Enterprise AI Chiefs Ditch Data Overhauls as Inference Costs Bite

推理成本重压下，企业AI主管放弃全面数据重构 大语言模型高昂的推理成本打破了传统软件的经济模型，迫使企业放弃耗时数年的数据清洗项目，转向通过严格指标评估来锁定高价值业务工作流。 专家组梳理了一条企业落地生成式AI的务实路线图，建议各机构从内部工作流着手，精确核算实际运行中的Token成本，并建立跨职能治理委员会。与会嘉宾质疑了“企业在部署AI前必须耗时数年清理数据湖”的假设，转而倡导采用就地检索、排序架构以及持续的智能体（Agent）评估框架。讨论中，发言者将AI定位为赋能增效的工具而非替代人力的手段，以缓解职场焦虑，并指出AI将在提升一线员工能力的同时催生全新的运营专业分工。 企业在引入生成式 AI 时应“从小处着手”，优先在沙盒环境中探索并验证内部员工用例，将其定位为生产力赋能工具，而非过早推向面向客户的外部场景。 (3:58) 西北相互（Northwestern Mutual）等大型企业的实践表明，设立横跨高管、法务、风险合规与变革管理的 AI 治理委员会，是统筹机构风险与对齐多年技术投资的关键保障。 (8:37) 客户支持是高回报的 AI 切入点，B2B 领域单次工单高达 40 至 60 美元的成本与完善的基准指标易于证明 ROI，且高自动化率正在催生负责治理和调优模型的“AI Agent 运维”新职能。 (16:48) AI 落地成功取决于能否直接嵌入 Slack、Zendesk 等员工日常工作流，同时企业应果断淘汰仍固守并发席位许可或缺乏现代 API 支持的落后供应商。 (19:51) 单凭 Token 计价测算无法反映真实的 TCO 总拥有成本，且企业无需等待耗时数年的全量数据治理，将遗留系统作为搜索与排序问题处理即可加速 AI 落地验证。 (25:47)

Enterprise generative AI adoption succeeds only when organizations prioritize high-ROI operational workflows, treat legacy data as an in-place search problem, and anchor rollouts in human-centered change management, according to leaders from Discount Tire, Northwestern Mutual, Maven AGI, and TxMinds. Managing the New Economics of LLM Inference (12:58) Sami Shalabi observes that generative AI fundamentally altered software economics because continuous model inference introduces heavy, recurring infrastructure costs that do not exist with conventional, near-zero-cost database queries. Enterprise teams must therefore impose strict economic filters, targeting dense operational pain points rather than automating marginal tasks. In customer support, for instance, standard B2B tickets cost $40 to $60, while complex engineering escalations can reach $1,000, creating an extensive historical paper trail of CSAT and resolution metrics where automated resolution delivers a clear, verifiable return on investment. Rajesh Krishna Bharathi emphasizes that Discount Tire prevents runaway costs by screening every proposed use case against revenue potential, cost savings, security impact, and whether AI is genuinely necessary. To avoid the misleading projections of paper token estimates, teams must map process flows end-to-end, account for every architecture line as an operational cost, and validate total cost of ownership through small sandbox proofs of concept before expanding to customer-facing channels. Replacing Data Lake Cleanups with Search Architectures (29:14) Sushma Niwalkar reports that the traditional enterprise assumption—where organizations must aggregate and clean all enterprise data into a pristine central lake before attempting AI deployments—has been discarded. Instead, targeted AI applications are now deployed concurrently with data engineering, using early vertical implementations to surface real operational deficiencies and actively shape corporate data strategy. Addressing monolithic legacy environments, Shalabi argues that data integration should be approached as an advanced search and ranking challenge rather than an ingestion bottleneck. Operating on an in-place 'graph of record,' systems can leave data within existing source systems and apply deterministic heuristics—such as document freshness timestamps—to arbitrate between conflicting documentation. This architecture transforms non-deterministic model outputs into verifiable responses by maintaining strict source provenance. Michael Giacometti adds that practitioners should serve as internal change agents, linking isolated departmental pockets of data to satisfy concrete operational needs without waiting for multi-year infrastructure overhauls. Establishing Empirical Metrics and Governance Councils (8:37) Northwestern Mutual managed enterprise risk by instituting a dedicated AI governance council that unites executive leadership, legal, compliance, risk, and change management. When deploying an internal knowledge assistant for financial advisors, baseline accuracy began at 60%, requiring sustained iteration and heavy subject matter expert involvement to reach 90% precision. Niwalkar stresses that risk mitigation and executive upskilling must accompany technical development from day one. Shalabi points out that without rigorous evaluation frameworks, organizations fly blind on subjective vendor claims. By implementing automated, agentic side-by-side evaluations, organizations can translate ambiguous performance targets into exact engineering requirements—demonstrating, for example, that lifting resolution accuracy from 73.5% to 90.6% requires integration with only three specific systems and field records. These empirical benchmarks deliver the concrete security and cost justification required to secure access from internal IT departments. Driving Adoption through Embedded Workflows and Augmentation (19:51) Panelists agree that technological capability cannot overcome poor user experience. Shalabi notes that forcing staff into external standalone portals stalls adoption; generative AI must instead be integrated into daily tools such as Slack, Microsoft Teams, Zendesk, or Salesforce. At commercial bank Rho, embedding a support bot directly within Slack led sales teams to organically adopt the tool for pre-sales technical research, demonstrating how native workflow integration dissolves traditional departmental silos. To overcome worker anxiety regarding job displacement, leaders must frame AI around task augmentation rather than head-count reduction. Niwalkar reports that Northwestern Mutual experienced zero AI-driven layoffs, instead transforming early advisor resistance into advocacy by demonstrating that AI removed two to three hours of manual administrative burden. Bharathi and Giacometti conclude that while repetitive tasks will automate away, organizations must continuously upskill frontline workers into emerging supervisory specializations such as AI agent operations. The panel emphasizes that long-term generative AI viability depends on aligning executive incentives with operational realities, ruthlessly pruning legacy vendors whose restrictive licensing or sluggish APIs impede integration, and maintaining active peer learning communities to navigate complex organizational change.

企业将生成式AI融入现有系统不仅是技术升级，更是一场涉及经济核算、数据架构与组织转型的系统重构。来自零售、金融和AI服务领域的从业者指出，企业应摒弃“等待数据完全理顺”的传统思维，优先切入高ROI的内部工作流，并通过建立严苛的评估框架与人机协同治理，实现从概念验证到企业级规模化生产的平稳过渡。 重塑软件经济学与小步快跑的落地路径 (3:58) Discount Tire人工智能卓越中心负责人Rajesh Krishna Bharathi主张企业在部署生成式AI时必须采取渐进式策略，坚持“从小处着手”。企业应优先锁定内部员工的使用场景，而非急于推向外部客户。在技术流程上，团队需预先绘制包含操作主体、AI交互层及业务产出的端到端流程图，并在沙盒环境中进行快速概念验证（PoC），以在解决模型一致性和外部交互前验证实际业务价值。Bharathi强调，纸面上单纯基于Token单价的测算具有欺骗性，系统流程图中的每一条技术链路与调用延迟都会产生成本，必须通过标准小规模部署来测算总体拥有成本（TCO）。 Maven AGI联合创始人兼首席技术官Sami Shalabi进一步指出，生成式AI本质上打破了传统软件的经济模型。在传统架构中，数据库查询和服务器调用的边际成本几乎可以忽略不计，但大语言模型（LLM）的推理成本高昂。以企业客户支持为例，B2B领域单次工单处理成本普遍在40至60美元，技术复杂的工程支持工单甚至高达1000美元。这种高成本与完备的跟踪指标（如CSAT客户满意度、首响时间）使得支持场景成为极少数ROI能够覆盖推理与算力开销的高价值落脚点。TxMinds副总裁Michael Giacometti亦警告，模型规模化带来的推理费用无法单纯依靠劳动力套利来抵消，企业实施团队必须向上管理管理层的预期，厘清效率提升与裁员指标之间的认知分歧。 跨职能治理机制与评估驱动的系统集成 (8:37) Northwestern Mutual人工智能高级总监Sushma Niwalkar分享了这家拥有168年历史的金融机构在治理层面的经验教训。该机构成立了由高管、风险合规、法律、人才发展及变革管理部门组成的AI委员会，统筹处理跨部门用例冲突与长期战略投资。在其为2.5万名财务顾问研发知识管理工具时，基线准确率最初仅为60%，通过业务领域专家（SME）深度参与人机协同循环与算法调优，最终将准确率提升至90%。Niwalkar透露，在针对45位高绩效顾问的早期测试中，由于系统偶发异常引发顾问直接向CEO投诉，暴露出变革管理缺位的严重后果，促使企业随后将高管培训与全员技能重构作为治理刚性要求。 在技术集成层面，Shalabi强调缺乏客观评估框架是企业AI落地中最普遍的短板。基于其在Google主导产品研发时投入巨资进行人工评估的经验，Maven AGI建立了基于智能体（Agentic）的自动化对比评估体系。清晰的量化指标使技术团队能够摆脱盲目集成几十个复杂遗留系统的困境，精准定位关键数据缺口。例如，系统可以通过基线分析明确指出，若要将准确率从73.5%提升至90.6%，仅需打通三个特定系统并获取特定用户记录字段。这种将指标与单次工单40美元成本直接锚定的量化依据，能以无可辩驳的商业论证促使企业IT安全部门主动开放关键系统权限。 倒置数据战略与基于搜索排序的架构解构 (35:15) 针对企业界长期存在的“数据未清理完毕便无法启动AI”的教条，Niwalkar指出当下企业认知已发生根本性逆转。传统上AI战略始终落后于数据战略，企业往往耗费数年试图将数据归整入完美的数据湖；如今，垂直向的AI实际落地反向推动并塑造了整体数据战略，实时暴露出底层真实的缺陷，倒逼数据工程团队与AI团队实现同步迭代。 面对数十年历史的单体遗留架构（Monolithic Systems），Shalabi提出应将生成式AI集成解构为“搜索与排序问题”（Search and Ranking）。Maven AGI构建的“记录图谱”（Graph of Record）允许数据保留在原始业务系统中，通过非生成式的工程启发式规则（如文档新鲜度元数据）在查询时动态筛选最新有效版本，而非耗时进行人工数据清洗。这种方法将非确定性的生成式大模型转变为具有严格溯源能力（Provenance）的确定性系统，使得生成的每一项回答均可回溯至权威业务记录，为各部门遗留系统中孤立的数据口袋（Data Pockets）提供了在不进行大规模重构前提下实现价值提取的路径。 嵌入原生工作流与重构岗位协同生态 (19:51) Shalabi结合产品管理经验强调，试图改变人类日常习惯极其困难，因此AI工具必须直接嵌入员工每天高频使用的原生界面（如Slack、Microsoft Teams、Salesforce或WhatsApp），而非强推独立的门户网站。商业银行平台Rho部署Maven AGI的案例表明，当针对Zendesk开发的客服辅助工具通过Slackbot形式延伸到跨部门协作场景后，销售团队自发将其采纳为理解复杂产品特性的核心工具，自下而上打破了支持、售前与客户成功之间的部门壁垒。 针对公众普遍担忧的AI取代人工议题，Niwalkar证实Northwestern Mutual并未因AI出现裁员，而是通过培养一线业务骨干成为“超级用户”（Power Users），以实际减负效应带动全员采纳。关于企业是否将AI作为修正2020年后科技行业过度招聘借口的观众提问，Bharathi与Giacometti回应指出，AI的发展逻辑类似五年前本地数据中心向云计算的迁移，其终极定位是基础使能工具而非人类替代品；人类在“连接上下文情境”方面的核心价值不可替代，未来的组织演进将减少低阶机械事务，转而催生专职维护、调优与监管AI系统的“AI智能体运营”（AI Agent Operations）等全新高阶岗位。 本场研讨表明，生成式AI在遗留系统中的有效集成依赖于对软件边际成本的清醒认知、搜索驱动的务实架构以及人机协同的治理机制。尽管技术变革伴随着劳动力市场重新洗牌的焦虑与过度反应，但专家一致认为，企业必须防范技术伙伴的生态落后风险，通过跨部门赋能与持续验证，促使员工完成从重复劳动者向系统变革者的角色跃迁。

## Background 该视频由行业会议频道 Ai4 于 2025 年 8 月 29 日发布在 YouTube 平台上。页面简介显示，本场圆桌研讨会汇集了来自 Northwestern Mutual、Discount Tire、TxMinds 和 Maven AGI 的多位资深高管与专家，由加州州立大学蒙特雷湾分校教授 Jill Hosmer-Jolley 主持，重点探讨如何将大语言模型与多模态生成式 AI 务实集成到现有企业系统与业务工作流中，攻克遗留系统互操作性、数据管理及规模化架构等技术与运营挑战。 ## TL;DR 本场研讨会深入剖析了企业落地生成式 AI 的务实路径，强调应从小规模内部员工场景切入，严格审视业务痛点是否真正需要 GenAI，而非盲目投入外部高风险场景。嘉宾们结合自身企业实践指出，AI 的核心定位是“赋能与增强”而非裁员替代，企业需构建涵盖风控、法务与高管赞助的跨职能治理委员会，并警惕规模化部署后剧增的 Token 推理成本。面对杂乱的企业遗留数据，专家建议打破“先花一年洗数据”的传统思维，通过构建记录图谱（Graph of Record）、引入新鲜度信号以及嵌入原生工作流（如 Slack、Teams、Salesforce），让高价值的垂直业务应用反向驱动数据治理与投资回报率（ROI）论证。 ## Atmosphere & style 整场研讨会充满务实、坦诚且极具一线实战气息的专业氛围，完全摒弃了泛泛而谈的空洞宏大叙事。嘉宾们毫不避讳地分享各自“交学费”的惨痛教训，从高净值投资顾问给 CEO 打电话投诉算法失误，到高管对“提效 30%”产生截然相反的战略误解，笑谈中满是真知灼见。现场互动尖锐而热烈，专家们在理性拆解资本市场“AI 替代人工”裁员借口的同时，展现出对技术周期律的从容洞察与对人类核心连接价值的坚定信心，让听众在接地气的经验倾囊相授中感受到深厚的技术共鸣与启发。 ## Key takeaways - [03:58](https://www.youtube.com/watch?v=gJ-wKhKhmA8&t=238s) 企业引入生成式 AI 应遵循“从小处着手”的原则，优先聚焦内部员工场景并严格自审是否真正需要 GenAI，避免盲目面向外部上线。 - [06:00](https://www.youtube.com/watch?v=gJ-wKhKhmA8&t=360s) AI 的核心定位是“增强”（Augment）而非“替代”人类，通过消除繁琐的重复性工作让员工专注高价值任务，从而缓解组织内部的失业焦虑。 - [08:31](https://www.youtube.com/watch?v=gJ-wKhKhmA8&t=511s) 大型传统企业落地 AI 需要坚实的高管赞助，并建立跨职能的 AI 治理委员会以确保用例透明度、战略对齐与跨部门成本分摊。 - [10:29](https://www.youtube.com/watch?v=gJ-wKhKhmA8&t=629s) 模型准确率的跨越式提升依赖高密度的敏捷实验与“人在回路”（Human-in-the-Loop）的业务领域专家（SME）持续校准。 - [12:58](https://www.youtube.com/watch?v=gJ-wKhKhmA8&t=778s) 生成式 AI 规模化落地面临显著的 Token 算力成本激增风险，推进者必须做好向上管理，拉齐效率提升与人员编制的预期。 - [15:55](https://www.youtube.com/watch?v=gJ-wKhKhmA8&t=955s) 生成式 AI 打破了传统软件调用的低成本经济模型，因此必须切入客户支持等单次交互成本高、ROI 易量化的高价值场景。 - [20:12](https://www.youtube.com/watch?v=gJ-wKhKhmA8&t=1212s) AI 工具必须深度原生嵌入员工现有工作流（如 Slack、Teams、Salesforce），而非构建孤立平台，以降低门槛并打破跨部门边界。 - [22:28](https://www.youtube.com/watch?v=gJ-wKhKhmA8&t=1348s) 企业应严格审视并管理供应商生态，果断淘汰跟不上技术步伐、死守传统按席位收费或缺乏高效 API 的落后合作伙伴。 - [25:49](https://www.youtube.com/watch?v=gJ-wKhKhmA8&t=1549s) 基于增收潜力、成本节约、安全态势与 AI 必要性建立严格的四维决策评估矩阵，并在 PoC 阶段输出清晰的架构与 TCO 成本模型。 - [29:14](https://www.youtube.com/watch?v=gJ-wKhKhmA8&t=1754s) 生成式 AI 打破了业务规模扩张与人力编制增长的线性依赖，将传统的客服支持等“成本中心”转型为提供高端服务的“收入中心”。 - [31:28](https://www.youtube.com/watch?v=gJ-wKhKhmA8&t=1888s) 通过构建“记录图谱”（Graph of Record）并引入新鲜度信号，直接将企业混乱分散的数据转化为搜索排序问题，绕过耗时漫长的数据清洗集中工程。 - [33:05](https://www.youtube.com/watch?v=gJ-wKhKhmA8&t=1985s) 建立基于 Agent 的量化评估框架，用确切的准确率提升与降本数据构建坚实的商业案例，以此打通 IT 系统的权限对接阻碍。 - [35:15](https://www.youtube.com/watch?v=gJ-wKhKhmA8&t=2115s) 传统“数据策略决定 AI 策略”的路径被颠覆，现代企业正通过构建具体的 AI 纵向业务应用反向驱动和发现底层数据策略的短板。 - [37:05](https://www.youtube.com/watch?v=gJ-wKhKhmA8&t=2225s) 面对历史遗留的庞大单体系统，应采用出处追溯（Provenance）机制保障生成可控，并将投入精准聚焦于评估体系指出的关键数据缺陷环节。 - [40:44](https://www.youtube.com/watch?v=gJ-wKhKhmA8&t=2444s) IT 行业每隔数年就会经历一轮技术重塑，从业者无需过度陷入 AI 取代人力的恐慌，而应主动学习并发挥人类跨领域“连点成线”的核心价值。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=gJ-wKhKhmA8&t=4s) 嘉宾介绍与小组背景 在 Ai4 大会的第三天早晨，主持人 Jill Hosmer-Jolley（加州州立大学蒙特雷湾分校教授，专注于神经多样性领导力与 AI）向参会者致谢并开场，随后邀请四位来自不同行业的重量级嘉宾依次进行自我介绍。 来自美国知名轮胎零售连锁 Discount Tire 的 Rajesh Krishna Bharathi 首先发言。他已在 Discount Tire 工作四年，此前在霍尼韦尔（Honeywell）深耕了 17 年，拥有长达 30 年的工程领导力与商业价值交付经验，目前正在公司领导 AI 卓越中心（AICE），积极探索生成式 AI 的前沿落地场景。接着是 Northwestern Mutual（西北相互人寿保险，美国排名前列的人寿保险与投资机构）的 AI 高级总监 Sushma Niwalkar。她负责赋能公司核心的 25,000 名一线理财顾问，通过构建 AI 工作流与工具，帮助顾问们与客户建立深厚的信任关系并提升展业效率。 随后发言的是来自 TxMinds（前身为成立于 2013 年的 Test and Experts）的 AI 与质量工程转型副总裁 Michael Giacometti。他分享了公司从质量工程、数字化工程逐步延伸至 AI 领域的历程，目前负责涵盖 AI 测试、AI 工程咨询、智能体框架（Agentic Frameworks）、流程编排以及组织政策与人才转型的七大核心业务体系。最后，Maven AGI 联合创始人兼 CTO Sami Shalabi 介绍了其第四次创业的成果：打造面向大型企业复杂工作流的客户体验（CX）AI 智能体。他指出，大企业处理一张客服工单的成本通常高达 40 至 60 美元，而 Maven AGI 构建的多模态智能体平台已在生产环境中稳定运行，并能自主处理超过 90% 的客户交互。在此之前，Sami 曾在谷歌任职 12 年，负责掌管 Google News 的全部产品与工程团队。 - **Jill Hosmer-Jolley**：加州州立大学蒙特雷湾分校教授、主持人 - **Rajesh Krishna Bharathi**：Discount Tire AI 卓越中心（AICE）负责人，30 年工程领导力老兵 - **Sushma Niwalkar**：Northwestern Mutual AI 高级总监，服务 25,000 名一线理财顾问 - **Michael Giacometti**：TxMinds AI 与质量工程转型副总裁，主导智能体框架与质量工程 - **Sami Shalabi**：Maven AGI 联合创始人兼 CTO，前 Google News 产研负责人，实现 90%+ 客户交互自主化 --- ### [03:58](https://www.youtube.com/watch?v=gJ-wKhKhmA8&t=238s) 生成式 AI 的落地路径：从内部小场景到敏捷 POC 主持人将首个问题抛给 Rajesh，请他分享企业引入生成式 AI 的切实方法。Rajesh 指出，虽然机器学习在很多企业已经不是新鲜事物，但在拥抱生成式 AI 时，企业必须牢记“从小处着手”（Start Small）的核心原则。最稳妥的策略是优先聚焦内部员工场景，理清端到端的业务流程，明确预期的商业价值和交付结果。 在具体推进中，首要任务是清晰定义用例中的“角色”（Actors）以及系统与 AI 的交互反馈。更关键的是，团队必须自问一个根本性问题：“这里到底需不需要 AI？这真的是一个必须用生成式 AI 解决的场景吗？”当业务用例和技术架构流梳理清晰、技术分层与推理成本评估完毕后，应当迅速进入沙盒环境开展快速概念验证（POC），以此验证实际产出的业务价值。只有在内部部门充分打磨好一致性与稳定性后，再逐步向外部客户场景推广。 ```mermaid flowchart TD accTitle: 生成式 AI 落地评估与实施闭环流程 subgraph S1["1. 业务定义阶段"] A["明确内部小场景与用例"] --> B["厘清角色与端到端业务流"] B --> C["定义预期商业价值与产出"] end subgraph S2["2. 技术评估阶段"] C --> D{"是否必须使用 GenAI?"} D -- 是 --> E["设计技术架构与模型分层"] D -- 否 --> F["采用传统规则或传统 ML"] E --> G["评估基础设施与推理成本"] end subgraph S3["3. 验证与推广"] G --> H["沙盒环境快速 POC 验证"] H --> I["打磨输出稳定性与一致性"] I --> J["推广至跨部门及外部客户"] end ``` - **实施准则**：优先从内部员工场景切入，跑通端到端流程，避免一开始就直接面向外部盲目上线。 - **必要性自审**：严格审视业务痛点是否真正依赖生成式 AI，防止“拿着锤子找钉子”。 - **闭环验证**：在厘清技术架构与算力成本后，依托沙盒环境快速推进 POC 验证。 --- ### [06:00](https://www.youtube.com/watch?v=gJ-wKhKhmA8&t=360s) 赋能而非替代：消除员工焦虑与业务技术协同 主持人 Jill 提及拥有 35,000 名员工的 Discount Tire 在门店柜台人员与顾客交互中的生动故事，询问 AI 如何为关键利益相关者带来收益。Rajesh 坦言，落地 AI 时普遍存在“AI 将取代人类工作”的焦虑与迷思。解决这一问题的关键在于从一开始就向各方明确：AI 的本质是“增强”（Augment）而非替代。 Rajesh 分享了他们与业务团队沟通的实操经验：直接询问员工在日常工作中是否愿意省去两到三个小时枯燥的重复性体力劳动。如果答案是肯定的，那么 AI 就是提升生产力的最佳赋能工具。Discount Tire 采取了深度协同策略：让业务骨干全程参与用例创建与技术流程架构讨论，同时为业务人员与技术团队同步提供分层培训，赋予他们普通用户和高级用户（Power Users）的系统访问权限。这种全员参与的方式不仅显著降低了员工的不安全感，更让他们主动成为推动 AI 工具迭代的超级用户。Rajesh 总结道："人类负责串联线索（Humans connect the dots），AI 只是赋能工具。" - **核心定位**：AI 是消除繁琐手工劳动的“增强型赋能工具”，而非替代人类员工的手段。 - **协同机制**：将业务利益相关者引入前期的需求定义与技术架构设计会议，增强归属感。 - **赋能体系**：提供多层次的工具访问权限与技能培训，将业务一线员工培养为 AI 超级用户。 - **业务价值**：借助结构化的 AI 工具支持，柜台员工能够为顾客提供更加精准、一致的门店体验。 --- ### [08:31](https://www.youtube.com/watch?v=gJ-wKhKhmA8&t=511s) 百年老店的 AI 变革：高管赞助与跨职能 AI 治理委员会 主持人将话题引向 Northwestern Mutual 的落地经验，Sushma Niwalkar 结合一家拥有 168 年历史的大型机构的转型历程，分享了深层次的组织治理智慧。面对外界对老牌企业“背负沉重历史包袱（Legacy with a capital L）”的刻板印象，Sushma 指出，Northwestern Mutual 之所以能在 2023 年生成式 AI 爆发伊始迅速启动变革，最关键的第一课就是获得了坚实的高管赞助（Executive Sponsorship）。 然而，随着高管层在看到巨大潜力的同时对潜在风险产生高度担忧，海量业务需求纷至沓来，公司迅速设立了专门的 AI 委员会（AI Council）。Sushma 指出，该委员会承担了三大核心使命： 1. **透明度**：确保整个组织自上而下清晰了解哪些高价值用例能够进入决策层视野并被优先推进； 2. **战略对齐**：严格保障 AI 项目与公司的多年期战略发展规划深度绑定； 3. **成本分摊与权责落地**：由于最终各业务部门必须将 AI 工具的建设与运行成本纳入各自的预算与投资结构中，委员会帮助业务部门提前认清终态蓝图与成本收益模型。 该委员会不仅囊括了技术人员，还深度吸纳了风险控制、法律与合规（CLIP Partners）、人才发展及变革管理等各职能专家，为大型企业在安全合规的前提下规模化落地 AI 构筑了稳固的治理底座。 - **组织第一课**：高管层的全力支持是推动企业级 AI 战略的核心驱动力。 - **委员会使命一（透明度）**：横向拉通高管层视野，公开透明地筛选与排序核心用例。 - **委员会使命二（战略对齐）**：确保技术研发紧密贴合企业的跨年度长期战略目标。 - **委员会使命三（成本闭环）**：促使业务部门主动理解并承接长期的 AI 工具运营与推理成本。 - **跨职能阵容**：涵盖风控、法务合规、人才培训与组织变革团队，确保技术合规与平稳过渡。 ### [10:29](https://www.youtube.com/watch?v=gJ-wKhKhmA8&t=629s) 敏捷实验、风险管理与变革管理的惨痛教训 Sushma Niwalkar 深入阐述了企业落地 AI 时必须掌握的核心能力。首先是**敏捷实验（Agile Experimentation）**与持续迭代。在西北互惠（Northwestern Mutual）为顾问团队构建首个知识管理工具时，最初面向 45 位用户时模型的基线准确率仅有 60% 左右。要将准确率从 60% 提升到 90%，必须进行高密度的技术实验，更离不开业务领域专家（SME）的紧密配合。Sushma 指着现场一位微笑的同事表示，若没有这些“人在回路（Human-in-the-Loop）”的专家持续校准，项目根本不可能成功。 ```mermaid flowchart LR accTitle: AI Knowledge Management Accuracy Improvement Loop subgraph Iteration["知识库迭代提升链路"] A["初始上线 (45位顾问)<br/>基线准确率 ~60%"] --> B["人在回路 (SME 业务专家)"] B --> C["敏捷技术实验与校准"] C --> D["模型优化与迭代<br/>最终准确率达 90%"] end ``` 紧接着是**风险赋能（Risk Enablement）**与**人才技能重塑（Upskilling）**。Sushma 强调，风险合规团队必须从概念验证（PoC）阶段一直到最终部署全程深度参与。谈到高管层对 AI 的学习热潮时，她幽默地透露，多位 C-suite 高管当天正在 MIT 参加 AI 培训，“我周五都不敢打开邮箱了，完全可以想象他们学完后会抛来怎样一堆硬核问题”。 在谈到变革管理（Change Management）时，Sushma 直言这是他们“交学费”最多的领域。当系统刚推送给 45 位高净值投资顾问时，由于顾问们极具主见且对响应速度与准确性要求极高，一旦回答不理想，他们甚至直接打电话向 CEO 投诉。Sushma 自嘲道，如今自己能在这里条理清晰地分享经验，全是因为当初“摔破了膝盖、擦伤了手肘”才换来的惨痛教训。 - **准确率跨越**：知识管理工具从初始 60% 准确率跃升至 90%，完全依赖 SME 专家的协同实验。 - **高管认知刷新**：高层领导前往 MIT 进修 AI，推动自上而下的技术审视与战略提问。 - **变革管理阵痛**：高价值顾问对系统瑕疵零容忍，倒逼团队将变革管理内嵌为标准流程。 --- ### [12:58](https://www.youtube.com/watch?v=gJ-wKhKhmA8&t=778s) 向上管理、文化适配与规模化阶段的 Token 成本隐患 主持人 Jill Hosmer-Jolley 将话筒交给 TxMinds 的 Michael Giacometti。Michael 指出，不论是落地哪种风味的 AI 或是 Agentic AI（智能体 AI），决定成败的关键都在于**企业文化（Corporate Culture）**。自上而下的战略愿景决定了 AI 将如何被使用。 Michael 强调，AI 落地人员本质上是“变革推动者（Change Agents）”。老员工的领域知识必须输入给 AI，同时用来培训初级员工接棒核心职责。AI 不会让人失业，而是带来岗位演进——“就像当年汽车淘汰马车时，马车夫转行去修汽车、补轮胎一样”。 ```mermaid flowchart TD accTitle: Alignment and Scaling Challenges in AI Integration subgraph CultureAndVision["愿景与文化分歧"] E1["高管 A 视角:<br/>提升 30% 生产力 / 效率赋能"] E2["高管 B 视角:<br/>借助 AI 裁员 30%"] CA["变革推动者 / Evangelist<br/>负责向上管理与预期拉齐"] CA --> E1 CA --> E2 end subgraph ScalingReality["规模化落地冲击"] POC["PoC / Demo 演示"] --> SCALE["规模化部署 (Scale)"] SCALE --> COST["Token 成本激增<br/>(劳动力套利无法解决算力开销)"] end ``` 谈及现实落地中的预料之外后果（Unintended Consequences），Michael 特别提到了**Token 成本激增**的冲击。很多人在演示和 PoC 阶段感觉良好，但一旦推向生产环境规模化运行，才惊呼“原来消耗 Token 这么贵！”靠传统的劳动力套利根本无法填补这种技术成本。此外，他还提醒大家必须做好“向上管理（Managing Upwards）”：面对同一句“提升 30% 效率”的口号，有的高管理解为全员赋能增效，而有的高管则将其直接等同于裁撤 30% 的员工。AI 推进者必须扮演好传道者与对齐者角色。 --- ### [15:55](https://www.youtube.com/watch?v=gJ-wKhKhmA8&t=955s) 软件经济学被打破：为何高价值的客户支持是最佳切入点 Maven AGI 的 CTO 兼联合创始人 Sami Shalabi 随后发表了极具技术与商业洞察的观点。他一针见血地指出，生成式 AI 从根本上**打破了传统软件的经济学模型（Software Economics）**。在传统开发中，工程师几乎不会去计较一次数据库调用的边际成本，但在生成式 AI 时代，每一次调用的成本都非常高昂（虽然比两年前便宜了许多，但依然不容小觑）。因此，企业绝不能盲目把 AI 铺到所有细枝末节上，而必须锁定**高价值用例（High-Value Use Cases）**以支撑其高昂的算力成本。 | 客户支持场景特征 | 传统指标与业务现状 | 引入生成式 AI 后的价值收益 | |:--- |:--- |:--- | | **单次交互成本** | B2B 平均 $40–$60；复杂技术产品高达 $1,000 | 自动化替代昂贵的工程师人力，ROI 极其显著 | | **数据成熟度** | 拥有丰富的 CSAT 评分、首次解决时间、首响时间等指标 | 具备海量量化基准，效果与商业回报（Business Case）极易衡量 | | **组织职责演进** | 传统处理繁琐重复工单的初级人员流失率高 | 催生新兴职能“AI Agent Operations”，由懂业务的专家持续管理与校准 AI | Sami 解释道，Maven AGI 之所以将切入点放在客户支持（Support），是因为该领域的痛点最为昂贵且过往的传统自动化工具收效甚微。普通 B2B 企业的单次客户沟通成本在 40 到 60 美元之间，而对于产品极其复杂的企业，由于需要软件工程师亲自下场排查，单次交互成本甚至高达 1,000 美元。同时，支持部门拥有完备的数据体系（CSAT、首次响应与解决时间），使得 AI 带来的投资回报率（ROI）能够被精确量化。 在谈及未来团队结构时，Sami 认为重复性琐事必将被完全自动化，未来企业需要招聘更高素质的人才来管理自动化率达 90% 的 AI 系统。市场上正在崛起一个名为 **“AI Agent 运维（AI Agent Operations）”** 的全新关键职能。最后，从产品经理的视角出发，Sami 强调“改变人类的行为习惯极其困难”，因此 AI 产品的核心原则是必须无缝嵌入到用户现有的工作流中，而不是强迫用户迁移或改变习惯。 ### [20:12](https://www.youtube.com/watch?v=gJ-wKhKhmA8&t=1212s) 深度嵌入工作流：驱动 AI 采纳与重构业务职能 Sami Shalabi 强调，推动生成式 AI 落地的核心瓶颈从来不是技术本身，而是用户体验与工作流集成。如果要求习惯了现有操作界面的员工跳出当前系统、打开一个独立的 AI 工具网站去完成任务再切回来，最终往往只会得到极低的用户采纳率。因此，Maven AGI 构建 AI Agent 的理念不是扔给企业一个孤立的 Chatbot，而是将对话体验深度原生嵌入到团队已有的日常工具中：如果一线外勤人员习惯使用 WhatsApp 和短信（SMS），系统就直接在这些渠道提供支持；如果团队重度依赖 Slack 或 Microsoft Teams，就部署专属的 Slackbot 与 Teamsbot；如果员工在 Salesforce 中工作，就在 Salesforce 内置 Agent Assist 辅助助手。 这种无缝嵌入不仅极大降低了使用门槛，还带来了深远的连带效应——它开始打破传统业务部门之间的边界。在客户体验（CX）领域，技术支持（Support）与客户成功（Customer Success）、售前（Pre-sales）与销售（Sales）的界限正逐渐模糊，企业得以从头到尾重塑全链路的客户旅程。Sami 分享了一个真实案例：商业银行平台 Rho 最初在 Zendesk 中部署了 Maven 的 Agent 辅助工具，客服人员在遇到棘手问题时习惯转到 Slack 进行内部协作；随后团队在 Slack 中上线了 Maven AI Bot 来协助排查问题，结果销售团队闻讯后也蜂拥而至，该 Bot 迅速成为销售人员了解产品底层运作机制最受欢迎的利器。这证明只要深度融入工作流，AI 就能在组织内部自发蔓延与学习。 ```mermaid flowchart LR accTitle: AI工作流原生集成与跨部门扩散路径 subgraph NativeWorkflows["原生工作流嵌入"] A["现场/外勤渠道<br/>(WhatsApp / SMS)"] B["协作办公平台<br/>(Slack / Teams Bot)"] C["核心业务系统<br/>(Salesforce / Zendesk)"] end subgraph AdoptionEvolution["采纳与职能扩散"] D["一线客服辅助<br/>(Agent Assist)"] --> E["内部协同解决复杂问题<br/>(Slackbot 协作)"] E --> F["销售与跨部门自主学习<br/>(售前/技术支持边界模糊)"] end NativeWorkflows --> D ``` --- ### [22:28](https://www.youtube.com/watch?v=gJ-wKhKhmA8&t=1348s) 评估合作伙伴生态与果断淘汰落后供应商 主持人 Jill Hosmer-Jolley 在抛出关于海量数据治理与一线就业影响的问题前，先请 Michael Giacometti 分享他的观察。Michael 引用了管理学畅销书《从优秀到卓越》（*Good to Great*）中的“刺猬理念”（Hedgehog Concept），提醒企业在推进 AI 落地时，必须对自身的合作伙伴与供应商生态进行全面审视，判断他们是否真的有能力跟上这轮技术变革的步伐。 Michael 直言不讳地告诫在场同行：“不要害怕淘汰（jettison）合作伙伴。”在推进质量工程与 AI 实施过程中，很多现有软件及平台供应商的跟进速度严重滞后：他们依然死守着传统的并发席位许可收费模式（concurrent seat license models）、迟迟不推出模型上下文协议（MCP）支持，或者其提供的 API 接口调用极其低效粗糙。企业管理 AI 转型时，不仅需要向上管理管理层、向下引导团队，还必须严格管理外部伙伴生态。如果供应商无法融入企业的 AI 愿景，企业必须果断做出替换决策。 --- ### [23:51](https://www.youtube.com/watch?v=gJ-wKhKhmA8&t=1431s) 一线员工的赋能与“超级用户”的自驱传播 针对 Jill 提出的“AI 落地是否引发一线员工恐慌并导致大规模裁员”的疑问，Northwestern Mutual 的 AI 高级总监 Sushma Niwalkar 给出了明确的回答：“目前还没有（Not yet）。”她指出，公司推行生成式 AI 的定位始终是“赋能工具”（Enabler）而非“替代品”（Replacement）。团队在筛选与落地所有用例时，核心考量都是如何帮助一线员工更高效、更轻松地完成日常工作，而不是削减岗位。 Sushma 回顾了早期项目落地的挑战：起初系统的用户采纳率非常低，很难直观体现工具的业务价值。破局的关键在于业务伙伴（Business Partners）与变革推动者（Change Agents）主动深入一线，面对面地向员工演示 AI 是如何帮他们减负、而不是抢夺饭碗。当一线人员真正体验到工具带来的便利后，组织内部迅速涌现出一批“超级用户”（Power Users）。这批早期采纳者成为了最好的内部代言人，通过同事间的口碑传播，打消了其他人的顾虑，自发带动了整个组织的采纳热情。 - **赋能优先于替代**：选型侧重于增强员工生产力，而非缩减编制。 - **化解抵触心理**：通过变革推动者一对一宣导其实际助益，消除被替代的恐惧感。 - **超级用户效应**：一线实际受益的员工自发成为推广中坚，推动全员采纳。 --- ### [25:49](https://www.youtube.com/watch?v=gJ-wKhKhmA8&t=1549s) Discount Tire 的用例四维评估矩阵与三支柱框架 Discount Tire 的 Rajesh Krishna Bharathi 随后深入阐述了企业面对蜂拥而至的 AI 需求时，如何建立科学的用例准入与落地机制。为了避免资源浪费，他们制定了四项严密的过滤标准，只有综合满足条件的项目才能获得最高优先级并提交给高管赞助人： 1. **增收潜力**：该用例是否能够直接或间接促进业务营收增长？ 2. **成本节约**：能否显著降低现有运营与人力成本？ 3. **安全态势**：是否能强化企业的整体安全与合规基线？ 4. **AI 必要性判断**：这是最关键的一环——该业务痛点是否“真的必须使用 AI”才能解决，还是传统规则系统即可搞定？ 在执行策略上，Discount Tire 遵循“先对内员工、后对外客户”的节奏，并构建了由三大支柱组成的实施框架：**治理支柱**（提供可复用的参考架构）、**敏捷 PoC/概念验证支柱**以及**全员教育支柱**。