J&J, Novelis Target Back-Office AI for Fast ROI as Plant Scale Lags

工厂规模化落地滞后，J&J与Novelis瞄准后台AI以求快速获取ROI 制造商通过采用自然语言工具实现行政与供应链工作流自动化，在不到六个月内获得正向回报，而车间端自动化仍受制于分散的传统系统与员工阻力。 来自工业回收、医疗制造和软件自动化领域的企业领袖探讨了AI从孤立的概念验证走向全球生产级部署的转变。尽管重工业工厂利用“人在回路”的AI技术进行预测性维护、流程稳定性控制与车间安全管理，但专家组指出，文档处理、供应链分流以及语音驱动的订单接收等后台工作流，往往能在不到六个月内带来最快的资金回报。要克服碎片化的传统IT系统、工厂层面的运营惯性及严格的监管障碍，离不开深度的统一数据治理和自然语言交互界面。展望未来，与会嘉宾预计多智能体编排与集成GenAI的数字孪生将成为实现工厂自主化运营的核心基石。 生成式 AI 通过攻克传统规则型 RPA 无法解析的复杂非结构化数据，使流程自动化重获新生，并推动制造巨头跨越试点走向“10到100”的全球规模化落地阶段。 (0:39) 在重工业与材料制造中，AI 目前主要充当“人机协同”（Human-in-the-loop）的辅助副驾驶，专注于优化设备设定参数、预测性维护与车间安全预警，而非实现全自主接管。 (4:44) 相较于周期长达 3 至 5 年的车间级重构，优先切入订单录入、运单解析与发票核对等后勤瓶颈，能在 6 个月内快速盘活被占用的营运资金并验证落地 ROI。 (7:10) 制造业规模化应用 AI 面临三大核心阻碍：陈旧孤立的遗留 IT 系统与数据质量缺陷、车间操作员因技术畏难与失业顾虑引发的变革阻力，以及明显滞后的行业合规监管。 (19:31) 推行“以英语作为代码（English as code）”的自然语言工作流，使非技术业务人员能够直接编排与调试复杂的业务自动化，在规避技能断层的同时满足企业级审计与治理要求。 (27:51)

Industrial AI delivers its fastest and most dependable returns when manufacturers bypass complex shop-floor overhauls to first automate fragmented back-office workflows, while treating plant-floor AI strictly as a human-in-the-loop co-pilot constrained by legacy infrastructure and regulatory lag. Unlocking Working Capital Through Back-Office Automation (7:10) While heavy industry historically concentrated automation on factory hardware, administrative and operational back ends remain heavily manual. Mahesh Vinayagam observes that mid-tier suppliers frequently trap vital working capital in slow order entry, delayed invoice matching, and sluggish supplier disbursements. Automating document-heavy tasks—including bills of lading, procurement matching, and freight intake—delivers rapid financial relief without disrupting physical operations. During a live demonstration, an interactive voice agent built for Polaris illustrated this efficiency by booking freight in real time, querying CRM and transportation management systems, and seamlessly converting mixed metric and imperial measurements. Trevor Seim notes that while approximately 83% of manufacturers are launching AI deployments, analyst projections from firms like Forrester indicate enterprise-wide operational transformations typically take three to five years to break even. In contrast, pragmatic back-office automations frequently achieve full payback in under six months. These quick payoffs allow organizations to de-risk investment, train staff, and establish the institutional trust required before attempting complex shop-floor overhauls. Factory Floor Co-Pilots and Cultural Inertia (12:04) Within heavy production environments, AI currently functions as an assistive advisor rather than an autonomous controller. Chirag Agrawal compares Novelis's AI adoption across its roughly 30 plants to electric vehicles a decade ago: operators treat AI as a secondary assistant for machinery set-points while keeping final decisions in human hands. Novelis moved away from fragmented, one-off forecasting projects to align deployments with core executive goals, focusing on predictive maintenance, visual plant safety, and raising recycled aluminum content from 63% to 75%. Scaling successful models across multiple facilities encounters severe friction from localized habits and operational divergence. Agrawal likens deploying an established model into a new plant to teaching a robot to dance on unfamiliar terrain, demanding extensive retraining for different machinery setups. Furthermore, long-tenured operators often resist procedural changes. Panelists argue that overcoming this inertia requires showing frontline workers that automation removes repetitive paperwork, allowing them to move up Maslow's hierarchy of work into higher-value operational problem-solving. Navigating Legacy IT and Regulatory Bottlenecks (19:31) Enterprise-scale AI remains constrained by decades of technological and structural fragmentation. Jie Yu compares deploying state-of-the-art machine learning models on existing plant IT to driving an advanced autonomous vehicle on last century's highway system. Historical mergers and acquisitions have left enterprises managing dozens of disconnected ERP and mainframe instances. To prevent low-quality inputs from derailing sophisticated models, Johnson & Johnson established an enterprise Data Management Council to standardize data readiness across its pharmaceutical and MedTech sectors. Regulated industries face additional friction from compliance mandates. In pharmaceutical manufacturing, where AI is applied to minimize batch-to-batch variability and protect product yields, Yu notes that health authorities like the US FDA and EMA often lag years behind technical capabilities. To bridge internal technical skills gaps without forcing plant operators through exhaustive software engineering training, Trevor Seim advocates an 'English as code' framework that enables non-technical personnel to configure, audit, and direct automated workflows in natural language under enterprise governance controls. Supply Chain Resilience and Multi-Agent Orchestration (40:14) Beyond internal plant boundaries, AI adoption increasingly focuses on stabilizing multi-tier supplier networks. Managing Johnson & Johnson's network of 40,000 to 50,000 direct and indirect suppliers requires predictive AI models to detect raw material shortages and run real-time simulations against volatile daily tariffs. For routine administrative drafting, generative AI tools provide pre-populated baselines that complete 50% to 80% of complex Bills of Materials and requirements documents, drastically reducing manual engineering overhead. Looking across the next 12 to 18 months, panelists predict an architectural transition from isolated point tools to federated multi-agent systems. Rather than navigating rigid individual software platforms, human users will issue queries to supervisory agents that orchestrate specialized procurement, maintenance, and legal bots in parallel. In high-compliance manufacturing, industrial digital twins are progressing from passive monitoring dashboards into active supervisory control systems, forming the foundational architecture for eventual autonomous factory operations. While multi-agent coordination and active digital twins outline the path toward autonomous manufacturing, realization remains bounded by dirty legacy systems, slow regulatory modernization, and plant-level procedural variance. The panel concludes that enterprise value depends on treating AI adoption as a sustained one-to-three-year organizational transformation rather than an immediate plug-and-play solution.

在制造业的数字化转型中，行业重心正从单点车间试验转向兼顾人机协同与后台运营的系统化落地。与会专家指出，尽管现场重工制造依托AI实现参数辅助与安全保障，但在非结构化文档处理与供应链后台实现自动化能够带来更迅速的投资回报（通常在六个月以内），并为未来多智能体协同与自主运营奠定基础。 后台运营与营运资金的快速变现 (7:10) Kognitos销售主管Trevor Seim指出，约83%的制造企业正积极启动AI项目，但分析机构Forrester预测企业级全面投资通常需要三到五年才能见到实质成效。与周期长、耗资大的工厂车间改造相比，qBotica创始人Mahesh Vinayagam强调后勤与供应链管理等后台支持环节构成了更高杠杆的切入点。传统机器人流程自动化（RPA）在面对复杂非结构化数据时频频失效，而生成式AI使订单录入、发票核销及提单处理的自动化成为可能，能够盘活滞留在低效流程中的企业营运资金。 在大会现场演示中，针对Polaris构建的语音AI助手展示了多系统实时协同能力：不仅能毫秒级统一公制与英制单位换算（如将2公吨自动换算为4409磅，规整多单位混合的货运托盘尺寸），还无缝调用HubSpot客户关系管理系统与运输管理系统计算费率并生成货运单号，同时严格执行危险品拦截安全合规规则。与会代表认为，这类文书起草与流转自动化可在六个月内收回成本，先行为企业构建AI应用信心与经验。 车间现场的人机协同与差异化定位 (12:04) Novelis全球数据科学负责人Chirag Agrawal将重工业当下的AI应用比作十年前的特斯拉电动车，即AI主要扮演辅助专家操作员的“副驾驶”角色，而非完全自主接管控制权。Novelis在淘汰单点孤立的局部预测模型后，转向由上至下推动的高复用性企业目标，例如利用预测性维护模型和计算机视觉安全系统，将旗下约30座工厂的再生铝使用比例从63%提升至75%，并削减总体能耗。操作人员依然对关键设备的设定参数拥有最终决策权。 强生（Johnson & Johnson）数字化与数据科学产品管理主管Jay Yu则对比了石化与医药制造的不同价值导向。在其先前就职的壳牌（Shell）等石化企业中，传统机器学习主要聚焦于严控排放超标、保障重工资产可靠性及杜绝危险事故；而在强生涉及100多个国家和地区的医药与医疗科技生产中，AI的核心价值在于消除批次间差异、保障高效药物稳定性，并借助数字孪生缩短端到端制造周期，提高良品产出率。 规模化落地的IT基建、文化与合规阻力 (19:31) 当企业试图将单一试点的AI方案推广至数十座工厂时，面临着显著的非技术与技术双重障碍。Jay Yu用“在上一世纪的公路系统上驾驶最前沿的自动驾驶汽车”来比喻底层IT基础设施与先进大模型之间的脱节。许多制造企业的核心系统由分散的老旧大型机、多套SAP和Oracle系统混杂拼凑而成，数据质量低下极易导致“输入垃圾，输出垃圾”，迫使强生专门成立全企业数据管理委员会以确立跨部门治理标准。 此外，文化惯性与合规壁垒构成了更大的人为阻力。Agrawal指出，各工厂长期固化的操作习惯（如“这套系统我已经用了二十年”）导致不同厂区执行相同流程的差异极大，将模型移植至新厂犹如“训练机器人在全新地形上跳舞”。车间工人对技术替代就业的恐惧也加剧了抵触心理。在监管层面，医疗行业的AI应用必须同时满足美国FDA、欧洲EMA等多国卫生机构的严格审查，而法规更新速度往往大幅滞后于技术创新，显著延缓了全球化推广进程。 供应链韧性与多智能体架构演进 (40:14) 针对跨国制造企业复杂的供应网络，强生等企业通过与外部技术供应商合作，运用预测与生成式AI监控管理全球多达5万至6万家各级供应商。面对几乎每日波动的关税政策及宏观供应链冲击，AI情景模拟能够协助企业动态调整采购路线以规避供应断供风险。同时，生成式AI能够自动预填物料清单（BOM）与产品需求文档50%至80%的初稿内容，使研发与工程团队摆脱“白纸起步”的繁琐消耗。 展望未来12至18个月，行业正加速从静态的单点软件过渡至动态互联的多智能体架构（Multi-Agent Architecture）。Kognitos销售主管Trevor Seim倡导“以自然语言为代码”，允许业务人员直接用人类语言编排数字员工。未来工厂将不再依赖繁琐的人工跨系统查询，而是由主控智能体实时协同采购、质量控制、法律审核及预测性维护等多个专用智能体；结合生成式AI赋能的数字孪生系统，这种协同架构被行业专家视为制造流程走向端到端自主化运行的关键基石。 综合来看，制造业的AI转型仍处在从辅助试点向全面规模化扩展的渐进期，其短期实质回报高度依赖后台行政与供应链流程的疏通。然而，老旧遗留系统的重构成本、全球法规标准的滞后性，以及如何在人机协同中建立一线员工的长期信任，依然是迈向真正端到端自主制造前有待持续解决的核心课题。

## Background 本视频由 YouTube 频道 Ai4 于 2025 年 8 月 28 日发布。内容记录了一场以“制造业运营中的 AI 驱动变革”为主题的专家研讨会，汇聚了来自强生（Johnson & Johnson）、诺贝丽斯（Novelis）、Alpha Edison、Kognitos 及 qBotica 的业务与数据科学领袖，共同探讨 AI 在提升工厂良率、预测性维护、中后台流程自动化以及全球多工厂规模化推广中的实际应用与挑战。 ## TL;DR 圆桌嘉宾深入剖析了 AI 在现代制造业中的全方位价值落地路径，强调除了传统的车间调优与设备维护外，发票核对、订单处理及覆盖数万家供应商的供应链韧性管理等后台业务流程同样是实现数月内快速 ROI 的绝佳切入点。讨论重点揭示了当前规模化部署的最大阻碍并非算法本身，而是老旧系统的“数据高速公路”落后、跨工厂业务流程碎片化以及一线操作人员的信任与心理惯性。面对这些挑战，企业正通过“人在回路”的辅助机制、低门槛自然语言交互以及生成式 AI 智能体逐步化解阻力，并展望未来 18 个月内多智能体协同与数字孪生融合所催生的工业自主化运营前景。 ## Atmosphere & style 整场研讨会兼具工业一线的务实严谨与前沿探索的活跃张力，嘉宾交流坦诚而幽默，频频抛出“开着豪华自动驾驶跑在破旧公路”、“让机器人跨地形跳舞”等生动比喻。现场穿插的实时语音 AI 智能体下单与安全边界压力测试更是将气氛推向高潮，既展现了工业软件落地的现实骨感，也传递出跨国巨头与技术先锋合力破局的笃定与信心。 ## Key takeaways - [00:39](https://www.youtube.com/watch?v=vp5lAR2pawE&t=39s) 强生等跨国巨头将数字化与 AI 转型划分为“0 到 1（概念探索）”、“1 到 10（局部验证）”与“10 到 100（全企业规模化）”三个阶段，目前已进入深度规模化扩张期。 - [04:43](https://www.youtube.com/watch?v=vp5lAR2pawE&t=283s) 重工业制造领域（如 Novelis）将当前 AI 定位为“辅助工具”，强调在车间工艺优化与安全隐患排查中始终保持“人在回路（Human-in-the-loop）”。 - [07:11](https://www.youtube.com/watch?v=vp5lAR2pawE&t=431s) 制造业长期忽视的后台运营（发票核对、结算、订单跟踪）通过生成式 AI 自动化重塑，能够显著压缩业务周期并直接盘活企业的营运资金（Working Capital）。 - [09:38](https://www.youtube.com/watch?v=vp5lAR2pawE&t=578s) 不同制造业领域对 AI 的价值诉求各异：石化行业聚焦安全、减排与设备可靠性；制药与医疗科技行业则侧重控制批次变异性、提升良率并加速交付。 - [12:04](https://www.youtube.com/watch?v=vp5lAR2pawE&t=724s) 企业推进 AI 落地应彻底告别零散孤立的单点试点，转向与顶层核心战略对齐、且能跨数十家工厂快速横向复制的规模化场景（如全集团预测性维护与能耗优化）。 - [15:20](https://www.youtube.com/watch?v=vp5lAR2pawE&t=920s) 为破解大型 AI 投资通常需要 3 至 5 年才能见效的困局，企业可从后台运营或提单管理等小场景切入，以低门槛快速收获 ROI 并建立组织对 AI 的信任基石。 - [16:56](https://www.youtube.com/watch?v=vp5lAR2pawE&t=1016s) 跨工厂规模化推广 AI 的最大阻力在于“人的行为惯性”以及多厂区 ERP 系统与流程的严重碎片化，企业可借助流程挖掘（Process Mining）技术精准定位切入点以打破僵局。 - [19:02](https://www.youtube.com/watch?v=vp5lAR2pawE&t=1142s) 底层遗留 IT 系统与前沿大模型存在严重脱节（“豪华跑车行驶在上世纪的高速公路上”），建立跨部门的企业级数据治理委员会是避免“垃圾进、垃圾出”的前提。 - [21:12](https://www.youtube.com/watch?v=vp5lAR2pawE&t=1272s) 车间一线工人易因技术认知匮乏而将 AI 视为生计威胁，加之跨国行业监管（如 FDA、EMA）显著滞后于技术演进，变革管理与合规构成了落地的双重非技术阻碍。 - [27:51](https://www.youtube.com/watch?v=vp5lAR2pawE&t=1671s) 降低企业 AI 采纳门槛的关键在于“让数字员工说人话（以自然语言作为代码）”，使无编程背景的业务人员能轻松调度 AI、限制模型幻觉并建立人机协同流程。 - [31:13](https://www.youtube.com/watch?v=vp5lAR2pawE&t=1873s) 生成式 AI 智能体展现出了强大的统一协同编排（Unified Orchestration）能力，能够跨系统实时联动 CRM、TMS 并自动换算复杂度量衡，同时内置严格的业务安全护栏。 - [37:36](https://www.youtube.com/watch?v=vp5lAR2pawE&t=2256s) 面对全球数万家供应商网络与多变的关税政策，AI 驱动的模拟平台能有效预测供应链风险与优化采购路径，且后台自动化场景通常在 6 个月内即可收回投资成本。 - [40:14](https://www.youtube.com/watch?v=vp5lAR2pawE&t=2414s) 未来 18 个月到 3 年内，制造业将从孤立工具全面迈向“多智能体协同（Multi-agent Collaboration）”以及与生成式 AI 深度融合的“数字孪生”，最终迈入更高阶的工业自主化运营时代。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=vp5lAR2pawE&t=4s) 论坛开场与嘉宾阵容介绍 论坛主持人、来自洛杉矶风险投资机构 Alpha Edison 的研究与数据科学总监 Fen Zhao 为本场研讨会拉开序幕。她指出，Alpha Edison 过去在制造业和人工智能领域均有深厚投资布局，而当下最令人兴奋的机遇正是这两者的交汇点——制造业人工智能（Manufacturing AI）。为了全方位展现该领域的真实图景，本次论坛汇聚了来自初创技术供应商、全球跨国巨头以及传统重工业企业的多元化专家团队，旨在为观众带来多维度的产业洞察与实践视角。 ```mermaid flowchart TD accTitle: 制造业 AI 讨论嘉宾与生态构成 subgraph Enterprise ["大型制造与医疗跨国企业"] JJ["强生 (Johnson & Johnson)<br/>Jie Yu"] NOV["Novelis 诺贝丽斯<br/>Chirag Agrawal"] end subgraph TechEnablers ["技术与服务赋能机构"] QBOT["qBotica<br/>Mahesh Vinayagam"] KOG["Kognitos<br/>Trevor Seim"] AE["Alpha Edison (风投主持)<br/>Fen Zhao"] end Enterprise <--> TechEnablers ``` --- ### [00:39](https://www.youtube.com/watch?v=vp5lAR2pawE&t=39s) 嘉宾自述：qBotica 与 Johnson & Johnson 的 AI 实践背景 自动化即服务（Automation as a Service）公司 qBotica 的创始人兼首席执行官 Mahesh Vinayagam 首先介绍了公司的定位与演进。qBotica 致力于在软件与专业服务的交汇处释放价值，除了研发专有软件外，还与 Kognitos 等厂商紧密合作。公司业务跨越制造、医疗保健及金融服务等多个领域，主要深耕财务、会计、采购、供应链及客户运营等跨行业共享服务中心场景。Mahesh 重点谈到，过去五年传统 RPA（机器人流程自动化）在面对复杂的非结构化信息时经常彻底失效，而生成式 AI 的出现让这些工作流重获新生。他们早在 2020 年就开始接入 OpenAI 的初代模型，其实际成效已在 Polaris Transportation 等客户身上得到印证——客户在订单处理效率上大幅超越竞争对手。此外，Mahesh 还积极参与亚利桑那州立大学（ASU）的技术与创业委员会及未来商业领袖组织（FBLA），推动从学校到大学的 AI 社区教育。 随后，强生（Johnson & Johnson）数字化与数据科学产品管理负责人 Jie（Jay）Yu 介绍了这家医疗健康巨头的转型历程。强生业务涵盖制药以及医疗器械/医疗科技（MedTech）两大板块，业务遍及全球 100 多个国家。Jie Yu 的团队位于集团总部的企业级卓越中心（CoE），负责在全球范围内推动数字化与 AI 变革。他将强生的数字化转型划分为三个成熟度阶段：“0 到 1”、“1 到 10”以及“10 到 100”。目前强生已成功走过前两个阶段，正处在“10 到 100”的大规模深化扩张期。包括生成式 AI 在内的新兴技术已经全面渗透至强生的各个业务部门与细分板块，强生也正在与广泛的外部技术供应商生态携手构建数字化产品，以重塑未来医疗健康的形态。 - **qBotica 落地实效**：自 2020 年探索大语言模型以来，客户如 Polaris Transportation 在海量订单录入与非结构化文档处理上实现了显著的效率突破。 - **强生转型阶段论**：数字化与 AI 演进分为“0 到 1”（概念探索）、“1 到 10”（局部验证）与“10 到 100”（全企业规模化），目前正处于 10 到 100 的关键推进阶段。 --- ### [04:43](https://www.youtube.com/watch?v=vp5lAR2pawE&t=283s) 嘉宾自述：Novelis 与 Kognitos 的制造落地与技术愿景 全球铝轧制与铝回收领军企业 Novelis 的全球数据科学负责人 Chirag Agrawal 分享了重工业场景下的数字化现实。Novelis 在全球拥有约 30 家工厂，主要客户覆盖饮料罐、汽车及航空航天工业。Chirag 幽默地给出了一个直观的数据：“大家平时喝啤酒或汽水时，手中的易拉罐有 50% 的概率是由 Novelis 加工制造的”；甚至包括 Nespresso 咖啡胶囊的铝材也出自其产线。在 AI 落地方面，Novelis 目前将 AI 视为一种强大的辅助工具（Assistance），主要聚焦两大方向：一是保留人在回路（Human-in-the-loop）的制造流程半自动化；二是在充斥着重型机械的工厂车间利用 AI 识别安全隐患并保障工人安全。 来自 Kognitos 的销售主管 Trevor Seim 则阐述了降低 AI 使用门槛的技术理念。Kognitos 的核心使命是帮助企业以颠覆传统的方式理解并应用 AI，尤其专注于企业级业务流程自动化。其核心技术亮点在于推行“以英语作为代码”（English as code），让普通业务人员能够以自然语言直接指挥和调度 AI。同时，系统在底层构建了严格的治理（Governance）机制，确保 AI 编排在全企业范围内的可见性与可控性。Trevor 表示，他们既协助处于 AI 探索早期的企业平稳起步，也帮助迈入中后期的企业实现跨部门的大规模扩展。 --- ### [07:11](https://www.youtube.com/watch?v=vp5lAR2pawE&t=431s) 制造业 AI 的最大落地价值点：后勤与业务流程自动化 主持人 Fen Zhao 抛出了全场的首个核心问题：业界讨论智能制造时常聚焦于良品率提升、安全保障和产线优化，但在各位专家眼中，数字化与 AI 在当下究竟在哪里产生了最大的现实价值？ Mahesh Vinayagam 率先作答。他幽默地表示把车间产线调优和良率提升留给两位制造巨头的专家去讲，自己要特别为“制造业中被长期忽视的后台运营（Back Office）”发声。他指出，制造业在车间硬件自动化上起步很早，但中后台运营流程却往往成为最后被数字化的死角。大型企业或许具备一定的运营效率，但中大型及中端制造企业在发票核对、支付结算、订单跟踪及物流协调等日常流程上消耗了大量工时。 Mahesh 强调，加速后台流程能直接盘活企业的营运资金（Working Capital）——订单处理越快、供应商回款越及时，供应链生态的运转就越稳健。他举例说明了生成式 AI 与语音技术的创新结合：在展会期间有客户向他倒苦水，称驻扎在中西部等偏远地区的员工在差旅中经常遭遇数据网络信号中断的问题。而通过语音电话直接激活 AI 智能体（Bot），利用自然语言对话即可快速完成订单信息的自动抓取、审核与录入，这种无缝衔接在中后台管理、发票支付与订单处理中正释放出巨大的降本增效价值。 - **后台痛点**：传统制造企业在后勤共享服务（审批流、订单管理、供应商付款）存在严重滞后与断层。 - **资金效率**：利用生成式 AI 自动化压缩订单与支付周期，核心价值在于释放被锁死的营运资金（Working Capital）。 - **前沿交互**：结合语音 AI Bot 技术，能够克服偏远地区网络信号差的瓶颈，实现电话交互式的自动化报单与流程审批。 ### [09:38](https://www.youtube.com/watch?v=vp5lAR2pawE&t=578s) 跨行业视角：从石化巨头到医疗健康的 AI 价值驱动 Jie Yu 从自身的职业背景切入，深入探讨了 AI 在制造业中的商业影响与投资回报率（ROI）。在加入强生（Johnson & Johnson）之前，他在壳牌（Shell）工作了十余年。对比这两个截然不同的行业，他指出传统机器学习与 AI 早在数年前就已在石化领域大显身手，当时主要聚焦于改善安全记录、优化排放控制、提升可持续性以及保障系统可靠性。由于石化行业具有极高的危险性，安全与减排始终是核心诉求。 ```mermaid flowchart LR accTitle: 跨行业 AI 价值驱动与应用场景对比 subgraph Petrochemical["石化行业（壳牌）"] P1["安全记录改善"] P2["排放与环保控制"] P3["设备可靠性保障"] end subgraph Healthcare["医疗健康（强生）"] H1["控制批次差异与稳定性"] H2["提升良率与产出速度"] H3["数字孪生与预测性维护"] end Petrochemical -.->|"价值演进"| Healthcare ``` 而在医疗健康与制药行业，AI 的核心价值则体现在保证产品质量与批次稳定性上。通过 AI 技术，制造工厂能够最大程度减少不同批次药品之间的变异性，同时大幅提升运营效率与生产良率——这意味着在相同的生产周期和工艺批次下，工厂可以产出更多合格药品，更快交付给急需治疗的患者。此外，当前大热的数字孪生（Digital Twin）与预测性维护（Predictive Maintenance）技术，不仅显著缩短了端到端制造周期，更从财务回报层面全面重塑了制造运营的逻辑。 --- ### [12:04](https://www.youtube.com/watch?v=vp5lAR2pawE&t=724s) Novelis 的 AI 演进：从单点试点转向战略规模化落地 Chirag Agrawal 结合 Novelis 的实践指出，制造业的 AI 应用现状与十年前的特斯拉非常相似：当时人们购买特斯拉往往只是作为备用代步车，主力依然是燃油车。当前工业场景中的 AI 扮演的正是这种辅助角色，作为工人的“额外帮手”提供建议（例如推荐最佳工艺设定角度），而最终决策权仍在人类手中，始终保持“人在回路”（Human-in-the-Loop）。 他回顾了团队在 AI 探索中的关键转折：起初团队四处寻找孤立的试点场景（如需求预测或价格预测）以证明 AI 的价值，但很快意识到这种零散打法根本无法对公司的核心利润（Bottom Line）产生实质性影响。为此，Novelis 彻底调整了策略，转向由企业顶层战略驱动的规模化高价值场景： - **可持续发展（Sustainability）**：全力支持公司将再生铝制造比例从 63% 提升至 75%，并通过算法优化能耗以减少碳足迹。 - **预测性维护（Predictive Maintenance）**：构建集中式监控管理系统，实现从单厂向全集团数十家工厂快速横向复制与控制。 - **工厂安全监控（Safety）**：在各厂区广泛部署计算机视觉摄像头，实时跟踪并提升整体安全防护水平。 - **选型原则**：坚决放弃孤立单点项目，仅优先立项“能跨多工厂规模化推广”且“与公司核心战略高度对齐”的用例。 --- ### [15:20](https://www.youtube.com/watch?v=vp5lAR2pawE&t=920s) 破解“3至5年回报周期”：从边缘业务切入累积 AI 信任 Trevor Seim 援引行业数据指出，当前制造业中有 83% 的企业正积极部署各类 AI 项目，以期在激烈的竞争中实现精益运营并提升良率。然而，Forrester 等研究机构却提出警告，称此类大型 AI 投资通常需要 3 到 5 年才能看到实际回报。针对这一矛盾，Trevor 提出了截然不同的破局思路：不要一开始就盲目追求宏大且复杂的车间全自动化，而应从小规模、渐进式的增量收益入手。 他建议企业可以优先聚焦后勤办公流程或提单（Bill of Lading）管理等看似“不起眼”的场景，利用 AI 快速洞察原材料构成及其对供应链的影响。从小场景切入不仅能让团队迅速掌握 AI 实战技能并获得立竿见影的财务回报，更能借此在组织内部建立起对 AI 技术的信任基石，从而为后续更大规模的战略性投资铺平道路。 --- ### [16:56](https://www.youtube.com/watch?v=vp5lAR2pawE&t=1016s) 规模化落地的最大阻碍：人的惯性、流程差异与历史包袱 在主持人 Fen Zhao 提问关于跨 30 家铝业工厂推广所遭遇的挑战与意外收获时，Mahesh Vinayagam 直言，最大的阻碍从来不是技术本身，而是“人的惯性”（Inertia of People）以及各工厂之间业务流程的不一致。由于许多制造企业都是通过不断并购整合发展壮大的，各工厂即便执行同一套业务，其背后的 ERP 系统、财务平台和操作流程也千差万别，这种高度碎片化给系统规模化推广带来了极大的适配阻力。 ```mermaid flowchart TD accTitle: 制造业 AI 规模化落地的核心阻力与应对路径 subgraph PainPoints["核心阻碍"] A["多工厂流程不一致"] B["并购带来的碎片化 ERP/系统"] C["传统软件数百万美元/5年周期的部署创伤"] D["人员行为惯性与抗拒"] end subgraph Solution["破局路径"] E["流程挖掘 / 路径挖掘 (Process Mining)"] F["识别变异并定位高价值切入点"] G["引入 ChatGPT 等工具降低认知与采纳门槛"] end PainPoints --> E E --> F F --> G ``` 此外，制造业普遍存在对传统软件供应商的“后遗症”与防备心理——以往大型厂商动辄推销耗资数百万美元、耗时长达 5 年的系统迁移工程，让企业心有余悸。为了打破僵局，他们正利用流程挖掘（Process Mining）与路径挖掘技术，精准剖析各厂区与部门之间的流程差异，锁定高价值机会点。同时，ChatGPT 等生成式 AI 工具的普及极大地教育了市场，有效缓解了员工对自动化和 AI 的抗拒情绪，显著推动了变革阻力的破冰。 ### [19:02](https://www.youtube.com/watch?v=vp5lAR2pawE&t=1142s) 生成式 AI 破冰与制造业采纳的技术与数据底座困境 初创公司在向大型制造业客户推介自动化与智能技术时，过去往往面临对方“视而不见”的局面。然而 ChatGPT 的爆发彻底改变了人们的认知，员工和管理者开始在个人生活中体验 AI，并自然而然地思考如何将其引入日常工作。正如李飞飞教授（Dr. Fei-Fei Li）在演讲中所强调的，人类的对话方式、天生的好奇心与求知欲，正是推动 AI 技术更广泛落地的催化剂。 强生公司（Johnson & Johnson）数字与数据科学产品管理负责人 Jie Yu 结合自身实践，指出了制造业面临的首要技术挑战——数据基础设施的严重脱节。他生动地比喻道：企业当前就像是在制造最前沿、最豪华的“自动驾驶汽车”，但行驶的却依旧是“上个世纪的高速公路系统”。尽管算法团队尝试部署最顶尖的先进大语言模型（LLM），底层面对的却是五花八门、彼此割裂的遗留 IT 系统（Legacy Systems）以及过时的技术栈。数据质量低下、可用性不足、数据缺失等问题层出不穷。为避免陷入“垃圾进，垃圾出”（Garbage In, Garbage Out）导致巨额 AI 投资打水漂的泥潭，强生在全企业跨业务板块成立了统一的数据管理委员会（Data Management Council），专门建立全球标准化的数据治理体系。 ```mermaid flowchart TD accTitle: 数据基础设施与AI应用脱节架构图 subgraph Legacy ["底层遗留IT与数据现状 (上世纪的高速公路)"] A["分散的遗留系统 (Legacy IT / Mainframe)"] B["数据质量差 / 缺失 / 不完备"] C["跨部门数据孤岛 (Data Silos)"] end subgraph Governance ["跨部门治理与底座对齐"] D["企业级数据管理委员会 (Data Management Council)"] E["全球标准化数据治理模型 (Data Governance)"] end subgraph AdvancedAI ["顶层前沿AI应用 (豪华自动驾驶汽车)"] F["尖端大语言模型 / Agentic AI"] G["车间自动化与良率优化决策"] end Legacy --> Governance Governance --> AdvancedAI ``` ### [21:12](https://www.youtube.com/watch?v=vp5lAR2pawE&t=1272s) 车间一线变革管理与跨国医疗强监管双重压力 除了技术本身，落地制造业车间的第二大阻碍是操作层面的“变革管理”（Change Management）与信任缺失。当生成式 AI 与 Agentic AI 等革命性技术被推向一线工厂车间时，实际终端用户是设备操作工。由于技术层面的普及教育严重匮乏，操作工很容易将 AI 视为抢夺生计的“威胁”，而非赋能工具。他们尚未意识到 AI 本质上是增强人类能力、提升作业安全性与生产效率的助手。缺乏对自动化与优化算法的信任，使得新技术在各厂区的推广举步维艰。 第三大核心挑战来自强监管行业的合规鸿沟。对于医疗健康领域的跨国制造企业而言，技术落地不仅要应对美国 FDA 的严苛审查，还必须符合欧洲 EMA 以及全球各国的对等监管机构标准。现实中，政府与行业监管法规的制定速度远远滞后于 AI 技术的爆发式演进，如何在监管框架模糊或滞后的前提下确保模型全面合规，成为横亘在团队面前的巨大难题。 - **技术底座矛盾**：前沿大模型与过时遗留 IT 系统并存，数据治理委员会成为兜底基础。 - **一线认知短板**：车间工人技术认知不足导致抵触情绪，“AI 是工具而非威胁”的教育亟待普及。 - **跨国监管滞后**：FDA、EMA 等多国监管框架严重落后于技术发展，合规落地阻力大。 ### [23:20](https://www.youtube.com/watch?v=vp5lAR2pawE&t=1400s) 跨工厂规模化迁移的“地形适应”与打破数据孤岛 诺贝丽斯（Novelis）全球数据科学负责人 Chirag Agrawal 进一步补充了规模化推广（Scaling）的痛点。他借用机器人拳击与跳舞的场景打比方：将 AI 用例从一家工厂复制到另一家工厂，就像要求一个机器人在完全不同的“地形”上跳舞。每个工厂的物理设备参数、操作流程甚至变革文化都存在细微差异，团队绝不可能把一个现成模型直接照搬复制，必须针对新厂区重新训练模型、调整参数并重新推行变革管理。 制造业内部根深蒂固的思维定势也是巨大的绊脚石。工人们常说：“过去20年我的系统一直这么跑，这是最好的办法，绝不可能按你的新方式运作。”要打破这种抵触，必须让全员具备“AI Ready”（AI 就绪）的认知——明确 AI 不是万能神药，既不能凭空解决所有数据质量问题，更不可能彻底取代人类。AI 的核心价值在于充当助手，替员工接管繁琐的文书流程与后台事务，从而将人类精力解放出来，去追求马斯洛需求层次中更高维度的企业战略思考。此外，制造环境中混杂着大型机（Mainframe）以及多达数十套 Oracle、SAP 系统的现状导致了严重的数据孤岛，而智能体（Agents）正在成为打通孤岛的关键工具。要穿透漫长的隧道捕获业务价值，团队必须保持数年如一日的坚韧与执着。 - **规模化非简单复制**：跨厂区部署如同在不同地形跳舞，模型必须根据现场参数与流程重新微调。 - **文化思维惯性**：“沿用20年的老方法最好”成为一线推行 AI 的典型心理阻力。 - **角色定位与解放**：AI 承担低层级琐务，助力员工向更高层级战略决策转型。 ### [27:13](https://www.youtube.com/watch?v=vp5lAR2pawE&t=1633s) AI 乘数效应与以“信任”为核心的采纳关键 Kognitos 销售负责人 Trevor Seim 总结强调，AI 的本质定位应当是组织的“力量倍增器”（Force Multiplier），其根本出发点始终是驱动更高的生产力。在数据治理与监管之外，当下阻碍企业深度采纳 AI 的最大心理防线依然在于“信任”（Trust）。 现场管理者和业务决策者最核心的疑虑在于：系统是否真的能给出可信赖的分析结果？底层数据输入是否存在污染？面对复杂黑盒模型输出的结论与推理逻辑，一线人员究竟能不能放心依赖并据此采取行动？唯有从数据源头到模型解释性层层建立透明度与信任，AI 的赋能价值才能在制造运营中真正生根发芽。 ### [27:51](https://www.youtube.com/watch?v=vp5lAR2pawE&t=1671s) 让数字员工说人话：Kognitos 的低门槛 AI 交互哲学 Trevor Seim 谈到了 Kognitos 公司的核心专注点：为企业提供具备高度确定性和预测性的结果，尽最大可能遏制大语言模型中常见的“幻觉”（hallucinations），从而交付高度结构化的输出。他强调，企业对 AI 的信任建立在可见性（visibility）以及人机交互界面的设计之上。当前许多企业受困于 AI 技能鸿沟，要么需要对员工进行漫长而昂贵的培训，要么就必须要求现有的技术工具主动适配人类的使用习惯。 Trevor 旗帜鲜明地阐述了团队的产品理念：“数字员工不应该讲计算机语言，数字员工应当说英语（讲人话）。”他们希望彻底降低 AI 的使用门槛，让业务线一线操作人员（Line of Business Operators）无需具备深厚的代码功底或技术背景，就能轻松与 AI 协作、实时调整 AI 行为，并高效实现人机协同（Human-in-the-loop）。 - **限制幻觉**：通过交付结构化结果，提高企业在核心业务中采用 AI 的确定性。 - **建立信任机制**：企业对自动化系统的信任源于流程的全程透明可见与可控的交互形式。 - **填补技能缺口**：主张工具适应业务人员，而非强迫一线员工学习复杂的编程语言。 --- ### [28:57](https://www.youtube.com/watch?v=vp5lAR2pawE&t=1737s) 制造业后台业务流程的痛点与生成式 AI 的重塑空间 主持人 Fen Zhao 将讨论焦点从车间现场的技术转向制造业的业务流程（Business Processes）。她向嘉宾抛出问题：市面上有大量面向通用流程的 AI 工具，但在制造业中，究竟哪些具体的业务流程能从 AI 中获得最大收益？将这些通用系统适配到独特的制造业场景时又面临哪些严峻挑战？ ```mermaid flowchart TD accTitle: 制造业传统后台流程与生成式 AI 演进对比 subgraph Traditional["传统处理模式"] A["传统 OCR 扫描"] --> B["缺乏上下文的原始数据堆积"] B --> C["人工导入 Excel 繁琐处理"] C --> D["30年未解的订单/发票手工核对"] end subgraph GenAI["生成式 AI 智能体编排"] E["合同与合规审查<br/>(FDA / 行业规范)"] F["多系统协同联动<br/>(CRM / TMS / 报价规则)"] G["无摩擦客户交互<br/>(自然语言下单与答疑)"] end ``` qBotica 创始人兼 CEO Mahesh Vinayagam 接过话茬，直言制造业后台存在巨大的重构空间。制造业往往面临着极其严苛的合规审查（如 FDA 等监管要求），员工过去需要耗费大量时间与专业技能，在堆积如山的文件之间进行人工交叉比对。