Syngenta CIO Applies Crop R&D Stage Gates to AI Portfolio Culling

先正达CIO借鉴作物研发门径机制淘汰AI项目 先正达通过叫停缺乏硬性EBITDA目标的项目，并借用化学与遗传育种管线的淘汰纪律，来扩大企业AI价值。 在本次主题演讲中，Jeremy Groeteke 针对企业级 AI 面临的“试点炼狱”问题，阐释了先正达如何将农业研发的门径管理框架应用于软件投资。为应对技术栈盲目扩张和虚荣指标，先正达制定了不可妥协的推进标准，包括硬性 EBITDA 指标以及业务与 IT 的共同所有权。演讲重点介绍了“Aiden”项目——这是蔬菜事业部实施的一个智能体供应链灯塔项目，该项目大幅削减了手动规划时间，体现了将智能体工作流跨研发、运营和商业领域打通能够释放真正的商业价值。 农业研发依赖极为严格的管线纪律和门径筛选淘汰机制，需要从数以万计的化学分子或植物杂交组合中筛选，最终推进单一商业化产品。 (0:04) 企业级AI往往难以规模化落地，原因是组织任由数千个概念验证项目肆意生长，既未施加筛选压力，也缺乏停止向劣质项目注资的组织魄力。 (0:04) 现代农业运营高度依赖高科技集成，例如自动化农田机械、物联网产量监测仪以及极速的“从田间到包装袋”供应链周转。 (0:04) AI和智能体工作流所能创造的最大企业价值，源于跨研发、生产、供应和商业运营实现端到端的系统打通，而非孤立地优化个别信息孤岛。 (5:08) 与周期长达8至10年的传统农业研发管线不同，数字化产品迭代迅速，若缺乏明确的业务指标，极易在内部或外部用户面前直接面临失败风险。 (5:08)

To overcome enterprise "pilot purgatory," Syngenta Group applies the stage-gate culling discipline of agricultural R&D to artificial intelligence, requiring initiatives to prove hard financial returns rather than lingering in unscaled experimentation. Applying Agricultural R&D Selection to Software (5:08) Syngenta operates across more than 90 countries, spending over $2 billion annually on R&D and connecting approximately 100 million hectares digitally. In traditional chemical discovery, researchers evaluate roughly 10,000 candidate molecules to produce a single commercial product, and geneticists cull thousands of plant crosses before release. Jeremy Groeteke, CIO of Vegetables and Flowers and Global Head of Digital Strategy, argued that digital portfolios require this identical selection pressure. While traditional R&D operates behind a firewall for 8 to 10 years before product launch, digital tools frequently iterate directly in front of internal users or growers. Groeteke asserted that the largest EBITDA margin improvements come not from optimizing isolated functional silos like sales calls, but from connecting workflows end-to-end across production and supply, R&D, and commercial operations. Culling Proliferation Through Strict Business Criteria (8:22) Following the emergence of generative AI, Syngenta held an enterprise hackathon within two months, adopting an initial "thousand flowers bloom" mindset. However, this experimentation caused tech stack proliferation, with teams deploying conflicting models, architectures, and codebases across an organization of 50,000 global employees. Capital constraints forced executive leadership to shift from open-ended exploration to rigorous portfolio culling. To advance past early stages, initiatives had to satisfy two mandatory rules. First, each project required an explicit EBITDA contribution to its business unit or faced immediate defunding. Second, projects had to be co-owned by a technical lead and a regional business leader with P&L accountability. To align vocabulary and expectations, leadership sent these executive pairs to an intensive week-long AI program at MIT. Lighthouse Automation in Biological Supply Chains (11:04) Syngenta concentrated investments into seven enterprise "lighthouse" projects, including Project "Aiden" in vegetable supply chain planning. Unlike static industrial components, vegetable seed inventory consists of living biological organisms with shifting quality specifications and expiration dates. To address change-management resistance among middle managers, the Aiden team targeted demand confirmation as an initial beachhead. Previously, a global team of 60 supply planners spent two to three weeks manually cross-referencing incoming orders against biological lots across 35,000 rows of spreadsheets. Implementing an agentic workflow combining foundational machine learning with automated validation reduced this process to a few hours. Planners now review system-highlighted exceptions each morning with human-in-the-loop sign-off, which has lowered days on inventory and freed working capital. Infrastructure Realities and Cross-Domain Digital Threads (14:05) Groeteke warned against justifying AI deployments through head-count reductions, stating that human domain experts remain essential to supervise automated agents. He also cautioned against "vibe coding" and prototypes built on isolated demo datasets, which frequently collapse when integrated into live transactional ERP systems or enterprise data meshes like Databricks. Enterprise leaders must disregard vanity adoption numbers to concentrate strictly on business value. According to Groeteke, the broader promise of enterprise AI lies in a unified "digital thread" where IT functions as a neutral intermediary between silos. For example, if field reports detect emerging pathogens such as Tomato Brown Rugose Fruit Virus, real-time commercial data can trigger R&D to advance disease-resistant seed lines through production stage gates—such as moving directly from PLC 3 to PLC 6—to accelerate commercial availability. Groeteke concluded that escaping pilot purgatory depends on moving past isolated proof-of-concept experiments and maintaining a disciplined focus on scaling high-impact use cases that deliver measurable economic value.

