JLL Ties Executive Bonuses to AI Adoption to Break Process Gridlock

JLL将高管奖金与AI应用挂钩以破除流程僵局 企业AI落地失败多源于割裂的组织孤岛与内部摩擦，而非模型局限，这促使企业在部署智能体之前重构高管激励机制并整合跨职能治理。 来自JLL、Zumba Fitness、NIX和Phenom的企业领袖分享了从基础AI实验走向规模化智能体部署的实用策略。专家小组强调，应优先关注高影响力的商业投资回报率（ROI），而非微小的效率提升，并通过将治理权统一归于单一负责领导者来打破组织孤岛。专家们详细阐述了具体的架构与运营框架——例如Phenom的四层系统、NIX的AI准入框架以及JLL的“离岸、自动化、智能体化”方法论——同时强调必须将高管薪酬与AI熟练度挂钩，并协助员工转型为更高价值的智能体编排者。 通用基础模型大约只能解决30%的企业用例，而其余70%则需要深度的垂直领域上下文和定制化的智能体控制架构。 (3:09) 企业应优先考虑直接的投资回报率和可扩展的商业价值，而非渐进式的员工生产力与效率提升。 (8:37) 相较于自主生成代码或内容，建立工作流治理、决策权归属和审批层级面临着更大的挑战。 (10:33) 现代智能体系统需要包含应用层、配置层、记忆/本体层以及编排层的四层架构。 (15:20) 分散在数十个团队和预算中的庞杂跨职能工作流，必须整合至拥有统一预算审批权的单一责任领导者之下。 (19:25)

Enterprise leaders deploying agentic AI emphasize that sustainable returns require restructuring fragmented organizational workflows, unifying cross-functional governance, and redesigning human roles rather than simply layering automation onto legacy systems. Re-Architecting Systems for Measurable Business Value (8:37) Panelists argue that enterprise AI initiatives must pivot away from narrow productivity metrics toward scalable, top-line business outcomes and domain-specific return on investment. Carlin Power notes that after JLL spent two and a half years developing proprietary tools to avoid vendor lock-in, the commercial real estate firm shifted to a hybrid strategy: adopting established vendor platforms while engineering proprietary agents where specialized domain context yields exponential value. Mahe Bayireddi of Phenom observes that general-purpose models address roughly 30% of enterprise requirements, while the remaining 70% require deep vertical context and customized frameworks. Modern agentic implementations also demand a shift from legacy two-layer software designs—traditionally limited to application and configuration—to four-tier architectures incorporating dedicated memory, ontology, and orchestration layers. In practice, systems like Zumba Fitness’s Kodu software engine illustrate this complexity by autonomously planning features, writing web and mobile code, clarifying project ambiguities across Slack and Jira, and generating staging previews prior to human deployment. Dismantling Departmental Silos and Governance Gridlock (19:25) Panelists emphasize that organizational politics and operational fragmentation pose far greater obstacles to AI deployment than algorithmic limitations. Power illustrates this friction through JLL's internal "request to resolution" process—an inquiry workflow that had fractured across 12 disconnected systems, 74 teams, 17 separate budgets, and roughly 240 executive stakeholders. To eliminate administrative gridlock and competing departmental priorities between the CIO, CFO, and CHRO, JLL consolidated the process under a single executive with dedicated cross-functional budget authority measured against firm-wide return on investment. Executive incentives must align with this operational consolidation. Power explains that JLL's chief executive officer tied global executive board compensation directly to mandatory AI transformation benchmarks, assessing leaders on enterprise fluency scores and the percentage of automated workflows within their business units. Bayireddi similarly asserts that technical deployments collapse when corporate governance is misaligned, advocating for frameworks that evaluate agent architectures across business unit P&L metrics, role hierarchies, and regional compliance structures. Mitigating Adoption Friction and Upskilling the Workforce (21:34) Transitioning staff from routine task execution to reviewing and orchestrating agent output generates significant cultural friction. Ralph Valdes highlights that employee time scarcity remains a primary impediment, as workers legitimately push back against training demands that interrupt daily responsibilities. To reduce adoption barriers, Zumba integrated Kodu directly into native environments like Slack and Jira rather than requiring staff to navigate unfamiliar dashboards. Valdes views AI as an augmentative instrument, cautioning that while AI itself will not eliminate roles, professionals who leverage the technology will replace those who do not. Fostering adoption requires systematically redefining departmental value propositions. Power describes guiding JLL's technical support personnel through an 80% reduction in routine ticket handling, shifting their core function into knowledge base managers and "agent-to-agent orchestrators." In technical domains, Max Ushchenko observes that software engineers are evolving into a "human orchestra" capable of handling five or more legacy roles simultaneously. Across broader operations, Bayireddi cites Baylor Scott & White Health, where screening agents reduced recruiter administration by 20% to 25%, and Thermo Fisher Scientific, which redeployed recruiters from routine candidate sourcing directly into hands-on onboarding. The Process Discovery Challenge and Standardization (39:46) Automating an unstable or poorly understood process yields counterproductive outcomes, with Ushchenko warning that AI magnifies underlying workflow defects rather than repairing them. Panelists acknowledge that leveraging AI for automated, top-of-funnel process discovery remains an unresolved industry challenge. While structured software codebases are readily mapped, autonomous tools cannot yet reliably uncover deeply buried tribal knowledge or fragile dependencies where a single unrecorded manual task could disrupt downchain operations. Consequently, enterprises rely on structured human frameworks before deploying agents. JLL utilizes a three-phase pipeline—transition, automate, and agentify—deploying Lean Six Sigma process engineers in Bangalore to map workflows within platforms like Mendix and Blueprint prior to introducing autonomous agents. Alternatively, Phenom addresses discovery by analyzing aggregated operational data across more than 700 enterprise customers, using vertical ontologies to pinpoint standard bottlenecks and prescribe automation targets without requiring clients to map internal processes from scratch. While agentic systems demonstrate tangible operational gains across talent acquisition, software development, and enterprise support, autonomous workflow discovery remains an unresolved industry challenge. Sustainable enterprise adoption ultimately depends on executive-led governance overhauls and disciplined process standardization rather than raw algorithmic capabilities.

