Cargill, Blue Diamond Tie AI Returns to Workflows Over Models

Cargill 与 Blue Diamond：AI 回报取决于工作流而非模型 企业 AI 部署失败多因组织与流程阻力而非技术限制，促使企业以 Token 遥测和工作流集成取代开空白支票式的盲目实验。 企业领导者认为，要实现 AI 的投资回报，必须摆脱僵化短期的财务核算，转向涵盖员工熟练度、12 个月运营改善以及多年专有数据壁垒的分层组合框架。在医疗诊断、农业制造及企业数据平台等领域，演讲嘉宾均强调，AI 部署极少因技术本身而失败，往往源于流程整合欠缺、第三方工具调用等隐性成本以及基础设施不足。表现优异的企业通过建立精细化的支出遥测、设定温和的模型路由护栏，并以自下而上、同行带头的工作研讨取代自上而下的硬性指令，从而弥合生产力差距，推动自发性落地采纳。 企业应采用多层次的风投框架来评估 AI 投资，涵盖员工回报率（Return on Employee）、12 个月内的业务流程改进以及长期的战略护城河。 (5:11) 沟通 AI ROI 需要因人而异：向一线员工强调个人效能释放，向高管展示底层财务收益，向董事会阐述长期战略差异化。 (7:30) 软件工程占企业 AI 总支出的 50% 至 80%，且团队间存在 20 到 30 倍的效率差异，是建立初始 ROI 基准最关键的领域。 (15:04) 精确核算 AI ROI 必须计入隐藏的运营成本，包括人类抢救失败自动化流程所耗费的时间，以及单次调用高达 20 至 30 美元的昂贵第三方 API 工具支出。 (19:00) 在强监管的医疗行业中需采取双轨策略：内部自主开发行政管理 Agent，而临床诊断则与第三方供应商合作，以规避成为 FDA 责任供应商的法律合规风险。 (20:45)

Enterprise leaders argue that extracting measurable return on investment from AI requires moving beyond blunt cost cutting to manage a multi-horizon portfolio anchored in proprietary data, spend telemetry, and workflow readiness. Across agriculture, healthcare, and software engineering, panelists emphasize that AI initiatives succeed not through standalone technology deployments, but by redesigning underlying business processes and imposing operational cost controls. A Multi-Tiered Framework for Enterprise AI ROI (5:11) Nithin Ramachandran (Cargill) and Steven Birgfeld (Blue Diamond Growers) advocate evaluating AI investments through a venture-capital lens rather than rigid, short-term accounting metrics. Ramachandran outlines a three-tier framework: leading indicators tracking employee literacy and tool integration ("Return on Employee"), operational metrics delivering cost reductions and productivity gains within twelve months, and strategic moats realized over three to five years ("Return on Future"). Both leaders emphasize that long-term competitive differentiation stems from proprietary corporate data and process architecture rather than foundation models, which rapidly decay into commodity parity. Effective deployment also requires tailoring the AI narrative across corporate tiers. Frontline workers focus on administrative relief that expands personal leverage, C-suite executives demand measurable bottom-line improvements within twelve months, and boards evaluate multi-year risk profiles and durable advantages. Ramachandran contends that enterprises must discard standalone "AI strategies," instead embedding reusable data assets into existing business goals through a "build once, deploy many times" philosophy adapted to regional regulations. Spend Telemetry, Engineering Baselines, and Hidden Denominators (15:04) Jason Cumberland (Revenium) attributes reported AI failures to an absence of measurement baselines. Across Revenium's enterprise clients, software engineering accounts for 50% to 80% of total AI spend, with a 20x to 30x cost variance per pull request or completed feature between high- and low-efficiency teams. Demonstrating ROI requires evaluating unit cost, throughput, and defect rates while planning token consumption alongside traditional staffing cycles. Accurate ROI tracking also requires capturing the full cost denominator, which extends well beyond direct inference fees. Cumberland details four foundational cost pillars: shared IT infrastructure, token consumption, human labor dedicated to supervising or correcting failed agent outputs, and external tool calls, such as financial KYC and identity checks costing $20 to $30 per invocation. To curb waste without stalling engineering velocity, Cumberland recommends gentle governance mechanisms, such as cost-effective default models for general staff and capped "flagship model budgets" for developers that fall back to standard models once exhausted. Operational Realities in Manufacturing and Regulated Healthcare (23:06) Practical deployments illustrate that AI returns diverge sharply depending on operational context. Birgfeld notes that Blue Diamond Growers achieved seven-figure savings on the factory floor through predictive maintenance, variety-specific yield analytics, and algorithmic workforce scheduling. Trend forecasting and automated content generation also reduced new product introduction cycles from eight months to three to four months. In contrast, exploratory pilots in crop forecasting and demand planning yielded negligible gains due to unpredictable weather patterns and market fluctuations. In healthcare, Dr. Reza Forghani (AdventHealth) outlines an operational divide driven by regulatory compliance and clinical risk. Administrative workflows, such as radiologist worklist orchestration and revenue cycle automation, offer straightforward ROI under HIPAA guidelines. Conversely, clinical diagnostic tools face complex FDA hurdles originally created for narrow, single-use algorithms rather than multimodal foundation models. To balance patient safety against regulatory overhead, health systems deploy internal models for administrative tasks but rely on specialized commercial vendors for diagnostic applications to avoid becoming an FDA vendor of record. Addressing Adoption Bottlenecks and Scaling Grassroots Culture (31:46) Reflecting on enterprise deployments, Ramachandran notes that project failures stem entirely from flawed business processes and cultural unreadiness rather than technical limitations. Automating customer service with rigid chatbots often degrades user retention to save minor transactional costs, whereas complex technical and scientific workflows require human-in-the-loop oversight. Corporate literacy must also shift from defensive risk policies—such as warning employees about data leakage—to active enablement that integrates AI directly into everyday workflows. To institutionalize proficiency, Cumberland and Ramachandran prioritize grassroots peer demonstration over static corporate courses. Cumberland highlights weekly, informal "desk-side AI happy hours" where super-users share live prompt techniques and workflows across technical and commercial teams. At enterprise scale, Ramachandran views structured corporate events as initial catalysts intended to spark "organic acceleration"—a tipping point where self-organized working groups and local lunch-and-learns sustain adoption without centralized mandates. Panelists caution that the era of corporate boards granting "blank checks" for AI experimentation is temporary. As advanced model usage expands, organizations must equip frontline managers with granular spend telemetry and robust digital infrastructure to justify token expenditures directly to financial leadership before scrutiny tightens.