先正达集团（Syngenta Group）数字战略全球主管兼蔬菜与花卉业务CIO Jeremy Groeteke指出，企业在推进人工智能时普遍面临“试点炼狱”（pilot purgatory）困境；为此，企业必须借鉴实体农业研发中对分子和杂交育种的严格阶段门控淘汰机制，以明确的财务指标与跨部门协同推动AI规模化落地。 借鉴实体研发管线的严格淘汰机制 (5:08) Groeteke首先纠正了外界对农业属于低技术行业的认知误区。作为年研发投入超20亿美元、数字连接全球约1亿公顷农田（全球可耕地约2.5亿至2.8亿公顷）的行业巨头，现代农业早已应用了物联网传感器与百万美元级自动驾驶农机设备，更在消费品供应链中支撑着高度周转（例如与百事合作在12小时内完成土豆从田间采收到制成薯片）。然而在软件与AI领域，企业普遍存在让未经充分验证的数字化产品直接在用户面前盲目试错的问题。 为了解决AI项目的无序扩张，先正达将其实体农化与育种管线的阶段门控（stage-gate）淘汰纪律引入AI软件投资中。在传统农化研发中，通常需要筛选评估约10,000个候选分子才能最终获得1款商业注册产品；植物育种专家同样会在严格的标准下无情淘汰数千个杂交种。传统实体研发往往在防火墙内历经8至10年才向市场交付成熟产品，而数字技术直接面对终端迭代，因此更需要在立项与推进阶段施加同等高压的筛选淘汰机制，避免缺乏明确商业指标的项目持续消耗资源。 设立EBITDA硬指标与业务技术协同制 (8:22) 在生成式AI爆发初期，先正达曾采取“百花齐放”的探索策略，并在ChatGPT推出后两个月内举办了全企业黑客松。然而，这种不受约束的实验迅速导致技术栈无序蔓延，不同团队各自采用互不兼容的模型、代码库与架构。面对约50,000名全球员工的庞大体量与资本约束，管理层决定从自由探索转向高压的资产组合筛选，明确决定“浇灌哪些项目、放弃哪些项目”。 为了落实这一纪律，先正达高层制定了两项不可妥协的硬性准入标准：其一，每个AI提议项目必须向所属业务单元证明明确的EBITDA财务贡献，缺乏清晰财务回报的项目会被立即终止资金支持；其二，推行业务与IT双负责人制，由承担损益（P&L）责任的区域业务领导者与技术负责人联合主管。为打破业务与技术团队之间的专业语言壁垒，管理层还将各项目负责人派往麻省理工学院（MIT）参加为期一周的高管AI集中培训，以确保双方在业务价值与技术实现上达成认知一致。 灯塔项目Aiden：供应链智能体应用 (11:04) 先正达将数字化投资收拢至七个旨在验证规模化商业价值的企业级“灯塔”项目。Groeteke亲自领导了蔬菜业务的端到端智能体供应链规划项目“Aiden”。他观察到，在企业AI落地中，高级管理层往往过于乐观，而中层管理团队则承受着剧烈的变革管理阵痛。为了打破这道“砖墙”阻力，项目组选择以“需求确认”作为单点突破口，迅速验证投资回报率。 与传统静态工业品不同，蔬菜种子属于具有动态质量规格与保质期特性的活体生物存货。此前，先正达由60名供应链计划员组成的全球团队需要花费2至3周时间，在多达35,000行复杂Excel表格中手动比对销售订单与生物批次库存。Aiden通过结合基础机器学习模型与自动化智能体核对工作流，将这项耗时数周的手工审计流程压缩至数小时内完成，计划员只需在每天早晨针对系统自动标出的异常项进行“人在环路”（human-in-the-loop）的核验。这一落地直接缩短了“库存周转天数”（days on inventory），显著释放被占用的营运资金，并让销售团队获得了实时的订单交付确定性。 规避工程陷阱与构建跨部门数字主线 (14:05) Groeteke告诫企业在规划AI战略时，切勿以“削减员工人数”为核心说辞，因为关键决策环节必须保留专家的人在环路监督；未来更现实的架构是专业AI智能体执行基础任务，而业务专家如同“主治医生”一般在上层监管多智能体运行。同时，企业必须警惕概念验证阶段的工程陷阱——依赖脱离实际业务的“氛围编码”（vibe coding）或临时云端样本数据构建的演示方案，一旦接入Databricks等企业级数据网格或事务型ERP生产系统极易崩溃。团队不应沉迷于用户采用率等“虚荣指标”，而应聚焦核心经济杠杆。 在Groeteke看来，企业级AI的最大价值不在于销售或局部预测等孤立职能部门的微调，而在于将生产供应、研发与商业运营端到端打通，形成流动的“数字主线”（digital thread）。在此体系中，IT部门应当扮演中立的“瑞士”角色。例如，当销售人员与种植户在田间发现番茄褐皱果病毒等突发病原体导致减产时，田间表现数据能够即时流转至研发系统，促使抗病新品种跨越生命周期阶段门控（如从PLC 3直跳至PLC 6）提速繁育，从而大幅加快种子商业化扩繁进程。 Groeteke总结指出，先正达数字化战略的核心是拒绝陷入无休止的概念验证，将资源集中在能够拉动核心财务杠杆的灯塔用例上。然而，将试点智能体嵌入复杂的底层ERP与实时数据架构时所面临的工程稳定性考验，以及人机协作模式下中层管理团队面临的持续变革阵痛，仍是该模式在全行业推广过程中需要审慎应对的核心挑战。

## Background 该视频由行业会议频道 Ai4 于 2026 年 8 月在 YouTube 平台发布，主讲人为先正达集团（Syngenta Group）全球 IT 与数字化战略负责人 Jeremy Groeteke。内容主要探讨企业如何借鉴农业科研中严苛的分子筛选与杂交淘汰机制，摆脱泛滥且低效的“AI 试点炼狱”，建立聚焦核心业务价值的 AI 投资组合治理框架。 ## TL;DR 先正达分享了其将农业研发中“万中选一”的筛选逻辑移植到企业 AI 治理的实践经验，强调创新的关键不在于广撒网做 PoC，而在于果断砍掉无法规模化落地的项目。公司通过设立明确的 EBITDA 财务门禁与业务技术双轮驱动机制，将分散混乱的探索收敛为 7 个具有实质损益影响力的“灯塔项目”。以供应链“需求确认”先锋用例为例，先正达利用基础机器学习结合智能体工作流（Agentic Workflows）与人在回路（Human-in-the-loop）机制，将耗时数周的手工数据匹配压缩至数小时并大幅释放营运资金。主讲人同时警惕依赖静态数据集的“氛围编程”陷阱，主张由中立的 IT 部门构建跨越研发、生产与销售的端到端数字线程，实现真实的商业回报。 ## Atmosphere & style 演讲整体风格务实、沉稳且极具实战洞见，主讲人以幽默风趣的开场迅速拉近与观众的距离，随后转入对企业数字化转型痛点的深刻剖析。叙事中充满农业科研与尖端科技交融的生动隐喻，既坦诚指出中层管理阻力与虚荣指标的危害，又以大刀阔斧的决断力展现出大型跨国企业在剥离炒作、追求底层经济效益时的严谨与魄力。 ## Key takeaways - [06:44](https://www.youtube.com/watch?v=3KbbRu1aucQ&t=404s) 农业研发依赖残酷的淘汰筛选机制（如筛选上万个分子才上市一款产品），先正达将这套“筛选-测试-晋级”的科研门禁逻辑完整移植到了企业 AI 项目管理中。 - [06:44](https://www.youtube.com/watch?v=3KbbRu1aucQ&t=404s) 企业在探索初期容易陷入技术栈泛滥与“试点炼狱”，必须具备果断停止注资（Defund）未达标项目的组织勇气，学会给少数高价值花朵浇水，让其余枯萎。 - [06:44](https://www.youtube.com/watch?v=3KbbRu1aucQ&t=404s) AI 项目晋级必须绑定硬性 EBITDA 财务回报指标，并由承担 P&L（盈亏责任）的业务 CEO 与技术负责人双轮驱动，拒绝沦为孤立的纯 IT 试验。 - [10:38](https://www.youtube.com/watch?v=3KbbRu1aucQ&t=638s) 先正达放弃对 50,000 名员工大水漫灌式的全员改造，转而集中优势资源打造 7 个具备显著业务影响力的核心“灯塔项目”（Lighthouses）。 - [10:38](https://www.youtube.com/watch?v=3KbbRu1aucQ&t=638s) 企业推进 AI 规模化落地面临结构性断层：高管层对前景极度乐观，而中层管理人员却在实际执行和适应中痛苦挣扎，构成需要重点攻克的变革阻力。 - [11:44](https://www.youtube.com/watch?v=3KbbRu1aucQ&t=704s) 先正达以“需求确认”作为滩头阵地，利用基础 ML 与智能体工作流将 60 位规划师耗时两到三周的人工 Excel 匹配压缩至数小时内完成，同时缩短库存周转天数释放营运资金。 - [11:44](https://www.youtube.com/watch?v=3KbbRu1aucQ&t=704s) 不要将写邮件或做总结等通用打字生产力提升作为核心 KPI，这类模糊效益极难反映在业务损益表（P&L）底线中。 - [14:05](https://www.youtube.com/watch?v=3KbbRu1aucQ&t=845s) 推进 AI 自动化不应以单纯裁员为假设，人类专家应扮演“人在回路”（Human-in-the-loop）的顶层监管与异常决策角色，统筹专业多智能体执行层协同运转。 - [14:05](https://www.youtube.com/watch?v=3KbbRu1aucQ&t=845s) POC 阶段基于静态数据集和“氛围编程”构建的 Demo 容易产生虚假繁荣，一旦接入实时 Data Mesh 或事务性 ERP 往往因真实负载而崩溃，团队必须警惕虚荣指标并勇于淘汰不达标产品。 - [17:02](https://www.youtube.com/watch?v=3KbbRu1aucQ&t=1022s) 通过构建跨越销售、生产供应与研发的端到端“数字线程”，能够将田间病害反馈与定向育种改良无缝串联，显著加速产品生命周期跨越与市场响应。 - [17:02](https://www.youtube.com/watch?v=3KbbRu1aucQ&t=1022s) 企业内部职能孤岛难以彻底消除，IT 与数字化部门应扮演“瑞士”般的中立角色搭建数据桥梁，确保关键数据在跨业务领域自由流动以驱动商业决策。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=3KbbRu1aucQ&t=4s) 破除“试点炼狱”：先正达在农业与数字化交汇处的定位 Jeremy Groeteke 担任先正达集团（Syngenta Group）蔬菜与花卉业务的 CIO，负责研发、商业、市场、生产与供应（P&S）以及集团全球数字化战略。在演讲开场，他幽默地调侃道，很高兴在拉斯维加斯看到大家没有泡在老虎机或酒吧，而是坐在这里探讨企业面临的共同难题——“试点炼狱”（pilot purgatory），即企业和供应商如何跨越 PoC（概念验证）与生产环境之间的鸿沟。 为阐明先正达的使命与业务广度，现场播放了一段短片。在充满风暴、干旱、病虫害等不可控风险的农业领域，农民的一年收成乃至生计脆弱不堪，进而关乎全球粮食安全。先正达通过在实验室与农田中融合生物学、化学、数据与 AI 技术，实现跨领域的农业突破，目标是在保护土壤和减少环境足迹的同时，帮助农田提高产量、养活全球增长的人口。 - 先正达集团全球业务覆盖 90 多个国家，年度 R&D 研发投入超过 20 亿美元。 - 其数字生态系统已连接全球约 1 亿公顷农田（全球可耕作农田总量约为 2.5 亿至 2.8 亿公顷，不含牧场）。 - 业务横跨植保产品、生物制剂、种子基因及数字化解决方案，深入消费端： - 墨西哥与百事公司（PepsiCo）合作的薯片（乐事、品客）供应链，从田间土豆采收到切片入袋可在 12 小时内完成。 - 市场上 90% 的抱子甘蓝（Brussels sprouts）以及 70% 的西瓜种质资源（germplasm）均源自先正达。 --- ### [03:01](https://www.youtube.com/watch?v=3KbbRu1aucQ&t=181s) 农业并非低科技：从传统 AI 到跨业务领域的智能体协同 很多人误以为农业是科技落后的行业，但 Jeremy 指出，早在 1996 年他的家族农场联合收割机上就已安装了测量产量的 IoT 传感器；如今田间价值数百万美元的农机已具备类似无人驾驶出租车的全自动驾驶能力。在实验室与 R&D 方面，机器学习、计算机视觉、全基因组预测模型早已深度应用，传统 AI 对先正达而言并不新鲜。 真正的突破在于生成式 AI（Generative AI）与智能体工作流（Agentic Workflows），它们能够打破研发（R&D）、生产与供应（P&S）、商业营销（Commercial）之间的孤岛。过去企业往往局限于单点优化（如仅优化 Salesforce 的销售拜访，或仅优化库存周转与营运资金的需求预测），而通过智能体打通跨业务域（Cross-domain）的数据与流程，才能在 EBITDA 和利润率（Bottom-line margin）层面释放巨大价值。 ```mermaid flowchart LR accTitle: 跨领域智能体协作架构 subgraph S1["传统单点孤岛"] A1["研发 R&D"] A2["生产供应 P&S"] A3["商业销售 Commercial"] end subgraph S2["协同整合"] B["智能体工作流 & 生成式 AI"] end subgraph S3["业务回报"] C["打通端到端流程"] D["解锁 EBITDA / 利润增长"] end A1 -.-> B A2 -.-> B A3 -.-> B B --> C C --> D ``` --- ### [06:44](https://www.youtube.com/watch?v=3KbbRu1aucQ&t=404s) 沿用科研的淘汰逻辑：从“百花齐放”到严格的门禁笑脸筛选 农业研发本质上依赖残酷的淘汰筛选机制：在植保化学领域，评估 10,000 个分子可能仅有 1 个能最终装瓶上市；在育种领域，成千上万次杂交中只有具备抗病性、抗脆性、耐旱性的精英种质才能晋级。先正达将这套严格的“筛选-测试-晋级”科研逻辑移植到了 AI 项目管理中： - **从失控到收敛**：ChatGPT 推出仅两个月，先正达就组织了黑客松，开启“百花齐放”（thousand flowers bloom）阶段。随后出现了技术栈泛滥（Tech stack proliferation）——不同的模型、LLM 与代码库各自为政。虽然这极大提升了组织的认知曲线，但在有限资本下，必须学会“给部分花朵浇水，让其余花朵枯萎”。 - **晋级标准与硬性指标**： 1. **硬性 EBITDA 目标**：各业务单元的 AI 项目必须给出明确的财务回报商业案例，无法达标者直接砍掉预算。 2. **业务负责人与技术负责人双轮驱动**：业务 CEO 与 CTO 必须紧密绑定，并一同被送往 MIT 参加为期一周的“AI 与商业”培训，确保双方使用同一种业务语言沟通。 3. **拒绝做成纯 IT 项目**：项目主导权必须由承担 P&L（盈亏责任）的业务领导者掌控，避免沦为单纯的 IT 内部试验而导致落地困难。 ```mermaid flowchart TD accTitle: 先正达 AI 项目筛选与淘汰漏斗 subgraph Exploration["探索阶段"] A["百花齐放 / 快速黑客松"] --> B["技术栈与模型广泛探索"] end subgraph Gating["门禁筛选 (Selection Criteria)"] B --> C{"硬性商业与组织指标"} C -- "未达 EBITDA / 缺乏业务主导" --> D["果断淘汰 / 停止注资 (Defund)"] C -- "达标" --> E["进入核心资源支持"] end subgraph Focus["灯塔工程"] E --> F["7 大灯塔项目 (Lighthouses)"] end ``` --- ### [10:38](https://www.youtube.com/watch?v=3KbbRu1aucQ&t=638s) 确立七大“灯塔项目”与直面变革管理挑战 先正达拥有约 50,000 名员工，无法对全员进行一次性的大水漫灌式改造（sheep dip the whole organization），因此公司集中优势兵力打造了 7 个具备显著 P&L 影响力的“灯塔项目”（Lighthouse Initiatives）。 Jeremy 本人亲自挂帅负责蔬菜业务的灯塔项目——代号为“Aiden”的端到端供应链计划协同重构。在推进过程中，他观察到企业变革管理中的普遍断层：**高管层（Senior Leadership）对 AI 的前景极度乐观，但中层管理人员（Middle Management）却在实际执行和适应中痛苦挣扎**。这种中层的阻力构成了推动 AI 规模化落地过程中必须直面并强行攻克的“坚硬砖墙”。 ### [11:44](https://www.youtube.com/watch?v=3KbbRu1aucQ&t=704s) 需求确认：先锋用例的选择与智能体工作流重构 先正达在推进 AI 落地时，第一步是寻找一个能快速向全公司证明 AI 价值的“滩头阵地”（Beachhead），而非泛泛而谈。他们直接切入了“需求确认”（Demand Confirmation）业务场景。农业供应链具有独特性——库存是有生命的生物产品（如种子），而非普通工业零件，存在质量随时间衰退的风险，因此需要实时与销售订单进行精准匹配。在过去，全球有 60 位规划师在数十个庞大的 Excel 表格中耗时两到三周，手动比对多达 35,000 行数据；而现在，通过部署 AI 智能体工作流（Agentic Workflow）与监督机制，这一流程被大幅压缩至数小时内自动完成。 这一变革并非全盘依赖大语言模型（LLM）或花哨的概念，传统的机器学习模型在其中依然发挥着巨大作用。现在规划师每天早晨上班时，繁重的数据匹配工作已经完成，只需扮演“人在回路”（Human-in-the-loop）的角色，专注于审核异常情况并进行决策确认。 ```mermaid flowchart LR accTitle: 需求确认业务流程升级对比 subgraph Old["传统模式（耗时 2-3 周）"] A1["销售订单流入"] --> B1["60 位全球规划师"] C1["生物库存变动"] --> B1 B1 --> D1["人工筛查 35,000 行 Excel"] end subgraph New["智能体工作流模式（数小时完成）"] A2["销售订单流入"] --> B2["基础 ML + 智能体工作流"] C2["实时库存数据"] --> B2 B2 --> D2["自动完成大批量匹配"] D2 --> E2["规划师人工复核异常项"] end ``` - **业务成果与财务影响**：除了节省海量人力时间，公司还成功缩短了库存周转天数（Days on Inventory），直接释放了原本沉淀在库存中的营运资金，在损益表（P&L）底线上体现出真金白银的价值。 - **销售端正向协同**：销售团队对订单履约有了极大信心，减少了因担心缺货而超额虚报下单的博弈行为。 - **度量指标的警示**：不要把“写邮件变快了”、“文档总结更好了”等通用生产力提升当作核心 KPI，这类模糊效益极难反映在业务损益表中。 --- ### [14:05](https://www.youtube.com/watch?v=3KbbRu1aucQ&t=845s) 规模化陷阱、人机协作模式与果断下线机制 在探索初期，“百花齐放”式做实验有助于组织学习，但企业必须清醒地认识到，最终只能将资源聚焦浇灌在少数能切实拉动营收或净利润的项目上。实施 AI 自动化不应以“裁员”（Headcount Reduction）为出发点，许多基于裁员假设的商业计划最终都被证明不切实际。人类的专业洞察在监督与异常处理中至关重要。未来的人机协作将形成层次化体系：基础智能体自主处理需求预测、配额分配及库存补给等具体任务，而人类专家（如同“博士/主治医生”）则在顶层统筹监督，确保多个 Agent 协同运转。 ```mermaid flowchart TD accTitle: 未来多智能体与人类专家协同架构 subgraph Oversight["人类顶层监管 (Human-in-the-Loop)"] H["人类专家 / 领域大师<br/>（统一监管与异常决策）"] end subgraph Agents["专业多智能体执行层"] A1["需求预测 Agent"] A2["配额分配 Agent"] A3["库存补给 Agent"] end H --- A1 H --- A2 H --- A3 A1 <--> A2 A2 <--> A3 ``` 企业在落地 AI 时往往面临“从 POC 到生产环境”的规模化瓶颈： - **概念验证（POC）的虚假繁荣**：许多团队在 AWS S3 中拉取一小块静态数据集做实验，依靠“氛围编程”（Vibe Coding）能做出惊艳的 Demo，但一旦接入生产环境的实时 Data Mesh（如 Databricks/Snowflake）或事务性 ERP 系统，系统便因无法承载真实负载而崩溃。 - **勇于“杀死自己的产品”**：产品经理常常把项目当成自己的孩子，很难客观看清缺陷。团队必须摒弃仅关注使用人数等虚荣指标（Vanity Metrics），紧盯底层经济效益；当项目无法达标时，要具备果断将其退役下线的勇气。 --- ### [17:02](https://www.youtube.com/watch?v=3KbbRu1aucQ&t=1022s) 打通端到端数字线程，以中立 IT 驱动跨部门价值 要判断哪些 AI 试点值得分配企业资源（"Deserve Acreage"），核心在于寻找能撬动 EBITDA 和损益表（P&L）的真正经济杠杆，而非盲目跟风通用的打字类生产力工具。通过将成熟的机器学习模型与自动化工作流相结合，打通销售、生产供应（P&S）与研发（R&D）全链路，构建贯穿全局的“数字线程”（Digital Thread），能释放出巨大的端到端商业价值。 ```mermaid flowchart LR accTitle: 跨部门数字线程闭环 Sales["销售团队<br/>（获取西红柿田间表现数据）"] -->|"反馈抗性下降 (褐皱果病毒)"| RD["研发部门 (R&D)<br/>（定向改良基因性状）"] RD -->|"产品生命周期跨越 (PLC 3 -> 6)"| Prod["生产与供应 (P&S)<br/>（快速扩繁制种）"] Prod -->|"现货订单实时确认"| Sales ``` - **全链路协同实战**：当销售团队在田间发现番茄品种因感染新兴的“番茄褐色皱果病毒”（Tomato Brown Rugose Fruit Virus, ToBRFV）而出现表现下滑时，该数据能实时回流至 R&D。研发团队可立即启动性状改良，生产团队将产品生命周期阶段从 PLC 3 跳跃推进至 PLC 6 并加速制种扩繁，迅速将抗性品种推向市场。 - **IT 组织的“中立国”定位**：企业内部的职能孤岛（Silos）难以彻底消除，但数据必须是流动的。IT 与数字化部门应当扮演好“瑞士”般的中立角色，在各个孤立部门之间搭建数据桥梁，让关键数据顺畅流转到做决策的人手中，并始终坚持“衡量真正关键的业务指标”（Measure what matters）。 ## Notable quotes > Most enterprises run AI the exact opposite way. They let a thousand pilots bloom, celebrate the experimentation, generate a flurry of proofs-of-concept — and then wonder why so little ever reaches production... — Jeremy Groeteke > The hard part of innovation was never starting. It's choosing what to kill. — Jeremy Groeteke > We went from sprawling, scattered experimentation to a focused set of high-impact initiatives, and the shift was less about technology than about organizational courage: the willingness to defund good ideas in service of great ones. — Jeremy Groeteke > [06:44](https://www.youtube.com/watch?v=3KbbRu1aucQ&t=404s) 在植保化学领域，评估 10,000 个分子可能仅有 1 个能最终装瓶上市；在育种领域，成千上万次杂交中只有具备抗病性、抗脆性、耐旱性的精英种质才能晋级。 — Jeremy Groeteke > [06:44](https://www.youtube.com/watch?v=3KbbRu1aucQ&t=404s) 给部分花朵浇水，让其余花朵枯萎。 — Jeremy Groeteke > [10:38](https://www.youtube.com/watch?v=3KbbRu1aucQ&t=638s) 高管层对 AI 的前景极度乐观，但中层管理人员却在实际执行和适应中痛苦挣扎。 — Jeremy Groeteke > [11:44](https://www.youtube.com/watch?v=3KbbRu1aucQ&t=704s) 不要把“写邮件变快了”、“文档总结更好了”等通用生产力提升当作核心 KPI，这类模糊效益极难反映在业务损益表中。 — Jeremy Groeteke > [14:05](https://www.youtube.com/watch?v=3KbbRu1aucQ&t=845s) 实施 AI 自动化不应以“裁员”为出发点，许多基于裁员假设的商业计划最终都被证明不切实际。 — Jeremy Groeteke > [14:05](https://www.youtube.com/watch?v=3KbbRu1aucQ&t=845s) 团队必须摒弃仅关注使用人数等虚荣指标，紧盯底层经济效益；当项目无法达标时，要具备果断将其退役下线的勇气。 — Jeremy Groeteke > [17:02](https://www.youtube.com/watch?v=3KbbRu1aucQ&t=1022s) IT 与数字化部门应当扮演好“瑞士”般的中立角色，在各个孤立部门之间搭建数据桥梁，让关键数据顺畅流转到做决策的人手中。 — Jeremy Groeteke ## Mentioned Jeremy Groeteke, 先正达集团 (Syngenta Group), 百事公司 (PepsiCo), 乐事 (Lay's), 品客 (Pringles), ChatGPT, MIT (麻省理工学院), Aiden, Salesforce, AWS S3, Databricks, Snowflake, 番茄褐色皱果病毒 (ToBRFV) ## Video description Jeremy Groeteke - Global Head of IT and Digital Strategy, Syngenta Group Not Every Molecule, Not Every Cross, Not Every AI Pilot Makes It to the Field Agriculture R&D runs on a hard truth: most of what you start doesn't make it — and that's the point. In crop protection, we screen tens of thousands of molecules to register a single product. In breeding, we make thousands of crosses to advance a handful of elite lines. Neither pipeline works because of the discovery; both work because of the discipline that comes after it — the stage gates, the multi-environment testing, and the ruthless culling of everything that doesn't earn its place in the field. We are very good at saying no to promising things that won't perform at scale. It's in our scientific DNA. Most enterprises run AI the exact opposite way. They let a thousand pilots bloom, celebrate the experimentation, generate a flurry of proofs-of-concept — and then wonder why so little ever reaches production, why budgets are spread thin across dozens of half-funded experiments, and why "we're doing AI everywhere" somehow hasn't moved the business. The hard part of innovation was never starting. It's choosing what to kill. At Syngenta, we stopped treating our AI portfolio differently from our science and applied the logic we already trust: broad exploration up front, then real selection pressure toward the few "lighthouse" use cases that demonstrably move the business — the ones worth concentrating talent, data, and capital behind until they reach full scale. We went from sprawling, scattered experimentation to a focused set of high-impact initiatives, and the shift was less about technology than about organizational courage: the willingness to defund good ideas in service of great ones. In this session, I'll walk through how we made that transition as a global agricultural enterprise — the selection criteria we use to decide what advances, the stage-gate discipline borrowed directly from our crop development model, the way we define and protect a lighthouse so it actually reaches scale, and the hard culling decisions (and internal resistance) that came with it. I'll be candid about what worked, what we got wrong, and what I'd do differently. You'll leave with a practical framework you can apply to your own AI portfolio on Monday morning — a way to decide which pilots deserve acreage, which to let go, and how to build the organizational muscle to tell the difference. Three takeaways: A stage-gate framework for deciding which AI pilots advance to scale and which get culled — before they quietly drain budget and attention. Why broad experimentation is the easy half of innovation, and how to build the discipline for the hard half: concentration and the courage to defund. How the screen-test-advance logic that governs both our chemistry and genetics pipelines translates directly to enterprise AI portfolio management. [Source](https://www.youtube.com/watch?v=3KbbRu1aucQ)

Jeremy Groeteke: Most enterprises run AI the exact opposite way. They let a thousand pilots bloom, celebrate the experimentation, generate a flurry of proofs-of-concept — and then wonder why so little ever reaches production... Jeremy Groeteke: The hard part of innovation was never starting. It's choosing what to kill. Jeremy Groeteke: We went from sprawling, scattered experimentation to a focused set of high-impact initiatives, and the shift was less about technology than about organizational courage: the willingness to defund good ideas in service of great ones. [6:44] Jeremy Groeteke: 在植保化学领域，评估 10,000 个分子可能仅有 1 个能最终装瓶上市；在育种领域，成千上万次杂交中只有具备抗病性、抗脆性、耐旱性的精英种质才能晋级。 [6:44] Jeremy Groeteke: 给部分花朵浇水，让其余花朵枯萎。 [10:38] Jeremy Groeteke: 高管层对 AI 的前景极度乐观，但中层管理人员却在实际执行和适应中痛苦挣扎。 [11:44] Jeremy Groeteke: 不要把“写邮件变快了”、“文档总结更好了”等通用生产力提升当作核心 KPI，这类模糊效益极难反映在业务损益表中。 [14:05] Jeremy Groeteke: 实施 AI 自动化不应以“裁员”为出发点，许多基于裁员假设的商业计划最终都被证明不切实际。 [14:05] Jeremy Groeteke: 团队必须摒弃仅关注使用人数等虚荣指标，紧盯底层经济效益；当项目无法达标时，要具备果断将其退役下线的勇气。 [17:02] Jeremy Groeteke: IT 与数字化部门应当扮演好“瑞士”般的中立角色，在各个孤立部门之间搭建数据桥梁，让关键数据顺畅流转到做决策的人手中。