企业推进智能体人工智能（Agentic AI）落地的核心难点在于跨职能流程重构与组织变革治理，而非模型算法本身；来自JLL、Phenom、Zumba Fitness及NIX的专家一致指出，企业必须从单纯关注效率转向追求明确的投资回报率（ROI），通过重构多层系统架构并对齐高管激励来化解落地阻力。 重塑架构与治理机制：化解跨部门割裂 (15:20) Mahe Bayireddi与Ralph Valdes强调，将AI简单“外挂”于遗留流程无法实现预期价值。传统企业软件多为应用与配置的双层结构，而现代化智能体体系需涵盖应用层、配置层、记忆与本体层及编排层共四层架构。项目推进受阻往往非因技术缺陷，而是由治理脱节及内部部门阻力所致。Valdes指出，代码生成技术已相对成熟，难点在于工作流治理中各环节的审批、设计与歧异仲裁；为此，Zumba将其软件工厂Kodu直接集成进员工日常使用的Slack和Jira中，以此降低工具切换门槛。 针对跨职能业务割裂，Carlin Power分享了JLL梳理“从请求到解决”流程的实践。该流程曾分散在12个系统、74个团队、17项独立预算及约240名高管之间，极易引发权责冲突。为消除推诿，企业打破了CHRO、CFO与CIO之间的条块分割，委任单一负责人统管端到端流程并享有专属预算，考核指标完全与公司整体投资回报绑定，取代了原先狭隘的局部职能指标。 锚定价值回报与自动化准入评估框架 (8:37) 在技术路径选择上，Carlin Power表示JLL已从早期专注自研大模型转向“自研与外购结合”的混合策略，借助成熟供应商底层能力的同时，仅在具备高商业价值的垂直领域定制专有系统。Power指出，大型企业重构系统的沉没成本极高，零敲碎打的微小工时节省不足以支撑巨额投入，AI项目必须证明其能带来指数级商业回报。此外行业经验显示，通用基础模型通常仅能覆盖约30%的企业用例，剩余70%高度依赖特定业务上下文与专业护栏。 为了精准筛选适宜智能体自动化的工作流，Max Ushchenko介绍了NIX建立的“AI资格评估框架”。该框架从三个维度进行前置量化：其一是人机协同机制（Human-in-the-Loop），厘清人工介入需全时监控、阶段抽查还是仅做终审校验；其二是开发维护成本与单次任务节省工时的对比，确保投产收益比健康；其三是任务触发频次，分析小时、日或月度运行规模，避免因任务低频导致投入浪费。 组织变革管理：打破时间稀缺与重塑激励 (21:34) 当员工角色从基础执行转向对AI成果的审核与协同，组织内部往往面临文化阻力。Ralph Valdes观察到员工技能升级的最大障碍在于“时间稀缺”——员工因全职业务处于饱和负荷状态，往往排斥为两个月后可能显现的收益而在当下放慢步调去学习新工具；因此，管理层需通过手把手流程推演与实际成效展示来降低启动阻力。Max Ushchenko则提出，现代软件工程师正在向兼顾五项以上职能的“一人交响乐团”演进，技术人员接到任务的首要认知步骤已转变为即时判断应由AI独立完成、人工处理还是人机协同。 为确保组织变革有效推进，Carlin Power指出文化调整必须由CEO自顶向下发起。JLL将全球执行委员会成员的奖金与AI赋能KPI强制挂钩，考核各业务单元的AI熟练度、流程自动化占比及平台采纳率。同时，管理团队引导各部门重塑业务定位，明确哪些工作被替代、增强或补充；例如其技术支持团队在两年前被预警80%传统工单将消失后，已成功转型为高维度的知识库专家及“智能体间编排者”（Agent-to-Agent Orchestrators）。 攻克流程发现盲区：分步演进与垂直本体 (39:46) 在实施智能体之前，全面厘清深层且隐蔽的手工工作流仍是行业普遍痛点。座谈专家承认，自动发现那些依赖基层个别员工非书面习惯、牵一发动全身的深层业务依赖项，目前尚无成熟的纯AI解决方案。为此，JLL采取了务实的混合策略，依托其在班加罗尔的业务服务中心，由精益六西格玛工程师通过Mendix和Blueprint等软件系统化映射L0至L2级流程，遵循“离岸过渡—低代码数字化—智能体化”的三阶段路径，以控制改造成本。 Mahe Bayireddi则展示了依托跨企业本体实现流程发现的方案。Phenom利用覆盖700多家企业客户及每18个月超10亿次交互的业务数据构建行业本体，先于客户掌握垂直领域的普遍瓶颈并据此施策。Bayireddi分享了两个实际成效：在面临护士退休潮与招募难的贝勒斯考特与怀特健康系统（Baylor Scott & White Health），自动化筛选帮助招聘人员每周节省20%至25%的行政时间；而在赛默飞世尔科技（Thermo Fisher Scientific），智能体管线加快人员导入后，原先从事重复搜寻的HR被动态转岗至现场入职管理工作。 Max Ushchenko与多位专家提醒，若基础业务流程本身残缺或未经优化，盲目叠加AI只会放大既有缺陷而无法解决根源问题。目前全自动提取非结构化、碎片化人工隐性知识仍是未解的技术难题，企业推进智能体转型的成功关键，依然取决于前期严谨的业务梳理、高管示范带动以及保持对可量化业务价值的战略定力。

## Background 该内容由 Ai4 频道于 2026 年 8 月 22 日发布在 YouTube 平台。视频记录了一场围绕“智能体 AI（Agentic AI）的工作流重构与变革管理”的企业高管圆桌论坛，汇聚了来自 Zumba、JLL、NIX 和 Phenom 等企业的技术与业务领袖，探讨企业如何重塑跨职能流程、对齐领导层与激励机制、应对组织文化阻力，并平稳推进自主智能体的落地应用。 ## TL;DR 论坛嘉宾结合 Ask JLL 领域智能体、Zumba 的 Kodu 智能体软件工厂、Phenom 的人力资源垂直大模型以及 NIX 的 SDLC 流程重构等一线实践，系统阐述了智能体 AI 的落地路径。讨论强调企业必须从单纯追求工时节省转向聚焦核心业务 ROI，通过设立单一流程负责人打破部门壁垒与高管冲突。同时，嘉宾们提出了智能体四层架构、AI 适用度评估模型与三步走演进路径，指出未来的核心竞争力在于通过高管激励对齐与人机协同，推动员工从基础工单处理者转型为智能体编排与监督者。 ## Atmosphere & style 整场圆桌讨论务实、锐利且极具实战沉淀，对话毫无空泛的口号，而是直击组织内部博弈、流程混乱与系统孤岛等现实痛点。四位嘉宾与主持人互动默契、风趣幽默，既有对“深埋组织六层之下”隐性单点依赖的生动比喻，也有对高管考核与流程重构的深刻反思，为观众带来兼具危机感与高可操作性的落地方法论启发。 ## Key takeaways - [03:09](https://www.youtube.com/watch?v=XOThNoSud90&t=189s) 通用上下文仅能解决约 30% 的基础业务场景，智能体必须深度融入垂直行业与各部门的专属上下文才能创造质的突破与真实 ROI。 - [08:37](https://www.youtube.com/watch?v=XOThNoSud90&t=517s) 大型企业在 AI 战略上逐步从全栈自研专有基础大模型，转向聚焦业务价值与 ROI，通过保留内部领域专家并引入成熟编排与外部技术实现平衡。 - [10:38](https://www.youtube.com/watch?v=XOThNoSud90&t=638s) 降低智能体落地阻力的关键在于“无缝融入既有工作场景”，如将决策确认和人机交互打通至 Slack 与 Jira，避免引入新的工具孤岛。 - [13:26](https://www.youtube.com/watch?v=XOThNoSud90&t=806s) 引入 AI 适用度评估框架（AI Eligibility Framework），通过计算自动化研发成本、后续人工审查维护开销及频次，科学筛选优先改造的流程。 - [15:20](https://www.youtube.com/watch?v=XOThNoSud90&t=920s) 智能体系统底层需具备编排层、记忆与本体层、配置层与应用层四层架构，并依赖涵盖行业、业务单元、地域、岗位与领导力结构的“五维超细胞”模型落地。 - [17:28](https://www.youtube.com/watch?v=XOThNoSud90&t=1048s) 很多企业未能获得 AI 预期的 ROI，核心根源不在于技术本身，而在于管理层未敢于拆解臃肿的层级架构及协调 CFO、CIO、CHRO 之间的目标冲突。 - [19:25](https://www.youtube.com/watch?v=XOThNoSud90&t=1165s) 解决跨职能流程治理壁垒的有效路径是设立“单一负责人”机制，由其全权统管流程决策并配备独立专属预算，只对公司整体 ROI 负责。 - [21:34](https://www.youtube.com/watch?v=XOThNoSud90&t=1294s) AI 不会直接剥夺工作，但掌握 AI 的人会替代不善用 AI 的人；员工核心阻力在于“时间贫困”，企业需通过早期内部标杆演示看到成效。 - [27:00](https://www.youtube.com/watch?v=XOThNoSud90&t=1620s) 变革推动依赖顶层设计，JLL 通过将高管的薪酬与奖金直接挂钩 AI 熟练度、自动化比例及智能体采纳度，形成硬性考核激励。 - [27:00](https://www.youtube.com/watch?v=XOThNoSud90&t=1620s) 通过重构业务价值主张实施“80/20 转型”，剥离 80% 机械工单后，员工被重塑为知识库专家与智能体编排架构师等高价值角色。 - [30:47](https://www.youtube.com/watch?v=XOThNoSud90&t=1847s) 蓝领与白领业务的 AI 诉求截然不同，管理层需根据业务扩张或收缩动态重配人力，如将制造领域的招聘专员敏捷转岗为入职培训管理。 - [34:18](https://www.youtube.com/watch?v=XOThNoSud90&t=2058s) 如果业务流程本身混乱无效，直接引入 AI 只会将缺陷成倍放大；企业必须遵循“先修复流程，再引入 AI”的基本原则。 - [39:46](https://www.youtube.com/watch?v=XOThNoSud90&t=2386s) 利用 AI 自动发现并记录工作流面临深层隐性盲区挑战，如深埋基层未被文档化的人工单点依赖，AI 尚难以敏锐捕捉与解析。 - [40:49](https://www.youtube.com/watch?v=XOThNoSud90&t=2449s) 工作流智能体化可通过高性价比的三步演进法推进：先通过精益六西格玛离岸梳理标准流程，再利用低代码平台自动化，最后进行智能体化改造。 - [42:05](https://www.youtube.com/watch?v=XOThNoSud90&t=2525s) 垂直 SaaS 厂商利用沉淀于 700 多家客户的跨组织经验构建底层行业本体，能先于客户诊断出流程断点，由外部逆向驱动企业自动化升级。 ## Detailed walkthrough 以下是根据视频转录稿 Part 1 整理的详细中文逐段梳理与深入解读： ### [00:04](https://www.youtube.com/watch?v=XOThNoSud90&t=4s) 开场致辞与与会嘉宾介绍 研讨会主持人以极高的热情拉开了本场论坛的序幕，对现场几乎座无虚席的盛况表达了由衷的感谢。在经历了一场精彩绝伦的主旨演讲后，主持人强调本场关于“智能体 AI（Agentic AI）”的专题讨论将为与会者带来极具价值的实战洞察。为了确保互动环节的高效与有序，主持人特别提醒现场听众提前构思提问，并说明在个人总结发言后将进入 Q&A 环节，届时会有工作人员递送麦克风。随后，四位来自不同领域的重磅嘉宾依次进行了自我介绍： * **Mahe Bayireddi**：Phenom 的 CEO 兼联合创始人。Phenom 是一家专注于人力资源（HR）领域的应用 AI 公司，致力于在智能体时代重新设计工作流程与人力运营，涵盖招聘优化、员工留存、人才流动性与职业发展等核心场景。 * **Max Ushchenko**：NIX 的 AI 与数据总监。NIX 拥有 30 年的市场经验，专门为客户提供软件解决方案、产品开发以及所需的技术专家资源。 * **Ralph Valdes**：Zumba Fitness 的首席技术官（CTO）。负责管理产品工程、质量保证（QA）、数据部门，目前全面主导 AI 战略，推动智能体在全公司范围内的落地与普及。 * **Carlin Power**：JLL（琼斯郎和 LaSalle，全球商业地产巨头）全球 AI 产品与企业赋能负责人。JLL 在全球 80 个国家拥有约 12 万名员工。Carlin 还在开场打趣道，作为明尼苏达州人，她非常感谢能来到拉斯维加斯威尼斯人酒店享受冷气。 ```mermaid flowchart LR accTitle: Panelists and Organization Structure subgraph Panelists["论坛嘉宾与代表企业"] A["Mahe Bayireddi<br/>(CEO @ Phenom)"] B["Max Ushchenko<br/>(AI & Data Director @ NIX)"] C["Ralph Valdes<br/>(CTO @ Zumba Fitness)"] D["Carlin Power<br/>(Global Head of AI @ JLL)"] end subgraph FocusAreas["核心专注领域"] A --> A1["HR 智能体 / 人力运营重塑"] B --> B1["软件工程 / SDLC 流程重构"] C --> C1["企业级智能体普及 / Kodu 工厂"] D --> D1["Ask JLL / 跨部门知识库智能体"] end end ``` ### [03:09](https://www.youtube.com/watch?v=XOThNoSud90&t=189s) 各大企业的智能体 AI 落地实践案例 论坛进入第一个核心议题，嘉宾们分享了各自企业或客户目前已经完成或正在推进的智能体 AI 项目： JLL 的 Carlin Power 首先分享了他们在一个月前刚刚发布的重磅产品——**Ask JLL**。这是一个汇集了企业内部各职能部门领域智能体（Domain Agents）的综合平台，覆盖人力资源、财务、技术支持、采购、研究、法务及合规等多个部门。这些智能体能够上下文调取并检索各自部门的知识库，不仅能回答员工的具体咨询，还能代表员工直接执行任务（例如提交休假申请、订购电脑及外设硬件、发起合规举报、申请法律合同等）。目前该系统主要面向内部员工，未来计划推出面向客户的智能体。 Zumba Fitness 的 Ralph Valdes 补充了两个方向的实践：一是利用 **Snowflake Agents** 对全公司的核心 KPI 及数据分析进行集中化管理，在保证数据可验证与准确的前提下实现数据民主化；二是推出了名为 **Kodu** 的“智能体软件工厂（Agentic Software Factory）”。Kodu 能够接收 Web 或移动端应用的需求，自主制定开发计划、编写代码、跨部门询问澄清问题，并在安全的环境中生成预览版本供人工测试与审批，最后完成发布。 Phenom 的 Mahe Bayireddi 则强调了他们在 HR 生态系统中的垂直深化。Phenom 的系统覆盖超过 700 家企业客户，每 18 个月触达逾 10 亿用户。他们构建了筛选智能体以及后端的劳动力规划智能体，帮助企业在业务收缩或扩张时，精准调整人才配置与资源规划。Mahe 特别强调，generic（通用）上下文只能解决 30% 的基础场景，唯有深度融入行业及部门的专属上下文，智能体才能带来质的飞跃。 NIX 的 Max Ushchenko 分享了一个医疗健康领域的软件工程案例。他们协助客户的软件工程师、业务分析师（BA）、项目经理（PM）和测试人员（QA）重构了软件开发生命周期（SDLC）。通过评估各个环节的“AI 适用度（AI Eligibility）”，在保留人工复核的同时，实现了自动化代码审查和安全风险评估。 - **Ask JLL 服务规模**：覆盖 JLL 全球 80 个国家约 120,000 名员工。 - **Phenom 数据体量**：服务 700+ 客户，18 个月内触达 10 亿+ 用户。 - **通用 vs 垂直场景效率**：通用上下文仅覆盖约 30% 业务需求，深度垂直上下文才是 ROI 的关键。 ### [08:37](https://www.youtube.com/watch?v=XOThNoSud90&t=517s) 智能体项目的战略选择：ROI 导向与自主研发平衡 在总结首个议题时，Carlin Power 为现场听众提供了极具操作性的落地建议。她指出，企业在决定自研智能体架构还是直接购买集成了编排层（Orchestration Layer）的商业成熟技术时，关键在于将业务与职能部门的专家与最适合自身预算、业务形态的技术栈进行匹配。 Carlin 透露，JLL 在两年前开启 AI 探索之旅时，为了避免被供应商绑定（Vendor Lock-in）并积累内部核心能力，率先自研了内部专有的 GPT 模型。但随着基础设施巨头（Labs）投入数十亿美元进行底层大模型迭代，JLL 果断调整了战略：保留内部的业务与领域专家，引入成熟的外部技术，并在特定高价值场景下补充低成本的自研专有技术。 更重要的是，JLL 在项目初期就做出了一个关键的方向性转变——**从单纯追求“效率提升与生产力释放”转向聚焦“业务价值与 ROI（投资回报率）”**。对于 JLL 这样庞大体量的企业而言，仅靠节省工时不足以支撑大规模投资，智能体项目必须能够带来指数级的业务回报。 ### [10:38](https://www.youtube.com/watch?v=XOThNoSud90&t=638s) 重构工作流与应对“人机交接”阻力 当谈到如何不只是将 AI“套用”在旧流程上，而是彻底拆解并重构以智能体为主导的工作流时，Ralph Valdes 拿 Zumba 的 **Kodu 智能体软件工厂** 举例：让智能体生成代码早已不是技术难题，真正的挑战在于——“智能体应该生成什么？谁来决定？遇到矛盾或歧义时怎么处理？由谁来进行设计审查与产品决策？” 为了克服员工对“权力被剥夺”的抵触情绪并推动项目落地，Zumba 采取了**“无缝融入既有工作场景（Meet people where they are）”**的策略： 1. **拒绝新开孤岛工具**：不强制员工登录特定的 IP 地址或新系统，而是将所有决策节点、交互和确认全部打通至员工日常使用的 **Slack** 与 **Jira** 中。 2. **智能上下文路由**：根据智能体涉及的代码模块和业务线，自动将问题与确认请求精准路由给对应的团队。 3. **降低交互门槛**：当智能体提出的问题不够清晰时，员工只需在 Slack 中回复“请解释得更清楚一点”，智能体即可重新组织语言，大大降低了用户的使用摩擦（User Friction）。 ```mermaid flowchart TD accTitle: Kodu Agentic Workflow subgraph Input["需求输入与规划"] A["需求提交 (Slack / Jira)"] --> B["Kodu 智能体拆解规划"] end subgraph Execution["开发与跨部门协同"] B --> C{"发现歧义 / 疑问?"} C -- "是" --> D["精准路由至对应业务团队 (Slack)"] D --> E["团队反馈 / 澄清"] E --> B C -- "否" --> F["自主编写代码与构建"] end subgraph Deployment["安全测试与发布"] F --> G["生成沙盒预览环境"] G --> H["人工 QA / 产品审查与批准"] H --> I["自动部署上线"] end end ``` ### [13:26](https://www.youtube.com/watch?v=XOThNoSud90&t=806s) AI 适用度评估模型与人机协同（Human-in-the-Loop） NIX 的 Max Ushchenko 深入探讨了如何科学判断一个工作流是否适合交由智能体接管。他们构建了一套被称为 **“AI 适用度框架（AI Eligibility Framework）”** 的评估体系，重点关注以下维度： 1. **人机协同模式**：判断该流程需要人类全程参与（Human-in-the-loop）、部分参与，还是仅在流程末端对最终结果进行验证。 2. **开发与运维成本占比**：评估构建该自动化流程所需的研发精力，以及后续人工检查和维护（Overhead）所消耗的精力。如果检查成本过高或运行费用昂贵，该流程就不适合自动化。 3. **可量化的影响**：精确计算某项工作流如果引入 AI，能节省百分之多少的时间，以及该流程发生的频次（按小时、天还是月计算），进而建立数学模型决定优先启动哪个项目。 主持人赞许道，与会嘉宾在剥离部分传统人力权限的同时，无一例外地强调了“以人为本（Human-Centered）”和保留人机协同，这极大地缓解了员工对 AI 的安全感危机与抵触情绪。 ### [15:20](https://www.youtube.com/watch?v=XOThNoSud90&t=920s) 智能体架构的四层模型与五维“超细胞”架构 Phenom 的 Mahe Bayireddi 提出了一个关于架构与变革管理的深刻洞见。他对比了传统系统与智能体系统的底层架构差异： * **传统系统**：仅包含“应用程序层（Application）”和“配置层（Configuration Layer）”。 * **智能体系统**：必须具备四层架构——**应用程序层**、**配置层**、**记忆与本体层（Memory & Ontological Layer）**以及**编排层（Orchestration Layer）**。 ```mermaid flowchart TD accTitle: Legacy vs Agentic Architecture subgraph Legacy["传统系统架构"] L1["配置层 (Configuration)"] --> L2["应用程序层 (Application)"] end subgraph Agentic["智能体时代系统架构"] A1["编排层 (Orchestration Layer)"] A2["记忆与本体层 (Memory & Ontological Layer)"] A3["配置层 (Configuration Layer)"] A4["应用程序层 (Application Layer)"] A1 --> A2 A2 --> A3 A3 --> A4 end end ``` Mahe 指出，ROI 无法凭空产生，它必须直接转化为业务影响，而业务影响取决于业务单元（BU）与职能部门之间的协同，这又与企业文化息息相关。Phenom 将此总结为 **“五维超细胞（5D Hypercell）”模型**： 1. **行业（Industry）**：企业所处的特定行业特性。 2. **业务单元（Business Unit）**：项目试图产生直接影响的具体业务部门及其 ROI。 3. **地理位置（Location）**：不同地区的招聘、留存与管理模式差异巨大（如北美与欧洲、亚洲的运营逻辑截然不同）。 4. **岗位/角色（Job/Role）**：重构所针对的具体职位。 5. **智能体架构与领导力结构（AG/Leadership Structure）**：**最关键的一环**。领导者的认知与组织架构直接决定了变革管理（Change Management）的成败。 ### [17:28](https://www.youtube.com/watch?v=XOThNoSud90&t=1048s) 组织壁垒、高管冲突与 ROI 的真正阻碍 在对话的最后，主持人提出了一个行业普遍存在的痛点：业务部门与职能部门之间往往存在天然的“高墙”，相互不理解对方的工作，AI 是否帮助打破了这种壁垒？ Mahe Bayireddi 的回答直击要害。他认为，这完全取决于企业的心理模型、行为习惯、战略以及领导风格（如 Jensen Huang、Jeff Bezos 或 Satya Nadella 截然不同的领导性格如何塑造企业的思维模型）。 Mahe 坦言，当前许多企业中普遍存在 **CFO（首席财务官）、CIO（首席信息官）与 CHRO（首席人力资源官）之间的剧烈冲突**——他们对 CEO 期望达到的目标无法达成一致。这种冲突本质上源于企业内部臃肿、扭曲的传统层级结构（Hierarchical Structure）。许多公司抱怨“智能体效果不好、无法带来预期 ROI”，但真实原因并不是智能体技术本身不行，而是**企业没有勇气去拆解和重组内部病态的组织架构与领导力结构**。如果不从根本上对齐组织结构与变革管理，再先进的智能体也无法释放出真正的商业价值。 <ElicitationsGroup message="如果您想深入探讨本场论坛提及的更多细节，可以选择以下方向："> <Elicitation label="详细拆解 Ask JLL 的知识库与领域智能体架构" query="请详细拆解 Ask JLL 的领域智能体（Domain Agents）架构，说明它们是如何与 HR、法务、财务等部门知识库对接并执行任务的。"/> <Elicitation label="深入分析 Zumba 的 Kodu 智能体软件工厂运行机制" query="请深入分析 Zumba 的 Kodu 智能体软件工厂，解释它是如何在 Slack 和 Jira 中实现需求拆解、跨部门询问和自动化开发的。"