在Ai4研讨会上，来自农业、医疗与技术度量领域的企业高管指出，人工智能的投资回报（ROI）不能局限于传统的单一财务公式，而需采用类似风险投资的分层评估视角，将重心转向工作流程重塑、专有数据护城河建设及深度的员工赋能。 风险投资视角的多元ROI衡量框架 (5:11) Cargill副总裁Nithin Ramachandran主张，企业不应依赖僵化的短期会计准则评估AI，而应借鉴风险投资的组合视角，跨越不同时间跨度分层追踪里程碑。第一层是作为前瞻指标的“员工回报”，核心在于评估员工的AI素养、工具渗透率以及日常工作习惯的建立；第二层是12个月内见效的“业务流程改进”，通过营运提效、排班优化和需求预测交付明确的CFO级财务指标；第三层则是着眼于3至5年的“未来回报”，其竞争护城河并非来自同质化的基础模型本身，而是源自企业专有的私有数据集与结构性业务优势。 推进AI项目时，价值叙事需要根据受众灵活调整：面向一线员工，重点在于消除繁琐重复的行政任务以释放个人潜能；面向CEO与CFO，核心落在可在一年内变现的底线收益与流程改进；面向董事会，则侧重于长远的技术差异化壁垒构建与跨期风险平衡。 场景分化：制造落地与临床合规的取舍 (8:26) Blue Diamond前首席信息官Steven Birgfeld阐述了农业协同场景下的分化实践。由于受到不可预测的天气因素干扰，农作物收成预测和客户需求感知未达预期；但工厂制造端取得了显著突破，通过预测性维护、品种出产率分析和工时调度优化，实现了高达“七位数”的成本节省；同时，利用风味趋势感知将加拿大番茄酱味杏仁等新品的上市周期从8个月压缩至3至4个月。 在医疗健康领域，AdventHealth医疗信息官Reza Forghani博士指出了行政与诊断两条截然不同的落地路径。影像智能排班与收入周期管理等非临床后勤场景能迅速展现清晰的量化回报，适合内部部署；而针对急性病变筛查等临床辅助诊断，由于FDA监管审批滞后且单点算法框架难以适应通用基础模型，医疗机构为避免承担合规主体责任和法律风险，更倾向于采购成熟的第三方商业方案。Forghani强调，AI应用绝不能沦为孤立的噱头，必须如同铺设重工业铁路一般，以扎实的底层数字化基础设施和无缝的工作流整合为基石。 软件工程支出管控与隐性成本计量 (15:04) Revenium联合创始人Jason Cumberland指出，在企业整体AI开支中，软件工程通常占据50%至80%，且团队间的代码产出或工单完成成本存在高达20至30倍的效率差距。早期生成式AI项目的无序试用导致缺乏基准度量，使企业难以证明实际成效。组织当前需要建立“Token预测”机制，从单位任务成本、交付吞吐量及缺陷率三项指标精细化评估工程产出。 核算ROI分母时，企业往往遗漏了两大隐性支出：一是人工介入纠正失败智能体流程所耗费的工时，二是自主智能体并发调用外部接口的开销（如金融领域每次身份核验或信贷查询成本可达20至30美元）。为控制开销，Cumberland主张采用“轻度提醒”而非僵硬的管控技术，例如为工程团队设定固定的“旗舰模型预算”（如每日或每周定额），一旦耗尽则自动平滑回退至标准模型，既能促使员工审慎选择算力层级，又避免了因完全阻断工具使用而导致的昂贵人力闲置。 赋能模式转型与组织自发加速机制 (31:46) 针对AI落地面临的摩擦，Ramachandran直言其主导的项目失败全部源于业务流程设计不合理或人员未准备充分，从未因为底层技术本身出现缺陷。例如在客户服务中盲目剔除人工会导致严重的客户抵触，而在复杂材料科学等高度专业且非确定性的领域，必须坚持“人机协同”。同时，企业内部的AI培训必须从以合规审查为主的防御性限制，转向引导员工在日常业务流程中主动应用工具的增长型思维。 推动AI文化扩散不能仅靠自上而下的行政命令。Cumberland倡导推行由一线同事主持、非正式的“工位AI茶歇”（AI happy hours），搭建低门槛的经验共享平台，让自发涌现的超级用户现场演示实操工作流；在大规模全球化企业中，Ramachandran认为真正的变革标志在于触发基层的“自发加速”——即员工开始自发组织跨部门研讨与学习工坊，使技术应用演变为自主驱动的飞轮。 与会者总结指出，企业董事会出于创新焦虑而放任AI开支“开空头支票”的阶段即将过去，财务问责与精细化计量正逐步向一线职能经理下沉。然而，随着多智能体自主调用复杂化、第三方API隐性成本攀升以及全球监管环境的动态变化，企业如何在保持创新灵活性与维持健康的单位经济模型之间取得长远平衡，仍是有待探索的持续课题。

## Background 本内容由 Ai4 于 2026 年 8 月 22 日发布在 YouTube 平台。页面简介显示，本次圆桌汇聚了来自 Blue Diamond Growers、Principal Venture Partners、Revenium、AdventHealth 以及 Cargill 的多位高管，共同探讨如何跨越不同业务场景与行业壁垒，实现可盈利、可衡量的实际 AI 投资回报（ROI），并分享成功案例、踩坑教训与规模化落地的核心法则。 ## TL;DR 圆桌探讨了企业如何跳出传统软件成本核算的思维局限，以类似风投的组合逻辑与“业务战略牵引”来重新定义 AI 的投资回报率（ROI）。来自农业、医疗和食品等行业的领袖分享了具体实践：从制造端精准预测带来的七位数降本，到医疗合规高墙下的审慎基建，再到将研发交付单位细化以治理 20 至 30 倍效能差距的实操方案。与会嘉宾一致强调，AI 项目失败的核心瓶颈往往在于人与流程而非算法本身，企业需要算清包含外部调用与人工兜底在内的隐藏成本，并通过预算问责与“AI 欢乐时光”等自下而上的文化机制，推动技术从概念验证迈向长期的业务护城河。 ## Atmosphere & style 整场研讨气氛务实严谨且充满实战洞见，彻底褪去了大模型初期的浮躁狂欢，转向冷峻的财务审视与深度的管理学剖析。四位来自不同领域的企业掌舵者在主持人穿针引线下侃侃而谈，既有医疗高管直面监管与高昂成本的清醒坦诚，也有大型跨国集团高管分享“实战伤疤”时的真知灼见。现场对话节奏紧凑、逻辑严密，充满大量极具实操价值的工业指标与度量方法，让听众深刻体会到企业 AI 从野蛮生长走向精细化运营与文化深层重塑的时代脉动。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=OG7QR4UETX0&t=4s) 传统软件 ROI 公式不再适用，AI 的真正价值不仅在于降本增效，更在于提升决策质量与赋予企业前所未有的新型能力，必须依不同时间跨度重构评估模型。 - [05:11](https://www.youtube.com/watch?v=OG7QR4UETX0&t=311s) 企业评估 AI 回报应借鉴 VC 投资组合逻辑，划分三层模型：领先指标（员工使用率与 AI 素养普及）、业务流程指标（硬性降本增效）以及基于专有数据的未来差异化回报（3-5 年战略护城河）。 - [07:31](https://www.youtube.com/watch?v=OG7QR4UETX0&t=451s) 推动 AI 需对不同利益群体定制叙事：基层员工强调消除杂务与能力扩张，高管关注 12 个月内确定的财务成效，董事会则聚焦战略壁垒与风险控制。 - [11:05](https://www.youtube.com/watch?v=OG7QR4UETX0&t=665s) 底层数字基建如同“修筑铁路”，缺乏稳健的网络、数据与算力底座，再先进的 AI 也是空中楼阁；机构必须在短期速赢与长期基建间维持动态平衡。 - [15:04](https://www.youtube.com/watch?v=OG7QR4UETX0&t=904s) 工程研发部门通常占据企业 AI 总支出的 50% 至 80%，且高效与低效团队间交付成本差距可达 20 至 30 倍，是企业精细化审计和降本增效的首要突破口。 - [16:56](https://www.youtube.com/watch?v=OG7QR4UETX0&t=1016s) “95% 的 AI 项目未见回报”源于企业在狂热期盲目跟风且缺乏前置监控；当前预算编制正从“人头预算（Headcount）”加速向“Token 消耗预测（Token Forecasting）”范式迁移。 - [19:00](https://www.youtube.com/watch?v=OG7QR4UETX0&t=1140s) 核算 AI 成本分母必须警惕四大隐形成本：底层基建分摊、实时 Token 消耗、人工抢救失败工作流的时间成本（Human Rescue Time）以及外部第三方 API 的高频并发调用费用。 - [20:45](https://www.youtube.com/watch?v=OG7QR4UETX0&t=1245s) 医疗核心临床诊断面临极其滞后且严苛的监管壁垒，这促使医疗机构将自研 AI 严格限定在行政运营领域，对核心诊断则更倾向采购获 FDA 突破性认证的成熟方案。 - [23:08](https://www.youtube.com/watch?v=OG7QR4UETX0&t=1388s) 蓝钻坚果通过制造端产出率与排班优化实现了七位数美元的硬性净节省，新品上市周期从 8 个月压缩至 3 个月，并将“全量准时交付率（OTIF）”作为加码或止损的硬指标。 - [28:34](https://www.youtube.com/watch?v=OG7QR4UETX0&t=1714s) 跨国多区域部署的核心命题是“不存在独立的 AI 战略，只有业务战略”，跨国企业必须提炼各区竞争优势交集，坚持“一次构建，多地复用”，并深度借力本土生态伙伴破除区域壁垒。 - [31:46](https://www.youtube.com/watch?v=OG7QR4UETX0&t=1906s) AI 项目落地失败几乎全部源于人员认知或业务流程准备不足，而非技术本身缺陷；绝不能为了节省微小成本而完全剥离人机协同（Human-in-the-loop），进而损害核心客户体验。 - [35:49](https://www.youtube.com/watch?v=OG7QR4UETX0&t=2149s) 治理研发效能差距应采取“温和推动”（Gentle Nudges）配合旗舰模型配额制，而非死板的限制规则，促使工程师建立模型成本意识的同时避免因停工等待产生更大隐形成本。 - [39:45](https://www.youtube.com/watch?v=OG7QR4UETX0&t=2385s) 企业 AI 预算管理必然经历四阶段演进：推动采纳、董事会近乎“空白支票”的保护期、用量激增后的成本警觉，以及最终基于全量遥测数据的精细化问责体系。 - [42:46](https://www.youtube.com/watch?v=OG7QR4UETX0&t=2566s) 大型跨国企业推进 AI 的本质是组织文化重塑，从自上而下的引导式演示与跨界输入，最终走向自发涌现；“内部自发组织的 AI 活动数量”是检验文化转型是否达标的终极指标。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=OG7QR4UETX0&t=4s) 圆桌开场与嘉宾背景介绍 圆桌主持人 Songyee Yoon 开启了当天的首场讨论，主题聚焦于企业界最关心的议题之一：AI 与投资回报率（ROI）。主持人指出，传统讨论 AI ROI 时，人们往往局限于效率提升（Productivity Improvement）或成本削减（Cost Cutting），但 AI 真正的价值远不止于此——它还能显著提高决策质量，并赋予企业以前无法实现的新型能力。然而，这些收益的呈现往往具有不同的时间跨度，若沿用传统的软件 ROI 计算公式，很容易导致评估失效和认知混乱。为此，本次圆桌邀请了来自不同行业的四位资深企业高管，共同探讨如何有效衡量 AI 投资并将其融入企业的顶层战略规划。 ```mermaid flowchart TD accTitle: 嘉宾阵容与各自核心AI投资方向 subgraph Panelists["圆桌嘉宾阵容与核心聚焦"] direction TB JC["Jason Cumberland<br/>(Revenium 联合创始人)"] --> JC_F["专注衡量企业内部 AI 使用的有效性与效率"] RF["Reza Forghani<br/>(AdventHealth 首席医疗信息官)"] --> RF_F["医疗基础设施建设、危重症优先标记及影像排程编排"] SB["Steven Birgfeld<br/>(Blue Diamond Growers 前 CIO)"] --> SB_F["围绕 SAP 生态构建自主环境，最大化种植者收益"] NR["Nithin Ramachandran<br/>(Cargill 资深副总裁)"] --> NR_F["投资企业级底层能力：人员素养、数据可靠性与流程定义"] end ``` 在自我介绍与年度最大 AI 投资的提问环节中，四位嘉宾分别阐述了各自的定位与发力点： - **Jason Cumberland**（Revenium 联合创始人兼 CPO）：Revenium 的业务定位较为独特，并不直接做业务部署，而是帮助企业精准衡量 AI 使用的功效与效率。就在参会前夕，《纽约时报》（New York Times）记者就 AI 对宏观经济的影响对他进行了专访，相关报道也刚好在会议开始前的周一刊发。 - **Reza Forghani**（AdventHealth 影像首席医疗信息官 CMIO）：医疗健康领域的复杂性极高，其最大投资在于底层基础设施建设，确保数字基建既保持前沿又能支撑未来的 AI 需求。在影像数字化领域，重点项目包括利用 AI 自动标记并优先处理危及生命的急重症扫描（充当“第二读片人”以防漏诊），以及面向放射科医生的智能工作列表编排（Worklist Orchestration）。 - **Steven Birgfeld**（Blue Diamond Growers 前 CIO / 顾问）：Blue Diamond 是全球最大的杏仁加工商。虽然他在过去八周刚转为退休顾问状态，但企业今年的核心依然是深化并拓展已有的丰富 ROI 用例，重点依托 SAP 生态系统打造自主环境（Autonomous Environment），并结合新兴 AI 工具箱为杏仁种植者（Growers）创造高价值回报。 - **Nithin Ramachandran**（Cargill 数据与 AI 资深副总裁）：Cargill 是全球最大的私营食品公司，此前他曾先后统领 3M 和可口可乐（Coca-Cola）的全球数据与 AI 业务。他指出，由于 AI 领域变化极快，昨天的“竞争优势”（Competitive Advantage）很快就会沦为“竞争均势”（Competitive Parity），因此嘉吉最大的投资并非押注在单一用例上，而是扎根于**企业级通用能力**（Enterprise Capabilities）——包括员工 AI 技能、数据可用性与可靠性、以及业务流程的清晰数字化定义，以此确保未来所有 AI 用例具备极高的可复用性。 --- ### [05:11](https://www.youtube.com/watch?v=OG7QR4UETX0&t=311s) 重新定义 ROI：风险投资视角下的三层评估模型 在探讨如何具体测算 ROI 与制定投资决策时，Nithin Ramachandran 提出，企业不能仅仅戴着首席财务官（CFO）的传统公式眼镜来看待 AI，而应当借鉴**风险投资机构（VC）的投资组合逻辑**（Portfolio Level Outcome）。 