[0:04] Jeremy Groeteke: pilot purgatory [0:00] Jeremy Groeteke: Not Every Molecule, Not Every Cross, Not Every AI Pilot Makes It to the Field [8:22] Jeremy Groeteke: thousand flowers bloom [5:08] Jeremy Groeteke: fail in front of the customer [11:04] Jeremy Groeteke: brick wall [14:05] Jeremy Groeteke: vibe coding [14:05] Jeremy Groeteke: Switzerland [19:05] Jeremy Groeteke: what matters at the end of the day

[0:04] Good afternoon everybody. Um, welcome to Vegas.  Looks like we got a few that aren't at slots table [0:09] or at the bar yet. So that's all right. Welcome.  My name's Jeremy Gretki. Really the two roles [0:16] I have at Senna Group is the CIO of our veg and  flowers business. uh oversee R&D, commercial, uh [0:24] marketing, uh PNS, production and supply as well  as do our digital strategy for the organization [0:30] uh across the globe. And so what we're going to  talk about today is is something that I think [0:35] many of you probably have faced uh or if you're a  vendor working with a customer uh pilot purgatory, [0:40] right? Trying to get the the new PC into  production and and business and why do we [0:45] make decisions we make? And so we've gone through  this as well and so we're here to share a little [0:50] bit of those learnings about that. And so again,  Senna Group is an agricultural company. I'm going [0:56] to give you just a brief video to kind of walk  through what Senna tries to do and accomplish [1:00] uh with our purpose here out here.  Everything is at stake. A storm, [1:08] drought, pests, a ravenous disease. That's  all it takes to destroy a year's work. [1:17] For a farmer, when there is so much you can't  control, the things you can control suddenly [1:23] become so important. When crops fail, a  farmer's livelihood is on the line. And [1:31] so is food security for all of us. But our world  holds answers in every field. And with innovation, [1:41] we can unlock them. With courage, collaboration,  and curiosity, we combine biology, chemistry, [1:49] data, and AI until a solution proves itself.  First in the lab, then out in the field, [1:57] side by side with farmers and growers. Together,  we create breakthroughs. Like when a crop makes [2:04] it through a harsh season. Like healthier plants  in healthier soil. More carbon kept in the ground. [2:12] More life at the field edge. Breakthroughs that  give us the power to scale fast, field by field. [2:20] And every step is done with one aim. To create  higher yields with lower impact on the farms that [2:29] feed us. In the soils that store our future. In  the hands of the people who carry the weight of [2:37] every harvest. With farmers and partners around  the world, we are working for a future where [2:45] agriculture can endure, increase to feed a growing  population, and improve the world we all share. [2:57] So, it's a general group. Uh you'll see  the play on the words there. Breakthrough [3:01] for farmers in every field whether that field is  computer science, uh biology, u soybeans, corn, [3:08] Brussels sprouts, whatever that is. We really  do believe that you can have breakthroughs in [3:13] all of those avenues. For those of you who may  not be familiar with the agricultural industry, [3:18] um often we get accused of not being very  technology savvy. Well, uh, in my own family farm, [3:24] we had IoT sensors and combines in 1996 that could  tell us how much yield was going on. And today, [3:31] you have million-doll machines that are driving,  no different than driving themselves, no different [3:35] than if you see the Tesla taxi in the trade show  booth. Uh, it can drive itself up and down the [3:41] roads in the fields. And so technology is there  in the field. Um, and an industry like ourselves, [3:46] we invest over $2 billion in R&D research. It's  nothing new for us. We've been doing machine [3:51] learning, computer vision. We operate across  90ome countries globally. We actually have [3:56] around 100 million hectares digitally connected  through our our digital ecosystem. Rough math, [4:03] there's about 250 million 250 80 million hectares  globally that are farmable, not counting range [4:09] and pasture, but farmable crop land. Um we offer  crop protection products, biological products, [4:16] genetics and digital solutions to customers  and supply chain around the world. And so we [4:21] actually touch a lot of the whole footprint. I  actually was in Mexico last week with PepsiCo uh [4:27] and potato production. So moving uh from potatoes  to your potato lace chips or Pringles. Fun fact, [4:33] that happens in about 12 hours or less from field  to bag uh that they can move their potato chips [4:39] through. Also, uh, so I'm in the veg side of the  business, so I'll give you some fun facts there. [4:44] 90% of your Brussels sprouts come from Senta. We  don't produce