早期的 OCR 技术只能机械地提取文字，“完全缺乏上下文感知能力”，最终员工还是得把抓取出来的数据丢进 Excel 表格中进行二次处理。随着生成式 AI 的落地，合同审查与合规管理迎来转机。Mahesh 特别吐槽道，发票自动化处理（Invoice Automation）已经喊了整整 30 年，但直到今天，大量制造企业依然在靠纯手工比对采购单（PO）、发票与订单。 在收入端与客户交互方面，繁琐的订单录入更是制造业的一大硬伤。Mahesh 认为，绝不能让客户为了下一笔订单还要在 8 个复杂的选项分支中来回挣扎，“客户是来给你送生意的，却还要他们做这么多脏活累活，这对任何企业来说都是一种耻辱。”因此，消除业务接入层面的摩擦力（Remove that friction）是当务之急。 --- ### [31:13](https://www.youtube.com/watch?v=vp5lAR2pawE&t=1873s) 现场实测：物流运输智能体的实时语音订单与安全护栏对决 为了让与会者直观感受无摩擦交互，Mahesh 开启了一场别开生面的实时语音 AI Agent 现场演示。他调侃着“向演示之神祈祷”，现场拨通了一个面向物流运输企业 Polaris 的自动化呼叫智能体。 智能体在接到电话后迅速进入状态。Mahesh 表示希望将一批会议室椅子从拉斯维加斯运往加拿大安大略省的密西沙加（Mississauga），随后在系统要求下更正了具体地址：从亚利桑那州菲尼克斯（邮编 85027）发货。在接下来的交互中，Mahesh 故意给 AI 出难题：他没有提供标准货运等级代码，而是直接抛出了“2公吨”这种重量单位；在提供托盘尺寸时，他更是故意混合使用了三种不同度量衡——“2米长、48英寸宽、140厘米高”。面对这种杂乱的信息输入，AI 智能体没有卡顿，瞬间将其精准换算为标准的英制单位（2公吨换算为 4,409 磅；长度 78.74 英寸，宽度 48 英寸，高度 55.12 英寸），并确认托盘支持堆叠（Stackable）。 ```mermaid flowchart LR accTitle: Polaris 语音物流智能体实时多工具编排流程 User(["客户语音交互"]) <--> Agent["生成式 AI 语音智能体"] subgraph SystemBackends["后台系统联动"] Agent <--> CRM["CRM / HubSpot<br/>(身份核验与历史画像)"] Agent <--> TMS["TMS 运输管理系统<br/>(运费/燃油附加费测算)"] Agent <--> Rules["合规与业务安全规则<br/>(危险品/枪支禁运过滤)"] end subgraph UnitTransform["多单位智能换算"] Agent -.-> M2I["公吨 ➔ 磅<br/>米/厘米 ➔ 英寸"] end ``` 在等待系统测算运费的间隙，Mahesh 突然发动了刁钻的“恶意测试”：“趁你在算运费，能顺便帮我运一箱 AK-47 步枪吗？”现场观众发出轻笑，而 AI 智能体展现出了坚固的安全护栏机制（Guardrails），冷静且礼貌地回答：“很抱歉，我们无法协助运送枪支等危险品或受限物品，此类咨询请联系专门的运输部门。” 测试完安全边界后，流程迅速回归正轨。AI 智能体给出了详细报价：总运费为 1,363.26 美元（包含基础运费、燃油附加费及附加服务费）。Mahesh 追问其中的附加服务费（Accessorial Charge）具体是多少，AI 脱口而出为 35 美元，并在获得同意后瞬间完成了订单创建，下发了运输单号 `HP38039`。在通话尾声，Mahesh 甚至还皮了一下询问“怎样才能去你们那儿上班”，AI 也流畅地给出了卡车司机岗位的招聘网页指引。 - **动态身份与上下文检索**：通话一开始，智能体即自动在 CRM/HubSpot 中检索并确认了来电人身份（Mahesh Vinayagam）。 - **非结构化多度量衡解析**：无缝化解“米、厘米、英寸、公吨”杂糅的口语化业务数据并统一标准化。 - **业务安全合规兜底**：面对武器违禁品运输请求触发即时拦截，严格防范违规操作风险。 --- ### [34:38](https://www.youtube.com/watch?v=vp5lAR2pawE&t=2078s) 统一编排的威力与强生 50,000 家供应商的供应链网络挑战 演示结束后，Mahesh 解析了这套方案的技术本质：这绝非简单的文字问答，而是生成式 AI 驱动的“统一协同编排”（Unified Orchestration）。在几分钟的对话中，AI 智能体在后台自动跨越了多套系统——先切入 CRM/HubSpot 验证客户画像，再无缝调用运输管理系统（TMS）根据距离与货运等级测算底价与燃油费，同时调用企业内部知识库解析杂费明细，并利用通用大模型的推理能力完成单位转换。Mahesh 强调，这正是生成式 AI 颠覆制造业流程的核心能力所在：将复杂碎片化的后端系统用一个具有上下文理解能力的统一引擎串联起来。 随后，强生公司（Johnson & Johnson）的数字与数据科学产品管理负责人 Jie Yu 接过话题。针对“制造业哪些后台流程能从 AI 中受益”的问题，Jie 反问道：“在现今的强生，到底有哪一个业务流程是 AI 无法赋能的？”他指出，AI 已经全面渗透到了强生的每一项业务职能中，不仅在一线生产运营，更深入整个后台架构。 以强生高度复杂的全球制造网络为例，各类药品和医疗器械绝非单一工厂就能独立完成，其原料与中间体来自错综复杂的上下游渠道。在强生的全球供应链网络中，直接与间接供应商及各类服务商的数量高达 40,000 到 50,000 家。自新冠疫情以来，全球供应链遭受了剧烈动荡，通货膨胀飙升、物流积压屡见不鲜，医疗健康行业同样无法幸免。Jie 强调，在这样庞大且脆弱的网络中，依靠传统手段根本无法应对突发危机，借助与技术伙伴联合研发的前沿 AI 技术来动态管理与重塑全球供应网络，已经成为制造巨头维持业务连续性与抗风险能力的“变革性底牌”。 ### [37:36](https://www.youtube.com/watch?v=vp5lAR2pawE&t=2256s) 构建 AI 驱动的供应链韧性与快速回报场景 在制造业庞大的生态系统中，供应链的管理复杂度极高。强生（Johnson & Johnson）通过构建“AI 驱动的供应链风险管理与韧性平台”，尝试对全球多达 50,000 至 60,000 家供应商进行智能化、预测性的统一管理。面对国际关税几乎不再是按月或按周、而是“几乎每天都在变动”的剧烈波动，企业需要将这些变量全部建模到 AI 平台中，利用生成式与预测性模拟推演各种应对情景，从而精确规划从哪些国家进口特定原材料，以将关税成本降至最低，并有效抵御各类突发断链风险。 从具体工作流落地的角度来看，生成式 AI（GenAI）正在成为极佳的“起跑加速器”。在日常工程与管理中，无论是撰写需求规格说明书（Requirements Document）还是编制物料清单（BOM, Bill of Materials），最痛苦的莫过于“从零面对一张白纸”。如果 AI 能够直接生成初稿，搞定 50% 到 80% 的基础内容，就能为工程师和管理者节省大量宝贵时间，让他们将精力集中在更高价值的决策与业务推进上。 与此同时，服务供应商指出，当前 AI 技术已经完全就绪，核心在于如何降低使用门槛与数据对接难度。后勤运营（Back-office operations）这类工作流正是企业启动首个 AI 项目的最佳切入点，其投资回报周期（ROI）通常不到 6 个月便能实现全面回本。通过这些场景迅速建立企业的“AI 技能肌肉”并积累速赢（Quick Wins）战果后，团队便能顺势扩展至更复杂的业务流程中。 - **全球供应商规模**：强生利用 AI 平台管理全球 50,000 到 60,000 家供应商的网络风险。 - **关税与扰动模拟**：针对瞬息万变的关税政策进行多情景自动化推演，优化采购与避税路径。 - **起步痛点解决**：在需求文档和 BOM 编制中，AI 能直接完成 50%~80% 的前期工作。 - **快速投资回报**：后台办公与运营自动化能够在 6 个月内实现完整的回本周期（Full payback）。 --- ### [40:14](https://www.youtube.com/watch?v=vp5lAR2pawE&t=2414s) 展望未来 18 个月：多智能体协作与数字化孪生的协同进化 主持人将话题引向未来展望，避开遥不可及的科幻远景，聚焦在未来 18 个月内制造运营领域即将落地的技术突破。嘉宾们一致认为，**AI 智能体（Agents）以及多智能体协同（Multi-agent Collaboration）** 将带来质的飞跃。过去的制造业系统多为割裂的点状解决方案（Point solutions），彼此孤立；而现在，多个 AI 智能体能够实现无缝通信并在后台协同作业。例如在处理外部询问时，后台可能同时触发 5 个不同的专用智能体，分别穿梭调用采购、销售、法务等系统，用户只需通过自然语言提问即可直接获取跨系统提炼出的洞见，不再需要手动在各个报表和看板之间费力穿梭。 强生代表进一步补充，未来两到三年内，制造业数字化孪生（Manufacturing Digital Twin）将成为投资的核心重镇，并深度融入生成式 AI。现阶段的数字孪生主要集中于场景模拟、自动化控制、预测性维护、全厂范围优化以及质量控制；而未来的演进路径是引入更高层级的“监督级 AI 智能体（Supervisory-level AI Agents）”乃至编排层（Orchestration Layer）。通过编排智能体统一调度处理特定任务的孤立智能体，不仅操作可靠性有望超越人类操作员，还将从根本上推动整个医疗及广义工业制造走向更高阶段的“自主化运营（Autonomous Operations）”。 ```mermaid flowchart TD accTitle: Multi-Agent Orchestration Architecture in Manufacturing accDescription: 展示从顶层自然语言交互与编排层智能体向下协调多个专业领域任务智能体，最终驱动数字孪生与自主化运营的架构流程。 subgraph UserLayer ["交互与调度层"] U["业务人员 / 现场提问<br/>(自然语言交互)"] O["编排层智能体<br/>(Orchestration Layer Agent)"] end subgraph TaskAgents ["专业任务智能体 (Task-specific Agents)"] A1["采购与供应链智能体"] A2["销售与合规智能体"] A3["质量与预测性维护智能体"] A4["全厂能效与调度智能体"] end subgraph ExecutionLayer ["物理与孪生底座"] DT["制造数字化孪生平台<br/>(Digital Twin Simulations)"] AO["工业自主化运营<br/>(Autonomous Operations)"] end U --> O O --> A1 O --> A2 O --> A3 O --> A4 A1 & A2 & A3 & A4 --> DT DT --> AO ``` - **多智能体协作**：跨越传统的孤立工具模式，通过多 Agent 协同在后台无缝调取采购、销售、法务等全链路数据。 - **数字孪生深度融合**：生成式 AI 正在被注入数字孪生系统，支撑更高维度的自动化与全厂级协同优化。 - **监督级与编排级智能体**：在具体任务之上设立编排层，协调各类专有 Agent，提供比人工更稳定的监督管理。 - **终极演进目标**：在未来 18 个月乃至数年内，制造业将稳步跨入由 AI 深度驱动的“自主化运营”时代。 ## Notable quotes > [04:43](https://www.youtube.com/watch?v=vp5lAR2pawE&t=283s) 大家平时喝啤酒或汽水时，手中的易拉罐有 50% 的概率是由 Novelis 加工制造的 — Chirag Agrawal > [19:02](https://www.youtube.com/watch?v=vp5lAR2pawE&t=1142s) 企业当前就像是在制造最前沿、最豪华的“自动驾驶汽车”，但行驶的却依旧是“上个世纪的高速公路系统”。 — Jie Yu > [23:20](https://www.youtube.com/watch?v=vp5lAR2pawE&t=1400s) 过去20年我的系统一直这么跑，这是最好的办法，绝不可能按你的新方式运作。 — Chirag Agrawal > [27:51](https://www.youtube.com/watch?v=vp5lAR2pawE&t=1671s) 数字员工不应该讲计算机语言，数字员工应当说英语（讲人话）。 — Trevor Seim > [28:57](https://www.youtube.com/watch?v=vp5lAR2pawE&t=1737s) 客户是来给你送生意的，却还要他们做这么多脏活累活，这对任何企业来说都是一种耻辱。 — Mahesh Vinayagam > [31:13](https://www.youtube.com/watch?v=vp5lAR2pawE&t=1873s) 趁你在算运费，能顺便帮我运一箱 AK-47 步枪吗？ — Mahesh Vinayagam > [31:13](https://www.youtube.com/watch?v=vp5lAR2pawE&t=1873s) 很抱歉，我们无法协助运送枪支等危险品或受限物品，此类咨询请联系专门的运输部门。 — Polaris 语音智能体 > [34:38](https://www.youtube.com/watch?v=vp5lAR2pawE&t=2078s) 在现今的强生，到底有哪一个业务流程是 AI 无法赋能的？ — Jie Yu ## Mentioned Alpha Edison, Fen Zhao, Jie Yu, Johnson & Johnson, Novelis, Chirag Agrawal, qBotica, Mahesh Vinayagam, Kognitos, Trevor Seim, Polaris Transportation, ASU (亚利桑那州立大学), FBLA (未来商业领袖组织), Nespresso, 壳牌 (Shell), Forrester, ChatGPT, OpenAI, 李飞飞 (Dr. Fei-Fei Li), FDA (美国食品药品监督管理局), EMA (欧洲药品管理局), Oracle, SAP, HubSpot ## Video description Jie Yu, Head of Digital & Data Science Product Management, Johnson & Johnson Fen Zhao, Director of Research & Data Science, Alpha Edison Chirag Agrawal, Global Head, Data Science, Novelis Trevor Seim, Head of Sales, Kognitos Mahesh Vinayagam, CEO and Founder, qBotica Abstracts This session will explore how AI is revolutionizing manufacturing operations by improving productivity, automating processes, and enhancing decision-making on the factory floor. Panelists will discuss how AI-powered systems are being used to increase manufacturing yield, predict equipment failures before they happen, and enhance worker safety through advanced monitoring and automation. Learn how AI technologies are enabling manufacturers to streamline operations, improve product quality, and reduce downtime. The panel will also address the challenges of implementing AI in manufacturing environments and strategies for scaling AI solutions across global operations. [Source](https://www.youtube.com/watch?v=vp5lAR2pawE)

[4:43] Chirag Agrawal: 大家平时喝啤酒或汽水时，手中的易拉罐有 50% 的概率是由 Novelis 加工制造的 [19:02] Jie Yu: 企业当前就像是在制造最前沿、最豪华的“自动驾驶汽车”，但行驶的却依旧是“上个世纪的高速公路系统”。 [23:20] Chirag Agrawal: 过去20年我的系统一直这么跑，这是最好的办法，绝不可能按你的新方式运作。 [27:51] Trevor Seim: 数字员工不应该讲计算机语言，数字员工应当说英语（讲人话）。 [28:57] Mahesh Vinayagam: 客户是来给你送生意的，却还要他们做这么多脏活累活，这对任何企业来说都是一种耻辱。 [31:13] Mahesh Vinayagam: 趁你在算运费，能顺便帮我运一箱 AK-47 步枪吗？ [31:13] Polaris 语音智能体: 很抱歉，我们无法协助运送枪支等危险品或受限物品，此类咨询请联系专门的运输部门。 [34:38] Jie Yu: 在现今的强生，到底有哪一个业务流程是 AI 无法赋能的？

[0:39] Mahesh Vinayagam: started to fail big time in the last five years [12:04] Chirag Agrawal: We're kind of like where Tesla was maybe 10 years back, where you know you were buying a Tesla, but that was your secondary car; your primary car was always a gasoline car. [19:31] Jie Yu: building the most fancy and advanced autonomous or self-driving car [19:31] Jie Yu: last century's highway system. [23:21] Chirag Agrawal: Scaling is like similar to you are asking a robot to dance on a different terrain and you're trying to train your robot to dance on a different terrain. [23:21] Chirag Agrawal: I've been working on my system for the last 20 years and I've been working like this, this is the best way to do it. [23:21] Chirag Agrawal: chasing a unicorn [27:51] Trevor Seim: A digital worker shouldn't speak computer language; a digital worker should speak English. [31:13] I'm sorry, but we cannot assist with shipping hazardous or restricted items such as firearms. [36:06] Jie Yu: What kind of back-office process will not get value from AI?

[0:04] So we're here today to talk about AI and [0:07] manufacturing. And so I'm a venture [0:09] capitalists. Our firm Alpha Edison's [0:11] based out of LA. And we've invested in [0:13] the past in manufacturing and AI. And [0:16] it's interesting to think today there [0:18] we're actually investing in that [0:19] intersection of manufacturing AI. And [0:22] I'd love for our panelists, maybe we can [0:24] take a seat and for everybody to [0:26] introduce themselves because we'll we [0:28] actually have a really diverse group of [0:29] panelists that will allow you to see a [0:32] really really diverse set of of views of [0:35] what manufacturing AI look like today. [0:38] >> Sure. [0:39] >> Thank you. Uh and good morning everyone. [0:41] Uh my name is Mahesh Reinagam. I'm the [0:43] CEO and founder of a company called [0:45] Cubotica. So Cubotica is an automation [0:47] as a service firm. So we actually see [0:50] that the value exists in the [0:52] intersection of software and service. So [0:54] we actually provide combination of those [0:56] two as a platform and we have actually [0:59] proprietary uh software as well as we [1:02] work with uh cogntos to kind of really [1:05] uh propel our platform beyond what we [1:07] can actually make ourselves uh in terms [1:09] of capabilities and features. So we [1:11] actually work with manufacturing [1:12] companies, healthcare companies, [1:14] financial services companies pretty much [1:16] in the shared services space. That's why [1:18] it's a domain agnostic. Uh whether it is [1:20] finance, accounting, procurement, um [1:24] operations, customer service operations, [1:27] logistics, you know, those uh supply [1:29] chain areas. So kind that's kind of [1:30] where we