/> <Elicitation label="对比 Phenom 智能体四层架构与传统系统架构的异同" query="请对比 Phenom 提出的智能体四层架构（编排层、本体层、配置层、应用层）与传统系统架构的异同，并解释其在实际业务中的作用。"/> </ElicitationsGroup> ### [19:25](https://www.youtube.com/watch?v=XOThNoSud90&t=1165s) 跨职能流程治理与“单一负责人”机制 在探讨如何落地 Agentic AI 之前，JLL 的 Carlin Power 提出了一个至关重要的前置准备步骤：对跨业务线和跨职能部门的复杂流程进行治理重构。她以企业中普遍存在的“请求到解决”（Request to Resolution，即员工发起需求并需要人工介入解决的流程）为例。在三个月前，这个流程极为臃肿混乱，分散在 12 个系统、74 个团队、17 个预算中心以及约 240 位高管级利益相关者手中。 ```mermaid flowchart TD accTitle: 请求到解决流程重构架构图 subgraph Past["改造前：分散混乱模式"] P1["12 个独立系统"] P2["74 个职能团队"] P3["17 个分散预算"] P4["240+ 位管理层 Stakeholders"] end subgraph Governance["治理重组：单一归口机制"] Lead["单一责任人 (Accountable Leader)"] Bud["独立专项预算"] Align["全公司 ROI 与商业价值目标"] end subgraph Present["改造后：敏捷运作模式"] Lead --> Bud Bud --> Align Align --> Out["打破部门壁垒（SOP/KPI 统一），高效驱动 Agentic 落地"] end ``` 为了彻底打破这种壁垒，JLL 采取了一项关键决策：将复杂的跨职能流程指定给单一负责人（Single Accountable Owner）。该负责人对整个流程的决策和组织重塑全权负责，从而避免了首席人力资源官（CHRO）、首席财务官（CFO）和首席信息官（CIO）之间的利益冲突。公司不仅为该负责人设立了独立的专属预算，还确保其考核只对全公司的整体效益与投资回报率（ROI）负责，彻底摆脱了过去各部门因各自 SOP、KPI 及预算限制而产生的内耗。 --- ### [21:34](https://www.youtube.com/watch?v=XOThNoSud90&t=1294s) 岗位角色演变与化解员工学习焦虑 主持人 Jill Hosmer-Jolley 随后将焦点转向组织变革中的员工转型：当员工从“执行工作的生产主体”转变为“监督与审查 AI 产出的人机协同者（Human-in-the-loop）”时，企业该如何推进。Zumba 的 Ralph Valdes 明确表示，自己并不认同“AI 会夺走所有人的工作”，AI 更是一种增强而非替代力量。但他强调，每个人都应该保持一种“健康的危机感”——“AI 不会抢走你的工作，但善用 AI 的人会夺走你的工作”。企业应鼓励员工为了自身的职业未来积极尝试与学习，因为无论身处哪家公司，这些技能都将决定未来的竞争力。 在推进培训与转型时，Ralph 指出最大的阻力在于员工的“时间贫困”：“我本身就已经全职满负荷工作了，现在你还要我花额外的时间忍受效率降低，去学习一个两个月后才可能有用的东西。”面对这种合理的抱怨，Ralph 建议企业内部的早期实践者积极提供协助，将自己的工作流和 Agent 演示给同事看，用直观的成效帮助大家看到“隧道尽头的曙光”，从而愿意承受初期的阵痛。 NIX 的 Max Ushchenko 则从软件工程视角补充道，AI 正在深刻改变工程师的工作性质。过去企业追求“摇滚明星般的资深工程师”，而现在借助 AI 工具集，普通工程师甚至需要成为能够同时胜任 5 项职能的“一人交响乐团”。因此，招聘门槛已全面转向要求具备 AI 实战经验，团队文化也转变为“接到任务的第一时间，就要思考该任务是该由 Agent 完成、AI 辅助完成还是由自己完成”。 --- ### [27:00](https://www.youtube.com/watch?v=XOThNoSud90&t=1620s) 高管激励绑定与重塑业务价值主张 针对员工转型过程中的压力管理，Carlin Power 指出，推动变革的核心往往不是 CHRO，而是来自 CEO 的顶层设计。在 JLL，CEO 将高管的薪酬激励机制与员工赋能和企业 AI 转型直接挂钩。Carlin 的团队曾历时数月与各部门领导人坐在一起，重新梳理并书写各自部门的核心价值主张（Value Proposition）。令人惊讶的是，很多业务部门甚至无法清晰表述自己当前的价值主张。 ```mermaid flowchart LR accTitle: 业务价值主张重构流程 A["梳理现状价值主张"] --> B["识别 AI 替代/增强环节"] B --> C["价值捕获分析<br>(吞吐量/新业务线/效率)"] C --> D["重写价值主张与员工角色"] D --> E["从底层支持转为<br>最大化投资回报的高价值团队"] ``` 团队通过系统性梳理，明确哪些环节会被 AI 替代、增强或重构，并将收益重新映射为业务吞吐量的增加、新营收线的拓展或服务质量的提升。通过奖励变革、而不是因效率提升去惩罚裁撤岗位，公司有效化解了员工的抵触与恐惧。Carlin 分享了一个具体案例：两年前她召集技术支持团队，直言“你们 80% 的日常工作未来将不再存在，我们可以坐以待毙，也可以重新定义自己的价值”。如今，这批员工已从被动的“工单处理者”蜕变为了“知识库专家”与“Agent 编排架构师（Agent-to-Agent Orchestrators）”，直接帮助企业最大化技术投资价值。 - **80/20 转型法则**：剥离 80% 的机械重复工作后，员工得以专注于 20% 最具挑战、最能提供情感价值与深度支持的核心难题，并将该价值成倍放大。 --- ### [30:47](https://www.youtube.com/watch?v=XOThNoSud90&t=1847s) 蓝领与白领差异化落地：从医疗到制造的用例 Phenom 的联合创始人兼 CEO Mahe Bayireddi 赞同 CEO 是核心驱动力的观点，因为资本市场正在倒逼企业建立全新的高性价比运营模式。他提出，企业的工作可以划分为 1 到 5 个层级（Zone 1 至 Zone 5）： - **Zone 1 与 Zone 2**：偏向一线员工（Frontline Workers，如蓝领操作工、一线护理人员）。 - **Zone 4 与 Zone 5**：偏向知识工作者（Knowledge Workers）。 ```mermaid flowchart TD accTitle: 医疗与制造场景中的 AI 流程差异 subgraph Healthcare["医疗场景 (以 Baylor Scott & White 为例)"] H1["痛点：护士退休多、毕业少、招聘极难"] H2["Agent 落地：自动化资质筛选全流程"] H3["成效：招聘官省 20-25% 时间，管理者省 3 小时/周"] end subgraph Manufacturing["制造场景 (以 Thermo Fisher 为例)"] M1["特点：制造员工招聘快、批量大"] M2["痛点：新人到岗后的培训与管理承载力不足"] M3["角色转移：招聘人员实时转向负责入职培训 (Onboarding)"] end ``` AI 在不同行业和工种中的自动化与 Agent 架构应用差异巨大，CHRO 需要动态识别哪些业务在缩减、哪些在扩张，并在内部灵活调配人才。例如在医疗健康领域，Baylor Scott & White Health 面临护士老龄化与极端用工荒，引入 Agent 自动化筛选基础设施后，每周为招聘专员节省 20%~25% 的时间，为业务经理每周节省至少 3 小时；而在 Thermo Fisher 赛默飞世尔的制药制造场景中，招聘速度虽快，但管理承载力有限，企业便将招聘专员的角色动态切换为入职培训管理人员（Onboarding），实现人力资源的敏捷重组。 --- ### [34:18](https://www.youtube.com/watch?v=XOThNoSud90&t=2058s) 嘉宾金句总结与核心建议 在进入问答环节前，主持人邀请四位嘉宾各自分享了一句最希望观众带回企业的核心总结： | 嘉宾 | 核心建议 | 详细内涵 | |:--- |:--- |:--- | | **Mahe Bayireddi** | 聚焦具体用例与业务结果 | 技术架构每年都在剧烈迭代（从 RAG 到图模型、Loop 工程再到 Agent Harness），只有锚定在稳定的业务产出上，才能在技术变迁中保持定力。 | | **Max Ushchenko** | 先修复流程，再引入 AI | 如果业务流程本身是混乱无效的，AI 不会解决问题，只会将错误和缺陷成倍放大。 | | **Ralph Valdes** | 拥抱变革，主动成为 Agent 的拥有者 | 未来 6 个月到 3 年内大量工作将被自动化，但也创造了巨大的新管理界面，员工应积极寻找并主动认领这些新机会。 | | **Carlin Power** | 领导者以身作则，发挥一线创造力 | 最有效的变革管理是领导者亲自示范行为模式，同时必须相信一线员工才能构想出最具颠覆性的市场竞争方式。 | --- ### [37:53](https://www.youtube.com/watch?v=XOThNoSud90&t=2273s) 观众互动：CEO 薪酬考核指标设计与流程挖掘 现场进入 Q&A 环节，观众 Arlon 提问 JLL 的 CEO 是如何通过薪酬机制推动全员采纳 AI 的。Carlin Power 透露，公司全球执行委员会（汇报给 CEO 及董事会的高管）有明确的业务 KPI 和薪酬 KPI，且薪酬考核具有一票否决权——未达成 KPI 将无法获得任何奖金： - **AI 熟练度（AI Fluency）**：硬性考察各业务线领导者的 AI 应用水平。 - **自动化比例目标**：针对各个具体业务板块和职能部门，制定明确的流程自动化百分比指标。 - **Agent 采纳度**：在统一平台内推动 AI 与 AI Agent 的实际使用深度。近两年来，每位高管的奖金考核均深度绑定了这些变革指标。 随后，Butterworth Advisors 负责人、前 CBRE 首席技术官 Nicholas Butterworth 提出了关于业务流程挖掘（Process Discovery）的挑战，指出在重构像 17 个系统并存的复杂工作流之前，如何有效厘清与发现这些底层流程往往是企业面临的最艰难一步。 ### [39:46](https://www.youtube.com/watch?v=XOThNoSud90&t=2386s) 利用 AI 发现与记录工作流的挑战与隐性盲区 小组讨论进入尾声，主持人抛出了关于变革管理漏斗顶端（Top of funnel）的关键问题：企业是否已经能够利用 AI 来主动发现并记录（discover and document）既有的工作流。Ralph Valdes 指出，这在当前仍然是一个极其棘手的挑战，团队目前还没有在这方面大量应用 AI。对于那些结构清晰的领域——比如软件工程和数据流——由于其工作流本身已经被高度文档化、透明化，因此识别和完成模式转换相对容易。 真正的痛点隐藏在那些深层且未被记录的人工节点中。Ralph Valdes 形象地描述了一种企业常见现象：在组织架构“深埋六层之下”（six levels deep）的某个基层员工，掌握着某项极其微小但关键的操作；如果这位员工每周没有在特定时间提交某项数据，整条下游业务链（downchain stream）就会彻底崩溃。现阶段的 AI 尚无法敏锐地挖掘出这种深层隐性依赖，而这正是企业未来亟需攻克的方向。 ```mermaid flowchart TD accTitle: 识别显性工作流与隐性盲区的对比 subgraph Explicit["显性/易捕捉领域"] A["软件工程与数据流"] --> B["已有清晰文档与规范"] B --> C["易于 AI 识别与流程转型"] end subgraph Implicit["隐性/深层盲区"] D["深处组织 6 层的关键岗位"] --> E["每周单点人工提交"] E --> F["AI 尚未能自动发现依赖"] F --> G["一旦遗漏引发下游链路崩溃"] end ``` --- ### [40:49](https://www.youtube.com/watch?v=XOThNoSud90&t=2449s) JLL 的三步走优化路径：离岸外包、低代码自动化到 Agent 化 针对工作流发现与梳理的问题，Carlin Power 分享了仲量联行（JLL）采取的“人机结合”策略。JLL 在印度班加罗尔建立了一个名为 JLL Business Services 的离岸中心，部署了大量精益六西格玛（Lean Six Sigma）流程工程师，结合低代码平台 Mendix 以及专用的六西格玛流程建模工具（如 Blueprint），自底向上地对组织内部所有 Level 0、Level 1 和 Level 2 流程进行全量绘制与梳理。 