Nithin 将 AI 的投资回报划分为三个循序渐进的评估阶段： ```mermaid flowchart LR accTitle: 类似VC逻辑的AI三层ROI评估体系 subgraph L1["第一层：领先指标 (Leading Indicators)"] L1_A["员工回报 (ROI on Employee)"] L1_B["AI 使用率与素养普及"] L1_C["业务流程接入度"] end subgraph L2["第二层：业务流程改善 (Process Metrics)"] L2_A["CFO 关注的硬性指标"] L2_B["降本 / 增效 / 劳动力优化"] L2_C["需求预测等标准化用例"] end subgraph L3["第三层：未来差异化回报 (Return on Future)"] L3_A["3 至 5 年长期壁垒"] L3_B["专有数据 (Proprietary Data)"] L3_C["企业核心护城河"] end L1 --> L2 --> L3 ``` 1. **领先指标（Leading Indicators / 员工回报）**：类似于早期社交媒体公司关注网络效应，第一阶段重点考察员工对 AI 的使用率、AI 素养以及业务流程的覆盖率。这是检验企业是否真正迈向“AI First”的先导信号。 2. **业务流程指标（Business Process Improvements）**：进入 CFO 真正关注的硬核财务指标，包括工作效率提升、劳动力优化、成本削减、收入增长以及需求预测等标准用例改善。 3. **未来差异化回报（Return on Future）**：关注 3 到 5 年后的长期战略回报。此类投资不单纯依赖 AI 技术本身，而是依托企业的**专有数据（Proprietary Data）**或独特业务壁垒构建不可复制的核心竞争力。 --- ### [07:31](https://www.youtube.com/watch?v=OG7QR4UETX0&t=451s) 针对不同利益相关者的多维度沟通策略 Nithin 进一步分析了在企业内部推行 AI 时，面对不同群体需要采取截然不同的叙事逻辑： | 利益相关者群体 | 沟通核心诉求 | 汇报侧重点 | |:--- |:--- |:--- | | **基层员工 (Employees)** | 赋能与减负 | 消除繁琐重复的日常杂务（Mundane Work），扩大个人的影响力范围（Span of Influence）。 | | **高管团队 (CEO / CFO)** | 短期成效与确定性 | 聚焦 12 个月内即可见效的流程改进、效率提升与成本压缩（尤其在上市公司面临业绩压力时）。 | | **董事会 (Board of Directors)** | 战略壁垒与风险控制 | 聚焦企业的长期差异化竞争优势，以及在规模化扩张过程中的风险可控性（Risk-conscious Scaling）。 | --- ### [08:27](https://www.youtube.com/watch?v=OG7QR4UETX0&t=507s) 农业与医疗健康领域的 ROI 落地实践 Blue Diamond Growers 的 Steven Birgfeld 与 AdventHealth 的 Reza Forghani 结合各自行业分享了具体的落地考量： - **农业加工与供应链优化（Blue Diamond Growers）**： Steven 强调，Blue Diamond 的核心使命是最大化种植者的回报。在与管理层和董事会沟通时，他们极力避免完全受困于短期 ROI，而是注重基础战略与架构的搭建。在具体指标上，除了通用的生产力提升外，重点关注行业核心 KPI——如**准时全额交付率（OTIF: On-Time In-Full）**与**杏仁出产良率（Yield）**。在员工层面，团队虽然对技术带来的潜在岗位变动抱有一丝担忧，但实际落地证明 AI 并非取代人力，而是为基层员工和高管层提供更优质的洞察以辅助业务决策，同时创造了大量的成本节约空间。 - **医疗健康与临床/运营双轨制（AdventHealth）**： Reza Forghani 指出，医疗领域的 ROI 必须拆分为两大维度来审视： 1. **临床与诊断支持（Diagnostic Clinical Support）**：受到极其严苛的监管要求约束。其 KPI 具有多因果性（Multifactorial），核心围绕患者安全、医疗质量、前瞻性预测治疗（Predictive Therapeutics）以及提升及时诊断效率展开，直接契合以消费者为中心的高质量医疗使命。 2. **运营与行政管理（Operational & Administrative）**：例如加速影像扫描获取时间（Scan Acquisition），此类用例具有更直观清晰的量化 ROI；此外还涵盖营收周期管理（Revenue Cycle Management）及纯行政流程的自动化改造。 ### [11:05](https://www.youtube.com/watch?v=OG7QR4UETX0&t=665s) 医疗场景下的 ROI 平衡：短期、长期与底层数字基建 在探讨医疗体系如何评估投资回报率（ROI）时，AdventHealth 首席医疗信息官 Reza Forghani 指出，关键绩效指标（KPI）应当明确划分为短期、中期与长期三种维度。在医疗领域，引入某项技术的初衷往往是为了改善患者体验或提高医疗安全性，毕竟医生的天职是救治患者；但与此同时，决策者必须保持清醒的财务意识——医疗服务极其昂贵，如果技术方案无法在财务上形成可持续闭环，医院最终将难以为继。特别是随着医疗信息系统（Informatics）和软件采购成本节节攀升，管理层必须直面尖锐的现实问题：“这笔费用最终由谁来买单？” 面对主持人关于“如何在短期高 ROI 项目与长远战略项目之间权衡取舍”的提问，Reza 分享了自己的决策哲学：必须在组合拳中保持动态平衡，既要有能快速见效、路径明确的务实项目，也要有立足长远定位的战略布局。他坦言自己虽然减少了临床执业时间，骨子里依然是一名一心为患者着想的医生，但医院高管必须严密把控资金的可负担性。在医疗 AI 的实际落地中，区分“吸睛小玩意（gadget）”与“真正可用的产品（product）”至关重要。临床一线的工作流融合度是决定成败的核心——一项技术即便看起来再令人兴奋，如果不够成熟、无法无缝融入医护人员的日常工作，引入后只会带来巨大的失望。 此外，Reza 将数字底层基础设施比作“修筑铁路”。他直言，许多医疗机构的 IT 部门依然困在过去的陈旧模式中：预算不足、被边缘化，往往在最后关头才被当作“事后补充”敷衍对待。如果缺乏健壮扎实的基础设施作为底座，所有令人心动的 AI 前沿构想都只是空中楼阁；若想在未来的数字化转型中保持长期竞争力，医疗机构哪怕面临成本压力，也必须优先啃下底层 IT 基建这块硬骨头。 - **多维 KPI 设定**：将临床指标（患者安全、就医体验）与财务指标（软件采购成本、支付方结算模式）深度挂钩。 - **跨职能落地团队**：融合临床一线最终用户、IT 技术骨干与财务管理人员，避免采购脱离实际工作流的技术噱头。 - **基建先行的“铁路理论”**：没有健壮的底层网络、数据通道与计算架构，再先进的 AI 算法也无法稳定运行。 --- ### [15:04](https://www.youtube.com/watch?v=OG7QR4UETX0&t=904s) AI 支出追踪的突破口：从工程研发与效能差异入手 针对许多企业尚未建立 ROI 评估机制、不知从何下手的困境，Revenium 首席产品官兼联合创始人 Jason Cumberland 提供了一套高度实操的破局路线图。他建议企业的当务之急是全面穿透账单与发票，搞清楚究竟是哪些业务部门正在推高 AI 支出。去给一个开销微不足道的边缘部门证明 ROI 毫无意义，即便项目顺利跑通也不会引起任何人的重视。企业必须按业务单元或团队将 AI 账单彻底拆解；如果当下还做不到精细化分摊，那么问题大概率潜藏在工程研发（Engineering）部门。 在 Jason 接触的所有企业客户中，工程研发部门普遍占据了 AI 总支出的 50% 到 80%。幸运的是，研发团队也是最容易套用标准化 ROI 模型的场景。软件工程的工作单元高度标准化且文档健全，无论是 Pull Request（PR）、交付的新功能（Feature），还是工单（Ticket），都拥有清晰的量化抓手。尽管工程师们常常争论任务复杂度的差异，但在庞大的研发团队样本下，这些差异最终都会被大数定律拉平。 更令人震惊的是团队之间的效能断层。Jason 透露，当他们深入客户现场、按“单次 PR 或功能聚合交付成本”对工程师团队的产出效率进行测算时，发现最高效团队与最低效团队之间的成本差距通常高达 **20 到 30 倍**。这意味着，管理层只要摸清各团队的真实行为模式，针对低效团队推行针对性的改进方案，就能在完全不打击创新活力的前提下大幅削减 AI 成本。当前很多企业都在屏息观望、祈祷支出不会失控，但当拐点来临、CFO 突然打电话问责时，如果没有前置的计量数据，团队根本无从应对。 ```mermaid flowchart TD accTitle: Engineering AI Cost Audit Flow accDescription: 识别高额 AI 支出、聚焦工程研发、量化交付单元并治理效能差异的落地流程 subgraph Spend["第一步：成本排查与部门穿透"] A["排查企业整体账单与发票"] --> B{"能否按业务单元拆分？"} B -- 否 / 缺乏细分 --> C["锁定工程研发部门<br/>(占总支出 50%~80%)"] B -- 是 --> D["找出 AI 消耗最高的业务单元"] end subgraph Metrics["第二步：量化研发交付单元"] C --> E["确定标准化交付计量单位<br/>(PR / Feature / Ticket)"] D --> E E --> F["对齐大规模工程样本<br/>抹平个体任务复杂度差异"] end subgraph Action["第三步：效能评估与成本控制"] F --> G["对比各团队单位交付成本<br/>(效率差距常达 20~30 倍)"] G --> H["重塑低效团队使用模式"] H --> I["在不影响创新的前提下有效降本"] end ``` --- ### [16:56](https://www.youtube.com/watch?v=OG7QR4UETX0&t=1016s) 拆解“95% AI 项目失败论”与 Token 预测新范式 针对主持人提到去年广泛流传的一篇 MIT 研究报告——声称“95% 的企业 AI 项目未能产生任何实质性 ROI”，Jason 给出了深层拆解。他认为这个耸人听闻的数字恰恰印证了缺乏度量带来的后遗症：去年是整个行业盲目跟风的“狂欢期（free-for-all）”，所有人都在不计代价地把 AI 往各种业务场景里塞，过程中没有任何前置的指标追踪。既然企业连基准数据和成效都无法度量，在财务账面上自然显得“全盘皆输”。 Jason 强调，企业证明 AI 的 ROI 必须回归最经典的工业与管理学三要素：**单位成本（Unit Cost）**、**吞吐量（Throughput，即单位时间内的产出数量）**与**质量（Quality）**。这三项核心指标本身并不新鲜，关键在于绝大多数企业在启动项目时根本没有埋设任何监控机制，导致项目收尾时大家只能陷入空洞的争论。如今成熟的企业不仅围绕 AI 使用建立了更规范的流程，还配备了度量监控系统，在决定技术选型时对开销预期已经心中有数。 更深刻的范式转移正在企业的组织预算编制中发生。过去，企业立项上线新功能时，通常需要规划人员编制（Headcount），经历长达两到三个月的招聘入职周期后才能逐步交付；而如今，AI 让项目启动速度大幅提升，**Token 消耗预测（Token Forecasting）**正在迅速取代单纯的人头规划，成为企业核心管理层必须掌握的专业硬实力。如果企业连今天的实际消耗都算不明白，就根本不可能预测明天上线新功能究竟要烧掉多少 Token。 - **ROI 证明的三大支柱**： - **单位成本**：完成单项业务任务的综合资源开销。 - **吞吐量**：固定周期内系统与团队所能处理并交付的任务容量。 - **交付质量**：产出结果的准确率、稳定性与业务可用度。 - **编制规划范式跃迁**：从以“人头预算与招聘周期（2-3个月）”为核心，转变为以“算力与 Token 消耗预测模型”为核心。 --- ### [19:00](https://www.youtube.com/watch?v=OG7QR4UETX0&t=1140s) 算清 AI 的“分母”：隐藏成本的四大核心维度 随着对话深入，主持人指出了一个普遍的财务困境：在评估 ROI 的“分母”（即投入的总成本）时，底层的 IT 基建到底该按多大比例分摊到具体项目中？在技术日新月异、工具频繁淘汰的当下，这种基础设施到底能服役多久？面对这种巨大的不确定性，企业往往算不清真实的成本账。 Jason 剖析指出，企业部署 AI 时如果想算清真实的投产比，必须全面监控并捕获四大隐藏成本维度，否则只会在虚假的高收益幻象中自欺欺人： 1. **底层基础设施（Infrastructure）**：承载业务流转所需的算力环境、网络底座与系统改造费用。 2. **模型 Token 消耗（Tokens）**：模型推理过程中直接产生的上下文吞吐费用。 3. **人类干预与“救援”时间（Human Rescue Time）**：这是最容易被忽视的隐形成本。部署过 AI Agent 的团队都知道，从第一次上线部署到真正能宣布成功，中间横亘着极其漫长、痛苦的人工干预链条。员工在系统报错、幻觉发作或流程卡死时去“抢救（rescuing）”失败工作流所花费的时间，本质上都是高昂的人力成本支出。 4. **昂贵的第三方工具外部调用（External Tool Invocations）**：在金融服务等严监管领域表现尤为突出。AI 智能体在执行任务时往往需要与外部系统交互，例如发起一次 KYC（了解你的客户）身份核验、调取个人信用报告或外部反洗钱数据，单次 API 调用的费用可能就高达 20 到 30 美元。 在传统的确定性软件工程时代，脚本每次执行都精准可控，调用次数被严格锁定在代码逻辑内；但在由大模型驱动的自主 Agent 世界中，10 个并发智能体可能会像扇形一样散开（fan out），为了完成同一个模糊任务各自进行多轮重复调用。到了月底，财务团队不仅会收到五家不同服务商寄来的天价调用账单，还会收到独立的 AI 大模型算力账单，而这两者之间由于缺乏底层追踪埋点（Instrumentation），根本无法相互关联归因，最终在企业内部留下巨大的财务黑洞。 | 成本维度 | 传统软件开发模式 | 现代 AI / Agent 部署模式 | 潜在财务盲区与风险 | |:--- |:--- |:--- |:--- | | **外部工具调用 (Tool Calls)** | 脚本单次线性执行，费用严格确定且可预期 | 多个 Agent 并行“扇出”，自主发起多轮外部 API 请求 | 单次 KYC/征信查询高达 20~30 美元，多智能体并发调用导致账单呈指数级膨胀 | | **人力时间投入 (Human Time)** | 固定的开发与运维排期，按人天预算结算 | 耗费大量高薪技术人员介入，频繁抢救执行失败的 AI 工作流 | 忽略了“人工修障与兜底时间”，虚增项目表面上的 ROI 效益 | | **模型推理算力 (Tokens)** | 无此项，仅按服务器实例资源计费 | 依赖预先建立的 Token 预测模型与实时消耗度量 | 缺乏实时监控，难以预测功能扩展带来的实际边际成本 | | **基础设施分摊 (Infrastructure)** | 机器设备折旧周期长且高度标准化 | 技术选型迭代极快，底层资产面临早期淘汰风险 | 难以公摊通用基建成本，各项目分母核算标准混乱 | ### [20:45](https://www.youtube.com/watch?v=OG7QR4UETX0&t=1245s) 医疗AI的双重困境：监管滞后与临床诊断的合规高墙 医疗行业的AI部署面临着其他行业未曾经历的监管重压，这直接重塑了其ROI（投资回报率）的测算逻辑。AdventHealth的首席医疗信息官（CMIO）Reza Forghani指出，技术演进的速度早已大幅甩开现有监管框架，当技术从单一、狭窄的弱AI应用迈向更通用的模型时，这种滞后性尤为致命。过去的医疗监管是基于“单一用途逐项审批”（approved one by one）建立的，根本无法适配能够处理复杂任务的新一代AI。例如在医学影像领域，直接由图像生成诊断报告草稿（draft text）被公认为下一波变革性跃迁（"step change"），但行业目前甚至没有一套现成的监管审批框架来规范这一流程。 面对既要保障患者安全又不能窒息创新的两难处境，医疗机构在实际推进中划出了一条泾渭分明的技术路径分水岭： ```mermaid flowchart TD accTitle: 医疗AI部署的路径分流与合规边界 subgraph S1["行政与运营场景"] A["业务需求"] --> B["企业内部基础模型 / AI Agent"] B --> C["HIPAA合规审查"] C --> D["快速落地与自主部署"] end subgraph S2["临床诊断场景"] E["诊断核心需求"] --> F{"自研还是外购？"} F -->|"自研门槛"| G["须成为登记备案厂商 (Vendor of Record) 获FDA批准"] F -->|"主流路径"| H["采购已获突破性认证的第三方成熟方案"] end ``` 这种合规壁垒决定了医院内部AI研发的边界。如果仅仅是将基础模型用于日常行政提效、流转自动化，只要符合HIPAA合规要求，机构完全可以自主搭建Agent；然而一旦触及核心临床诊断，几乎没有哪家医疗集团愿意将自己变成必须直面FDA极其漫长严苛审查的“登记备案厂商（vendor of record）”。这导致医疗领域的深水区创新目前依然高度依赖外部专业供应商，也延缓了高价值临床AI的变现与规模化复制。 - 监管新风向：年内已有一些AI医疗方案获得了突破性医疗器械认定（breakthrough device designations），为创新商业化破冰带来了曙光。 - 战略权衡：其他行业可以肆意在核心业务尝试全栈自研，但医疗机构受限于FDA认证与合规成本，内部生成式AI更多局限在非临床流程。 --- ### [23:08](https://www.youtube.com/watch?v=OG7QR4UETX0&t=1388s) 蓝钻坚果的AI试验场：制造端斩获七位数节省与番茄酱杏仁传奇 蓝钻坚果（Blue Diamond Growers）的IT与服务副总裁Steven Birgfeld分享了这家拥有百年历史的坚果巨头如何推行极其激进的多线AI试验。面对外界对其涉足AI领域的惊讶，Steven坦言团队最初的策略就是“抛开短期ROI的执念”，先全面探索AI究竟能在哪些业务环节真正产生杠杆效应。在这场广撒网的实验中，不同业务板块的表现呈现出巨大反差： 有些曾被寄予厚望的方向遭遇了现实的冷水。农作物收成预测（crop forecasting）受制于不可预测的“大自然母亲（Mother Nature）”，模型表现很难质变；需求感知（demand sensing）与需求计划在预测客户下单时同样收效甚微，并未真正“触动业务指标的针尖（moved the needle）”。然而，试验在制造与生产端却迎来了爆发式的回报。通过部署预测性维护、产出率管理（yield management）以及品质监控系统，蓝钻能够精准预测不同品种杏仁的产出率及匹配制程。结合排班、合规与加班工时优化的劳动力规划系统，蓝钻坚果在制造端直接拿下了高达“七位数”（百万美元级）的成本缩减。 ```mermaid flowchart LR accTitle: 蓝钻坚果AI多线探索与业务产出对比 subgraph LowROI["试错与微效场景"] A["农作物产量预测"] --- A1["受制于天气自然因素"] B["客户需求感知与计划"] --- B1["预测偏差未能拉动业务"] end subgraph HighROI["高杠杆突破场景"] C["制造端产率与维护"] --> C1["七位数成本削减"] D["劳动力排班与合规"] --> C1 E["市场趋势洞察"] --> E1["加拿大番茄酱风味杏仁爆款"] end ``` 更具戏剧性的是市场营销端的突破。AI通过解析海量消费趋势数据，预测亚洲或加拿大等特定市场的下一轮风味潮流，蓝钻据此直接定义并推出了风靡加拿大的“番茄酱风味杏仁”（ketchup flavored almond）。从新品风味研发到利用AI自动化生成宣传图文内容，营销端不仅省下了外包成本，更为企业抢占货架建立了先发优势。Steven总结了多业务线试验的三大核心支撑： - 业务共创：业务团队必须亲自入场并充当项目的核心拥护者（advocate）。 - 衡量基准：正如Jason所强调的，在启动第一天就必须确立基准测量指标，以便能量化变化。 - 统一本体：蓝钻引入了能自动生成本体论（automated ontology）的数据工具，将散落各处的数据孤岛拉通对齐，解决了多系统数据匹配的底层难题。 --- ### [27:02](https://www.youtube.com/watch?v=OG7QR4UETX0&t=1622s) 从8个月到3个月：以交付与底线回报衡量AI成效 在多学科业务的迭代推进中，蓝钻坚果建立了一套清晰的“加码或止损”评估模型。当主持人问及究竟依赖哪些关键数据信号来决定对某个AI项目持续追加投资还是果断抽身时，Steven明确列出了两大硬性准绳： 首先是履约核心指标“全量准时交付率”（OTIF, On-Time and Full delivery）。这一前台指标是制造业、良率控制、供应链与物流综合效能的集中体现，AI如果不能直接拉升这一指标，其战略价值就会大打折扣；其次则是无可辩驳的底线财务收益（bottom-line numbers），即能否像制造车间那样沉淀出七位数的硬性净节省。 最直观的飞跃体现在新产品的上市周期（time to market）上。在引入这套AI驱动的趋势挖掘与内容生成工具前，蓝钻一款新品从概念到铺货通常需要漫长的8个月左右；而通过AI工具链的赋能，这一周期被极限压缩到了3至4个月，效率提升了一倍以上。这种在开发周期和实际耗费成本上的“双重压缩”，构成了蓝钻坚果继续重仓AI部署的决策底气。 --- ### [28:34](https://www.youtube.com/watch?v=OG7QR4UETX0&t=1714s) 跨国部署的方法论：“不存在AI战略，只有业务战略” 掌舵嘉吉（Cargill）AI与数据团队、并拥有前3M公司丰富跨国落地经验的高级副总裁Nithin Ramachandran，将讨论引向了全球化落地的终极拷问。在单一大区试点AI与推向多国市场完全是两个物种：不同辖区面临迥异的监管法规、各区域的业务本质差异巨大、底层的数字基建水平更是千差万别。 面对这种错综复杂的局面，Nithin抛出了一个提纲挈领的断言：“首先必须从脑海中抹去的概念就是‘AI战略’——根本不存在所谓的AI战略，只有业务战略（There's no AI strategy. There's only a business strategy）。”他尖锐地指出，如果脱离业务去谈AI，技术就会沦为“一只到处找钉子砸的锤子”（a hammer searching for a nail），不仅徒耗资源，更无法实现价值闭环。 针对全球多区域部署，Nithin梳理出了一套系统化的落地执行框架： - 提炼竞争优势交集：首席AI官的第一要务是横跨所有大区，寻找不同市场在竞争差异化上的“共同交集”（例如销售端赋能、供应链需求管理等痛点），以此作为全球创新的聚焦点。 - 全球化视角与复用性：全球化ROI的核心命脉在于“一次构建，多地复用”（build once, deploy many times）。绝不能为每个地区重造轮子，而要在统一范畴内进行标准化创新。 - 拥抱本土生态伙伴：基建与合规的区域孤岛无法单靠总部系统强行覆盖，例如几乎所有跨国企业在中国市场（China）的运营环境都与欧美截然不同。此时必须引入深度理解本土语言、文化与监管格局的生态伙伴，才能在尊重区域独特性的同时达成架构复用。 只要牢牢抓住“业务战略牵引、聚焦共性优势、坚持资产复用、依靠伙伴落地”这条铁律，跨国企业在极其复杂的全球版图上推进AI投资时，就能构建起真正可持续的胜势。 ### [31:46](https://www.youtube.com/watch?v=OG7QR4UETX0&t=1906s) 失败教训与人机协同：AI 部署为何不能单纯取代人类 讨论转向了 AI 落地过程中“远超预期”与“高开低走”的真实案例。Nithin Ramachandran 分享了自己积累的诸多“实战伤疤”（battle scars），并直言不讳地指出：在所有失败的 AI 项目中，**没有一次是因为底层技术本身不过关，失败全部源于人员（People）或流程（Process）层面的准备不足**。 ```mermaid flowchart LR accTitle: AI 项目失败归因与人机协同框架 subgraph S1["失败归因分析"] A["技术能力 (Technology)"] -->|并非主要瓶颈| B["项目成败"] C["人员准备度 (People)"] -->|核心瓶颈| B D["业务流程准备度 (Process)"] -->|核心瓶颈| B end subgraph S2["落地指导原则"] B --> E["客户体验优先<br/>(非单纯省几毛钱)"] B --> F["赋能与扩张员工<br/>(Employee Expansion)"] B --> G["保留 Human-in-the-loop"] end ``` 在业务流程方面，最大的误区之一在于企业盲目上线 AI 并理所当然地认为客户会喜欢。Nithin 坦言自己作为消费者时，极其反感在急需人工客服时被困在与 AI 聊天机器人的死循环（recursive loop）中。企业往往为了节省流程中的“几毛钱”而牺牲了客户体验，但真正的核心资产是客户本身。因此，绝不能在关键节点完全移除人类（Human-in-the-loop）。 在人员与组织文化方面，企业最初的 AI 素养培训（AI literacy）往往偏向消极的风控合规——告诫员工哪些数据不能输入大模型。然而真正的痛点在于，企业必须教会员工如何将 AI 融入到日常工作流中（into the flow of work）。推动 AI 融入业务流程的目的不应是单纯的裁员（workforce reduction），而是“员工能力扩张”（employee expansion），以增长型思维引导员工发掘业务机会。 针对主持人关于“客服领域何时适合完全自动化”的追问，Nithin 区分了两种业务场景： - **确定性/交易型场景**：例如移动银行 App 的基础业务，目标明确且规则清晰，可以通过 Chatbot 实现自动化以减少线下网点访问。 - **非确定性/专业探索场景**：例如 3M 开发的材料科学产品“Ask3M”，涉及深度材料科学的复杂非确定性问题，AI 极难独立闭环，必须依靠专业人员在回路中协同。 --- ### [35:49](https://www.youtube.com/watch?v=OG7QR4UETX0&t=2149s) 用“温和推动”解决 20 倍效能差距与 Token 成本控制 主持人将话题转向 Jason Cumberland，询问此前提到团队之间高达 20 倍的生产力与 ROI 差距从何而来，以及如何在不借助外部力量的情况下识别和衡量这种差异。 Jason 指出，改变工程师行为最有效的方式是**“温和推动”（gentle nudges），而非强制死板的限制规则（strong guard rails）**。当企业毫无指引地直接推行 Claude 或 Codex 等工具时，只有极少数直觉敏锐的工程师能自主做出正确的模型路由（Model Routing）选择，大部分人则缺乏成本意识。缩小这一差距的核心在于对模型调用的精细化分级与控制： - **普通知识工作者（非研发）**：通过工具默认配置进行调整，设定合理的默认模型。顶级旗舰模型仅对特定申请者开放，采取“按需申请、理由合理即批准”（permissive on request）的策略，避免全员无门槛滥用。 - **研发工程师与旗舰模型预算制**：针对自由实验导致的成本失控，Jason 建议为工程师设定固定的“旗舰模型配额”（Flagship Model Budget，例如每天或每周 1,000 美元）。当顶级模型（如 Anthropic 或 OpenAI 旗舰模型）的额度耗尽后，工程师必须切回仍具备极高能力的标准模型继续工作。 Jason 强调：“没有什么比工程师停工等待更昂贵的了，那比所有消耗的 Token 都要贵得多。”固定预算的机制就像一次按键前的温和提醒，促使工程师主动思考当前任务是否真有必要动用旗舰模型，从而在不中断开发进度的前提下大幅降低企业的整体账单。 --- ### [38:23](https://www.youtube.com/watch?v=OG7QR4UETX0&t=2303s) 技能提升的最佳实践：非正式的“AI 欢乐时光”与用量遥测 在探讨如何对员工进行 AI 技能培训与技能重塑（upskilling）时，Jason 破除了寻找“灵丹妙药”的幻想，认为唯一的路径就是直接躬身入局（dive in）。 目前各业务部门内部通常都已涌现出个别能把 AI “玩出魔法”的业务尖子。将这些先进个人组织起来，向同岗位的普通同事展示其具体工作流与提效秘诀，是知识扩散最有效的方法。为此，Revenium 内部推行了 **“AI 欢乐时光”（AI Happy Hours）**： - **组织形式**：并非去酒吧喝酒，而是员工坐在工位上举行的非正式小组研讨会。 - **主持角色**：由业务骨干（Peer）而非管理层（Manager）主持，每周定期交流真实用例与踩坑经验。 在管理层层面，领导者应充分利用各大 LLM 供应商后台提供的用量遥测数据（Telemetry），分析各团队调用频次最高的技能与功能，挖掘新颖的高价值用例，并将其在全公司范围内推广复现。 --- ### [39:45](https://www.youtube.com/watch?v=OG7QR4UETX0&t=2385s) AI 预算演进曲线：从董事会的“空白支票”到管理层问责 Jason 总结了多数企业在 AI 预算管理上所经历的生命周期演进（Continuum）： ```mermaid flowchart LR accTitle: 企业 AI 预算与采用生命周期演进 A["阶段一：全力推动采纳<br/>(Driving Adoption)"] --> B["阶段二：创新保护期<br/>(董事会'空白支票')"] B --> C["阶段三：转折临界点<br/>('天哪，别用那么猛')"] C --> D["阶段四：精细化问责<br/>(基于遥测系统量化 ROI)"] ``` 1. **推动采纳阶段**：鼓励全员探索，消除冷门群体的认知门槛。 2. **创新保护阶段**：许多企业董事会因担心阻碍创新而给予团队近乎“空白支票”（blank check）式的预算支持。 3. **转折点（Transition Point）**：随着员工使用熟练度激增，Token 账单水涨船高，管理层开始惊呼“用得太多了”。 4. **管理问责阶段**：当单人单周产生 10,000 美元调用费时，管理层必须给出明确的产出评估。 为了让各级管理者在面对 CFO 的成本质询时不必临阵编造答案，企业必须建立自动化的遥测与度量系统，将 Token 预算的问责机制常态化下沉至基层主管手中。 --- ### [41:40](https://www.youtube.com/watch?v=OG7QR4UETX0&t=2500s) 现场互动：AI Happy Hour 的推广落地 在进入现场问答环节后，台下观众首先询问了关于“AI 欢乐时光”（AI Happy Hour）的具体执行细节：企业在最初推行这一机制时是如何逐步扩展规模的，员工对此类分享形式的接受度和参与意愿又是如何演变的。 ### [41:51](https://www.youtube.com/watch?v=OG7QR4UETX0&t=2511s) 初创团队的周五 AI 分享会：营造自发驱动的探索氛围 初创公司的 AI 落地与探索往往源自天然的好奇心。在分享内部实践时，Jason Cumberland 提到他们最初只是在工程团队内部发起了一个小型的 AI 经验分享会，但出乎意料的是，全公司上下几乎所有人都主动参与了进来——不仅销售和市场部门积极到场，连 CEO 也亲自列席。为了保证会议效果，组织者通常会在会前邀请一到两位成员准备分享案例。而在实际推进中，收集案例几乎毫不费力，因为在当前的工具迭代速度下，“每周都会有人体验到某种全新的、感觉像魔法一样的技术突破”。组织者只需精心策划两三个开场故事，后续的讨论热度和分享势头就会自然而然地被带动起来。 为了营造最轻松开放的交流环境，这类分享会固定安排在周五上午的中段进行。这一时间点既能保证团队处于一周中最放松的状态，又能跨越时区——美国团队喝着早间咖啡，欧洲团队则正好在下午端着啤酒上线交流。实践证明，只要给员工提供一个展示和交流的舞台，大家对于分享自己用 AI 做出的亮眼成果往往抱有极高的热情，这种自下而上的分享机制很快便能在组织内部生根发芽。 - **跨职能参与**：原本仅面向工程师的分享会自发扩展至销售、市场及最高管理层（CEO）。 - **轻量化策展**：每次会议预先安排 1–2 个“魔法时刻”（Magic Moments）案例抛砖引玉，靠团队惯性驱动后续自由讨论。 - **跨时区轻松氛围**：周五上午举行，兼顾欧美时差（早间咖啡与午后啤酒），降低交流心理门槛。 --- ### [42:46](https://www.youtube.com/watch?v=OG7QR4UETX0&t=2566s) 十万人巨头的文化转型：从引导式赋能到自发涌现 与敏捷灵活的初创企业相比，拥有超千亿美元营收规模、员工人数超 10 万的大型跨国企业（如嘉吉公司 Cargill）在推动 AI 变革时，其核心本质在于“组织文化的系统性重塑”。Nithin Ramachandran 指出，在如此庞大的组织体量下，AI 转型无法单纯依赖自上而下的命令，而是需要设计多层次的活动矩阵来引导并点燃全员的自驱力。 ```mermaid flowchart TD accTitle: 大型企业 AI 文化转型的多层推进机制 subgraph TopDown["自上而下引导与输入"] A["引导式活动<br/>(展示具体实用场景)"] B["跨界交流活动<br/>(引入白皮书与外部专家)"] end subgraph BottomUp["自下而上自发涌现"] C["有机加速<br/>(员工自发午餐会/兴趣小组)"] end TopDown -->|"注入初始动能"| BottomUp BottomUp --> D["终极衡量指标：自发性组织活动规模"] ``` 大型组织通常将 AI 文化建设拆解为三个阶段性动作：首先是“引导式活动”（Guided Events），通过实操演示直接向员工展示 AI 能帮他们解决哪些具体工作；其次是“跨界交流活动”（Cross-pollination Events），通过引入行业白皮书、邀请外部前沿演讲嘉宾拓宽业务视野；而最关键的目标则是实现“有机加速”（Organic Acceleration），即促成员工自发围绕 AI 组织午餐分享会或专项小组。对于大型企业的数字化掌舵者而言，衡量 AI 推广是否成功的终极指标正是“组织内部自发涌现的 AI 活动数量”。所有顶层设计的铺垫和推动，本质上都是为了注入足够的初始势能；一旦这种底层的自发活力开始自主运转，就意味着大规模文化转型已经达成目标。 在 panel 的尾声，主持人向全体与会嘉宾的精彩洞见致谢，并邀请现场观众以热烈的掌声结束了整场研讨。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=OG7QR4UETX0&t=4s) 昨天的“竞争优势”（Competitive Advantage）很快就会沦为“竞争均势”（Competitive Parity）。 — Nithin Ramachandran > [11:05](https://www.youtube.com/watch?v=OG7QR4UETX0&t=665s) 这笔费用最终由谁来买单？ — Reza Forghani > [15:04](https://www.youtube.com/watch?v=OG7QR4UETX0&t=904s) 最高效团队与最低效团队之间的成本差距通常高达 20 到 30 倍。 — Jason Cumberland > [16:56](https://www.youtube.com/watch?v=OG7QR4UETX0&t=1016s) 去年是整个行业盲目跟风的“狂欢期（free-for-all）”，所有人都在不计代价地把 AI 往各种业务场景里塞，过程中没有任何前置的指标追踪。 — Jason Cumberland > [28:34](https://www.youtube.com/watch?v=OG7QR4UETX0&t=1714s) 首先必须从脑海中抹去的概念就是‘AI战略’——根本不存在所谓的AI战略，只有业务战略（There's no AI strategy. There's only a business strategy）。 — Nithin Ramachandran > [28:34](https://www.youtube.com/watch?v=OG7QR4UETX0&t=1714s) 一次构建，多地复用（build once, deploy many times）。 — Nithin Ramachandran > [31:46](https://www.youtube.com/watch?v=OG7QR4UETX0&t=1906s) 在所有失败的 AI 项目中，没有一次是因为底层技术本身不过关，失败全部源于人员（People）或流程（Process）层面的准备不足。 — Nithin Ramachandran > [35:49](https://www.youtube.com/watch?v=OG7QR4UETX0&t=2149s) 改变工程师行为最有效的方式是“温和推动”（gentle nudges），而非强制死板的限制规则（strong guard rails）。 — Jason Cumberland > [35:49](https://www.youtube.com/watch?v=OG7QR4UETX0&t=2149s) 没有什么比工程师停工等待更昂贵的了，那比所有消耗的 Token 都要贵得多。 — Jason Cumberland > [41:51](https://www.youtube.com/watch?v=OG7QR4UETX0&t=2511s) 每周都会有人体验到某种全新的、感觉像魔法一样的技术突破。 — Jason Cumberland ## Mentioned Steven Birgfeld, Songyee Yoon, Jason Cumberland, Reza Forghani, Nithin Ramachandran, Blue Diamond Growers, Principal Venture Partners, Revenium, AdventHealth, Cargill, 3M, Coca-Cola, New York Times, MIT, FDA, HIPAA, SAP, Claude, Codex, Anthropic, OpenAI, Ask3M ## Video description Steven Birgfeld — VP, Information Technology and Services, Blue Diamond Growers Songyee Yoon — Founder and Managing Partner, Principal Venture Partners Jason Cumberland — CPO & Co-founder, Revenium Reza Forghani — Chief Medical Information Officer (CMIO), AdventHealth Nithin Ramachandran — Senior Vice President, AI and Data, Cargill What does real AI ROI look like in practice? This panel brings together executives who have successfully delivered profitable AI deployments across diverse business problems and industries. From cost reduction in operations to revenue growth in customer-facing functions, panelists will share candid stories about where AI has delivered measurable returns, and where it has fallen short. Learn about the common traits of successful use cases, the pitfalls to avoid, and the lessons that can help your organization replicate ROI at scale. [Source](https://www.youtube.com/watch?v=OG7QR4UETX0)

[0:04] Nithin Ramachandran: 昨天的“竞争优势”（Competitive Advantage）很快就会沦为“竞争均势”（Competitive Parity）。 [11:05] Reza Forghani: 这笔费用最终由谁来买单？ [15:04] Jason Cumberland: 最高效团队与最低效团队之间的成本差距通常高达 20 到 30 倍。 [16:56] Jason Cumberland: 去年是整个行业盲目跟风的“狂欢期（free-for-all）”，所有人都在不计代价地把 AI 往各种业务场景里塞，过程中没有任何前置的指标追踪。 [28:34] Nithin Ramachandran: 首先必须从脑海中抹去的概念就是‘AI战略’——根本不存在所谓的AI战略，只有业务战略（There's no AI strategy. There's only a business strategy）。 [28:34] Nithin Ramachandran: 一次构建，多地复用（build once, deploy many times）。 [31:46] Nithin Ramachandran: 在所有失败的 AI 项目中，没有一次是因为底层技术本身不过关，失败全部源于人员（People）或流程（Process）层面的准备不足。 [35:49] Jason Cumberland: 改变工程师行为最有效的方式是“温和推动”（gentle nudges），而非强制死板的限制规则（strong guard rails）。 [35:49] Jason Cumberland: 没有什么比工程师停工等待更昂贵的了，那比所有消耗的 Token 都要贵得多。 [41:51] Jason Cumberland: 每周都会有人体验到某种全新的、感觉像魔法一样的技术突破。

[8:26] Steven Birgfeld: maximize that return to our growers [11:05] Dr. Reza Forghani: Who's going to pay for this at the end? [11:05] Dr. Reza Forghani: None of this would work if you don't have a robust infrastructure. I look at it like the railroads before you even do all the, you know, fancy things you have to have, you know, core infrastructure that can support it. [28:36] Nithin Ramachandran: Build once, deploy many times [28:36] Nithin Ramachandran: There's no AI strategy... There's only a business strategy. [28:36] Nithin Ramachandran: a hammer searching for a nail [31:46] Nithin Ramachandran: I don't know about you guys, but I hate being on the receiving end of an AI customer service bot when I just want to get to a human being. [35:48] Jason Cumberland: the most effective way to change behaviors is with gentle nudges, not strong guardrails. [35:48] Jason Cumberland: there's nothing more expensive than an engineer who can't work... that's more expensive than all the tokens. [38:23] Jason Cumberland: at your desk, not that fun, but they're really effective. [39:48] Jason Cumberland: manufacture an answer every time the CFO sends a question down the chain.