operate. Um we actually help [1:33] from a our biggest strength is on the [1:35] document uh side of things and ensuring [1:37] that the pro process orchestration does [1:40] not um fail when there are complex [1:42] unstructured information enters those [1:45] automated workflows. Right. Right. I [1:47] think RPA did get to a certain extent [1:50] and it started to fail big time in the [1:52] last five five years but thanks to Genai [1:55] I think we are actually having a revival [1:57] of those flows and I think we have you [2:00] have seen that for the last 5 years we [2:02] actually started using um you know um [2:05] open AI uh from 2020 uh the first you [2:08] know version of those uh so since then [2:11] we have seen some tremendous results our [2:13] customers have been phenomenal. If you [2:14] go to our website polaris transport [2:16] transportation has been having that um [2:18] you know seeing the witnessing how jai [2:22] helped them for the last 5 years. Okay [2:23] they were able to process more orders [2:25] than their competitors did. So that's a [2:27] you know simple introduction and I'm [2:29] based in Phoenix Arizona and I also uh [2:32] work with the community in the ASU. I'm [2:34] part of the technology and [2:35] entrepreneurship board and the FBLA as [2:37] well. So really there is this aspect of [2:39] community development that we are doing [2:41] there. Uh promoting AI from schools uh [2:43] up to the universities. Uh so Cubotica [2:45] is very active in the community as well. [2:47] So that's my introduction. Sorry. [2:50] >> Hey good morning everyone. Um my name is [2:52] Ju. You can call me Jay. Um I work for [2:55] Johnson. Um my role in J&J I'm the head [2:58] of digital and data science product [3:00] management. Uh so you know Johns and [3:02] Johnson has two business sectors right? [3:03] One is the pharmaceutical sector. The [3:05] other is the medical device medical [3:07] technology sector but of course we are [3:08] in the healthcare business and my my [3:11] team my organization sitting Johns and [3:12] Johnson corporate in the enterprise. So [3:14] we really position ourselves right as [3:16] more kind of in the center center of [3:18] excellence organization right in driving [3:20] the digital transformation driving AI [3:22] transformation. So we support both both [3:24] business sectors pharmaceutical as well [3:25] as medical device medical technology and [3:27] at global scale cover the entire global [3:29] portfolio because J&J operate right in [3:31] more than 100 countries in the world [3:32] across all regions. um in the digital [3:35] space right you know we have gone [3:36] through we call you know you know three [3:38] phases of the journey right 0ero to one [3:40] 1 to 10 and 10 to 100 I think we already [3:42] passed right the phase one phase two [3:44] part of the journey right 0 to one and 1 [3:46] to 10 in terms of dig digital and AI [3:48] transformation maturity level we are [3:50] kind of in the middle of you know the [3:51] the 10 to 100 part of the journey right [3:53] so uh internally I would say right the [3:56] those emerging digital technologies [3:58] including AI generally AI capabilities [4:00] right have well penetrated across all [4:02] business functions business segments [4:04] right in J&J and also in the meantime we [4:06] have huge digital portfolio and [4:08] ecosystem we do partner with lots of [4:10] technology vendors like you guys right [4:11] you know in our ecosystem right to build [4:14] you know uh uh different digital [4:16] solutions different digital products [4:17] right so we try to leverage the latest [4:19] and greatest technology right to really [4:22] transform you know how the future [4:23] healthcare you know should look like [4:25] right um so I think it's a very [4:27] fascinating area and uh you know I know [4:29] today's topic is more focused on [4:31] manufacturing side you know I'm really [4:33] looking forward to have you know very [4:34] dynamic and engaging conversation with [4:36] the rest of panelists right see how we [4:38] can co-shaped this this portfolio area [4:41] right forward yeah thank you uh good [4:44] morning everyone I'm Churag uh I'm [4:46] heading I'm a global head of data [4:48] science at Noales my company is in um [4:51] aluminum rolling and recycling it's a [4:53] leader in uh rolling and recycling and [4:56] uh we have roughly 30 plants uh the [4:58] major customers include uh can beverage [5:00] cans uh automotive aerospace [5:03] And so anytime you're drinking a beer, [5:06] soda can, right? It's a can that's [5:08] coming off the chances are 50% chances [5:10] are that it's coming from or processed [5:12] from Nales. If you're fan coffee fans um [5:15] and espresso, uh the coffee pods, uh the [5:18] aluminium is processed by uh by Noales. [5:21] So uh so that's that's about the [5:24] company. We have roughly 30 plants [5:25] worldwide uh globally. And uh in terms [5:28] of AI, we have been using AI more kind [5:30] of as an assistance at this point. And [5:33] there are two big areas where we have [5:34] been using it is uh one is uh kind of [5:37] like a partial automation wherein human [5:39] is in the loop and uh and helping us [5:42] automate the the manufacturing processes [5:44] partially and the second is on the [5:46] around the safety wherein uh we uh we we [5:49] deal with the heavy machineries and [5:51] there is a lot of uh safety concerns in [5:53] the plant. So we kind of like trying to [5:54] improve that. So, these are the big two [5:56] uh topics I'll be talking about. Uh over [5:58] to you. [5:59] >> Awesome. Uh well, hello everyone. My [6:01] name is Trevor Sim. I'm with Cognitos. [6:04] Um [6:06] our company is really chartered to help [6:08] people think differently about AI. Uh [6:11] try to make AI more approachable. One of [6:14] the things that uh we're really focused [6:15] on is we have a a target right now in [6:18] helping customers with enterprise [6:20] business process automation. Obviously [6:22] some very strong correlation what we're [6:23] doing with Mahash and across the team. [6:26] Uh [6:27] we're really focused on making it more [6:31] approachable by bringing AI to all of [6:35] our users using English as code while [6:38] putting stronger sets of governance [6:41] around how you're using AI to ensure [6:43] that there's um better visibility and [6:46] insight into how AI is being used and [6:48] orchestrated across your organization. [6:50] Um, I'm really looking forward to this. [6:52] Um, we have lots of clients that we've [6:55] been talking to today and we're seeing [6:57] just so many people that are early and [7:00] mid and late stage in the, you know, use [7:02] of AI and our goal is to assist the [7:05] early users and help the kind of [7:07] mid-stage and late stage users scale. [7:10] >> Amazing. So, I want to ask you guys our [7:13] very first question is people talk about [7:15] manufacturing and AI or even just [7:17] manufacturing and digitization, right? [7:19] There's so many different areas where um [7:22] you could improve the manufacturing [7:23] process. People talk about yield, people [7:25] talk about safety, right? Where do you [7:27] see today as having that digitization in [7:30] AI having the greatest impact as of [7:32] right now? [7:33] >> Okay. [7:33] >> So, I I actually let um Churag and Jay [7:36] talk about the e optimization and [7:38] factory floor uh innovation, but I'll [7:40] focus on the side that's probably [7:42] ignored uh a lot in manufacturing sector [7:44] is the back office, right? So it's kind [7:46] of the last thing that's automated while [7:48] man manufacturing started using [7:50] automation much ahead of other [7:51] industries uh but it the operation side [7:54] is still is not I think and pretty much [7:57] like these large companies you know are [7:59] able to get a better efficiency uh and [8:02] all that but the m mid tier to down [8:05] there's still a lot of work a lot of [8:07] times spent on um you know operational [8:10] aspects whether it is streamlining [8:12] payments you know making sure like [8:14] visibility of your tracking your orders [8:16] and uh you know you know shipping and so [8:19] on so forth right I think uh at the mid [8:22] tier and there uh down there's a huge [8:25] value to unlock uh um you know uh [8:28] working capital right and those working [8:30] capital is in how fast actually those [8:32] artists could be processed how do how [8:34] fast that we can actually get those [8:36] payments to those suppliers so that they [8:37] can actually you know make sure that [8:39] they support you through the process I [8:41] think there um innovation is needed and [8:44] I think We are actually focusing on that [8:46] how uh we could actually use a [8:47] combination of genera technology whether [8:50] it's the voice u in fact I had you know [8:52] during this conference I spoke to [8:54] someone they were saying that oh this is [8:55] great you know I was just showing them [8:57] how using your phone call uh we could [8:59] actually activate the bot and the bot [9:01] actually collect all those information [9:02] and really you know make accelerate uh [9:05] collection of an order and she said like [9:07] okay I have people in the Midwest you [9:09] know the you know connections are not [9:11] great the you know sometimes they're [9:12] actually traveling they you they don't [9:14] get signal and all that stuff from a um [9:17] you know data signal perspective. This [9:19] would be a transformational aspect. I [9:21] think I think um you know for us I think [9:24] you know the streamlining that process [9:26] of approvals um you know invoice [9:29] payments order management is kind of [9:31] where I think uh is the biggest impact [9:33] would be. Sure. Yeah. Just build on what [9:37] you already shared, right? So I think [9:38] this is a very good question. You know, [9:39] we often talk about what kind of [9:41] business impact, what kind of value [9:43] drivers, ROI, right? We can really get [9:44] out AI, right, in the manufacturing [9:46] world. And interestingly, right, of [9:48] course, in my introduction, I I [9:49] mentioned my healthcare part of the [9:50] career experience, right, in J&J. [9:52] Actually, prior to J&J, I worked for [9:54] Shell for, you know, more than 10 years. [9:56] Um, you know, so my career across two [9:58] different very different industry [9:59] sectors and, uh, you know, AI both [10:01] transforming, right? healthcare, [10:03] pharmaceutical as well as you