Rajesh 为想要开启生成式 AI 之旅的企业梳理了一条实用路径： ```mermaid flowchart TD accTitle: Discount Tire 生成式 AI 用例评估与 PoC 流程 A["收集业务需求并召集利益相关方"] --> B{"四维决策树过滤<br/>(增收/降本/安全/AI必要性)"} B -->|未通过| C["归档或采用传统技术解决"] B -->|通过| D["进入敏捷 PoC 阶段"] subgraph PoCArtifacts["PoC 三大核心交付物"] E["用例交互图<br/>(Use Case Diagram)"] F["技术架构流图<br/>(Technical Flow Diagram)"] G["全生命周期成本模型<br/>(TCO Cost Model)"] end D --> PoCArtifacts PoCArtifacts --> H["小规模规范化试点部署<br/>(Canonical Deployment)"] ``` 在敏捷 PoC 阶段，团队必须产出三项核心交付物：用例图、技术流程图以及成本模型。Rajesh 特别强调，技术架构流图中的每一条调用连线在现实中都是真金白银的成本（包括模型 Token 消耗与周边基础设施开销）。实际运行中经常会遇到纸面成本模型所未能预估的“隐形成本”，因此必须先进行小规模的规范化试点（Canonical Deployment），校准实际开销后再做全局推广。在数据层面，Rajesh 建议系统必须紧贴“单一事实来源”（Source of Truth），如果核心数据已在特定平台上，应优先利用该平台原生的 AI 组件，以加速交付并避免数据割裂。 --- ### [29:14](https://www.youtube.com/watch?v=gJ-wKhKhmA8&t=1754s) 突破线性人员扩张瓶颈与职能融合 Sami Shalabi 顺着数据管理与一线岗位的话题做进一步剖析。他指出，许多传统业务职能（尤其是客服与支持部门）在过去一直面临着严峻的“线性扩张困境”——人力成本与客户规模严格呈线性正相关增长。 每当企业 CEO 制定出激进的业务增长目标时，CFO 往往会陷入极度焦虑。因为按照历史模型测算，当客户量大幅增加时，支持团队的编制可能必须从 1,000 人被迫扩张到 7,000 人，这种规模的人力膨胀在财务和管理上都是不可持续的痛点。而生成式 AI 的引入，正在从根本上打破这一线性依赖关系，让业务规模扩张与人力编制增长实现脱钩。 ### [29:56](https://www.youtube.com/watch?v=gJ-wKhKhmA8&t=1796s) AI 时代的组织变革：将成本中心转化为利润中心 Sami Shalabi 指出，客服等传统业务部门长期面临“无法规模化”的瓶颈。生成式 AI 的引入移除了大量繁琐的重复性工作（busy work）。根据 Maven AGI 的观察，虽然极少数客户会选择裁员，但绝大多数企业是将节省下来的时间用于增加与客户的交互频次。客服人员逐渐转型为客户成功（Customer Success）专家，整个部门也顺理成章地从纯粹的“成本中心”（cost center）升级为创造价值的“收入中心”（revenue center），从而在全局范围内提供高端的“白手套”（white glove）服务体验。 从技术团队管理的角度来看，工程组织的传统架构正面临前所未有的重塑。在过去（如在 Google 工作期间），产品经理与工程师的比例可能高达 1:20；而随着 AI 研发工具带来爆发式的生产力提升，传统的 Product vs. Eng vs. UX 比例已被彻底打碎。领导者目前面临的核心抉择在于：是聘用更多产品经理来匹配研发速度，还是进一步扩大工程师的生产力规模？这种新型工程组织的构建模式仍在积极探索之中。 Sami Shalabi 强调，企业的核心策略应当是**利用 AI 节省出的资源去驱动更大的商业价值，而非仅仅做简单的降本裁员**。 ### [31:28](https://www.youtube.com/watch?v=gJ-wKhKhmA8&t=1888s) 破解“数据杂乱”困局：记录图谱与新鲜度信号 大型企业在落地 AI 时最常抱怨的问题是：“我们的数据太乱了，散落在各处，根本不知道哪些可用，请等我们花一年时间把数据清洗干净再谈 AI。”然而，对于软件服务商而言，等待一年并不现实。Maven AGI 在架构设计上摈弃了“先将所有数据集中导入数据湖再进行清洗”的传统范式，因为这种做法缺乏足够的人力支撑，难以落地。 团队选择将该难题转化为一个高阶的搜索与排序问题。相较于基础的 RAG（检索增强生成）技术，他们构建了**记录图谱（Graph of Record）**： * 系统不再强制要求所有数据流动并存储至集中平台； * 直接精准指向企业数据所在的原始位置； * 结合搜索工程技术，大幅提升了系统的响应质量与价值落地速度（time to value）。 在解决企业知识库冲突（如 SaaS 产品新旧文档并行导致矛盾）方面，系统引入了**“新鲜度信号”（Freshness Signal）**。当用户的查询或意图进入系统时，算法会自动识别并优先匹配最新的文档，从而以最轻量的方式解决了传统数据清洗的难题。 ```mermaid flowchart TD accTitle: 数据架构演进对比图 subgraph Traditional["传统模式（周期长、人力成本高）"] T1["混乱分散的企业数据"] --> T2["数据集中导入数据湖"] T2 --> T3["人工/机械化数据清洗 (耗时约1年)"] T3 --> T4["AI 模型训练与调用"] end subgraph GraphOfRecord["Graph of Record 模式（高效率、实时响应）"] G1["分散的原始数据源"] G2["用户意图 / 检索请求"] --> G3["记录图谱 (Graph of Record) 索引"] G1 -.->|指针直接指向| G3 G3 --> G4["结合新鲜度信号 (Freshness Signal) 排序"] G4 --> G5["输出确定性高的准确回答"] end ``` ### [33:05](https://www.youtube.com/watch?v=gJ-wKhKhmA8&t=1985s) 量化价值与 ROI：用 Agentic 评估框架打通 IT 阻碍 对于真实商业场景的落地，缺乏评估框架（Evaluation Framework）就意味着“盲人摸象”。只有建立明确的量化指标，才能持续优化系统性能并构建坚实的 ROI（投资回报率）论证。Sami 提到，过去在 Google 时，团队为了追求高精度的排序算法，曾在人工评估（Human Evaluations）上投入巨资；而现在 Maven AGI 研发了基于 Agent 的自动评估系统，不仅降低了成本，还使企业内部的缺失环节变得清晰可见。 在企业级应用中，接通庞大的 API 和系统往往会遇到来自 IT 部门的阻力——IT 团队绝不会为了“给接口而给接口”。基于清晰的量化指标，企业可以精细化地评估成本与收益。例如，在每笔客服工单成本为 40 美元的背景下： * 系统能够精准计算出，要将准确率从 **73.5%** 提升至 **90.6%**，必须获取哪 3 个特定系统的访问权限以及用户记录中的哪几个具体字段； * 这套明确的降本增效数据（如节省数亿美元）能为 IT 部门奉上一份具有强说服力的“银盘上的商业计划书”（business case on a silver platter），将 AI 项目直接推向高优先级。 - **73.5% 到 90.6%**：Maven AGI 演示中准确率提升的具体指标示例。 - **$40/ticket**：单次客户交互/工单的典型处理成本量化标准。 ### [35:15](https://www.youtube.com/watch?v=gJ-wKhKhmA8&t=2115s) 颠覆传统路径：由 AI 纵向应用反向驱动数据策略 主持人和 Sushma Niwalkar 对 Sami 的观点表达了高度认同。主持人特别提到，她一直倡导将 AI 节省下来的时间转化为“白手套”高端服务，而不是简单地裁减员工，这在 Northwestern Mutual 这类保险与金融服务领域尤为关键。 Sushma 补充指出，行业过去的惯性思维是“数据策略决定 AI 策略”，即必须先搞好数据治理才能开始做 AI。但如今这一模式已被彻底颠覆：**企业正在通过构建具体的 AI 纵向业务应用（AI Verticals）来反向驱动和塑造数据策略**。在实践过程中，AI 团队能够敏锐地捕捉到数据链路中的破损与短板，从而带领数据合作伙伴（Data Partners）共同推进迭代，实现了 AI 与数据建设的并行协同发展。 ### [37:05](https://www.youtube.com/watch?v=gJ-wKhKhmA8&t=2225s) 针对单体与历史遗留系统的 GenAI 改造实践 在现场互动环节，一位来自拥有三千多年历史（或数十年历史）传统大厂的观众提出了挑战：公司的系统属于极度庞大的单体架构（Monolithic），数据在早期并非核心资产，如今正处于向模块化拆分的过渡期。在这种缺乏现代化数据基础设施的传统单体系统中，应该如何引入生成式 AI？ Sami Shalabi 对此给出了针对性的解法： 1. **接受现状而非推倒重来**：绝大多数企业都面临单体架构和数据混乱的现实，盲目追求大而全的数据改造项目往往耗时费力。 2. **将非确定性问题转化为确定性搜索问题**：利用搜索与排序技术（Ranking Problem），系统能够跨越结构化与非结构化数据直接定位答案。 3. **建立出处追溯（Provenance）机制**：在生成答案的同时提供可追溯的引用来源（Referenceability），使领域专家（Domain Experts）能够清楚知道数据来自何处，从而让不可控的 AI 生成变得精准可控。 4. **基于评估数据实施精准投资**：通过 AI 评估体系监控系统在何处偏离轨道，精准定位数据缺陷，将资源集中在影响最大的关键环节，以最快的速度实现商业价值。 ### [39:43](https://www.youtube.com/watch?v=gJ-wKhKhmA8&t=2383s) 打破数据孤岛与推动内部变革 当面对像梅西百货（Macy's）这样从传统零售商演变为事实上的科技公司的大型企业时，数据往往分散在各个孤立的部门和业务角落中。Michael Giacometti 指出，许多团队虽然在局部使用数据，却根本不知道如何将这些数据点串联起来。对于身处组织内部的员工而言，这正是扮演“变革推动者”（change agent）的最佳契机。具体路径在于：首先精准锁定具有商业价值的用例，进而识别出实现该用例需要调取哪些分散的数据集；接着主动与掌握这些数据的人沟通，用 AI 和自动化方案帮他们减轻工作负担，把他们也拉入变革推动者的阵营。 通过这种自下而上的赋能，企业不仅能优化内部流程、让各项任务以更快的速度推进，还能直接转化为更卓越的外部客户体验，带动更高的销售量。Michael 建议，从业者在开展 AI 项目时不妨先在身边做一番摸底调查，了解周围的人都在用什么工具和数据，找到既能切实帮到他们又能与公司大方向保持一致的切入点，这样才能在组织中赢得真正的买账与支持。 ```mermaid flowchart TD accTitle: 推动企业内部 AI 变革与数据连接的执行路径 subgraph S1["1. 锚定与调研"] A["调研内部工具与数据使用现状"] --> B["锁定契合业务战略的高价值用例"] end subgraph S2["2. 跨部门连接"] B --> C["定位分散的数据孤岛"] C --> D["赋能数据持有者：简化其日常工作"] end subgraph S3["3. 价值交付"] D --> E["内部：提升执行效率，推动战略目标"] D --> F["外部：改善客户体验，促进业务增长"] end ``` --- ### [40:44](https://www.youtube.com/watch?v=gJ-wKhKhmA8&t=2444s) 裁员叙事背后的真相与人类的核心价值 来自 ECMC 的现场观众 Daniel Ninsky 抛出了一个尖锐的提问：他极度怀疑 AI 并没有像各大公司宣称的那样取代了那么多的工作岗位；所谓“AI 替代人工”更像是一种易于被资本市场接受的公关借口（"market palatable excuse"），其真实目的是为了修正 2020 年科技行业过度扩张时期“只要简历上写着工程师就招进来”的人员冗余（glut of hiring）。Daniel 认为这种夸大的叙事正在损害公众对 AI 的信任，让大家过度陷入失业恐慌。 Rajesh Krishna Bharathi 对此深表赞同，并再次强调了“人的因素”（human factor）不可替代。人类最擅长的是跨领域的思考与“连点成线”（connecting dots），而 AI 从本质上说只是一种使能工具（enabler）。Rajesh 回顾了科技行业的周期律：整个 IT 领域每隔大约 5 年就会经历一次彻底的迭代，5 年前是云计算（Cloud），如今是生成式 AI。当初云计算刚兴起时，许多人也曾极度抗拒，把数据紧紧攥在自己的本地数据中心里，“就像孩子抱着泰迪熊不放一样”（"holding on to the data like their teddy bears in data centers"），但现在云架构早已成为常态。展望未来，从实习生到资深工程师，所有人都将具备 AI 能力；从业者不必焦虑于被取代，而应该主动去学习、考取认证、在实际工作中上手试用，并成为新技术的布道者。 - **行业周期律**：IT 领域大约每 5 年就会迎来一轮技术重塑（从数据中心到云计算，再到如今的生成式 AI）。 - **人机分工定位**：人类负责全局洞察与连点成线，AI 充当增强工具与效率杠杆。 - **应对策略**：积极学习、考取相关认证、在日常实践中运用并成为团队内的技术布道者。 --- ### [43:06](https://www.youtube.com/watch?v=gJ-wKhKhmA8&t=2586s) 终场总结：理解技术周期，走出象牙塔 在小组讨论的尾声，Michael Giacometti 为整场圆桌会议做了总结发言。他直言关于技术替代人力的恐慌，大家“早就看过了这部电影”（"we've seen this movie before"）：每当颠覆性新技术出现并承诺大幅提升效率时，管理层往往会过度反应、匆忙裁员，但不久后又会以全新职能的形式把人才重新招募回来。面对这种周期，技术人员与业务人员最核心的任务不仅是理解和消化 AI，更要学会向外传授与布道，在企业内部积极做好向上管理与跨部门协同，推动公司达成战略目标。 Michael 同时鼓励在场的参会者打破纯技术的思维局限，多就一些非技术层面的实际工作挑战展开交流与互助，比如在组织中如何与严苛的上级沟通协作、如何推动跨团队共识等。最后，专家团队对所有参会观众表达了感谢与开放交流的意愿，整场研讨会在热烈的掌声中落下帷幕。 ## Notable quotes > [06:00](https://www.youtube.com/watch?v=gJ-wKhKhmA8&t=360s) Humans connect the dots, AI is just an enabler. — Rajesh Krishna Bharathi > [08:31](https://www.youtube.com/watch?v=gJ-wKhKhmA8&t=511s) Legacy with a capital L — Sushma Niwalkar > [12:58](https://www.youtube.com/watch?v=gJ-wKhKhmA8&t=778s) Change Agents — Michael Giacometti > [22:28](https://www.youtube.com/watch?v=gJ-wKhKhmA8&t=1348s) Don't be afraid to jettison partners. — Michael Giacometti > [23:51](https://www.youtube.com/watch?v=gJ-wKhKhmA8&t=1431s) Not yet. — Sushma Niwalkar > [25:49](https://www.youtube.com/watch?v=gJ-wKhKhmA8&t=1549s) Source of Truth — Rajesh Krishna Bharathi > [33:05](https://www.youtube.com/watch?v=gJ-wKhKhmA8&t=1985s) business case on a silver platter — Sami Shalabi > [40:44](https://www.youtube.com/watch?v=gJ-wKhKhmA8&t=2444s) holding on to the data like their teddy bears in data centers — Rajesh Krishna Bharathi > [43:06](https://www.youtube.com/watch?v=gJ-wKhKhmA8&t=2586s) we've seen this movie before — Michael Giacometti ## Mentioned Ai4, Discount Tire, Honeywell, Northwestern Mutual, TxMinds, Test and Experts, Maven AGI, Google, Google News, Jill Hosmer-Jolley, Rajesh Krishna Bharathi, Sushma Niwalkar, Michael Giacometti, Sami Shalabi, Daniel Ninsky, ECMC, MIT, Salesforce, Slack, Microsoft Teams, WhatsApp, Zendesk, Rho, Macy's, Good to Great ## Video description Sushma Niwalkar, Sr. Director - AI, Northwestern Mutual Jill Hosmer-Jolley, Faculty, California State University Monterey Bay Sami Shalabi, CTO & Co-Founder, Maven AGI Michael Giacometti, VP AI & Quality Engineering Transformation, TxMinds Rajesh Krishna Bharathi, Segment Strategist, Discount Tire Abstract This session will offer a detailed, step-by-step look at the practicalities of integrating generative AI such as LLMs and multimodal AI models into existing business systems and workflows. Panelists will share insights on overcoming technical challenges, such as interoperability with legacy systems, managing data flows, and scaling AI infrastructure. Learn how businesses are handling the complexities of aligning AI deployments with operational needs, ensuring seamless integration, and delivering tangible business value through well-executed AI implementations. [Source](https://www.youtube.com/watch?v=gJ-wKhKhmA8)

[6:00] Rajesh Krishna Bharathi: Humans connect the dots, AI is just an enabler. [8:31] Sushma Niwalkar: Legacy with a capital L [12:58] Michael Giacometti: Change Agents [22:28] Michael Giacometti: Don't be afraid to jettison partners. [23:51] Sushma Niwalkar: Not yet. [25:49] Rajesh Krishna Bharathi: Source of Truth [33:05] Sami Shalabi: business case on a silver platter [40:44] Rajesh Krishna Bharathi: holding on to the data like their teddy bears in data centers [43:06] Michael Giacometti: we've seen this movie before

[3:58] Rajesh Krishna Bharathi: Do you have the need for 2 to 3 hours of manual work in your daily life of work? If they don't want to do that, then AI is going to be a best enabler to increase their productivity. [3:58] Rajesh Krishna Bharathi: Humans connect the dots. AI is just an enabler. [8:37] Sushma Niwalkar: We learned the hard way. We had, you know, scraped knees and bruised elbows before we got change management on. Today, it's very much part of our process. [12:58] Sami Shalabi: When we build software today, you don't think about the cost of a database call. But now you have this thing where if you use it, it's stupid expensive... As a result, you have to go after high-value use cases. [19:51] Sami Shalabi: Try to change human behavior. Good luck. [22:28] Sushma Niwalkar: When frontline workers are actually using the tool and they see the value that it provides, it generates a lot of power users, and these power users become really the voice for us in the organization to talk about how actually it's making their lives much more simpler and making our lives easier. [25:47] Rajesh Krishna Bharathi: do you need AI for this use case? [25:47] Rajesh Krishna Bharathi: every line in your technical flow diagram is a cost [29:14] Sami Shalabi: One of the most common things we hear from every customer, every any enterprise level customer is, 'Our data is a mess. It's all over the place. We don't know what's good. We don't know what's bad. Come back to us in a year when we've when we clean up the data' and we're like, 'Oh my god, okay, we don't want to wait a year because we want to sell them our solution.' [43:06] Michael Giacometti: we've seen this movie before