[0:04] Good morning everyone. Welcome to the panel,  first panel of the day I guess. Um and [0:10] uh AI and ROI. I think when we talk about when  we think about RORI and AI, we often talk about [0:17] uh productivity improvement or cost cutting. But  the real benefit also comes from better quality [0:25] decision making as some of the kind of a new type  of things that we were not able to do without AI [0:31] etc which also comes in at in different time  scale. So I think we when we think about how to [0:36] measure ROI can get often get confused very easily  because like how the kind of traditional the [0:43] formula that we're using for calculating ROI for  software doesn't really work. So I'm really glad [0:49] that we we c have like this four very experienced  uh executives from all sort uh all different [0:57] industries have been working on deployment of ARI  thought a lot about how to measure it and how to [1:04] integrate into the strategy making. So I think  it's we can we can have a very rich discussion [1:10] based on their real experience. So first of all  we'll get start with in round of introductions. [1:15] Please introduce yourself, your role, and  your biggest AI investment that company's [1:21] making this year. Hey everyone, my name is Jason  Cumberland. I'm one of the founders of a company [1:26] called Revenium. So I have a little bit different  take. Uh these guys will tell you about uh putting [1:30] AI deployment inside of their companies. But what  Revenium does is actually helps companies measure [1:34] the efficacy and efficiency of ROI or of AI use  inside their companies. um and got real for me as [1:42] to how like how pervasive this question is getting  just before this conference started. I got a call [1:46] from a reporter at the New York Times that wanted  to talk about the effect of AI on the economy. Um [1:51] and that article landed on Monday, just before  this conference started. So today you'll hear [1:55] a bunch of specific examples, uh but if you want  a broader take on how this is all affecting the [1:59] economy in general, uh that article is from Monday  and it's a good read. Go ahead. Hi everyone, I'm [2:05] Resza Fori. the chief medical information officer  for imaging at Advent Health. Um when we for our [2:11] company guess in healthcare things are complex. So  one of the major investments is in infrastructure [2:17] to make sure not only our digital infrastructure  is cutting edge but also we are AI ready for [2:22] future. If you think about imaging specifically  where a lot of AI action has going going because [2:30] we have been digital for so long. Our major  projects one has been on using AI to flag [2:38] and prioritize uh critical conditions on certain  types of scans life-threatening conditions which [2:44] has had benefits both in terms of timely diagnosis  and actually even a second reader effect to make [2:50] sure they don't get overlooked. And some of the  other ones are operational. Specifically for us is [2:56] uh intelligent uh work list orchestration for  radiologists that if you know the nuances can [3:02] be very impactful for our operations. Yeah. Hi,  I'm Steve Burkefeld. I'm the CIO of Blue Diamond [3:10] Growers, the largest almond processor in the  world. Um but more recently, last eight weeks, [3:15] I've been retired now. So still doing advisory  work for them and some other companies. [3:20] Um so you know with blue diamond we had built  a lot of different use cases on ROI. So the [3:26] real initiative this year is to continue that  journey expanding it and then more so you know [3:31] we leverage the SAP ecosystem. So working closely  with SAP on their autonomous environment to kind [3:38] of build that out and then augment accordingly  with these additional AI tool sets that we've [3:44] we've put in place. Looking for high value for  the the grower. Definitely. Thanks Steve. Hi [3:51] everybody. Uh Nathan Ramachandran. Um I serve as  the senior vice president of data and AI uh at [3:57] Cargo Foods. Uh it's the largest privately owned  food company in the world. Uh and prior to that [4:02] um I led I was the executive leader responsible  for data and AI globally for uh 3M uh and cola [4:09] company prior to that. Um so I've been through  the cycle quite a few times in terms of the [4:13] transformation journey with data and AI. Um,  in terms of making, you know, kind of where [4:18] we're kind of making our largest investments,  uh, honestly, it's not on a specific use case, [4:23] but it's honestly about into the enterprise  capabilities. Um, because the space is moving [4:29] extremely fast and use cases are changing on a  day-to-day basis. What used to be competitive [4:35] advantage turns into competitive parity pretty  quickly um, in in this in the world of AI. So, [4:41] we're still looking at the base fundamentals,  which is how do we make sure that our people [4:46] are equipped um to work with AI? How do we make  sure that we have the foundations from a data [4:51] availability and data reliability standpoint? And  of course, from a process definition standpoint, [4:56] do we really understand the processes in our  company um so that we can actually digitize [5:01] them and make them more intelligent? But it's  all about that enterprise capability that drives [5:05] the reusability uh when we are kind of you know  trying to activate AI use cases. That's great. [5:11] Yeah. Let's get a little bit more specific when  we talk about AI and ROI. Can you walk us through [5:18] like when you when you think about ROI return on  investment and like when we make an investment [5:24] decision like what does that like what are the  steps you take? How does that kind of can you [5:29] walk us through like what you think about like  how to gather information and how to get to that [5:34] uh get the number that that will help us make  a kind of final final decision as an executive? [5:38] Okay. Um yeah, sure. Um so when we when we kind of  look at look at ROI, right, traditionally I think [5:46] like you know when you look at it from a CFO lens,  you're looking at it from like what's my formula, [5:50] right? Um and and I think you the way you