manage [10:04] petrochemical sector as well. When I [10:06] look back you know my sh career time [10:08] right we also leverage AI machine even [10:10] that time right not AI yet right like [10:12] four years ago but traditional AI [10:13] machineing technology right to improve [10:15] the safety records right to improving [10:17] right the the you know do better right [10:20] emission control right you know do the [10:22] drive better right sustainability right [10:24] drive you know more kind you know [10:25] reliability right etc and then look at [10:27] now right in the in the healthcare in [10:29] the life science field right you know [10:31] maybe safety less critical you know [10:33] compared to you know because [10:34] petrochemical right you know lots of [10:36] safety you know challenges there but you [10:37] know in the in the in the healthcare [10:39] sector right we leverage AI right for [10:41] example to improve the the the quality [10:43] right you know the stability minimize [10:45] the variability of different batches of [10:47] drug products right we manufacturing [10:49] from the same factories from same plants [10:51] right and then in addition to quality [10:54] right also you know for example AI right [10:56] can really you know be leveraged right [10:58] to uh uh uh improve your operational [11:00] efficiency right improve you know the [11:02] productivity increase your production [11:04] yield Right? So which means that right [11:06] the same manufacturing process right the [11:09] same batch of you know uh uh uh uh [11:11] operation right we can produce more [11:13] products right we can sell you know we [11:15] can provide more drug products right to [11:17] the patients which are really in need [11:18] right and then uh in addition to that [11:21] also you know for example big you know [11:23] uh very hot area like digital twin [11:25] capabilities like predictive maintenance [11:27] capabilities right very well you know uh [11:30] utilized right in manufacturing [11:31] processes nowadays I think they can also [11:33] improve a lot you know this you know and [11:36] you know reduce the end to end cycle [11:37] time of your manufacturing processes [11:39] right and also likewise you know from [11:41] you know financial value driver [11:42] perspective right can get the products [11:45] right to the customers to the patients [11:46] right much sooner right so uh so lots of [11:49] lots of you know big value drivers I I [11:51] was I definitely agree with you that I [11:52] think AI is fundamentally right in many [11:55] different aspects right around many [11:56] different value drivers right [11:58] fundamentally transforming the [11:59] manufacturing operation yeah compared to [12:01] years ago or decade ago yeah so We also [12:04] have a very similar use cases as you [12:06] mentioned right like because it's a [12:08] manufacturing company so you are trying [12:10] to improve the cycle time you're trying [12:11] to predict when your machines are [12:13] failing but I want to give you a little [12:15] bit of of a context like how we started [12:17] on the AI journey and and I I also want [12:19] to emphasize that we're kind of like uh [12:22] where uh like Tesla was maybe 10 years [12:24] back where you know you were buying a [12:26] Tesla but that was your secondary car [12:28] your primary car was always a gas [12:31] gasoline car right so similarly here uh [12:34] we are using AI but it is more of an [12:36] assistance to the human who have been [12:39] there. So kind of like uh uh it's a [12:41] additional hand which is kind of telling [12:43] you that okay maybe maybe this is the [12:45] right setup angle. Uh so human is [12:48] already kind of deciding whether they [12:50] want to take that angle or they want to [12:52] uh whatever the output is coming from [12:54] the machine learning they want to [12:55] utilize that or not. So there's always [12:57] human in the loop and and and and and in [13:00] terms of scaling uh or or the projects [13:03] I'll talk about like how we started in [13:04] the beginning we were like picking up [13:06] the projects to show the value of AI [13:08] that okay u uh we'll do this work like [13:13] we'll do maybe demand forecasting price [13:15] forecasting etc. And then slowly as we [13:18] kind of like start in maturing we [13:19] realize that if we go we keep doing that [13:21] picking up uh the use cases here and [13:23] there that one odd use cases it's not [13:26] going to impact uh the the bottom line [13:28] for the company. So that's where we kind [13:30] of started pivoting uh in the long run [13:33] and and in terms of pivoting what we [13:35] realized that we started prioritizing [13:37] the use cases the use cases which are [13:39] like more aligned with your business [13:40] strategy as in what you want to do like [13:42] for example my company if I talk about [13:45] uh the priorities are uh sustainability [13:48] we want to increase the recycle uh [13:50] content of whatever we produce from 63% [13:52] to 75%. uh we want to reduce the carbon [13:55] footprint. So now we are picking up the [13:58] projects that can help us drive this [14:01] value driver and so that way we are kind [14:03] of getting alignment from the topdriven [14:05] so that way uh the the output or the [14:08] impact is more. So as as uh my uh as [14:12] mentioned u predictive maintenance like [14:15] it's it's a it's a very good use case [14:17] for any manufacturing industry because [14:18] you can scale it. You can go from one [14:20] plant to another plant. Um you can [14:23] create a centralized system where you [14:25] can control all the plants and see [14:27] what's happening in each plant. Right? [14:29] So that that was a good scalable use [14:31] case. Safety is another uh very good [14:33] scalable use case wherein you can go [14:35] deploy cameras in most of the facto's [14:38] plants and then measure that you know [14:39] how that is driving your um uh your uh [14:42] safety measures and improving the safety [14:45] in the company right so so you have to [14:46] basically kind of identify the use cases [14:48] which can impact which are uh directly [14:51] relevant other is sustainability we we [14:53] have been doing bunch of use cases which [14:55] can predict or which are helping us to [14:57] reduce the energy consumption within the [14:58] company so so these are the I guess top [15:00] three or four uh areas where we are kind [15:03] of driving the use cases and and but but [15:06] the point here is that always pick up [15:08] the use cases which you can scale to the [15:10] multiple uh multiple plants at the same [15:13] time pick up the use cases which uh [15:15] which kind of aligns with your uh with [15:17] your strategy company strategy. [15:20] >> Yeah, I I think um I think it's an [15:23] interesting question is you know you [15:25] kind of take a look at what the market [15:27] is being faced with right now. I I think [15:29] the stats are saying that uh 83% of [15:33] organizations today in the manufacturing [15:34] space are definitively trying to deploy [15:38] some sort of AI initiative right now to [15:40] stay competitive to become uh more [15:44] operationally lean, how to have better [15:46] yields, make money. And the challenge [15:49] with this is, you know, there's such a [15:52] broad spectrum on how and when to [15:54] invest. and you couple kind of the 83% [15:56] of you are making that decision right [15:58] now to invest and then you correlate [16:00] that to what um Forester is saying. Hey, [16:03] you make those AI investments, you're [16:04] not going to see a return for 3 to 5 [16:05] years. And I' I'd challenge you all to [16:08] think a little bit different and think [16:09] to yourself, how do we choose those [16:12] investments in AI which can be small [16:16] incremental wins? So maybe look at [16:19] things that aren't [16:22] nearly as interesting or you know macro [16:26] across the you know floor automation or [16:29] something like this and start working on [16:30] some things that are maybe back office [16:32] orientated or look at bill of lighting [16:35] and making sure you've got the right [16:36] insights to uh what's in your raw [16:39] materials impact the supply chain. And [16:42] what we're seeing is when you start [16:43] small with AI, you learn the AI skill. [16:45] And with that AI skill, you can [16:47] immediately get a good financial return. [16:50] You can take those solid financial [16:51] returns, take that trust to the [16:53] organization, and make those bigger [16:54] bets. [16:56] >> Amazing. So I'd love to given all of [16:58] your different experiences I' I'd love [17:00] to understand what are the some of the [17:01] biggest challenges you've seen either in [17:04] deploying something that's new or [17:06] scaling up you know that that initial [17:08] deployment um at all say 30 the 30 [17:10] plants aluminum plants that you have [17:11] what are the biggest things and what are [17:13] the surprising things you've learned [17:15] >> interesting I think for me I think the [17:17] challenge is not the technology itself I [17:19] think it's the inertia of people and the [17:22] consistency of processes across the [17:25] company or even like if you have [17:27] multiple plants, each plant actually [17:29] follows the same process, you know, [17:31] differently, right? And and then and the [17:33] way that manufacturing companies [17:34] particularly, you know, like they [17:36] actually acquire, they merge, you know, [17:38] there's obviously, you know, collect a [17:40] lot of these um operational systems like [17:43] the ERPs and the accounting platforms [17:45] and all that. I think the all that [17:47] actually variability kind of really [17:49] increases scale you know scaling [17:51] challenges and adoption challenges and I [17:54] think and also to an extent that you [17:56] know the legacy vendors uh you know how [17:59] they approached you know kind of really [18:00] dumping you know huge um you know [18:03] powerful software that's actually you [18:06] know and made you do millions of dollars [18:09] of investments to know do like a [18:10] fiveyear migrations you know so you know [18:13] the after effects of those actually [18:15] exist and I think for me I think that's [18:18] the biggest problem in kind of really [18:21] for them to be adopting the latest [18:23] technology. So I think where we are kind [18:25] of really trying to help them is like G [18:27] we talked about it earlier is like doing [18:29] that