them as a farmer, but we produce the [4:49] genetics that farmers buy to do that. Seven out  of 10 watermelons you eat come from our germ plasm [4:56] in Sententa. And so, uh, we touch directly right  at the consumer level, not only in the corn and [5:02] soybeans and and rice production, but every deep  roots into agriculture. So what does that all have [5:08] to do with AI at an AI conference is the fun part,  right? Uh most people don't put those two worlds [5:14] together, tech stack and technologies. One of the  key things, like I said, we've been doing this [5:19] for quite some time, especially in the labs and  our R&Ds and we have machine learning, computer [5:24] vision, whole genome prediction models. You know,  AI in the traditional sense is nothing new for us. [5:31] the generative AI and the agentic world that many  are talking about today is the new technologies [5:37] and we're really starting to connect uh across our  ecosystem from lab to customer in the field to our [5:44] supply chains that ship and move seeds and crop  protection products around the world and I would [5:49] argue this is really relevant for any business  you're in some of the bigger unlocks that we see [5:54] from IBIDA topline bottom line is moving across  your domains what I mean by domain is production [6:01] and supply R&D and your customers. We've been so  siloed in solving, hey, how do I make a sales call [6:09] much better through Salesforce or that nature? How  do I do demand forecasting better for inventory [6:14] management or working capital? The reality is  and especially now with the Gentic uh workflows, [6:20] we can connect these across these domains and  actually unlock more IBIDA bottomline margins [6:26] versus the single domains. And that's really  what we see is looking at in our case our domain [6:31] is agriculture and the subdomain is the different  pillars of PNS, R&D and commercial and then how do [6:37] we move agentic AI and generative AI into those  components and so we'll talk about that. So one [6:44] of the few things that we've always known in our  world we do thousands of molecule discoveries and [6:51] not very but very few of those make the field or  the final jug that we sell to a customer. could [6:56] be 10,000 molecules that get evaluated for disease  control, but only one makes it to discovery or to [7:02] production. Very similar in the pharma industry  as well. Genetics is no different. We'll have [7:07] multiple crosses. We have hard selection criteria  for advancing those germ plasms, whether they have [7:13] the right disease tolerance, brittle, disease  resistance, uh drought tolerance. So, you have [7:18] a very strict selection criteria to move the next  thing forward. And what we've done is taken some [7:24] of that to our AI initiatives as well. What is  our selection criteria to move that pilot from [7:29] phase one to phase two. It's not just usability  and and user adoption. It's did it provide value [7:35] to the business. And so you can see a funnel  there on the right that try to start to walk [7:39] through that. But the reality is is you really got  to think through your selection process. And this [7:45] is I think challenging for a lot of us. Um, in our  own industry, we always get in this debate that [7:51] digital is nimble and moving much faster than our  traditional R&D organization that takes 8 to 10 [7:56] years to develop a product. The reality is they do  the same thing. It's who they do it with. Our R&D [8:02] organization does it behind kind of the firewall  and not until it's a final product does it go to [8:08] the customer. In our digital world, we work with  the customer day one and we have to fail in front [8:13] of them and that doesn't always go over well.  whether that's my internal supply chain planners [8:18] or a true farmer that I'm working with. And so  we have to think about that on your selection [8:22] criteria. So what we did is we started out  probably like many of you or if you're a vendor [8:27] working for somebody um the thousand flower booms  when uh chatbt came out what uh two and a half [8:34] three years ago now uh we had a hackathon like  two months after it launched to really have the [8:40] team see what was possible with this technology  and we started really what we call the thousand [8:45] flowers boom methodology. we let everybody kind  of go after ideas and and really going and then [8:50] we had tech stack proliferation. Everybody  was using different models, different LLMs, [8:56] uh different code bases. Um and became chaos  but a great learning experience. I don't think [9:03] our organization would be where it's at from a  learning curve without going through this step. [9:08] But the reality is when you only have a mount x  amount of capital that you have to invest some [9:15] flowers you need to water and some you need  to let die and that's really what we did. So [9:19] going through we put in hard kind of requirements.  