need  to look at ROI from an enterprise perspective [5:57] is is from the perspective of like a venture  capital firm, right? Which is do you really see a [6:03] portfolio level outcome um that's coming out, you  know? So when we talk about the ROI definitions, [6:09] we talk about it very similar to how we would  have looked at social media companies back in [6:14] the day when your first set of investments was  into network effects. So very similarly when we [6:18] look at the enterprise we look at it are we seeing  leading indicators um from a a return on employee [6:25] perspective. So that's are we seeing utilization  of AI? Are we seeing literacy uh in the workforce [6:30] of AI? Are we seeing increased utilization in  business processes? So that's that initial phase [6:35] which are the leading indicators that tell you  that you know your company's becoming AI first [6:41] and then you have your business processes. This is  now the metrics that the CFO actually cares about [6:46] which is am I seeing the productivity, am I seeing  you know uh workforce optimization, am I seeing [6:53] reduction in costs um or growth in revenue. So  these are all business process improvements. This [6:58] is the standard stuff you'd see demand forecasting  things of that nature. And then finally what we [7:03] really look at is what differentiates the company,  right? And that's where we start looking at return [7:08] on future like do we see AI investments that  are going to give us a return 3 years to 5 [7:14] years in future and those are not grounded in  AI technology those are grounded in things like [7:19] proprietary data or some form of differentiation  in the company. So that's a lens that we wear and [7:25] and and and you know the conversation shifts as  we talk to different stakeholders in the company. [7:31] How does it change from like when we talk to  employees versus executives and to the board [7:36] members? Um so from an employee perspective the  conversation is really more on what it can do [7:42] for them. Right? So it is about taking out mundane  day-to-day mundane work from their day-to-day and [7:47] allowing them to increase their span of influence.  When we're talking to the executive team so a CEO [7:53] a CFO especially in public companies like you want  results immediately. So we're talking about you [7:59] know the process improvements um the things that  deliver return in less than 12 months. When we're [8:05] talking to the board it's a completely different  discussion because the discussion is about what do [8:10] we do in things from a differentiation standpoint  and what do we do in things from a risk conscious [8:14] scaling standpoint. So it's about balancing the  risk and and long-term sustainability of rewards. [8:20] Um so the storyline shifts as you talk from one  stakeholder to another. That's great. Steve, [8:27] can you share about your your approach? Yeah, as  mentioned, our mission is to really maximize that [8:32] return to our growers. So, uh, naturally looking  at what those those business value percent pieces [8:39] are. Um, really in talking to the exec level and  the board itself, it's been as well, we can't [8:46] totally focus on short-term ROI. We really really  got to focus on getting the strategy in place and [8:51] the components in place so that we can build  further on this. Um so um it's predominantly [8:58] you know productivity is the given but where  we've been focusing are what are those measurable [9:03] outcomes everything from cost savings to um better  performance metrics the you know we're measured [9:12] in the industry by being on time in full you know  increasing the yield of the almonds as well. We're [9:18] doing a lot of work in that space to to make sure  we we're maximizing those components of it. Um, by [9:24] the same token, you know, the lower the employees,  the team members are really excited to see what [9:30] this is doing. Um, a little nervous about how it  impacts their job. But in every case, we've seen [9:36] um not replacing their job, but giving better  insights to them at the team level as well as [9:42] the executive level of of making better business  decisions. Um, we're seeing a lot of cost savings [9:48] opportunities that are coming out of that as well.  Um so so again just what's that maximum maximizing [9:54] the value to the growers is that total total  effort. Yeah. So from our perspective there's [10:00] uh overlap is somewhat of what Nitton already  mentioned in terms of enterprise approach. I would [10:04] say in healthcare sometimes I find it useful to  break things down more is it a diagnostic clinical [10:11] support versus operational and administrative and  you know the former has very specific regulatory [10:18] requirements that impact adoption. Our KPIs  are typically multiffactorial, right? So, [10:23] it's going to be of course patient quality and  safety in the future, predictive therapeutics. [10:28] They're still sort of not fully mature. And then  there are things that you know just go to timely [10:35] diagnosis. How can you know what things align with  our mission of delivering highquality consumer [10:41] focused care that's timely? On the operational  side, sometimes it's just things such as making a [10:48] scan be acquired faster. Those some of those have  actually more clear ROIs. Um, and then there are [10:56] other operational aspects or you know healthcare  has its diagnostic part and everything from [11:00] revenue cycle management and other things that  have pure administrative work and there how do [11:05] we enhance those? When we look at again KPIs, it's  usually I like to look at short, intermediate, and [11:13] long-term because they're very different. And of  course, you know, for us, sometimes we may bring [11:17] something because it enhances patients experience  because we are there to treat patients, right? [11:22] Something that produces safety, but I'm conscious  and depending on what level you're talking about [11:27] it at the end, we also have to be sustainable.  Healthcare is expensive and I do think it the onus [11:33] is on us to take that. So I also like to look more  and more at those especially as the informatics [11:40] and software costs are getting higher and higher.  