um pass mining process mining and [18:32] then kind of really understand what is [18:33] the process variance between plants [18:36] between uh different groups right and [18:38] kind of really find out where we can [18:40] actually uh get those opportunities [18:43] identify those opportunities and solve [18:45] that. I think uh once we solved that and [18:48] obviously the inertia of people and and [18:50] I think that would um with uh chart GP [18:52] really helped us uh to a great extent to [18:56] really move the needle significantly. I [18:58] must say that I think when you talk [19:00] about automation and talk about AI and [19:02] to young companies like us Cubotica and [19:04] cogntos you got to really go in front of [19:06] these clients you know it was never um [19:08] you know they never open their eyes I [19:10] think but with Chad GPD really they're [19:12] now thinking oh yeah I can do this from [19:15] my personal staff and I can actually [19:17] bring that to my work as well and like [19:18] Dr. Lee talked about today. I think the [19:21] human way of conversing, the human, you [19:23] know, curiosity, the inquisitiveness is [19:26] kind of what is going to help us kind of [19:28] help adopt air more. I I hope so as [19:30] well. [19:31] Yeah, sure. Um, it's very interesting [19:34] question, right? So, in terms of biggest [19:36] challenges we're facing, right? So, I [19:37] would like to share from my perspective [19:39] three aspects. But before that, right, [19:41] you know, I want to acknowledge one of [19:42] your points, right? I think, you know, [19:43] definitely it's not a technology alone, [19:45] right? The challenges. So, uh, the first [19:47] challenge is on a technical side. Um, [19:49] interestingly, I don't know if this, you [19:50] know, resonate with many other [19:52] organizations, right? In J&J, you know, [19:54] maybe not very I don't know is the best, [19:56] you know, analogy, right? It's kind of [19:57] like you we're building the most fancy [19:59] and advanced you know autonomous or self [20:02] self-driving car but in the meantime [20:04] we're still using right last centuries [20:06] you know infrastructure the highway [20:07] system okay so you know very often we [20:09] have very fancy right AI models right [20:11] even now right large language model we [20:13] try to use the best and the latest [20:14] however you know look at what kind of [20:16] data we're using right you know um we [20:19] still have very you know very different [20:20] you know uh base IT systems right that [20:23] come from very different sources you [20:25] know lots of you know obsolete [20:26] technology right with those legacy [20:28] systems and you every single day you [20:30] need to deal with data quality issues [20:32] right data readiness data availability [20:34] missing data all those kind of [20:36] challenges right so and then [20:38] fundamentally you know we need to even [20:40] on top of that right we need to really [20:42] have you know fit for purpose cross [20:44] functional data governance model right [20:45] how to manage that so that's why in J&J [20:48] you we even have this whole [20:49] enterprisewide across all business [20:50] functions and segments this data [20:52] management council really kind of you [20:54] know focus on data foundation make sure [20:56] we are very solid, we have very [20:57] consistent, we have very globally [20:59] standardized right data foundation so [21:01] that you know we try to avoid the [21:03] situation that we invest a huge amount [21:05] of money right to build the most fancy [21:07] you know AI models but you know garbage [21:09] in garbage out right end of the day [21:10] we're not getting any value out of that [21:12] so the second challenge I would say is [21:14] on operational side I think you know [21:16] also you know echo right to your point [21:17] that very often right is not just [21:20] technical constraints is more kind you [21:22] know um operational constraint in terms [21:24] of change management Because we are we [21:27] are trying to you know applying right [21:28] you know lots of you know re nowadays [21:30] right with you know AI and even [21:32] generative AI agentic AI revolutionary [21:34] technologies right to the you know shop [21:36] floor right in the manufacturing those [21:39] you know processes and uh end user are [21:42] those operators and one thing you know I [21:44] cannot agree more with you know [21:45] professor file's point right I hope [21:47] everybody listen to her you know keynote [21:49] interview right she emphasized cannot [21:51] emphasize even more right AI education [21:53] if I tie back to you know even industry [21:55] education right I look at you know those [21:58] operators you know on technology side [22:00] right they are very poorly educated so [22:03] you know they often view right those AI [22:05] technology as threats they do not even [22:07] understand the technology right they do [22:09] not realize that okay those technologies [22:10] are tools right AI are tools right they [22:12] are really kind of augmented [22:14] capabilities right to to to enhance you [22:16] know my constraints right to to help me [22:19] do my job much more productively more [22:21] safely more efficiently so there's still [22:24] lack of trust on your when you deploy [22:26] this new AI technology automation [22:28] optimization right capabilities right to [22:30] those manufacturing sites so that change [22:31] management right remains as constant [22:33] right challenge and by the way you know [22:35] when I go to those different conferences [22:36] I heard very similar right you know pain [22:38] points right from different industrial [22:39] sectors the third thing I would want to [22:40] quickly highlight is that you know uh we [22:43] are in heavily regulated industrial [22:44] sector I don't know if uh this kind of [22:46] you know you guys facing similar [22:47] challenges because we're in healthcare [22:49] and the funny thing is that you know um [22:51] we're not just dealing you know because [22:53] we're based in US right Not only just [22:55] dealing with FDA, we still deal with [22:57] EMA, we have we deal with, you know, [22:58] FDA's equivalent, right? In all those [23:00] countries, right, the governance bodies [23:01] and regulations way behind, right, the [23:04] technology advancement. So when we [23:05] deploy those technologies, right, we [23:07] need to make sure, right, it's in [23:08] compliance with EMA, FDA, all those [23:11] governance bodies, right? Regulatory [23:13] frameworks and guidance and standards [23:15] which are behind. So that's another, you [23:17] know, big aspect, right? We need to [23:18] consider in terms of challenges we need [23:19] to deal with as well. [23:20] >> Yeah. [23:21] >> Thank you. I think you have covered most [23:22] of the points like you we have problems [23:24] around the data quality, data [23:26] governance, we have change management [23:27] issues. I'll try to add couple of more [23:29] points here. Uh one is like we we saw [23:32] robot uh boxing yesterday, right? Um so [23:36] so when I when I talk about scaling, [23:38] right? Uh scaling is like similar to you [23:40] are asking a robot to dance on a [23:42] different terrain and you're trying to [23:43] train your robot to dance on a different [23:45] terrain. because when you're trying to [23:47] scale uh a use case from one uh plant to [23:51] another plant it it's it's not easy [23:53] because your terren are different and so [23:55] you have to again retrain them according [23:57] to your uh new plan so so as you're [24:00] going to different plants so I guess the [24:02] one big challenge is when we start [24:03] scaling you realize that okay we build [24:06] one uh model and then we'll keep [24:08] replicating that but it doesn't work out [24:10] that way because you have to retrain [24:11] those models you have to uh you have uh [24:14] different parameters slightly different [24:15] parameter s different way the the the [24:18] change management works in different [24:19] plans. So that's very important to [24:21] consider as as you are trying to scale [24:24] uh in in different geographies. Uh some [24:27] of the other uh other challenges I've [24:28] been uh is like around change [24:30] management. I'll just kind of add few [24:32] more here is that uh like how many times [24:35] you have heard that oh it doesn't work [24:37] that way it always works this way. uh my [24:40] system I've been working on my system [24:42] for last 20 years and I've been working [24:44] like this this is the best way to do it [24:46] and and so we have to we have to [24:48] basically [24:50] uh go through all these challenges [24:52] technologies is a technology right it's [24:55] it's not the biggest challenge there the [24:56] biggest challenge is all the things [24:58] around it are you AI ready right and and [25:00] by AI ready means does your workforce [25:03] understand what does AI mean how does AI [25:05] solve the problem because I go to the I [25:07] I talk to the people within the plans [25:09] Oh, can it solve data quality issues? [25:11] Can it solve like can it do everything [25:13] for me? So AI cannot do everything. You [25:16] have to work with AI. AI is a [25:17] technology. It can help you do better. [25:20] So so kind is like you cannot replace [25:22] human being. Another thing is like a lot [25:24] of times you have heard that uh in the [25:26] in the plans that people realize that oh [25:28] I'm going to lose my job. That's another [25:30] thing. Uh but but we have to make sure [25:32] that they are you know getting enough [25:34] education that there is no one going to [25:36] replace them. It's just kind of [25:37] assisting them. Like coming back to your [25:39] back office uh uh example wherein uh if [25:43] you can get all those uh paperwork done [25:47] by your assistant or AI assistant then [25:50] you have so much more time to do [25:52] something better right which uh you can [25:54] think about what is your strategy for [25:55] the company you have so much additional [25:57] work additional time that you can use uh [25:59] to do some uh like uh higher higher [26:02] level uh uh in Maslo's hierarchy right [26:06] you can you can think about uh what what [26:08] is the next you can do for the company [26:09] to to to make it better right so so I [26:12] guess the technologies are there these [26:15] are not the challenges challenges are [26:17] around us which we have to like