[0:04] Thank you everyone for coming. We are [0:06] very thrilled to see you here this [0:08] morning on day three of a very amazing [0:11] conference. Um I'm Jill Hosmer Jolly. [0:14] I'm a professor at several universities [0:16] but primarily California State [0:18] University Monterey Bay and I specialize [0:20] in neurodiversity leadership and AI. And [0:23] I'm going to turn it over. [0:25] >> Yeah. Uh I'm Raj Barti. I'm from uh [0:28] Discontire. I've been with Discontay for [0:31] 4 years. Before that, I was with [0:32] Honeyville for 17 years. Um I'm a [0:35] veteran in uh engineering leadership. I [0:38] have done um you know business value [0:40] deliverables um throughout my 30 years [0:43] of career and um in um discontently [0:47] leading an AICE [0:49] um we are exploring a lot of generative [0:51] AI cases. So thanks for coming um good [0:54] morning to all of you and great to have [0:56] you. So looking forward to a great [0:58] conversation. [1:01] >> Thank you Raj. Uh I'm Sushma Nalkkar and [1:04] I'm from Northwestern Mutual. It's an [1:06] insurance company and an investment [1:08] company and we're the number one [1:09] insurance company in the United States. [1:12] Um and I lead uh AI for the field [1:14] function. Basically uh advisers about [1:17] 25,000 advisers form core to our company [1:20] in terms of how they actually connect [1:22] with our nines and build trusted [1:24] relationships and my job is to enable [1:26] them with AI workflows and tools so that [1:30] they can do their jobs better. Michael, [1:32] >> thank you very much. My name is Michael [1:34] Jacomedi. I am the vice president of [1:36] artificial intelligence and quality [1:38] engineering transformation at TX. TX [1:41] minds was originally test and experts [1:43] founded in 2013 where we specialized in [1:46] quality engineering services later [1:48] evolved to digital engineering and then [1:49] layered on AI on top as the head of AI [1:52] basically I've been building out seven [1:54] offerings everything ex going from [1:56] testing of AI to AI with testing to AI [1:59] engineering and advisory and lately [2:02] focusing on agenic frameworks [2:03] orchestration and the policy and people [2:05] changes that need to happen because of [2:06] it. uh started off in QA. I think that's [2:09] a great leadin for what we're doing in [2:10] AI and uh very happy to be here. Thank [2:12] you for having me. [2:14] >> All right, nice to meet you all. Sammy [2:17] Shalabi, founder and CTO of a company [2:19] called Maven AGI. We build AI agent for [2:22] customer experience focused on very [2:24] complex workflows, focused on large [2:27] enterprise. Um the problem we we solve [2:31] is that customer experience is really [2:32] really challenging for most most large [2:34] enterprises. Uh um most of the customers [2:38] we talk to spend between $40 to $60 per [2:40] support ticket. Uh we're building AI [2:43] agents that actually automate that [2:45] entire experience. So we build AI agents [2:46] that operate in every imaginable [2:48] modality. You can track with it. You can [2:50] you can talk to it. Uh we we also do [2:52] agent assist in every imaginable [2:54] ticketing system. And then we also build [2:56] a platform uh that allows you to manage [2:59] these AI agents at scale. We're at a [3:01] place where uh uh uh we're deployed, [3:04] we're live, we're in production. and our [3:06] AI agents are handling over 90% of uh [3:10] customer interactions autonomously. [3:13] Prior to Maven, Maven's my fourth [3:15] company and I also spent 12 years at [3:17] Google where I used to run all of [3:18] product and engineering for Google News. [3:22] So the brilliant organizer of this event [3:25] Daniel has put together a fabulous panel [3:28] here where we have two people from [3:30] retail organizations and they'll talk [3:32] about their implementation both the [3:34] challenges rewards and then we have [3:36] fabulous people who help a number of [3:39] other uh organizations. So we'll go back [3:42] and forth also we will open up for [3:44] questions. We have a very intelligent [3:47] person here in the front row from Macy's [3:49] and he's going to ask our first question [3:51] in about 15 or 20 minutes. So please [3:54] prep your questions if you have anything [3:56] for us. All right, I'm going to start [3:57] with Raj. [3:58] >> Yeah. And so you know all of this uh [4:01] generative AI a lot of you are familiar [4:04] probably with machine learning. Um it's [4:07] not new. It's something that you know is [4:09] already there. You probably are you know [4:11] experiencing that in your companies. But [4:13] as we embark on this generative AI, [4:15] couple of things that uh you know keep [4:18] in mind is start small. Um start with uh [4:22] internal use cases. Uh try and see if [4:25] you can uh do some things with uh your [4:27] employees [4:29] and u be able to get uh sort of a a [4:32] process flow really to give an [4:34] understanding of where you want to start [4:36] and where you want to end. What are your [4:37] outcomes? What are your business values? [4:40] key is to also get the use cases. Start [4:44] off with a use case. Um it's very [4:46] important to understand uh the actors [4:49] the system in in this case the [4:51] generative AI and what is that um the [4:55] actors is going to get back from the [4:56] system. Um and then of course you know [4:59] get to a technology flow. Um why is [5:03] there a need for AI? Um is is this a [5:06] case for AI? So in generative AI [5:09] obviously if you need that uh the reason [5:11] I'm saying that is as you get the use [5:13] cases you get your technical flow and it [5:15] it gets into the the model of uh the [5:18] technical layers coming in for AI then [5:20] you get to the cost angle of it once you [5:23] understand the cost of it um try and get [5:26] to a sandbox for a rapid P. So I think [5:29] that would you know clearly understand [5:31] uh the value that you're going to [5:33] generate with the P for the generative [5:35] AI cases and of course you know there [5:37] are multiple cases that we are exploring [5:39] um that um internally we want to do [5:42] first with a lot of our departments and [5:45] then there are a lot of cases for the [5:46] external uh customers also that we want [5:49] to do after we iron out um the [5:52] consistency part of it uh in AI. So I'll [5:56] add to my panelists if they want to add [5:58] anything. [6:00] >> So uh before we move to Sushma I I do [6:02] you told me a fascinating story [6:04] yesterday. Yeah. Discount uh tires has [6:07] 35,000 employees and it but a lot of [6:10] outlets most of you may be familiar with [6:12] it. Um can you talk about um how the AI [6:16] benefits the two stakeholders that are [6:19] key in that relationship the [6:20] counterperson and the customer? [6:23] >> Yes. um as we embark on this um journey [6:26] right um there is a myth that you know [6:29] AI will replace people right so to deal [6:34] with that I think you know it's not [6:35] going to replace you probably hearing [6:37] this consistently from the speakers u is [6:40] that it's going to augment and that [6:43] something that you know you have to [6:44] right out of the gate talk to your [6:46] stakeholders [6:48] and um the key thing is to ask them back [6:51] you know do you have um the need for 2 [6:55] to three hours of manual work in your in [6:58] your daily life of work right so if if [7:01] they don't want to do that then AI is [7:02] going to be a best enabler to increase [7:04] their productivity uh it's important to [7:06] also have the business stakeholders in [7:09] the room when you are creating the use [7:11] case that I was talking about and also [7:13] make them part of the technical flow [7:14] design discussions that way they [7:17] understand they are part of the equation [7:19] um another thing that we did is to try [7:21] and get them training not just the [7:24] technical people also the business [7:26] people part of the training and then [7:28] give them access to tools that are in [7:30] multiple layers uh as end users and [7:33] power users it's important for them to [7:35] be part of this journey uh because they [7:37] are the ones that are going to use it [7:39] from an IT point of view um it's also [7:42] going to bring lot of uh that rapidness [7:44] that I was talking about to build that [7:47] um and so it eliminates a lot of that um [7:50] anxiety they have uh and and then make [7:52] them part of the journey. So I think [7:54] that helped us and by the giving them [7:56] access they become the power users of [7:58] the AI technology and be able to use [8:00] that use case and make it work for [8:03] whatever they can do with that um [8:05] technology. So that really helped. Um, [8:08] end of the day, I want to say this that [8:11] humans connect the dots. AI is just an [8:14] enabler. [8:16] >> And I really loved one of the things [8:19] that you said yesterday was that the [8:21] customer experience is more exact and [8:24] consistent when the employee has access [8:27] to a a well-developed AI. and Sushma [8:31] that's something that we I heard from [8:33] you when you talked about rolling out to [8:35] your agents. Can you talk about that? [8:37] >> Yeah, sure. So, um it's a little bit of [8:40] a story. So in the beginning of 2023, I [8:43] mean when when generative really um [8:45] began. Um so I'm from Northwestern [8:48] Mutual as I mentioned before and we are [8:50] a 168-year-old company and and I and I [8:53] can see everybody going you know that's [8:54] a that's legacy with a capital L but uh [8:58] fortunately for us you know we have a [9:00] lot of executive sponsorship in terms of [9:02] actually building capabilities and tools [9:04] and really want to push that agenda [9:06] forward. So when I look back in our [9:08] journey, I think the first lesson that [9:10] we learned was the importance of [9:12] executive sponsorship in actually [9:14] driving an EI agenda forward. So that [9:16] was the first thing and then as we went [9:18] forward, we realized that you know they [9:21] were not only very uh enthralled by the [9:23] promise that it provided but they were [9:24] also really worried about the uh risk [9:26] associated with it. So obviously we had [9:29] a lot of use cases coming and then we [9:31] had to set up a council like a lot of [9:32] you probably would have had a council [9:34] and the objective of that council was [9:36] three-fold right the first one was to [9:38] make sure there was transparency across [9:40] the organization around what use cases [9:42] really rise to the top across our seuite [9:45] that was number one number two is uh [9:48] making sure that you know we were [9:50] aligned to the strategic