at the time we  have a lot of beverages uh customized beverages [8:21] how the recipe should work and we have a lot  of machinery how to troubleshoot right so a [8:25] lot of things is taking away a lot of cognitive  load from our barista. So the business use case [8:30] at a time is to how can we simplify work to bring  people on board quickly and have a consistent kind [8:37] of performance and allow us to deliver it again  elevated differentiated store experience right [8:44] so that was use case and we started with I know we  partner with Microsan um we leverage open AI model [8:50] and put on the gallery around this and again  we start with very small kind of area limited [8:56] user use cases and then limited stores and get a  feedback from the uh the the people over there. [9:03] Then we finetuning make sure accuracy. Then we  go to one store, 300 store, final store and then [9:08] chai 11,000 store last November. So that was one  example and the other example is uh Home Depot. So [9:16] this is around customers. So if you think about  Home Depot is a very unique kind of space. It's [9:21] a home improvement market. It's not general not  like a Walmart, Target. can walk in, you know, [9:26] relatively the product knowledge around what you  want to buy, you know, but if you walk into Home [9:30] Depot, if you're seasoned DIYs, you know, pros we  call, then you know what you're doing. But many [9:36] of us don't. Like personally, I don't I'm not a  handyman. I'm intimidated by that store because of [9:42] all the sheer size of the product offering, right?  So, how do you as a customer be able to find what [9:48] you needed and confident enough to know walk away  with it? So we come up with this idea of magic [9:53] apron which basically is a shopping assistant. If  you think about our com we have three million SKs [9:58] in the docom space there's a lot of uh product  to discover with. So we started with the use [10:04] case of discovery how we simplify discovery for  our customer. Um and again like I said earlier [10:10] you have to think about platforms where you start  about a use case knowing that use case potentially [10:14] be able to have enterprisewide implication.  So you have to think about platforms. So in [10:19] this case we build out the platform gradually you  know adding all the pieces the accuracy the the [10:25] um the performance the the cost efficiency uh  around all the things security right so you got [10:31] go that going and then and that's where we started  with pilot and then we go limited um deployment [10:38] now it's in all stores I'm sorry not all.com and  the app we also introduced into the stores so you [10:44] as you walk in the store you have an in store  mode you can use that as well now we have not [10:49] not stop there because again like I said for you  to really have a transformation are really going [10:54] to be lasting and successful you have to think  about step one step two so in this case we have [10:59] a platform in place now what a use case we can  pile onto it so this is what I call the snowball [11:04] effect you got a win going you got a home run  let's make a second home run the third home run [11:08] right so and later this year early this year later  last year we introduced uh for pro so think about [11:17] how you plan your project pro is different from  DIYs. DIY is like just I have this thing to fix. [11:23] I need to buy a particular piece. I need to find  a particular piece. But pro is more more complex. [11:28] They are building a house. They're renovating  a house. They're building a entire uh apartment [11:33] block. Right? So so they have a different needs.  So we build this um we call blueprint takeoff. So [11:40] you take a residential blueprint and we be able  to take that blueprint and convert it into a set [11:46] of material quantity price inventory and that  become the list of purchase for the pro and we [11:54] are creating this we call basically the the master  sheet if you will for all the boutique materials [11:59] for these projects that's the kind of the next  step now we take that further to say okay how [12:05] we think about in our call center experience like  a lot of times home depot experience is not stop [12:11] at purchase. You have a lot of post purchase work  associated with it, right? You have a lot of thing [12:16] to deliver to your uh to your house or to your  site. You have you may want to change quantities. [12:22] You may find that you're missing something.  You need to add to it. So, how do we create [12:26] that kind of experience? So, so we line up with  our call center and use that as part of that call [12:31] center experience. So now you have an ecosystem  of a customer experience that build up on one [12:37] platform had a snowy effect right from that point.  Yeah. I mean it's pretty incredible what you guys [12:42] have created. Um I wonder if you could advise  some of the um businesses in the room who don't [12:48] have that same wealth of data that and years of  customer data that Home Depot has to draw on. What [12:53] sort of tools can they be looking at? How can they  build their own customer service bots and customer [12:59] service tools without that depth? That's a good  question. So if you look at from that perspective, [13:04] you know, so for a small mediumsiz company or  something they're just starting up. Um I would say [13:10] number one thing have a key principle in mind. You  must find a way harvesting your own data because [13:16] the thing that differential in the end is going to  be your proprietary knowledge. The language model [13:21] doesn't matter the the infrastructure doesn't  matter. The even orchestration doesn't matter. [13:25] You can you can get all those things pretty much  free down the road, right? So so what different is [13:30] your data. So you have to find a way to harvesting  papar data. So if you started use a use case you [13:36] can work with some kind of Starbucks or or some  kind of a vendor make sure you structure such as [13:41] that you can you can own your data versus the  data gave to that particular company that you [13:46] start accumulating this knowledge and become a  foundation point. That makes sense. Um, and one [13:52] thing that I hear from a lot of business owners  and people who are rolling out these tools is that [13:58] they have different standards for personal versus  maybe a departmentwide tool versus something that [14:04] consumers are going to touch. So, how do you  think through what the sort of standards are [14:08] for each one? Yeah, that's a great question.  