在具体落地上，JLL 制定了一套清晰且高性价比的三步演进路径： 1. **离岸迁移（Offshore Transition）**：先将梳理明确的标准流程移交至班加罗尔的离岸服务中心。 2. **流程自动化（Automate via Mendix）**：借助 Mendix 低代码工具将规则化步骤自动化。 3. **Agent 化改造（Agentify）**：在自动化流程之上注入智能体能力，实现端到端的 Agentic 运作。 Carlin Power 坦言，市面上或许有采用全职等效员工（FTE）或昂贵咨询方案的做法，但企业必须对股东资金负责，这种结合精益流程工程与低代码工具的阶梯式推进，是目前性价比最高的实现路径。 ```mermaid flowchart LR accTitle: JLL 流程重构三步演进模型 A["班加罗尔精益六西格玛团队<br/>Mapping L0/L1/L2 流程"] --> B["第一步：离岸迁移<br/>(Transition Offshore)"] B --> C["第二步：低代码自动化<br/>(Mendix Automation)"] C --> D["第三步：智能体化改造<br/>(Agentify)"] ``` --- ### [42:05](https://www.youtube.com/watch?v=XOThNoSud90&t=2525s) Phenom 基于跨客户本体与 AI 的逆向诊断与对客赋能 Phenom 首席执行官 Mahe Bayireddi 提出了截然不同的企业级软件服务视角。他主张企业软件厂商必须在客户自己分析之前，先一步对客户的业务进行深度解剖，因为“如果你不能比客户更懂他们所在的垂直赛道，企业软件就失去了存在的根本意义”。 Mahe Bayireddi 阐释了其背后的核心机制：单一企业内部的视野往往受限于自身上下文，而 Phenom 能够聚合跨 700 多家企业客户沉淀的数据与业务上下文，并将这些洞察凝练注入到产品底层的数据本体（Ontologies）中。当他们走向客户时，系统早已深度理解该客户的业务断点，并能通过广泛部署的 AI 精准指出哪些环节存在优化和自动化的迫切需求。这种诊断不是为了技术炫技，而是帮助企业直面当下的业务压力；借助 AI 进行这种大规模上下文提炼与诊断，成为了推动业务指数级扩张的核心驱动力。 - **跨客户上下文融合**：打破单家企业的数据孤岛，利用 700+ 客户的跨组织经验构建行业核心本体。 - **主动式业务诊断**：在对客交付前，利用 AI 预先测算潜在效率漏洞，由外部驱动企业内部变革。 --- ### [43:13](https://www.youtube.com/watch?v=XOThNoSud90&t=2593s) 圆桌尾声与线下交流 主持人 Jill Hosmer-Jolley 对四位嘉宾充满见地、幽默且富有实操价值的分享表示高度赞赏。由于论坛时间限制，现场未能开放更多观众提问，主持人对在场观众的耐心表示感谢，并幽默地提醒大家：“我们优秀的嘉宾们接下来会走到走廊，大家可以在过道里随时‘围堵’他们继续交流。” 圆桌讨论在热烈的掌声与致谢中正式落下帷幕。 ## Notable quotes > [03:09](https://www.youtube.com/watch?v=XOThNoSud90&t=189s) generic（通用）上下文只能解决 30% 的基础场景，唯有深度融入行业及部门的专属上下文，智能体才能带来质的飞跃。 — Mahe Bayireddi > [10:38](https://www.youtube.com/watch?v=XOThNoSud90&t=638s) 智能体应该生成什么？谁来决定？遇到矛盾或歧义时怎么处理？由谁来进行设计审查与产品决策？ — Ralph Valdes > [17:28](https://www.youtube.com/watch?v=XOThNoSud90&t=1048s) 企业没有勇气去拆解和重组内部病态的组织架构与领导力结构。 — Mahe Bayireddi > [21:34](https://www.youtube.com/watch?v=XOThNoSud90&t=1294s) AI 不会抢走你的工作，但善用 AI 的人会夺走你的工作。 — Ralph Valdes > [21:34](https://www.youtube.com/watch?v=XOThNoSud90&t=1294s) 我本身就已经全职满负荷工作了，现在你还要我花额外的时间忍受效率降低，去学习一个两个月后才可能有用的东西。 — Ralph Valdes > [21:34](https://www.youtube.com/watch?v=XOThNoSud90&t=1294s) 接到任务的第一时间，就要思考该任务是该由 Agent 完成、AI 辅助完成还是由自己完成。 — Max Ushchenko > [27:00](https://www.youtube.com/watch?v=XOThNoSud90&t=1620s) 你们 80% 的日常工作未来将不再存在，我们可以坐以待毙，也可以重新定义自己的价值。 — Carlin Power > [34:18](https://www.youtube.com/watch?v=XOThNoSud90&t=2058s) 先修复流程，再引入 AI。 — Max Ushchenko > [34:18](https://www.youtube.com/watch?v=XOThNoSud90&t=2058s) 最有效的变革管理是领导者亲自示范行为模式，同时必须相信一线员工才能构想出最具颠覆性的市场竞争方式。 — Carlin Power > [42:05](https://www.youtube.com/watch?v=XOThNoSud90&t=2525s) 如果你不能比客户更懂他们所在的垂直赛道，企业软件就失去了存在的根本意义。 — Mahe Bayireddi ## Mentioned Ralph Valdes, Carlin Power, Max Ushchenko, Mahe Bayireddi, Jill Hosmer-Jolley, Nicholas Butterworth, Zumba Fitness, JLL, NIX, Phenom, California State University Monterey Bay, Butterworth Advisors, Ask JLL, Kodu, Snowflake Agents, Slack, Jira, Mendix, Blueprint, Baylor Scott & White Health, Thermo Fisher Scientific, Jensen Huang, Jeff Bezos, Satya Nadella ## Video description Ralph Valdes — Chief Technology Officer, Zumba Carlin Power — Global Head of AI and Enterprise Product Engagement, JLL Max Ushchenko — Director of AI and Data, NIX Mahe Bayireddi — CEO & Co Founder, Phenom Jill Hosmer-Jolley — Faculty, California State University Monterey Bay Agentic AI reshapes how work gets done. This panel explores how enterprises are re-architecting workflows to integrate autonomous agents, from front-office operations to back-office processes. Panelists will highlight how to identify the right use cases, redesign cross-functional processes, and redefine employee roles to ensure smooth adoption. The discussion will also cover change management strategies, including leadership alignment, workforce training, and cultural adjustments, that enable agentic AI to be embraced across the enterprise. [Source](https://www.youtube.com/watch?v=XOThNoSud90)

[3:09] Mahe Bayireddi: generic（通用）上下文只能解决 30% 的基础场景，唯有深度融入行业及部门的专属上下文，智能体才能带来质的飞跃。 [10:38] Ralph Valdes: 智能体应该生成什么？谁来决定？遇到矛盾或歧义时怎么处理？由谁来进行设计审查与产品决策？ [17:28] Mahe Bayireddi: 企业没有勇气去拆解和重组内部病态的组织架构与领导力结构。 [21:34] Ralph Valdes: AI 不会抢走你的工作，但善用 AI 的人会夺走你的工作。 [21:34] Ralph Valdes: 我本身就已经全职满负荷工作了，现在你还要我花额外的时间忍受效率降低，去学习一个两个月后才可能有用的东西。 [21:34] Max Ushchenko: 接到任务的第一时间，就要思考该任务是该由 Agent 完成、AI 辅助完成还是由自己完成。 [27:00] Carlin Power: 你们 80% 的日常工作未来将不再存在，我们可以坐以待毙，也可以重新定义自己的价值。 [34:18] Max Ushchenko: 先修复流程，再引入 AI。 [34:18] Carlin Power: 最有效的变革管理是领导者亲自示范行为模式，同时必须相信一线员工才能构想出最具颠覆性的市场竞争方式。 [42:05] Mahe Bayireddi: 如果你不能比客户更懂他们所在的垂直赛道，企业软件就失去了存在的根本意义。

[0:04] Jill Hosmer-Jolley: Our plan is to have each one of them answer some questions and then we will roll into a Q&A... get your questions ready. [8:37] Carlin Power: We were focused a lot more on the value and the ROI of the agents that we were delivering and not so much on the productivity and efficiency that I think where everybody started. [10:33] Ralph Valdez: Getting an agent to write software is easy... What should it generate? Who decides? How do you handle discrepancies? Who approves something? These are the things that really make it more of a challenge. [15:20] Mahe Bayireddi: Agents have no value in it because you have to make decisions in such a way so that you can make the moment so that the impact do exist... until unless you can really rip a part of where the structure aligns you cannot really get the best ROI at all. [21:34] Ralph Valdes: AI is not going to take your job. Someone who learns AI is going to take your job. [21:34] Ralph Valdes: I already work full-time all the time, and now you're asking me to work extra time to be slower so that I can learn something that might help me two months from now. [26:59] Carlin Power: 80% of your job is not going to be a job anymore. So, we have two choices. We can brace for impact or we can decide what our new value proposition is going to be. [34:18] Max Ushchenko: AI will only magnify the problem but will not fix it.