priorities um [9:53] that was really important our multi-year [9:54] strategy and the third piece of it was [9:57] at the end of Today the individual [9:58] business functions actually absorb the [10:00] costs of the tool that you're building. [10:02] So in their investable structures have [10:04] to actually absorb that cost and they [10:06] need to understand what the end state [10:07] looks like. So we make sure that the [10:09] council solved for all of that. That was [10:11] lesson number two. We really needed that [10:13] and our council included members from [10:15] risk our clip partners as well as you [10:18] know talent and development change [10:19] management all of these people. So that [10:22] was the second lesson that we learned [10:24] and uh the third I think part of the [10:27] lesson that we learned was kind of has a [10:29] few shades within it. The first was the [10:32] uh need for experimentation and the need [10:35] for agile experimentation is extremely [10:37] important. So we built our first use [10:39] case as I said I you know I enable [10:40] advisor and the first use case we built [10:42] was a knowledge management tool and when [10:44] you build that when you start with 45 [10:47] users your baseline is pretty low. you [10:49] are in like 60% baseline in terms of [10:51] accuracy to go from 60 to 90%, you [10:54] really need experimentation and [10:56] iteration and you needmemes. Some of [10:58] mymes are here with me. I mean we could [11:01] not have done it without them. Um I I [11:03] can see one of them here smiling but [11:05] they are extremely important. The human [11:07] in the loop to actually make this work [11:09] is extremely important. So [11:11] experimentation both technically as well [11:13] as with humans is extremely important. [11:16] And then the other piece that I really [11:18] want to call out is risk enablement. [11:20] Making sure all your risk partners are [11:22] with you every step of the way from [11:24] conceptualization to actual deployment. [11:27] That's the second part that I did wanted [11:28] to share in terms of lesson number [11:30] three. And the last part I really want [11:32] to talk about is talent and upskilling. [11:35] uh a lot of my um seuite um you know [11:39] leaders are at MIT today to learn and I [11:42] do not want to open my email on Friday [11:44] because I can imagine the set of [11:46] questions they've gone to MIT to learn [11:47] about AI. So I think what I really want [11:50] to walk away uh or at least share with [11:53] everybody in terms of what we've learned [11:54] is one is executive sponsorship uh [11:57] agility in terms of both experimentation [12:00] and organizational agility in terms of [12:02] change management. So when I talk about [12:04] change management that's was the least [12:06] indust un understood for us when we went [12:08] to do the use case because we rolled it [12:10] out to 45 advisors and suddenly we found [12:13] that you know there was a lot of noise [12:15] people calling up the CEO himself [12:16] because they're a very vocal very [12:18] opinionated ext multi-millionaire [12:21] advisers right they want answers quickly [12:23] and they want answers correctly so when [12:26] that doesn't happen you know they call [12:27] up the CEO so now I'm here rattling off [12:30] all these bullets saying that you know [12:32] we had all these process We didn't we [12:34] learned the hard way. We had, you know, [12:35] scraped uh knees and bruised elbows [12:38] before we got change management on. [12:40] Today, it's very much part of our [12:41] process. So, uh that's what I wanted to [12:43] kind of share with this group in terms [12:45] of the lessons we've learned in our AI [12:47] journey. And those are a lot of lessons. [12:50] I really appreciate you going through [12:52] those. And I'm going to turn it over to [12:54] Michael now because I can see the wheels [12:56] turning in Michael's mind. [12:58] >> They never stop. So the the key here is [13:00] that the great experiences and stories. [13:04] One of the things that we've been able [13:05] to do is not just implement AI, agentic [13:08] AI, all flavors AI for all sizes of [13:10] organizations, but one of the things [13:13] there are a couple things that ring [13:14] true. The first is that AI requires [13:18] strong culture. In order for AI to be [13:21] implemented, doesn't matter if you're [13:22] going from the use case, the [13:23] technologies you are using, it has to [13:25] align with the corporate culture. The [13:26] division from the top is going to [13:27] dictate how the AI is implemented and [13:29] how it is used. The second thing is that [13:32] we become change agents. When we're [13:34] leveraging AI and implementing it, it's [13:36] going to change the way we work. But all [13:38] of us in this room have domain [13:39] experience. That domain experience needs [13:41] to be shared with the AI. And then what [13:43] we need to do is take that domain [13:44] experience and train those at the entry [13:46] level so they can take over our roles [13:48] and have domain experience. So nobody [13:50] loses their job over AI. We're just [13:51] finding new positions. It's like when [13:53] the horse and carriage were outdated by [13:56] the car, people became mechanics and [13:58] repaired tires. Same idea. The third [14:01] thing is that there's a law of intended [14:02] and unintended consequences that come [14:04] with this. People are very quick to [14:05] understand the proof of concept. They [14:07] want to get to the use case. They want [14:08] to do the policy and practice alignment, [14:10] but then there's going to be the [14:12] intended and unintended costs here. One [14:15] of the things that I knew but really [14:18] validated here and heard it was the [14:20] concept of wow, I didn't know tokens [14:22] cost that much when they were [14:24] implementing the the the AI at scale. [14:26] It's like when you make that jump [14:29] from demo and P to actually starting to [14:32] scale it, there are going to be a lot [14:34] more costs and labor arbitrage is not [14:35] the way to save that cost. And so one of [14:38] the things that we need to do as a [14:39] collective team here is put our minds [14:41] together, put the cultures that we have [14:42] at our companies together, put our, you [14:44] know, what we've learned and integrate [14:46] that in AI. Think with those as either [14:48] co-workers or you being a human manager [14:49] of a digital asset. One of the other [14:52] things that we've also learned here is [14:55] it's our responsibility to manage [14:57] upwards. I apologize for all the sea [14:58] sweeten here, but one of the things that [15:02] is being done as you're implementing AI [15:03] is managing upwards. One of the lessons [15:06] we've learned is that if you talk to two [15:08] executives, one's going to have a [15:10] mindset where 30% efficiency means [15:13] implementing AI, agentic AI, pick your [15:15] flavor. [15:17] The other executive is going to say 30% [15:19] efficiency means leveraging AI to remove [15:22] 30% of the workforce. [15:24] Two executives, two organizations, two [15:27] different visions, and it's up to you to [15:28] help align it. And so, you've got to [15:30] become an evangelist. You've got to [15:31] become a change agent. And that's the [15:33] one thing I've noticed when implementing [15:34] AI across the board. [15:37] >> Well, I love that you made that point [15:39] because in the first keynote this [15:41] morning, we definitely heard about the [15:43] value of keeping the humans involved. [15:45] And my personal thing is not to lose the [15:47] people along the process. Um Sammy, it's [15:51] going to be hard to follow these three, [15:52] but I know you and you can do it. So, [15:55] off you go. [15:55] >> All right. Um, one of the things about [15:59] generative AI is when you start actually [16:02] doing it, it broke something fundamental [16:05] in software. It actually broke software [16:07] economics. When we build software today, [16:10] you don't think about the cost of a [16:12] database call. But now you have this [16:15] thing where if you use it, it's stupid [16:18] expensive. Two years ago was crazy [16:21] expensive. It's still it's gotten much [16:22] much cheaper. So as a result [16:26] you have to go after high value use [16:28] cases. When we talk to our many of our [16:30] customers there's a lot of use cases. I [16:32] mean this stuff can automate everything. [16:34] It can improve many many many things but [16:39] picking a use case that is high value is [16:43] actually a critical part because it [16:45] justifies its cost. That's I think the [16:46] the first part. In our case, when we did [16:48] our analysis and looked at the problem, [16:51] we ended up going after support for a [16:53] couple of reasons. One, I was shocked [16:56] how expensive support was. [16:59] The average support ticket in most B2B [17:02] companies is $40 to $60 per interaction. [17:04] Like that number is crazy. We have some [17:06] customers who have very complicated [17:08] products are spending $1,000 per [17:10] interaction because engineers are are [17:12] handling that. And when you start with [17:14] that, that's an area where historically [17:17] automation and tooling has never [17:19] actually solved that problem. So uh uh [17:24] so so we kind of focused on that as [17:26] being kind of initial entry because you [17:28] there's very clear ROI. And the other [17:30] thing about support it is it's probably [17:32] one of the functions that has the most [17:34] amount of data. Everybody's there's like [17:37] seat scores, time to first resolution, [17:39] time to first answer. It's like a it's [17:42] like a myriad of a massive amount of [17:44] data. So when you start to introduce it, [17:46] it has very clear uh uh uh ROI and you [17:51] can pretty much build a very clear [17:52] business case around the overall impact [17:54] and the results that you have. That's I [17:56] think the first part is like you have to [17:58] go after high value use case because [18:00] this stuff is expensive. There's a lot [18:01] of opportunities. Um the second part and [18:04] this one uh I I I didn't appreciate [18:07] until we started to do it more is [18:10] imagine walking into an organization and [18:12] pretty much saying like this AI stuff [18:14] and everybody it's going to replace the [18:16] human beings the person on the other [18:17] side is like holy holy moly I'm going to [18:20] lose my job and is not going to be a [18:24] change agent and so one of the things [18:27] that we thought deeply about is well [18:29] what we we're spending I mean we thought [18:31] about it from a product perspective But [18:33] from a communication perspective, we [18:35] actually needed to create a vision for [18:37] what's the future here. And when you [18:40] kind of think yes, the reality is all [18:42] the busy work's gone. All the the the [18:44] like the low-level uh simple stuff is [18:48] going to get automated away. But the [18:49] future of for example customer [18:51] experience is not about just humans in [18:54] the loop, etc. All the busy stuff is [18:55] gone. But uh uh and uh entry level high [19:00] school graduate type folks are probably [19:03] that that world is changing. But the the [19:05] future of that organization is one is [19:06] you're going to have to hire college [19:08] graduates because there's going to be [19:10] someone having to manage this AI. So you [19:12] build this AI you get it to 90% [19:15] automation rate. It's amazing but the [19:17] world changes the products change. So [19:18] you need some way to manage this stuff [19:20] over the long term. And as a result we [19:23] spend a lot of time uh sometimes at the [19:25] seauite in publicly traded companies [19:28] like okay you have to think about the [19:29] org differently you have to think about [19:30] the incentive structures differently uh [19:33] it's a very there is a new world uh [19:36] we're calling it AI agent operations [19:38] that's a new function that's starting to [19:40] emerge in in in in organizations and [19:43] we're seeing this happen is like the [19:45] actual the customer experience person [19:46] who's had the first deployment ends up [19:48] being the AI expert that's I think one [19:50] one core [19:51] The next thing is and this is this comes [19:53] just from kind of my product background [19:56] is that human change is hard. Oh my god. [19:59] Try to change human behavior. Good luck. [20:02] And if you are trying to build a product [20:06] in the AI space, you have to integrate [20:09] in people's workflows. It's like if [20:12] they're used to doing a thing their way [20:14] and you tell them go to this other UI [20:16] and do the AI and come back, you're [20:18] going to get very very little adoption. [20:20] So, a big part of this that I don't [20:22] think many people appreciate, the user [20:24] exper the tech is incredible, [20:26] >> but the user experience is critical, [20:29] which is why when we built our AI [20:30] agents, it's not here's a chatbot or [20:32] here's a website. We've deeply [20:34] integrated the actual conversational [20:37] experiences into people's workflows. If [20:39] you have field agents and they're using [20:40] WhatsApp and SMS, we got a solution for [20:42] you. If your folks are in Slack, we got [20:45] a Slackbot. If they're in Teams, we got [20:46] a Teamsbot. If you're using Salesforce, [20:49] we have an agent assist experience [20:51] inside of Salesforce. So being deeply [20:54] integrated into people's workflows is is [20:56] critical for driving adoption [20:58] and it starts to kind of unlock uh uh um [21:03] uh new kinds of experiences. So as this [21:06] stuff gets deployed and and and you [21:08] start to see the results, one secondary [21:11] thing that starts to happen, the thing [21:13] we're seeing in customer experience is [21:15] the actual functions are starting to [21:17] blur. What's the difference between [21:19] support and success? What's the [21:21] difference between pre-sales and sales? [21:23] And and it's you you start to get to a [21:25] place where you're reimagining the [21:27] actual experience from soup to nuts. [21:29] It's not about support. It's about the [21:31] entire customer experience. And we we're [21:33] seeing this consistently. We deploy this [21:36] stuff with with uh uh with for example a [21:39] true story uh row the the uh the the [21:43] bank uh uses us. We started kind of in [21:46] in Zenesk uh uh and we had an agent [21:48] assistant. The agents were using it in [21:50] there and one of their behaviors was if [21:54] you can't figure it out, they jump into [21:56] Slack and then they kind of work [21:57] together. So then we did then they [21:59] deployed their Slackbot the the we [22:01] deployed the Maven AI slackbot and it [22:03] started they were using it there to kind [22:05] of resolve issues and collaborate. Then [22:07] the sales people got wind of it and then [22:10] and it became the the most popular tool [22:12] that the sales folks were using to kind [22:14] of figure out how the product works. So [22:15] the this stuff is very but being part of [22:17] the workflows actually results in more [22:19] of the stuff spreading and having [22:20] learning. But you got to really realize [22:23] that you have to be part of people's [22:25] workflows. It's because human change is [22:27] is is really hard. [22:28] >> That's great. Thank you, Sammy. Now, I'm [22:31] gonna give uh these two a little heads [22:34] up that I'm going to ask them a question [22:36] in a moment, but I'm going to go to [22:37] Michael. But after Michael is done, one [22:40] of the things I'd like the two of you to [22:42] address that Sammy raised is number one, [22:44] an excessive amount of data that's [22:47] available. How do you manage that data? [22:48] And the other thing, and we heard this [22:50] again in the keynote and you brought it [22:52] up, is your frontline workers are now [22:55] using the AI systems. Can you address [22:57] after we hear from Michael, can you [22:58] address um how that's impacted [23:00] employment levels? And before we go to [23:03] you, let's hear from Michael. [23:04] >> I'm going to be really quick. Uh so [23:06] essentially, one of the things when [23:07] people were talking brought up to mind, [23:11] good to great, it's a great book. They [23:13] had the concept of a hedgehog. One of [23:14] the things that you're going to need to [23:16] do while you're implementing AI is [23:17] evaluate your partner ecosystem and see [23:19] if they're going to go along for the [23:20] ride for you. Do not be afraid to [23:22] jettison partners. One of the things [23:23] we've noticed, especially as we started [23:25] implementing aic, is that some of the [23:27] vendors that we are using for products [23:29] for platforms aren't keeping up with the [23:31] curve. They have concurrent seat license [23:32] models. They have they're not putting [23:34] out MCP for a while. They have their API [23:36] calls are garbage. One of the things I [23:39] would say is while you're evaluating [23:40] that, make sure you're evaluating your [23:42] partners. It's not just the idea of [23:44] upward management outward or downward. [23:46] You're also managing the partners. [23:47] They're either along the ride in the [23:48] vision for you or they're not. And if [23:49] they're not, be prepared to make a [23:50] change. [23:51] >> Excellent. Thank you, Sushma. [23:54] Did you want to uh can you talk to us [23:56] about data management and your frontline [23:58] workers? Did you see a lot of I mean, [24:01] I'm sure they had some stress when you [24:03] told them you were going to be [24:04] implementing AI. Did it change their [24:06] employment levels? Did we did you see [24:09] the kind of job losses that people are [24:11] forecasting? [24:13] Not yet. Not yet. I mean that would be [24:16] the short answer. I think the way we've [24:19] approached it has been an enabler versus [24:22] a replacement. And so in terms of the [24:25] way we are picking up use cases, it's [24:27] basically how do you empower each of [24:30] these frontline workers to do their jobs [24:32] better and more effectively and more [24:34] efficiently. uh in terms of all the use [24:36] cases that we have [24:38] >> and so when they see that happening with [24:39] their co-workers it probably makes them [24:41] more eager to embrace AI because now [24:44] they don't feel threatened. [24:45] >> Yes. And one of the things that we do do [24:47] is I think you know we talked about user [24:49] adoption user experience um it's [24:52] extremely important in the beginning we [24:54] had very little user adoption experience [24:56] for us to really see the impact of the [24:58] tool and you know again it was our [25:00] business partners our change agents who [25:02] actually went around you know talking to [25:04] them about how it can actually help them [25:06] versus take their jobs away and that [25:08] really really helped us and and the one [25:11] thing that I would want to say from a [25:14] from My use case um experience is that [25:18] when frontline workers are actually [25:20] using the tool and they see the value [25:22] that it provides, it generates a lot of [25:24] power users [25:25] >> and these power users become really the [25:27] voice for us in the organization to talk [25:30] about how actually it's making their [25:31] lives much more simpler and making our [25:33] lives easier. [25:35] I love that enabling the power users or [25:37] what we call first adopters early [25:39] adopters to to share that excitement and [25:41] everybody in this room you all are early [25:43] adopters right we're still so early in [25:45] this process Raj can you address that at [25:47] discount tire [25:49] >> yes um you know a couple of our panel [25:51] members um had embarked on you know the [25:54] prioritization of use cases right I want [25:57] to just qualify that a little bit um at [25:59] least what we did is you know write a [26:01] criteria for that um which is you know [26:05] it's almost about four things. One is [26:08] you know does it increase your revenue [26:10] because use cases are going to pour it. [26:12] So these criteria is going to be useful [26:13] for you to vet out what makes the cut [26:16] improve revenue does it have cost [26:18] savings in it and then um does it [26:21] increase your security posture and then [26:24] most importantly you want to know [26:25] whether do you need AI for this use [26:28] case. So that's the four things that we [26:30] are trying to filter out and if it's a [26:32] combination of everything gets a higher [26:34] priority. So I think that's going to [26:37] create you know a sort of a really you [26:40] know value ad for going back to the [26:42] executive sponsors and saying it's [26:44] meeting all the criteria this is [26:45] something that we want to embark on and [26:47] by doing so I think you know keep in [26:49] mind that internal first external [26:52] probably after we wet out internal [26:54] employees based use cases and then we we [26:58] really drew a actually a process [26:59] mechanism for it we created three [27:01] pillars one is about governance the [27:04] first pillar it's uh you know trying to [27:06] get out reference architectures that you [27:08] can reuse. The second pillar was rapid [27:11] you know piloting and proof