What's your IBIDA target? So we gave a hard IBIDA [9:26] target to all the business units to meet through  their AI initiative. If you couldn't get to that [9:32] from a business case standpoint you were basically  cut and you didn't get funding. Did you have a [9:37] business stakeholder combined with the CTO? So in  our world, we have regional business structures, [9:44] North America crop protection and and IT  structure as well. So I'm using the terms, [9:48] the titles figuratively, not literally, but  basically we had the CEO and the CTO go together [9:54] once they had the business case. Actually sent  them to MIT for a week-long training course on AI [10:00] and business to both speak the same language.  If you've ever been around it and business, [10:06] they don't always speak the same language. Uh,  and we needed to get those two people on the [10:11] line together. So, we put those two criteria  together as well. Who holds the business line [10:17] profitability? And that's your business leader.  Don't make the AI pilot an IT initiative. Um, [10:24] it generally comes across poorly. Uh, and actually  change management is tough. And so if you see [10:31] these uh problems on the right uh you need to kind  of red up the red flag uh and really move forward. [10:38] And so with that then we built out and and landed  on seven lighthouse initiatives we called them [10:45] um to kind of show the organization the way we're  around 50,000 employees. So we didn't think we [10:51] could sheep dip the whole organization in one one  go. So that's why we chose very specific targets [10:58] on P&L impact and then ultimately resulting  in these seven lighouses. I lead one of these [11:04] lighouses in the vegetable business uh that you  see on the screen and we called it Aiden. Uh [11:10] everybody got into naming. I think we all thought  we were marketers for a second. Uh but anyway, [11:15] we called it Aiden and what it is is an end toend  supply chain planning uh workflow or overhaul. [11:22] One of the things I would tell you is we had  the IBIDA targets, we had the business owner, [11:26] the CTO all coming together, but the change  management, one of the things I see is actually [11:31] senior leadership is very optimistic on AI, middle  management is really struggling with it. Uh, [11:38] and so that change management becomes a really  big brick wall you kind of got to push through. [11:44] And so in this one of the things we did really  quickly is what's the beach head that we can [11:49] prove that we can work with AI and prove to the  organization that there's value here instead of [11:54] saying we're going to go into. So we went straight  into demand confirmation. So in our business we [11:59] have orders coming in from sales team. We have  inventory. Our inventory is a live product. It's [12:04] not a widget. It has biological implications. So  it can fall out of quality. And so we don't we [12:10] have to match that kind of in real time. We had 60  planners around the world working in Excel files [12:16] matching inventory of quality against the orders  coming in. We now put in an agentic workflow and [12:22] an AI uh oversight. Not everything is gener or  large language models. Some of the basic machine [12:28] learning technologies are still hugely valuable  in this. And so don't get you're wrapped around [12:34] the axle on uh large language models and agentics.  Sometimes the basic stuff works really well. But [12:40] we took those 60 planners doing two to three weeks  of manual 35,000 lines of Excel files, manually [12:48] looking at and matching down to hours. Now our  planners walk in in the morning, everything is [12:53] done. There's a human in the loop, which is the  planner. They audit the exceptions and improve [12:58] and move on. And so huge time savings. We've also  now been able to reduce our days on inventory. So, [13:05] back to KPIs, we're reducing days on inventory,  and we can track that back to working capital [13:11] that's tied up in inventory now and prove actually  what we're doing from a bottom line standpoint. [13:17] We're also giving the sales team more confidence  because now they know their orders are fulfilled, [13:22] fulfilled. They're not uh ordering more than  they're actually selling. Every salesman out [13:28] there has a way to play the game uh and because  they're more confident in that. And so it's a [13:34] real world example of where you can take AI  and the Gentic works and start automating and [13:40] bringing true value to your business and track  through KPIs, not just I wrote my emails faster, [13:46] I summarize or document that better. And that's  only just general productivity gains and it's [13:51] really hard to put into your KPIs or into your  P&L uh as we go forward. So, I'd encourage you [13:57] to look at those um uh different ways you  score and metric uh the system. Again, I'm [14:05] not saying don't do a thousand flowers blooming to  understand. It depends on where you're at in your [14:10] journey. If you need to get your organization to  educate and understand how to do this stuff, yes, [14:16] just realize you have eventually got to only  water a few of those plants that are actually [14:22] going to deliver business value to to you. whether  that's bottom line, top line, more sales growth, [14:27] whatever that metric it is for you. Um, the other  thing I would tell you is is is picking on my [14:33] supply planners is what we did. Don't go after it  just to automate and potentially do a headcount [14:39] reduction. I think a lot of business plans are put  on at the high level headcount reduction savings [14:45] because we're going to automate this workflow.  