[0:04] Hi everyone. I am so excited to see a nice full  almost room. I appreciate you all taking the time. [0:13] What an exciting keynote we've had and um and  making this session a priority for you really [0:19] is wonderful. We have seriously thank you. We have  seriously amazing panelists here for you. Our plan [0:29] is to have each one of them answer some questions  and then we will roll into a Q&A just to let you [0:37] know to get your questions ready when I ask them  for their final thoughts before Q&A that's when [0:43] you can start thinking what you're going to ask.  And the wonderful moderator who was just up here, [0:47] she is going to bring you a microphone. So just so  everybody can hear your question, hold off until [0:54] she gives you a microphone. Uh, all set everybody  good? Great. I'm gonna go ahead and start with [1:02] Mahi and uh have you all just take just a couple  of minutes to just tell them what they who you are [1:08] and what they can expect from you. I'm Mahi Bay,  CEO and co-founder of Phenom. Uh, we don't Phenom [1:14] is an applied AI company focused on HR domain.  Our primary focus is how can we redesign work [1:21] and people operations in the agentic era. uh  we actually do things which are very unique in [1:27] terms of how can we make the recruiting a lot more  better, retention more effective but also how we [1:34] make the talent mobility and talent development in  a much more effective format. Yeah. Hey everyone, [1:42] great to meet you. Uh my name is Max Rushko. I  am director of AI and data techniques. Uh we have [1:48] been in market for 30 years helping our clients  building their software solution products uh [1:54] providing their needed expertise and resources. So  in a nutshell thank you. My name is Rafael Valdez. [2:04] I am the CTO of Zumba Fitness. I run product  engineering, QA, data, and now I am in charge [2:13] of AI and ensuring that agents spread throughout  the organization. Um, we that's it just that no [2:22] problem. But, um, yeah, excited to be up here.  Awesome. Hi everyone. I'm Carlin Power. I work for [2:28] JL, a global commercial real estate company that's  about 120,000 employees in 80 countries. I run all [2:36] of product AI product and enterprise enablement.  Um I'm from Minnesota and so I'm thankful both [2:43] to be here and for the air conditioning at the  Venetian. That's excellent. You know what? I think [2:49] we'll start with you first. Sure. Um so let's  kick off with a two sentence in Oh, sorry. My my [2:56] mistake there. Wrong line. Um tell us about an AI  initiative you've either completed or one that's [3:02] currently in progress. and uh yeah details to help  them all really get it. Happy to do that. So um we [3:09] actually just about a month ago released a product  that we call Ask JLL. Um it is a collection of [3:15] domain agents um across all of our corporate  functions. So HR, finance, uh technology, sourcing [3:22] and procurement, uh research, legal, ethics. Um,  and basically what they do is they the the teams [3:33] uh put together all of their knowledge bases  and these agents go and contextually reference [3:38] um those those knowledge bases to answer questions  and and take action on an employes's behalf. So [3:45] right now they're employee facing. We're working  on client-f facing agents, but they'll do anything [3:50] from uh uh scheduling time off to ordering um  laptops to or or other peripherals and equipment [3:59] um helping people file ethics complaints uh  request legal contracts. So, anything that you [4:06] have that's enterprise level um you know content  or or domain expertise, we have have centralized [4:14] and output agents against. That is amazing. I'm  so impressed always. I mean, everybody at this [4:21] conference gives me information I want to go home  and use right away. Um Ralph and Max and Mahi, [4:27] do you have some similar similar experience any  either of you would like to share on this? Yeah, [4:32] sure. Sure. I uh so we have we have something  similar. I I'll give three brief I promise to [4:41] be brief examples. Um, we use Snowflake agents  for a lot of of what it is that you're describing [4:47] in order to capture and centralize our our data  analytics from across the organization from every [4:56] major KPI point in a place that we know the data  is verifiable, correct, and that everybody can [5:02] use. And it's a way to to democratize that out. At  the same time, we've started to roll out different [5:10] agentic workflows. Uh, one example that that  I've been very involved in um is something we're [5:16] calling Kodu, which is a aentic software factory,  which pretty much anything that needs to be [5:23] developed on any one of our our web platforms or  native mobile application platforms can be given [5:28] to Kodu and Kodu will do the job of figuring out a  plan and how to build it. ask the right people any [5:34] questions it has and any type of of discrepancy it  finds. Start working on it. As it's working on it, [5:41] it'll send out questions to the relevant  groups for for things. It'll send out previews [5:45] so people can try it and test it in a safe  environment before anybody sees it publicly. Um, [5:52] and then approve it and and get that out. And  we're actually rolling the same thing out now [5:57] for information that's brand new that needs to  go into the knowledge bases. Um so yeah that's [6:04] that's great. Mahi what's been your experience? So  we actually deploy agents within the HR ecosystem [6:11] for recruiting. We have agents like screening  agents. We actually built for like we have over [6:16] 700 companies who use our product. Over a billion  users use our product every 18 months. And like [6:23] we have in-depth understanding about like how  people really get retained, how they stay within [6:28] the company, how resource planning hap actually  happens, how workforce planning agents can really [6:33] build in the backend infrastructure. So that  where a company is shrinking and where a company [6:37] is growing, how do you really readjust your  talent and the resource planning in an effective [6:42] format is what our tools actually does. But the  most important thread is we are a very vertical [6:48] focused. Uh the point is context is everything.  uh like the generic context there are 30% of the [6:55] use cases which we normally really solve but there  are very specific context which you have to really [7:00] go for and there the efficiency and the impact  what you can really bring in is a very different [7:06] equation so how do you build agent harnesses at  each use case is what our expertise is within HR [7:12] ecosystem well that's amazing I just have to say  thank you um Carlin did you want to close out this [7:20] question with any additional observ observations.  Um now go to Max too. Oh go ahead Max Max. Thank [7:26] you. So our exper expertise mostly connected with  the engineering the building software. So our [7:34] like something worth sharing. So we have partnered  one one of our clients in the healthcare and they [7:39] wanted to enable their software engineering team  their business analyst project managers quality [7:44] assurance with AI not just only AI tools but  rebuild their software development life cycles [7:50] using AI. Uh so what was our uh main goal to  analyze their workflows how they people do the job [7:58] uh like specific workflows for software engineer  like writing the code maybe deploying setting up [8:04] CI/CD other stuff and across all roles understand  what are the AI eligibility for those workflows [8:11] and how we will model like model model them with  engine AI tools. So we have started like I would [8:19] say four months ago we already have done a lot  of those discovery phases I understand about [8:24] the workflows and also started implementations  already with the covering the code reviews uh [8:29] covering the security risk assessments uh using AI  that's excellent thank you and final thoughts on [8:37] this yeah I think just just to close and because  this is supposed to be a practical session right [8:42] of like what are you actually supposed to do with  all this I think the advice that I would give you [8:46] is that I I don't think that it matters so much  as like if you're going to build agents yourself [8:50] in house and you have your own development  teams versus you go and buy technology that [8:55] has a lot of that infrastructure and a lot of  that orchestration layer already built. I think [9:00] what you need to do is make sure you are pairing  the right experts and your businesses and your [9:03] functions with the right technology stack that  makes the most sense for your business, your [9:09] your budget and then the impact. And I think the  one thing that we did really really differently [9:13] at the beginning that I would also advise you  on is we were focused a lot more on the value [9:20] and the ROI of the agents that we were delivering  and not so much on the productivity and efficiency [9:25] that I think where everybody started. And not that  that's a bad thing that there just wasn't enough [9:29] in it for us to just do it for efficiency. We're a  massive company and so the value has to be really [9:37] um really clear and it has to pay off  exponentially for us to invest and and I [9:43] love that. And when did you make that decision and  sort of how is it panned out for you? Um I mean [9:49] we started our AI journey about two and a half  years ago. Um and we started building a lot of [9:55] the stuff proprietary because we one didn't want  better vendor lock in and uh we really felt like [10:02] we needed to build some of the expertise in house  in order for us to make those decisions the build [10:07] by decisions and so you know we just spent time um  experimenting we built an internal proprietary GPT [10:15] um and then we got to the point we're like all  right well the labs are investing billions of [10:19] dollars in this stuff so maybe we can't really  keep up so let's keep the experts and the and [10:23] and the domain experts in the different business  lines and corporate functions and let's bring in [10:28] the tech that we need and then supplement with  lowcost proprietary tech where that made sense. I [10:33] love that combination. That's brilliant. I'm going  to start off with Ralph on this one. instead of [10:38] just bolting AI onto current processes. Um the uh  how have you actually ripped apart and redesigned [10:46] a workflow specifically a workflow to give an  agent the lead role and and and maybe also go [10:51] into like what kind of resistance did you face?  Sure. Yeah, it's a great question. Um so I'll use [10:58] the Kodu example. Um getting an agent to write  software is easy. That's pretty much a solved [11:04] problem in terms of just getting it to generate  code. What should it generate? Who decides? How do [11:13] you handle discrepancies? Who approves something?  Who does the a design review versus a product [11:21] decision review? These are the things that really  make it more of a challenge. And so what we did [11:27] is we did made a few decisions that really helped  um start because this product only launched just [11:34] a month a month ago. So let me be clear about  that. But what we did is we made some decisions [11:38] that really helped it get adoption going. Number  one, we made sure we met people where they were [11:44] at. We didn't want to create a tool that required  people to go to another place to interface with [11:50] another tool to No. Most of our company works in  Slack. It works in Jira. Nobody likes working in [11:59] Jira. I'm sorry if anyone from Atlassian is here,  but nobody enjoys that process. So everything [12:05] goes through Slack. Every decision, every touch  point. We have mapping routes out based on what [12:11] the code is doing and what business line it's  touching so that the right teams get the right [12:16] context and the right touch points in order  to figure out how to answer the questions. [12:21] Um, sometimes AI is not the clearest at  explaining what it is that it's asking you. Um, [12:28] but it's great. You can just tell it, hey, give  me a clearer explanation and it'll come back, [12:33] gives them a clearer explanation. We really tried  to focus on user friction and user experience in [12:38] order to drive the adoption. Cuz if it would have  been just some tech tool that technology owned and [12:43] ran in a closet somewhere and you had to go to  some IP address, that would never have worked. [12:49] instead we tried to make sure we met people they  were at. So I really think that that's one huge [12:54] takeaway that I would recommend. And I love your  point of helping people meet the technology where [13:01] they are comfortable. I mean one of the beautiful  things about AI really is that it's the first time [13:06] we've a actually been able to say to a software  I mean I can't say to you know people software [13:11] Excel is like I don't understand it. Could you  be more clear? You know instead I'm going to [13:14] a YouTube video or something. And I love the way  you organize that to make your employees not feel [13:19] insecure as they approach that. It's amazing.  Very helpful. Um Max, how about you? So from [13:26] our standpoint, the main part about understanding  like where we need to apply or not apply AI though [13:32] this is up kind of like we call it AI eligibility  framework. So we assess uh like how much impact [13:39] of AI will be for this workflow or not. So is it  still will need human being be in a loop all the [13:46] time part-time like at the end of the process  just validating the result. Uh so how much will [13:52] how how much effort will will be spent to build  those automation and how much effort will be as [13:58] overhead to checking it and if really for some of  the processes we understood that it's like it will [14:04] not help it will not save the time it will just  also might be expensive in some cases. So that [14:10] that's what we assessing as well. And another like  standpoint of view of that is uh like immeasurable [14:18] impact everybody talked about Roy company you  you calculate it in different ways. Uh but we [14:25] also spend okay how much like picking one workflow  how percentage of the time might be saved uh how [14:33] often this process happening monthly daily hourly  etc. So me kind of creating this calculation model [14:41] and then just deciding from where to start.  