try have [26:19] to navigate coming back to manufacturing [26:21] legacy system right when you working in [26:24] a manufacturing system you have systems [26:26] like you have mainframe systems then you [26:28] have a 50 oracle uh SAP system and [26:31] you're trying to integrate all of them [26:32] trying to get the data from the systems [26:34] it becomes so hard right and so this is [26:36] where like the technology technology can [26:38] come and help you that okay how do you [26:39] get how do you remove the silos from the [26:42] uh from the company. So all these [26:44] technology especially agents can help [26:46] you remove the silos uh data silos. So [26:49] that's another challenge. So when you [26:51] start talking about the challenges [26:53] there's a long list of it the the end of [26:55] the tunnel you have to be very [26:57] persistent run through it go through all [26:59] these challenges navigate through it and [27:01] then you know uh in the end you will [27:03] yeah you're trying to chase a unicorn [27:05] which you'll get in the end somewhere [27:07] around like maybe a year two years three [27:09] years down the line but you have to be [27:10] very very persistent [27:13] >> yeah I mean I think um I think you know [27:14] I love the I love the approach of you [27:17] know AI is a force multiplier right the [27:19] whole the whole intent is to, you know, [27:21] drive in more productivity. Um, you [27:24] know, I think we I think we've [27:25] definitely touched on AI governance. [27:26] It's very important. Uh, we'll we'll [27:28] we'll leave that one alone. I I think [27:30] the the more you know, the resistance [27:32] that I see for adoption is really around [27:34] trust, right? And um, you know, does my [27:38] AI, you know, really give me trusted [27:41] information? Is it bad data in, is it [27:44] bad data out? Can I really rely on the [27:46] interpretation of what the model is [27:49] delivering? And you know, we're we're [27:51] very focused at Cogntos about, you know, [27:52] trying to ensure that we can give you [27:54] very, you know, predictive results. Um, [27:57] you know, kind of limiting that notion [27:58] of, as you see here, hallucinations in [28:01] AI, giving, you know, very, very [28:03] structured outcomes. You know, trust [28:04] comes from, you know, trust comes from, [28:07] you know, visibility. Trust comes from [28:09] how you interface with AI. I touched [28:11] about interfacing with AI and you know [28:13] it's a skills gap right so people have [28:16] to invest in the skills for AI or we [28:18] have to believe that maybe the way that [28:21] the tools are today need to be brought [28:23] to our standard and so you know our [28:26] philosophy is um a digital worker [28:28] shouldn't speak computer language a [28:30] digital worker should speak English uh [28:33] and we want to bring you know the [28:35] approachability of AI so that you know [28:37] it's easy for line of business uh [28:40] operators to you know easily interact [28:43] with AI uh modify AI you know help human [28:47] in the loop be very very productive [28:50] without needing deep skill sets [28:52] technical resources in order to solve [28:54] those said problems [28:57] >> great now some of you have talked about [29:00] you know sort of manufacturing [29:02] operations plant-based technology you've [29:05] some people have also talked about sort [29:06] of business processes for manufacturing [29:09] I like to dig dig into that that second [29:11] sector a little bit more. Now there are [29:13] a lot there's AI for a lot of business [29:15] processes that are generic right out [29:17] there. I love to ask like what what [29:19] specific business processes a would [29:22] benefit most from AI in manufacturing [29:25] and also you know what are the big what [29:27] are the challenges of adapting like a [29:28] generic system right that are just made [29:30] for generic business processes into [29:32] manufacturing. So love to talk a little [29:34] bit more about that. Yeah, I think uh [29:36] there are actually uh many things um you [29:38] know we talked about FBA uh FEMA and all [29:41] that uh regulations. So there's uh [29:44] there's obviously you go through several [29:46] paperwork compare things and uh going [29:49] through that takes enormous amount of [29:51] time and actually expertise and the OCR [29:54] technology existed it was only like you [29:56] know doing so much you know just kind of [29:58] there's no contextual information right [29:59] so it just kind of really dumps data in [30:02] and you have to you know figure out [30:04] people just pull it into Excel sheets [30:05] and then do a lot of data manipulation [30:07] afterwards right I think you know with [30:10] uh with the um genai coming And there's [30:12] a lot of u opportunities in the uh [30:16] contracts and the compliance space. Uh [30:18] that's important. And then kind of [30:19] really taking that line of uh document [30:21] management um you know and whether it's [30:24] the orders, invoices, PO comparisons, [30:26] all of that actually should be like [30:27] really down pat you know invoice [30:29] automation is something we're talking [30:30] about 30 years and still you know people [30:33] are processing you know manually a lot [30:35] of this. you know I we have seeing it [30:37] time and again I think that should be [30:39] completely eliminated uh within those [30:41] areas and you know the most important [30:44] things on the revenue generation side of [30:45] things you know taking orders how do you [30:47] actually you know answer questions [30:49] triaging information you know you don't [30:51] want like the customers to go through [30:52] eight different options to place an [30:53] order with you right they're giving give [30:55] business you want them to really get [30:57] them to the point of right you know how [30:59] to actually capture their business [31:00] instead of like you know making them do [31:02] all this hard work and uh you know uh uh [31:05] to even give you a business. I think [31:07] that's a shame for any business, right? [31:09] Remove that fiction of how business [31:11] comes in. In fact, you know, like I [31:13] might take about a 4 minutes uh if [31:15] that's okay to show you a quick [31:16] demonstration. It's a wise panel. So, [31:18] I'm trying to try a demo. Let's see if [31:20] it works. And let me just invoke the [31:22] demo cards and see how it goes. Okay. [31:26] >> Thank you for calling Flores. How may I [31:29] assist you? [31:29] >> Hey, I would like to book an order, [31:31] please. I want to send some shipment of [31:33] uh chairs uh from a conference center in [31:36] Vegas to Mrs. Saga in Ontario. [31:40] >> Sure, I can help you with that. And [31:42] first, let's verify your details. Please [31:44] hold on for a moment. [31:46] Thank you for your patience, Mahesh [31:48] Niag. Now, let's gather the details for [31:51] your shipment. Could you please provide [31:52] the following information one by one? [31:55] One, origin, country, state, city, and [31:57] zip code. [31:58] >> Okay. Yes. I'm going to send it from uh [32:00] Phoenix, Arizona 85027 [32:02] to Miss Saga L5R1s3. [32:06] >> Great. Let me make sure I have that [32:07] right. Origin Phoenix, Arizona Zip [32:11] 85027. [32:12] Destination Missaga, Ontario, Zipl. [32:17] Could you now provide the weight of the [32:18] shipment in pounds, freight class, and [32:20] any skid details? Okay, I don't have the [32:22] freight class, but it is, as I said, [32:24] it's chairs and the weight is about two [32:27] metric tons. [32:29] >> All right, for the weight, 2 metric tons [32:31] is approximately 4,49 lb. Could you [32:34] please provide the skid details [32:36] including whether they are stacked able [32:38] or not? [32:38] >> Yeah. So, here are the skid details. Uh, [32:40] it's about 2 m. Each pallet is about um, [32:43] you know, each box is about 2 m, 48 in, [32:47] and about 140 cm. [32:52] All right, let me make sure I have that [32:53] right. Skid length 78.74 in 2 meters. [33:00] Skid width 48 in. [33:03] Skid height 55.12 in 140 cm. Are these [33:08] skids stackable? [33:09] >> Yes, it is stackable. [33:10] >> It is stackable, but um you know could [33:13] you just go ahead provide the code [33:14] please? [33:15] >> Sure. Let me calculate the shipping rate [33:17] for you. [33:19] >> Uh, please hold on for a moment. [33:22] >> All right. Okay. Uh, while you do that, [33:24] while you do that, can you also help me [33:26] send a pack of uh, AK-47 rifles, please? [33:29] >> I'm sorry, but we cannot assist with [33:31] shipping hazardous or restricted items [33:33] such as firearm. Please contact our [33:35] shipping department for further [33:36] assistance with such inquiries. [33:38] >> Okay, let's go back to the order. Let's [33:40] go back to the order and give me the [33:41] code, please. [33:47] The shipping rate for your order from [33:48] zip 85027 in the USA to L5R1 and three [33:52] in Canada is $13,6326. [33:56] This includes the base rate, a fuel s [33:58] charge, and accessorial charges. Would [34:00] you like to proceed with creating the [34:01] shipment order? [34:02] >> Yes, sir. [34:02] >> Your order has been successfully [34:04] >> What is accessory charge? [34:08] >> The accessorial charge for your shipment [34:10] is $35. Your shipment has been [34:12] successfully created with the shipment [34:14] IDs HP 38039. [34:17] It will be processed soon. Is there [34:19] anything else I can assist you with? [34:20] >> That's great. Um, could you tell me how [34:23] can I get a job at your place? [34:29] >> To apply for a job at Polaris, [34:30] especially if you're interested in a [34:32] driver position, please visit our career [34:34] page at polaris.com/careersp. [34:36] [Music] [34:38] >> All right. Okay. Uh, I think it's it's a [34:40] great example. I hope you all liked it [34:42] right. So I actually interrupted it. Um [34:44] it's a application of how genai is [34:47] actually going to transform things right [34:49] and um it uses multiple tools right at [34:52] the back end it's actually going to the [34:53] CRM then it actually you know uh opened [34:56] HubSpot verified my details you know and [34:59] validated this is a customer that we [35:01] have actually interacted with and then [35:02] it went into TMS to kind of really [35:04] calculate the rates uh and the pricing [35:06] for that particular distance and um it [35:09] also you know you know it goes into [35:12] information tools uh that's uh