of concept [27:14] and then third pillar is the education [27:16] which you know some of the panel members [27:17] touched upon. All of this we created a a [27:20] sort of a process chart that feeds in [27:24] from all of this. I'll give you a quick [27:26] rundown which can be a sort of a tip to [27:29] get you guys started in generative AI [27:31] journey. you know pretty much you start [27:33] with you know in um use you know getting [27:36] the stakeholders in the room which I [27:37] mentioned and um then making sure you [27:40] waited through that uh decision tree [27:42] process for prioritizing if that use [27:44] case is value ad once you get into that [27:46] uh rapid PC mode uh we really created [27:50] sort of three things out of that um [27:53] build area one is the use case diagram [27:55] you can use rational for it second is a [27:58] technical flow diagram you can use lucid [28:00] chart whatever you EA composer and then [28:03] the cost model is a important thing [28:06] before you start trying to do a pilot [28:09] cost is you know like uh panel member [28:12] mentioned tokens it's also about every [28:16] line in your technical flow diagram is a [28:18] cost you need to be able to say what's [28:20] the entirety of your uh you know total [28:23] cost of ownership before you try and [28:25] build the P even after you build the P [28:29] what we found out is that you know what [28:31] you have in paper in the cost model is [28:33] not what the AI is doing. It has other [28:35] hidden costs that you will embark. So [28:37] start small with you know canonical [28:40] deployment and and then try and see if [28:43] the cost is something that you can [28:44] factor before you project for the [28:46] entirety. From a data standpoint what we [28:48] did is you know we tried to stay close [28:50] to where the source of truth is not [28:53] trying to uh go away from the [28:56] centralized data models. We are trying [28:58] to still evaluate but uh stay close to [29:01] your you know source of truth. If you [29:03] are on a platform that is having the [29:05] source of truth try and stick to their [29:08] AI uh components and so you can build [29:11] faster. So a couple of tips. [29:14] >> Wonderful. Thank you so much Sammy. So [29:19] I mean I'll I'll I'll cover kind of two [29:21] parts the data and and and and the the [29:23] one thing that we have observed and this [29:26] isn't just in customer experience but [29:27] across fun the functions are blurring. [29:31] So for example in customer experience [29:33] uh the the general challenge is that [29:35] it's a a a business unit that actually [29:39] scales linearly with the number of [29:40] customers. So you get more customers and [29:42] then the CFO is always unhappy because [29:45] the CEO is trying to grow the business [29:47] and then you do the numbers and you're [29:48] like, "Okay, we're going to go from [29:49] 1,000 to 7,000 people based on the [29:52] historic model." So it's it's it's it's [29:54] a common problem that you hear because [29:56] this function hasn't scaled. The good [29:58] news is is that the the AI stuff removes [30:00] all the busy work and we've seen most of [30:03] our customers do one one of two things. [30:05] Some do some workforce reduction, but [30:08] and very rarely do we see that. most of [30:10] the time is that what they're what they [30:11] end up doing is they're using the save [30:15] time to actually increase the number of [30:17] customer interactions. So they become [30:18] more like success folks. So the function [30:21] turns from a cost center to a revenue [30:24] center because now they're increasing [30:27] they're actually can are capable of [30:29] doing white glove experiences across the [30:31] board. So so that's kind of like ma [30:33] mainly on our customers. talk a little [30:36] bit because I'm I'm the CTO so I run I [30:38] run the product and edge team at at at [30:40] the company. We're seeing an very [30:42] interesting dynamic too. So anyone who [30:44] runs an engineering world there's a [30:46] historical ratio of product versus edge [30:48] versus UX. All that now is out the [30:50] window. When we were at Google it was [30:52] like 20 to1 for product. And right now [30:54] because the just got so productive that [30:58] ratio is completely off. So you take [31:00] that okay how do you address that? Do [31:01] you just hire more product managers or [31:03] do you scale up the engineers? And this [31:04] is an area where to be honest I think [31:06] we're still trying to figure out what [31:08] the future of engineering organizations [31:10] look like. [31:11] So so that's kind of like my advice is [31:15] actually use the savings to drive more [31:19] business value as opposed to it's just a [31:21] cost reduction and try to think about it [31:22] in terms of a cost center to a revenue [31:24] generating center because you can [31:26] actually achieve that. Um, on the data [31:28] side, one of the most common things we [31:32] hear from every customer, every any [31:35] enterprise level customer is, "Our data [31:36] is a mess. It's all over the place. We [31:40] don't know what's good. We don't know [31:41] what's bad. Come back to us in a year [31:44] when we've when we clean up the data and [31:47] we're like, "Oh my god, okay, we don't [31:49] want to wait a year because we want to [31:50] sell them our solution." So we actually [31:52] architected our product in a way such [31:55] that it is not about let's put [31:57] everything into a data lake and then [31:59] clean the data because that's just not [32:01] reality. That doesn't happen. People [32:03] don't have the appropriate staffing. And [32:05] so the way we thought about that is that [32:08] you can apply approaches that aren't [32:10] generative AI that actually can address [32:12] that. So in our case, we actually [32:14] flipped it into a search problem. a like [32:17] a I mean rag is probably the simplest [32:19] form but we've taken it to like a few [32:21] level notches higher where we're [32:23] actually we call it a graph of record [32:26] where we're actually pointing to your [32:27] data wherever it sits as opposed to [32:30] expecting it to flow into our overall [32:32] system and there's engineering [32:34] techniques to achieve that that give you [32:36] the high level quality and very high [32:37] time to value because you don't have to [32:38] clean the data and I'll give you like a [32:40] simple example uh with SAS products [32:44] there's a lot of product documentation [32:46] The most common problem is there's [32:48] conflicting documentation because the [32:49] stuff changes the old docs are. So for [32:52] example, we introduce a signal called [32:53] freshness. [32:55] So when the thing comes in when a in an [32:59] intent or a query comes in, we pretty [33:00] much find the most recent document [33:02] because it's the most likely answer. [33:03] That's I take the first part. The second [33:05] part is is if you're thinking about [33:07] creating systems that solve real [33:10] problems, real use cases, real value. [33:12] One of the biggest gaps we found in this [33:16] in this world is that you need an [33:18] evaluation framework to actually measure [33:21] if the stuff's working. [33:24] And that and the beautiful thing is if [33:27] you have a metric, you can move it. If [33:30] you don't have a metric, you're flying [33:32] blind. If you have a metric, you can [33:34] actually build an ROI case. And so one [33:38] of the very first things we did at Maven [33:41] and this is actually like we didn't [33:43] invented Maven like when I was at Google [33:45] I used to spend an obscene amount of [33:46] money on human evaluations obscene [33:48] amount of money where I had a team of [33:51] data scientists I have the best AI I [33:53] have the best ranking algorithm and [33:54] you're like I can't deal with best I [33:56] need like real data so we used to run [33:58] these side by side so we actually [33:59] created an agentic version of that and [34:01] that pretty much gave us the metric the [34:03] beautiful thing about a metric is that [34:06] once you have it. You can actually start [34:08] to see the gaps in enterprise. There is [34:10] so much data, so many APIs, and when you [34:13] start to embark on a project, you're [34:14] probably looking to integrate into an [34:16] insane amount of systems without the [34:19] data. And so when you have the data, you [34:22] can start to see where the gaps are and [34:25] then you can actually focus on the [34:26] things that have impact. So when we kind [34:28] of work with our customers, we can [34:30] clearly articulate to go from 73.5 to [34:33] 90.6, six, we need access to these three [34:36] systems and these fields on the on on on [34:39] the user record. And because there's [34:41] there's a metric, you can actually look [34:43] at if you're spending $40 a ticket per [34:46] or per customer interaction, you can [34:48] actually put a dollar amount around it [34:49] and you can actually build a very clear [34:51] ROI use case which solves a whole bunch [34:54] of issues with our IT friends [34:56] because now you're coming in and and [34:58] clearly saying like this is going to [35:00] save this many hundreds of millions of [35:02] dollars. Now this is a priority and you [35:05] can actually build you pretty much put [35:07] the business case on the silver platter [35:09] because people don't like giving access [35:10] for the sake of access. It needs to be [35:13] grounded in a business case. [35:15] >> Thank you Sammy and uh follow Sammy on [35:19] LinkedIn because you can see he is eager [35:21] to help you with your implementation. Um [35:24] before we go to the audience questions [35:27] I'm going to uh two things. is one of [35:30] the things I love your comment about [35:31] white glove treatment with the extra [35:34] time because if anybody's uh gotten [35:36] caught into a conversation with me, you [35:37] know that that is what I talk about all [35:39] the time. How do you use this save time? [35:42] Don't cut the employees. Um and then [35:44] thinking about white glove treatment. Of [35:45] course, Sushma is brilliant and [35:48] wonderful. And where do you want white [35:49] glove treatment? It's with your [35:51] insurance and financial services people [35:53] for sure. So, um, I'm going to turn it [35:55] over to to you and then we'll roll out [35:57] to questions from the audience. [35:59] >> I just wanted to, you know, really echo [36:02] what Sammy said about data being a mess. [36:04] And one thing that I have realized over [36:06] time is that typically AI strategy [36:09] followed data strategy. And, you know, [36:11] everybody was used to having this notion [36:14] that, you know, you have to get your [36:16] data in pristine order before you can [36:18] build anything related to AI. But I [36:21] think that's gone out of the window to [36:23] what you're saying and we're really we [36:26] are actually influencing the data [36:27] strategy through the AI verticals that [36:29] we are building because we're really [36:30] understanding where the gaps are and [36:32] building that data strategy for them. So [36:34] that's something that I've discovered [36:36] along the way and part of the influence [36:37] that we were doing organizationally [36:40] is building bring our data partners [36:43] along in in this journey versus saying [36:46] versus them asking us to wait till the [36:48] data is in place. So that's something [36:50] that you want to have concurrently. [36:52] Excellent. Thank you so much. Okay. So [36:55] do you have a microphone that you're [36:57] going to bring around? Um so [37:01] does anybody have any questions? You [37:03] have a question for us, right? Comment. [37:05] >> So yeah, I understand the part of the uh [37:08] data thing is coming from the as part of [37:11] it when we starting out the AI strategy, [37:13] right? So being my company is been there [37:15] for a legacy company been there for a [37:18] couple of year like couple of decades [37:19] more or three decades more and when we [37:21] are moving out of it right when [37:22] initially that data is not the main [37:24] source and it it it became part of it [37:26] initially in recently uh things so how [37:29] you want to handle it like in case if it [37:31] is there for more than the gates of [37:33] company like the large it's not a [37:36] monolithic right everything is [37:37] monolithic now we are going to [37:38] modelizing this applications how we are [37:41] thinking in that way like how we want to [37:43] take it to the gen involved in that part [37:45] like when you if it is a monolithic and [37:48] how you want to implement that part [37:49] here. Um, so in in in our case [37:55] pretty much because we flipped it into a [37:57] search problem, [37:59] monolithic is not a problem because it [38:01] be we've turned it into what we call a [38:03] ranking problem is like how do you [38:05] actually understand structured and [38:06] unstructured data such that you're able [38:09] to identify the appropriate [38:11] like where the answers are. And the the [38:15] interesting thing is once you go down [38:17] this path of it's becoming a search [38:19] problem, it actually turns a [38:20] non-deterministic system which is genai [38:22] and machine learning which is like you [38:24] don't know where stuff's coming from to [38:26] a very deterministic system because [38:29] every time you get an answer you can [38:31] actually get we call it in ranking [38:33] provenence which is the referenceability [38:35] back to where the stuff is coming from [38:38] which gives you tremendous insight when [38:42] the folks who are domain experts look at [38:44] the data are able to identify. So when [38:46] so when you say it's like it's a [38:48] monolithic thing, it's like that's the [38:49] reality. You're not the only one. [38:51] Everybody's in this mess and you have [38:53] tons and tons of data. So it's it's so [38:56] one approach is like let's go clean it [38:58] up or just my my my perspective is [39:01] embrace it like make it part of how [39:03] you're thinking about it as opposed to [39:05] let's do a massive transformation and [39:07] then use the evaluations the data to [39:11] actually help you triage where you need [39:14] to invest. This is actually what like [39:16] the the other half of our platform is is [39:18] one where when things are going off the [39:21] rails will actually say why and in some [39:24] cases help you identify the gaps because [39:26] this data management problem and and [39:28] it's a it's it's a very overwhelming [39:30] problem if you're looking at it from [39:33] like looking at it holistically. trying [39:35] to kind of focus the use cases and then [39:38] thinking about how to approach it with [39:39] impact is a much faster way to value [39:43] because it's very easy to get lost in [39:45] things that don't have value. [39:47] >> I think Michael wants to share some of [39:48] his uh expertise with you as well. [39:51] >> Yeah, thank you. So basically the other [39:52] thing too is you know looking at it [39:54] taking a step above looking at the [39:56] situation where you're at Macy's if you [39:58] look at the history of the company has [39:59] evolved from just a retailer to now you [40:01] could classify as a technology company. [40:03] That means data is being used in [40:04] pockets. people are using it for certain [40:06] reasons. They don't know how to connect [40:07] the dots and you as a change agent can [40:09] help identify those areas. So just pick [40:12] out the use case. Once you identify the [40:14] use case, what are the pockets of data [40:15] that you need? Talk to those people with [40:17] data because now you can make their [40:19] lives easier. Get what you want out of [40:20] it. Make them change agents. And really [40:22] what you're doing is you're satisfying [40:24] your internal and external customers. [40:25] It's a it's more delightful customer [40:26] experience where they buy more volume, [40:28] but internally you're able to do things [40:30] faster to progress the mission of Macy's [40:32] to become what it needs to be. So that's [40:34] the other thing to take advantage of too [40:35] is take a survey as who's using what [40:37] around you. Find out use cases that [40:39] going to help them that align to the [40:41] strategy and help promote what needs to [40:43] be done. That's going to buy. [40:44] >> Hello Daniel Ninsky from ECMC. Uh not to [40:48] be conspiratorial, but I am so deeply [40:52] unconvinced that AI is taking as many [40:54] jobs as companies are claiming. Uh, I [40:58] feel like especially for engineers, it's [41:01] a market palatable [41:04] excuse to correct for the glut of hiring [41:08] anyone that said engineer in their [41:10] resume in 2020. Uh, and I feel like it's [41:15] hurting the trust and like belief that [41:20] AI can be helpful. I feel like it's [41:22] creating a narrative that AI is going to [41:24] create or take people's jobs when I feel [41:27] like at the scale it's not taking as [41:30] many jobs as people are attributing it [41:34] to. What is your thoughts? [41:36] >> I'm going to start with Raj and then go [41:38] to Michael with this one. [41:39] >> Yeah. So I was alluding that human [41:42] factor, right? So humans are good in [41:44] connecting dots. Um AI is definitely an [41:47] enabler, right? From that aspect I think [41:50] u depending on where you are you know [41:52] whether you're a student you're an [41:53] intern you're in a early engineer you're [41:56] a you know senior engineer what wherever [41:58] you are in the journey I think um in IT [42:02] or in the industry we've seen technology [42:06] revamp itself right every 5 years it's [42:08] something or the other 5 years back it [42:10] was cloud today it's AI you know people [42:12] are trying to see whether AI is you know [42:15] needed can it help will replace me. All [42:18] that is there. Same happened with cloud. [42:22] 5 years back people were like you know [42:24] holding on to the data like their teddy [42:26] bears in data centers. Not anymore. So I [42:29] think if you fast forward right I'm [42:32] giving a nonlinear answer but if you go [42:34] fast forward yourself into future AI is [42:36] going to be there. All of us will be AI [42:39] enabled right we'll be using it [42:42] including interns including engineers [42:44] everybody is going to be there. Humans [42:46] are going to be good good at connecting [42:47] dots and they are never going to be [42:49] replaceable, right? So I think the [42:51] important factor is to be aware of how [42:53] to use it, how to train yourself, how to [42:57] make sure you get certifications [42:59] um and then use it, try it and and then [43:02] you know enable and be the evangelist. [43:05] That's what I would say. [43:06] >> So we have one minute left and I am [43:08] going to hand it over to the amazing [43:10] Michael to close us. [43:13] So first off, thank you for the [43:14] opportunity for being on the panel. [43:16] Great teamwork, great, you know, it's [43:18] going to be a long long panel [43:20] afterwards. Uh, so one of the things I [43:22] wanted to say is, you know, thanks for [43:23] attending and I hope you learned a lot [43:24] about practical effects of implementing [43:26] AI, but also the effects on the company [43:28] and how you can play a role. The idea is [43:30] that we've seen this movie before. New [43:32] technologies have been introduced. All [43:34] of them said we're going to make it more [43:35] efficient. Executives will overreact by [43:37] cutting jobs, hiring them back in [43:39] different roles. We've seen the movie [43:41] before, but the goal here is that [43:43] everybody needs to be very, you know, [43:45] you're not just learning AI, but you [43:47] also have to learn it, absorb it, and [43:49] then go teach and preach AI. You got to [43:52] be the change agents. You got to be the [43:53] ones that if the company has a mission [43:55] to get to a certain level, get them to [43:57] that certain level through AI, through [43:58] your practitioners, through your ability [44:00] to manage internally or upwards, [44:01] downwards, internal, external. [44:04] Always keep learning. But at the same [44:05] point in time, one of the things I hope [44:07] this motivates, I saw a lot of people [44:08] taking notes in here. What's the one way [44:10] that you guys can collaborate, share [44:12] your experiences? You guys have your own [44:13] panel discussions and come away with [44:15] some practical takeaways that may not be [44:17] technologically oriented and maybe I [44:19] have a really awful supervisor. How do I [44:20] deal with them? That's a great [44:23] discussion to have and that's one that [44:24] needs to happen in conferences like [44:25] this. So, with that in mind, you know, [44:27] you've got four resources up here, five [44:29] resources up here that would love to [44:30] talk to you more, love to have you, love [44:33] to talk to you, love to speak with you, [44:34] and more to the point, love to learn [44:35] from you. We don't know everything. So [44:38] you guys probably know more than we do. [44:40] So [44:40] >> thank you. Thank you so much to each of [44:42] the panelists. Thank you to AI4 for [44:45] having us. Daniel and our moderator, our [44:48] >> Let's give our lovely panel a round of [44:50] applause. That was awesome. [44:52] [Applause]