What we've seen is that's not going to be reality. [14:50] the human in the loop is still too powerful at  this point and too impactful to make sure that [14:56] this is needed. I do see for in the future that we  will probably have uh a variation that comes along [15:02] because again our demand planning is a beach head.  We're now moving into demand forecasting whereas [15:07] allocation and inventory replenishment and we  expect kind of many agents will handle those [15:13] independently and a master human in a sense will  oversee those agents and make sure that they're [15:18] operating in conjunction. I I've seen heard some  talks you know the agents go through much like a [15:23] a teenager adult you know professional they grow  up in learning we kind of see a similar approach [15:29] I think more in the hierarchy we will have basic  agents doing basic tasks but the human will be [15:35] the oversight or the PhD doctor making sure things  are happening correctly uh they can still a little [15:41] bit go off they're on their own scalability is  a key piece uh too many times in the PC's you'll [15:48] you'll create a mini data set in S3 or whatever  you've got and then when you go to production it [15:54] fails because you've hooked it up to your live  data mesh or your uh transactional S3 bucket [16:00] maybe or transactional ERP data set depending on  what you're solving and the system crashes. This [16:06] is another thing your teams have to have clear  road maps of what is your plan to get to scale. [16:12] uh we've got a data mesh data bricks and it's not  they don't care if it's snowflake or whatever but [16:16] what's your plan to make sure that data is live  and through this uh too many times the vibe coding [16:22] um gets us great demos and great possibilities  but actually doesn't get us to production [16:29] uh when we plug and hook up all the servers and  the APIs uh to this. So really think about that [16:35] and don't be afraid to retire. Uh, I've worked  with a lot of product managers over the year and [16:39] it's sometimes it's really hard to kill your kids.  Um, when you're a product manager and you grow [16:44] something up, sometimes it's hard to see the fault  in the product you've got. And again, that's why [16:50] what is the KPI you're tracking? Is it a vanity  metric that you're looking at of just usabil [16:55] usership or is it truly bottom line that you're  changing within your business? Uh, so again, [17:02] which pilots deserve acreage? Where do you go? I  would encourage you to look for economic levers [17:07] in your business or your client's business. Where  is it you can actually show value? I'm I'm not a [17:14] fan of chasing the pro the the human productivity  initiative, right? Chat GPT cla everybody's going [17:20] to solve that. You as a business owner need to  understand what's actually going to move your [17:25] IBIDA, your P&L or not. And again, some of the  traditional technologies, don't throw them out, [17:32] right? uh we've got some great machine learning uh  mechanistic models for demand forecasting that all [17:38] we really need to do is make sure that there's  an automated flow and it talks to it and this [17:43] is why I say connecting your sales team to your  PNS to your R&D and creating a digital thread [17:52] across that will actually unlock more of your  potential value because now if my sales team [17:58] sees I've got a product coming in we can confirm  the order they have confidence to sale. I see the [18:04] product performance of that tomato. That goes  back to my R&D organization that says, "Hey, [18:09] wait a minute. We have a product performance drop  because a new disease came in, brown rugos. Maybe [18:14] I need to improve that trait performance." R&D  now can immediately put that in my production [18:19] team. I can do a jump from PLC 3 to six, do a  seat increase, and get that to the customer in [18:26] a faster time frame. But if I can't coordinate  that communication between that um you're not [18:32] going to be able to unlock that value and we all  know departments are siloed but data is fluid and [18:38] I think that's a key piece especially the IT  organization is a little bit like Switzerland [18:43] you sit in the middle of all those groups and the  IT organization or digital organization can help [18:49] those groups talk via data. Um they still will  be in their silos. I'm not going to say you're [18:54] going to solve corporate uh silos, but you can  get the data to move to the individuals that [19:00] need it to make the right decisions. But you  got to track it. You got to look at the KPIs. [19:05] Um major what matters uh at the end of the day.  So with that, I'll close and ask any questions.