You know, I think that what one of the things [14:47] that's really interesting about that is uh how  each one of you so far and you're next. You're [14:53] the closing on this one. Um how you have taken uh  while you're taking something away. I mean a lot [15:02] of this question is really about when you take the  power away from the person and give it to the AI, [15:06] how does that person react? But you know you when  you comment in an organization that this is we're [15:11] keeping the human in the loop like you mentioned  where this is very human- centered. You meet them [15:16] at the middle that kind of thing. I think it does  make it a lot easier. I really appreciate your [15:20] comment that way. Thank you. So the our point  of view is workflows are designed based on the [15:25] systems what we have before and the workflows  right now what you have to design is the newer [15:30] systems what you're really building. If you look  at the the last generation systems, there are only [15:34] two things which are important. Your application  and your a configuration layer which is what [15:38] really makes the workflow work. But now you have  an application, you have a configuration layer, [15:44] you have a memory an onlogical layer and then you  have an orchestration layer. And what happens is [15:49] your use cases what you have to pick up for which  use case you want to fine-tune is very important. [15:55] Normally in general what happens is what are the  few like what are the areas where I can really [15:59] bring ROI really effectively but the ROI comes  only when you can really make a business impact [16:05] straight but business impact is in the hands of  business units and the functional units of the [16:11] company really operates to really interconnect. So  how do you really understand the business impact [16:15] of a business unit based on a business strategy  align with the people operations and that connects [16:22] with where your culture is. So what we call that  as a hypersell it's a five-dimensional element [16:27] which industry you belong to then you basically  have to look at which particular business unit [16:33] do you want to create an impact for and what is  the ROI you can really generate then you have [16:37] to look for which particular location because all  locations are not same what I mean by that is the [16:46] way you really recruit or retain or really grow an  employee specifically in North America is entirely [16:53] different from what you do in Europe that's  different from Asia. So how do you understand [16:57] that? Then you look into which particular job or  a role you are really looking at and then what is [17:02] the AG structure? The AG structure of a leader is  very important for an ROI to really come out. ROI [17:08] cannot come out because you really like it. Uh  it's not because your agents are really doing [17:12] it. Your AG structure has to be well thought out  because your change management is so effective. [17:18] So, I'm going to dive into that a little bit  more because I I need a little more info and [17:23] hopefully everybody else will benefit from this  answer. So, you talked about you have four layers, [17:29] right? And that's excellent. And one of the other  challenges which I see in every industry honestly [17:35] is the difference between the business units  and the functional units. And have you found, [17:40] and it's okay if you haven't, but I'm super  curious, um, that the the advent or or placement [17:47] of the AI processes between the four layers  and between the critical function and critical [17:53] business units, there's traditionally been a  sort of a wall between those units that they [18:01] don't necessarily understand each other's work.  Has the AI helped you them further understand the [18:08] work of the other people around them? It depends  on the personalities and the hierarchy do matter [18:13] and the company's culture do matter because what  is a company about? A company is all about mental [18:19] models, behaviors, strategy and leadership  styles. Each company differs dramatically. [18:25] What is the difference between Jensen and Jeff  Bezos or Satya? their personalities of how mental [18:32] models are actually built. So the AG structure  is predominantly connected to that and if the AG [18:38] structure is not well connected and if the leaders  are lousy your ROI doesn't exist. It doesn't [18:44] matter whatever you really deploy. Agents have no  value in it because you have to make decisions in [18:49] such a way so that you can make the moment so  that the impact do exist. And I constantly see [18:54] right now there is a conflict between the CFO the  CIO the CHRO for what the CEO wants to get out [19:03] of as an outcome. And this conflict is because  of the hierarchical structure within a company [19:08] which nobody want to talk and they're really  saying like agents are not really performing [19:12] that is not per it's not that your ag structure  is convoluted in a very ugly format until unless [19:19] you can really rip a part of where the structure  aligns you cannot really get the best ROA at all [19:25] in this market. And now that leads us to Carlin  because you probably have some very specific No, [19:32] not at all. Um, so I want to talk about  like a pre-step to getting there. Um, [19:39] and a decision that we made that I think was  really helpful, which was there's there's a lot [19:44] of processes that cross a lot of business lines  and in corporate functions, right? And so let's [19:48] talk about a specific one which I call request to  resolution. Anytime an employee has a request and [19:54] some human in the loop has to resolve something  within a company or for a client. Um so right [19:59] now oh so three months ago um that sat in um uh  12 systems uh 74 teams, 17 budgets um and about [20:14] 240 stakeholders at like the leadership level.  Um that process now has one accountable owner. [20:21] So um so we took a lot of these massive crossf  functional processes and we assigned a leader in [20:27] the organization to be accountable to the for the  decision-m of that process and that organizational [20:33] redesign um so that it wasn't a conflict between  the CHRO and the CFO and this and the CIO which is [20:40] very real also in our um uh organization but was a  tactical way we could start to break things apart [20:49] um make sure we understood the impact act, but  that the person that was making the decisions [20:53] was accountable to the benefits and the ROI to the  firm and not to the individual area that they were [20:58] accountable to that had different SOPs, different  KPIs and and different budget constraints. We also [21:05] pulled um a budget out separately to be able to  empower that leader um to be able to make some of [21:14] the decisions that the individual teams or leaders  weren't empowered to do before. So I think before [21:20] you get there, which is a really big step for  where most people are today, that is what advice I [21:27] would give you. And that is amazing advice. I hope  everybody wrote that one down. Um, so the next [21:34] one I am going to start with Ralph again on this  one. Um, but uh but we're going to save Mahi for [21:42] the end because he's a he's a key part of this.  And one of the things that we know about these [21:47] challenges is that each person's job is changing.  And so where they were they were the human doing [21:55] the work, now they're the human in the loop and  reviewing. So when it so in your workforce, and [22:01] the reason I'm ending with you, Mahu is because  you have an overview of many companies here. Um, [22:07] as you've approached this process, are you  redesigning job descriptions? Are you getting [22:11] resistance to that? are you putting them into  further training? How are you moving people from [22:16] production to reviewing the production of an AI?  So, it's it's a very valid point. Um I I I'm not [22:27] one of those people that sits there and says uh AI  is going to take everyone's job or anything like [22:33] that. I I've I've worked very closely with I for  a long time now and I I see it more as an augment [22:39] than than a replacer. But I do think that there  should be a healthy amount of fear for everyone. [22:47] And what I mean by that is the you know I didn't  say this has been said a thousand times but AI is [22:54] not going to take your job. Someone who learns AI  is going to take your job. And that's the message [23:01] that we try to impart to everyone is experiment as  much as you can in order so that you don't do it [23:09] for the company, do it for yourself because you're  going to need these skills whether you're here or [23:14] you're somewhere else. These skills are the ones  that are going to make a difference going forward. [23:18] And so we have invested a lot of time in training  and in trying to get people there. The biggest [23:26] friction point that I find, and it is entirely  valid, is everyone's busy. So, everyone sits there [23:34] and says, "The number one complaint I get, and  it's it's hard to disagree with because it's it's [23:38] very fair. I already work full-time all the time,  and now you're asking me to work extra time to be [23:47] slower so that I can learn something that might  help me two months from now." And the feedback [23:55] I always give them, it goes back to I will help  you. Someone will help you. Please always ask for [24:01] help. You do not have to go this alone. I highly  recommend for all of you, the ones of you who are, [24:08] and I imagine it's most people here cuz you're  at this conference. If you're really into AI and [24:13] you have a workflow and you have demo it to your  peers, show it to people, walk them through it, [24:19] tell them you're available to answer any questions  they have. and just try and show people what is [24:27] possible and what works is so that they see the  light at the end of the tunnel so they're willing [24:34] to go through that slight early pain in order to  get the learnings from the long term. Excellent [24:40] point. Everybody is busy. Thank you Max. Yeah. So  from our standpoint that I believe that software [24:48] engineer is really one of the areas that impacted  by AI a lot and in our domain I would say it's [24:57] not even that you should learn AI but it's not  even a question who use it better more efficient [25:04] uh could build with it and something that he  might never not even known before. So roles [25:11] changing and evolving is something new. Sometimes  it's just change of a title honestly and maybe [25:17] just adding a few rows in the job description. But  sometimes and it's more for startups and smaller [25:24] company. Right now it's like a human orchestra  like doing five or different jobs that this [25:30] person wasn't able to do before. Uh so from our  standpoint it's really changing the job nature. [25:38] uh it's very pushy on the efficiency of one human  being. Now it's like there was like previously was [25:46] like rockstar uh senior software engineer right  now like everybody should be kind of rockstar [25:52] and doing this thing. So that's that's influenced  us a lot. But on the other point we that's it's [25:59] when we like looking seeking for uh new people we  already expecting them to have some head hand head [26:06] hand head hand head hand head hand head hand head  hand head hand head hand head hand head hands-on [26:06] experience and sometimes if it's like senior role  deep uh experience with AI tool set AI basics [26:15] something like that and what we do also for our  workforce as well because we are our main value [26:22] for our clients is to have experienced and uh well  trained engineers. So we we put a lot of effort, [26:30] a lot of investment into education and training  and building a culture uh of like thinking about [26:38] the task how I can do so as soon as you get  the task you should think like from the very [26:44] beginning. So should it be done by some agent  by I me so it's like changing of your vision [26:52] uh of how you do the job. Thank you so much.  And before we go to you, I'm going to ask you, [27:00] you think of your uh CHRO, right? And and how  is that person specifically addressing sort of [27:08] maybe some examples of how they're addressing the  stress that your employee base might be having [27:14] um during this rapid change? I actually  don't really think it's our CHRO. Um a lot of [27:21] um the empowerment our employee base feels around  AI is act actually starts with our CEO. Um and so [27:30] his his expectations and the way that he set kind  of the compensatory structure of our executives is [27:37] very much aligned to the empowerment of employees  and the transformation that our company needs to [27:42] go through to be able to work with AI. And so  I think that's a fundamental difference in in [27:48] in the shift that we had to have culturally  is that you kind of had to start at the top [27:54] with the executives, align the the incentives and  their compensation to the behavioral change that [28:02] um that he wanted to see in our organization. And  then as you go down tactically in the layers, um [28:10] what it became was my team was responsible for uh  educating all of our leaders on how to essentially [28:17] rewrite their value proposition of like what  their function or their business unit delivers [28:23] and competes against in the market today. And you  would be I mean I certainly was surprised maybe [28:29] you wouldn't be of how many people couldn't really  even articulate their current value proposition. [28:34] And so we we sat with leaders for months like  writing out their current value proposition, [28:40] mapping out like what is going to be replaced,  enhanced or augmented with AI across every single [28:45] thing that you are responsible for and how is the  market going to react to that and then remapping [28:50] all of that to a value capture exercise where it's  like is that more throughput is it more lines of [28:55] revenue is does it open more markets for us? does  it improve efficiency so that we can uh spend more [29:00] time on delivering better quality um services  to our clients? Um and then rewriting the value [29:07] proposition based