So, at Home Depot, we have this mental model [14:13] uh to really segment the use case into kind of  three buckets, if you will. The bucket number one [14:19] is we call personal productivity. So this is about  really raised kind of bar in terms of AI literacy [14:26] uh for person and know who work in that company.  Um and in the case you using Microsoft copilot you [14:32] you use the Google uh pro uh Gemini pro right  the goal is to understand how to use the AI to [14:38] effectively as your day-to-day work. There's no  II behind it. It's really about literacy razor [14:43] bar right then you have a second kind of segment  which is I call the team productivity and think [14:49] about the things that you want to do within a  department or across several teams there's a very [14:54] clear business outcome you want to accomplish  like I want to automate some kind of process [14:58] whether it's a accounting closing or some kind of  a procurement right so you have a outcome in place [15:04] in this case you really need find something that  is low uh local no code platform so our citizen [15:10] developer uh in those B department can really use  that platform to build out that process. And then [15:16] the third category is the enterprise cate category  is where it's either enterprise productivity right [15:21] whether with the supply chain side or whe whether  it's a finance side or it's a really creating the [15:27] revenue growth in a customer experience side in  this case back to what I already said you really [15:32] need a robust platform ideally it's something  you build upon something foundationally good [15:38] right whether it's a has a cost performance  security right a scale implication in place [15:44] Um and that's how we think about this and how  we kind of roll our stuff. Now across all these [15:50] there's a one piece kind of underlying across all  these is a governance is knowing what agents you [15:56] deploy whether you're sing personal agent team  agent enterprise agent be able to managing that [16:01] uh in in a safe environment. Got it. Um and on  that note there's been a lot of talk in the ether [16:07] about openw weight models open source tools. Um  what's your what's your take? Um I would say we [16:14] always want to start from the problem we're going  to solve what business problem want to solve and [16:19] go backward. Um and depend on what that is you  again like like mentioned earlier the framework [16:24] right you start with the B problem then look at  the techn technology feasibility associated with [16:29] the problem is it tools from closed source versus  open source or is it and look at the criteria [16:36] around you know cost performance so on so forth  you select on there I would say the one thing I [16:41] would say always going to be uh important is I  always suggest you create abstraction layer in [16:47] front of all these vendor related or whether open  open weight or closed weight models because you [16:53] will know it's moving so fast to what's good today  may not be good tomorrow. So you want to be able [16:58] to flip uh back and forth. I think I would say  one of the the goal I would I would say for our [17:05] um you know whoever doing the work going forward  is that if you can stay on one platform and [17:10] tomorrow if that platform either go bankrupt or  has uh you know challenge issues you'll be able to [17:15] flip to a different platform within a day or week.  you need to building that kind of model. Um, and [17:21] obviously you've led a lot of projects, done a lot  of things. Is there anything that you think right [17:28] now is never going to make it out of pilot or  is there anything that makes you think, oh, this [17:31] is something that should not should not um move  forward? Oh, that's a tough question because never [17:37] say never. [laughter] I was with the technology we  have right now. I would say I would tell you this [17:44] one. the one that I've been contemplating quite a  bit and it's going to be a tricky one. Um it's the [17:49] voice uh conversation conversational AI um back  to my early days um you know 2019 we start that [17:57] journey already with the voice AI for QSI if you  think about the place you want in the QS industry [18:03] where you can take out labor uh is to leveraging  send out the you know kind of automation right [18:08] taking the order from customer be able to crossell  from the point it's a great use case it is so [18:15] difficult to crack even today with a virus of AI  I and just so many kind of uh variations you have [18:23] to deal with um and the people how they how you  interact with uh the brands different so it this [18:30] is the one I would say if somebody crack it it  is a billion dollar opportunity all right maybe [18:34] that's a challenge for the room okay we have  12 seconds left I think we did that perfectly [18:39] thank you all for joining us and thank you Mimi  Chen for your very helpful answers thank you.