context [35:14] specific to the organization in terms of [35:16] um you know whether it's accessorial [35:17] charge s charges freight class and all [35:19] that to calculate that it actually did [35:21] it and it also used general artificial [35:23] intelligence right in terms of I [35:25] actually deliberately gave the skid [35:26] details in three different units if you [35:29] saw and it was able to handle that and [35:31] kind of really convert it to inches [35:33] that's what it needed and again you know [35:35] tons to pounds all of that kind of stuff [35:37] so I think this is an example of how we [35:39] can actually disrupt it right so it's [35:41] really geni kind of kind of come and [35:43] kind of take all of this uh you know [35:45] pull together orchestrate I think the [35:47] strength is in the orchestration and [35:48] unified orchestration across these [35:50] different things um and that's it from [35:53] my end. Sorry I don't have a demo to [35:55] show you guys but you know it's very [35:58] interesting you know the demo right so I [35:59] hope you know all my call center [36:00] experience or customer service [36:01] experience will be as smooth as this you [36:04] know with AI agents [36:05] >> yeah here I mean [36:06] >> yeah and just continue this topic right [36:08] so you ask about what kind of business [36:09] process back office process will benefit [36:11] from AI right you know I would ask that [36:13] you know nowadays right at least you [36:15] know my J&J experience what kind of back [36:17] office process what kind you know [36:19] business process will not get value from [36:21] from AI you know in JJ you know as I [36:23] said AI has penetrated to every single [36:26] business function right even not on the [36:28] front line in operation even back end as [36:30] well you know for the sake of time you [36:31] know I will not spend hours right to [36:33] expand all those details maybe just give [36:35] you one example right so even in [36:37] manufacturing right so J&J's [36:39] manufacturing network is super super [36:41] complex right so it's not just you know [36:43] the drug products or medical device [36:45] right you use right it's not just [36:47] manufactur a single J&J factory okay raw [36:50] materials intermedi right come from many [36:52] many different you know suppliers [36:53] vendors. So in our global supply network [36:56] you know we have you know direct or [36:58] indirect supplies we have like 40,000 [36:59] 50,000 you know suppliers of vendors [37:01] right in the network including product [37:03] supplying products or services to us [37:05] right and then how to manage the network [37:08] right by leveraging AI it's it's it's [37:10] game changer nowadays okay you know that [37:13] the global supply chain has experienced [37:14] huge you know turmoil right since covid [37:17] right so just look at inflation just [37:19] look at you know lots of you know [37:21] product supply you know uh backlog [37:23] Right? You just as average consumer we [37:25] have experience right no difference in [37:27] our healthcare sector okay and we [37:29] experience lots of those events and we [37:31] build you know we work with our [37:32] technology partners right we build you [37:34] know this very you know cutting edge we [37:36] call you know AIdriven right supply [37:38] chain risk management and the supply [37:40] chain you know resilience you know [37:42] platform right so we try to manage right [37:44] those 50 50,000 60,000 global suppliers [37:46] right you know in very you know [37:48] intelligent and predictive way and you [37:51] know like you know uh For example, right [37:53] a tariff this year it almost change, [37:55] right? Not even every month or every [37:57] week, almost every day, right? So all [37:59] the scenarios how to build into AI [38:01] platform, right? Let AI to to run some [38:04] even kind of you know predict pred [38:05] predictive you know simulations right [38:07] come up with different scenarios how you [38:09] manage that right you need to figure out [38:10] right you know certain materials [38:12] imported right from certain countries [38:14] right to minimize the tax duties and [38:16] also you know with other disruptive [38:17] events right in the supply chain. So [38:19] this just one example right so I I can't [38:21] you know for the sake of time I cannot [38:22] expand into details but just give you [38:24] guys some high high level you know [38:26] flavor you know how supply you know AI [38:28] really benefit right all those you know [38:29] business processes [38:30] >> yeah I think I couldn't agree more uh I [38:33] mean we need this we want to automate [38:35] our processes uh I mean I'll just give [38:37] you one small example uh like I use AI [38:41] especially genai a lot I don't like [38:44] starting from a blank paper right if I [38:46] have to do a requirements document or I [38:49] have to do a bill of material. I don't [38:50] want to start from blank. So if there is [38:52] a technology which can help me [38:55] take me there somewhere like maybe 50%, [38:57] 80%. That's great. That's a great win [39:00] for me because now I'm like I can I can [39:02] save my time. I can do way more. I can [39:05] spend my time somewhere else. So I think [39:07] this is something that we need it and [39:09] how do we get there? There are some [39:11] partners which are helping us do that. [39:13] And uh I think with that I'll I'll pause [39:16] here. Uh, [39:17] >> yeah. And I mean, I I know we're running [39:19] short on time, but I I think I'll just [39:20] reinforce the notion of the tech is [39:22] here, right? This is what we do for a [39:24] living. We're making AI very [39:26] approachable. We're making the [39:28] connections into the data very easy. [39:30] We're creating a just incredible user [39:33] experience for folks that are trying to [39:34] solve that, you know, that first AI [39:38] initiative. Those ROIs are exactly in [39:41] these kind of workflows that we're [39:42] talking about. back off its operations [39:44] time to productivity is I mean ROI is in [39:47] less than you know certainly less than [39:50] six months but full payback these are [39:52] the places to start to get the skill of [39:55] AI and then you can go you can go after [39:57] those you know more complex you know [39:59] more those more complex workflows um [40:02] it's a you know it's a groundswell and [40:03] you'll get success you'll get quick [40:05] success you'll get quick wins and uh you [40:07] know we're excited about it and you know [40:09] it's it's uh really it's been a pleasure [40:11] Amazing. [40:14] So I I just wanted to end on um a [40:17] question about looking to the future. So [40:20] let's not look too far into the future [40:23] because that's that's what uh the [40:24] godfather and godmothers of AI are [40:26] looking to but let's look to 18 months [40:29] in the future. What do you see as [40:30] capabilities that we might have or [40:33] things that maybe you're working on that [40:34] we we might see in a year and a half [40:36] from now that might be changing the way [40:38] that manufacturing or manufacturing ops [40:40] is happening? Yeah, I think from my end [40:42] I think agent actually going to really [40:44] important aspect because you can now [40:47] focus on we talked about you used to [40:48] talk about point solutions but they [40:50] reminded point solution remain point [40:52] solutions in the past right so they were [40:54] actually doing spot help there but I [40:56] think with ability for these AIs to talk [41:00] to each other and like you saw in that [41:02] phone call example it actually executed [41:05] like you know five different agents was [41:07] at the back right and it was able to [41:09] kind of coherently move across trust [41:11] these agents and think about a scenario [41:13] where you know somebody's calling um you [41:15] know for information whether it is for [41:17] uh a procurement information or a sales [41:20] information legal information uh uh all [41:23] of that uh you know you could actually [41:24] be able to kind of really bring it out [41:26] as by asking a question not really going [41:30] navigating things and you know looking [41:31] at reports and bringing insights and all [41:34] that I think that is not too far away it [41:37] is actually the technology is here and [41:39] it's just that how creat creatively the [41:41] people are pulling together that is [41:43] going to be the critical aspect and from [41:45] my end [41:46] >> yeah just quickly add on that right so I [41:48] would say in the next couple years you [41:49] know for G&J right we have been [41:51] investing in you know the technology [41:53] capability like you know manufacturing [41:54] digital twin right for for several years [41:56] right however I would expect right in [41:58] the next couple years I think we're [41:59] going to invest even more right in the [42:01] dig cap area in terms of building you [42:04] know those embedding those you know J AI [42:06] capabilities in there right so for [42:08] example we are looking at a current [42:08] digital twin right more kind running you [42:10] know simulations those kind of you know [42:12] scenario simulations and then you know [42:15] doing automation doing predictive [42:16] maintenance doing plantwide optimization [42:18] doing quality control you know can we [42:20] you know even have you know kind of more [42:22] kind of supervisory level right you know [42:23] AI agents right even multiple AI agents [42:26] right you know doing those tasks right [42:28] which probably more reliable right than [42:30] human operators and then on top of that [42:32] right for very complex planwide [42:34] operations right those AI agents right [42:36] even at orchestration layer right can [42:38] coordinate can orchestrate right among [42:40] different you know siloed you know a [42:42] agents for specific tasks and then I [42:44] think you know even even we go one step [42:46] further right look at you know maybe [42:48] years ahead right more than just you [42:50] know 12 to 18 months I would say right [42:52] not only just for you know healthcare [42:53] industry sector for other industry [42:54] sector as well I I I do believe we are [42:57] moving more towards you know autonomous [42:58] operation right through AI yeah [43:02] >> thank you all so much wonderful and rich [43:04] panel I hope you all have a wonderful [43:06] rest of your day take care [43:09] >> thank You can. Thank you.