off of that. Um, so I think we  avoided some of that. Like not that there wasn't [29:14] fear in the organization, but we avoided some of  the longstanding fear and resistance from making [29:21] sure that people were rewarded for how they used,  implemented, and transformed their business with [29:28] AI and were not punished for it by like replacing  jobs and roles. Um, and just like as a tactical [29:36] example, you know, I used to run product support,  so technology support for the company. And I sat [29:42] my whole team down two and a half years ago and  I said, "Hey, I'll like 80% of your job is not [29:47] going to be a job anymore." So, we have two we  have two choices. We can brace for impact or [29:52] we can decide what our new value proposition is  going to be. And I think um you know now they're [29:59] knowledgebased experts and they're agent to agent  orchestrators and like to go from reactive ticket [30:05] support as a means to an end and a necessary evil  to being able to tell my CEO that these people now [30:15] um help us maximize the value of our tech  investment is a pretty significant transformation. [30:20] And I love that you said that that last section  because I think that that's that is key. What [30:25] they need to hear is like look that's the 8020  rule. We're going to take 80% of the sort of wrote [30:31] exercises off of your job. Now you get to take  that 20%. Which is really rewarding, right? The [30:36] more difficult issues and the people who really  need help and then we're going to expand that [30:40] to 100. That makes that job so much more cool,  isn't it? 200 or 2,000, you know. Yeah, exactly. [30:47] Um so Mahi those have been some great examples but  you have the view of a lot of companies what have [30:53] you seen that's different than what we're hearing  here so the one point Carolyn said which is about [30:58] the CEOs are actually the the primary drivers I  agree with that the primary reason why CEOs are [31:04] the drivers because the market is pounding there  you have to bring a different level of operating [31:11] model which can really save money and your growth  has to be there if that doesn't happen, the market [31:18] won't spare you. And they're saying the technology  is the impetus to really make it happen. And now [31:23] that does actually filtering through the whole  log structure. But every company is like we really [31:29] think about work in multiple zones. Zone 1 2 3 4  and five. One and two or more knowledge worker. [31:36] One and two are more like frontline workers and  four and five or more knowledge workers. The way [31:42] AI is helping in frontline worker to a knowledge  worker is day and night. But no company is [31:48] completely knowledge workers, completely frontline  workers. You might tilt towards 7030 or 8020 or [31:54] 9010 your your thing changes. Now what happens  is how do you really really deploy in these [32:00] segments in effective format. Now there is another  trajectory what we also have seen by industry how [32:07] much you can automate and how can how much you can  use agentic infrastructure do vary tremendously [32:13] and now based on that you basically have to see  within a my company where my complete realignment [32:20] has to come in that's where CHRO comes into  picture our point of view is every CHRO can really [32:25] see where the company is shrinking where the  company's growing where the company is stable in [32:29] that how do you move people between the company in  an effective format and what the skills what the [32:35] process is and what the systems have to be rebuilt  and as long as people are doing it they're getting [32:40] an effective results I'll give an example uh a  bunch of them like take a healthcare uh bailis [32:46] cotton white is one of the companies who use  our product they have multiple healthare systems [32:50] within Texas and what happens is they have to hire  people nurses but if you look at nurses today the [32:58] number of people who are graduating are less  than the number of people who are retiring and [33:03] the country is having an aging population. And now  if you really look at that nurse hiring and nurse [33:08] retention is all-time lowest like it's a very  complicated equation. Now how much you can really [33:13] automate is like you can do all the screening  infrastructure in a very effective format and you [33:19] can really save at least 20 25% of the time of a  recruiter straight away per week and a manager you [33:27] can save at least 3 hours. Yeah. So what is the  infrastructure you will deploy for that particular [33:33] task. Now if you apply the same thing for  manufacturing which is thermopisher is one of our [33:38] customer but in thermopisher 20% of their actual  employee base in manufacturing pharmaceutical [33:45] manufacturing but in pharma manufacturing what  happens is you can hire a lot more people more [33:50] efficiently and fast it's all about time and cost  but what happens is once they show up do you have [33:56] a capacity to manage it right and now you move  recruiters from doing recruiting job to actually [34:02] onboarding job. So, you're switching jobs on the  fly based on the use case you have at hand and [34:08] the complexity you're really building. That is  very helpful and thank you for the example of [34:12] the nurses because I think we can all relate to  that one. Um, so we only have 10 minutes left. [34:18] We're going to move into Q&As's but before we do  uh you know we're competing with in your memory [34:24] with all the other sessions today. So I've asked  the panelists to think of that sort of one two [34:30] sentence statement they want you to remember when  you are traveling home and getting ready to go [34:35] back in the office and I'm going to go ahead and  start with you just take like 30 seconds so the [34:43] one most important thing is focus on the use case  and outcome so effectively uh where agent hardness [34:49] can really sit. The problem I'm really seeing  right now is if you come to AI for the last two to [34:55] three years, every year the concept is completely  switching. You rag from like graph models to what [35:03] you really call as loop engineering to graph  engineering to right now harnesses. You're [35:08] constantly switching. And if it is constantly  switching from a technological landscape, [35:13] what is the one thing which you can really think  through in an outcome basis that you can be stable [35:18] upon and then you are really constantly adjusting  is a very critical thing to focus on and focus is [35:24] difficult with this much information in 3 days  I have to or four days if you count Monday. Um [35:29] yes so Max what's your takeaway? Yeah. So this  idea have been like have been heard through all [35:35] the panel but the if the process that you try to  automate or try to change is broken itself or it's [35:42] not mature AI will only magnify the problem but  will not fix it. So before you starting applying [35:50] agent, genai whatever you technology you wanted  to make it better optimized if the process itself [35:58] doesn't work like it should be or doesn't provide  the value AI will not help you in that matter. I [36:04] actually have to tell you I love that one. I am  definitely taking that one back. Fix the process [36:09] first before you start adding the AI into it or  at least not fix the process but identify the [36:15] weaknesses. Excellent. Thank you, Ralph. It's a  high bar to follow, Ralph. Sorry about that. For [36:22] me, change is coming. I loved your example of the  8020 because there is so much that's going to get [36:30] automated over the next 6 months to 2 3 years. But  there's going to be so much opportunity that that [36:37] creates for people who are flexible and willing  to go out and learn and manage and own new things. [36:42] there's going to be an entire new surface of work  that someone's going to need to be the owner of. [36:48] And so my recommendation to everyone and your  peers and everything else, find those things. [36:54] Find the way that you can bring an agent into the  organization and you can own it and look for the [37:01] opportunities that going to get created from all  of this change because it's it's going to be out [37:04] there. Excellent. I love that. And finally, um  lead by example. Um the the the most effective [37:14] change management I have seen is when leaders  role model these behaviors for their teams. And [37:20] I know that that seems kind of elementary, but  you would be surprised at how often that isn't [37:25] happening. Um and that's easier for us because  obviously our incentives are aligned to that. But [37:31] you your teams are going to come up with the most  transformative ways to compete in the marketplace. [37:37] Like if you if you leave that all up to yourself  or the leadership teams, you're not going to [37:41] achieve what you're really capable of. I love  that. Ask the people who are actually there, [37:46] include them. H that's fabulous. Those are great  takeaways, but now it's open to the crowd. And [37:53] we have our first question. This panelist, we're  here. Arlon, I'm wondering if you can share what [38:02] the CEO did from a compensation standpoint to get  AI adoption going across the organization. Yes. [38:08] So um our um each one of our global executive  board members which are essentially the people [38:14] that report up in through our CEO and our um  and our board of directors have a set of of [38:21] business outcome KPIs and then compensatory  APIs and KPIs and um they cannot uh uh get [38:31] paid any of their bonuses if they don't meet their  compensatory KPIs. So our CEOs set AI fluency KPIs [38:42] um set the percentage of automation that they  expected in each of the business lines and and [38:47] corporate function corporate functions and then  also around just uh a general adoption of AI and [38:54] AI agents within the platform. So there's all  different permutations of those based on like [39:00] which corporate function or business unit it was.  But um every single one of their compensatory API [39:06] KPIs for the last um two years have been around  AI and and transformation. Excellent. Thank you. [39:19] Thanks. Okay. Uh Nicholas Butterworth, Butterworth  Adviserss. I'm a former former CTO. I was at CBRE [39:27] for a number of years. Oh man. Uh and uh turned  transformation consultant. So, you know, before [39:33] you can redesign a workflow, you have to discover  it. Um, and sometimes business process discovery [39:40] is not the easiest thing. In your example where  you listed the 17 systems and the real thing, [39:46] you guys been able to leverage AI in ways to  discover and document the workflows first, [39:53] you know, at the top end of the funnel of the  change process. Who wants to take that one? Well, [40:01] I we haven't been able to leverage a ton of AI  for that. It's a difficult challenge to solve [40:07] for sure. I think in the places where it's  easy software, like you know, to your point, [40:14] software engineering has been transformed so much.  It's been very easy to identify those workflows. [40:20] Those workflows are clearly documented. They're  out there. I would also say the same for data. [40:25] All of our data flows have been very easy to to  transitionalize. The things that are hidden are [40:30] more the that employee who is six levels deep who  has that one function that if they don't submit [40:38] that one thing at one time a week the whole  downchain stream breaks what we have not found [40:44] that with AI yet but it's a great I think it's  something that that we need to solve. It's a great [40:49] question. Um so we we have um uh but but it's  both human and and agents. So we have an offshore [41:00] function that we call JLo business services  and it's essentially based in in Bangalore and [41:04] um a lot of our lean six sigma six sigma process  engineers sit in in Bangalore and we also use a [41:12] um a workflow tool called Mendix and then also um  I'm going to get this wrong uh blueprint something [41:21] it's a six sigma software um and uh so there is a  whole team of people um that have their lean six [41:30] sigma certifications have this Mendix software um  and then um are actually mapping every single one [41:37] of what we call our level 01 and two processes  within the organization and the first thing [41:43] that we do is we transition it offshore and then  the second thing that we do is uh is automate it [41:48] through Mendix and then the third thing that we  do is a genify it so um that's the cheapest way [41:53] we have figured out how to do that I think there's  way more expensive ways to do do that with FDS but [41:59] um you know we can't spend all of our shareholders  money so so we actually do a bit different like [42:05] uh we we actually analyze a company way before a  company analyzes the reason is this if you go to [42:12] a customer or a prospect if we don't understand  that particular vertical better than them we are [42:19] no use that's what enterprise software has to  fundamentally change so how do you know better [42:26] than what they really know within that specific  function is because the context is not a single [42:32] company. The context within a company there are  multiple contexts and each context we can really [42:38] really look from all the 700 customers on top  of it whatever the data we can curate so that [42:43] we really fill in the first the product with  the ontologies. So when we go to a customer, [42:49] the customer we already know the customer so deep  and we can really show them which areas they have [42:54] to optimize and automate and why they should do  it. Not because there is something really coming [43:01] out. It's because there's a business pressure you  have that you have to really reveal right now. And [43:08] whenever we go with that and we use AI for that  in an extensive format that actually helped us to [43:13] scale a lot. I have had such a good time listening  to the four of you. Oh my goodness. Uh but we have [43:20] unfortunately run out of time and I I know we have  other questions over here and I appreciate your [43:27] patience. Um but our amazing panelists will be  meandering their way to the hallway. So, you know, [43:34] feel free to corner them. Thank you everyone.  Thank you each of you. You've been amazing.