We really have to ask those hard questions have [11:45] clear KPIs and also things that are who's going to  pay for this at the end because we won't be able [11:50] to deliver effective patient care if we can't be  financially viable. Yeah. I mean I think it's kind [11:57] of very multifaceted kind of nature of this ROI  is is very striking and I think that uh because [12:05] of that when we make a choice like when we make a  choice which project to fund first the the project [12:13] with the highest short-term RARI might not be the  highest the longer term RARI project like how do [12:19] you make that kind of the balance and tradeoff  when you're making a decision? That's an excellent [12:24] question I think and it will likely differ on  the healthcare organization and your mission. [12:30] I'm very fortunate in a being an organization  that I I believe has been run very effectively [12:36] but also we have a long-term vision can tell you  my personal approach is to have multiple and also [12:43] have things that are more straightforward and have  things that really position us for the long term. [12:50] Um and that combination is important because  it goes back to what I mentioned is you know [12:54] I still practice although less I'm a physician  and that part of me just wants to treat patients [12:59] but sometimes I see that people uh you know  especially in the executive level they can be [13:05] switches well who's how we going to afford this  and that's really the way we look at it. I would [13:10] say for success in healthcare you need end user.  So you need to know what matters right. So some [13:17] things may look very exciting but they may be  immature and you may bring them in and actually [13:22] be quite disappointed. I would say early on one  of the things I focused on was workflow like the [13:28] difference between a gadget and a product. We're  getting much better at that now because as these [13:32] things are becoming expensive the vendor said  but I see that's one thing especially in the [13:36] clinical domain you want things to be effective  they have to be used so you need now that's not [13:42] the only expertise you need because you need the  other expertise but I would say that is one key [13:46] thing but also you need to counterbalance that it  how does this serve our long-term mission and how [13:52] does it become sustainable so again having that  multiffactorial uh multid-disiplinary team and [13:58] I would say the One other essential thing is  our digital infrastructure. So I'm very proud [14:05] of our IT team we deliver. But I can tell you in  healthcare and sometimes it's just because of the [14:12] cost. A lot of its I almost feel they operate on  the past model. They're underfunded. They're like [14:20] an afterthought. Our it is not an afterthought.  I'm I'm actually part of it. They know of projects [14:26] even when I start thinking about them. And I  do think that past model is not conducive to [14:33] the upcoming digital transformation. And think of  it all the exciting AI things. None of this would [14:38] work if you don't have a robust infrastructure.  I look at it like the railroads before you even [14:43] do all the you know fancy things you have to have,  you know, core infrastructure that can support it. [14:49] And I would say that's one thing that's not  easy because it's cost. But I do think it's [14:54] uh important for health care to invest in that  if they want to be successful and competitive in [14:59] the long run. That's good. I have a couple more  questions, but I'm going to move on to Jason. [15:04] Yeah. Yeah. So I I think I I'd try to help you all  by giving you a general roadmap that we talk to [15:09] clients about in terms of if you're if you're  not measuring ROI at all and you want to know [15:12] where to start. The first thing is looking across  your billing or invoices and understanding which [15:18] groups inside the company are driving the most  spend because going and trying to prove ROI on [15:22] a group that doesn't contribute meaningfully to  spend nobody cares about even though you can you [15:26] know reach the finish line and say you've done  it. And if you're not sure that's the starting [15:29] point is trying to get your spend broken out by  business unit or team. And if you can't do that [15:34] um then I would say your problem is probably in  engineering because that's generally the answer. [15:38] Um engineering in across all the customers we work  with represents at the low end 50% of AI spending [15:43] at the high end about 80%. And so we usually  start there and engineering is lucky and it [15:48] fits into an easy ROI model. You can measure the  unit of work out of engineering pretty easily. Um [15:53] it's generally a poll request. It's a feature.  It's a ticket. It's something that is clearly [15:57] documented. And there's always a little bit of  argument about simple versus complex work in [16:02] engineering, but the point is across a big group  of engineers, it sort of averages out. And so, [16:07] um, to tell you how diverse the problem is, on  average, when we go into a company and we measure [16:11] the efficacy of engineering teams in terms  of cost per feature poll request aggregated, [16:16] there's usually a 20 to 30x difference between the  least efficient team and the most efficient team. [16:21] And so just by knowing that and being able to  identify, you know, the teams that you want to [16:25] go target with different behavior patterns, you  can meaningfully reduce cost without impacting [16:29] innovation. And that's always everybody's fear is  how do we how do we roll out some kind of sensible [16:33] cost controls around AI without impacting  innovation. And for now, everybody's kind of [16:37] holding their breath hoping that, you know, not  that the spending isn't too much. But everybody [16:42] reaches that tipping point where you have to put  in some kind of controls. And if you don't start [16:45] the measurement process now when the point comes  when your spending is out of control and the CFO [16:50] calls you don't actually know where to go. So that  would be my general starting point advice. And Pa [16:56] I think that uh one of the most circulated article  last year was the MIT research one and saying that [17:02] 95% of the all this AI project will not return any  any kind of meaningful return on investment. Um I [17:11] mean I think I'm I'm sure that this there's a lot  more to kind of um to understand to that to that [17:18] kind of claim. So like please help us understand  how can we read that type of headline and what's [17:24] underneath what are the kind of nuances that  we have to keep in mind. I think it links back [17:28] to the same problem is that last year it was a  free-for-all. We all did it. You went and you used [17:32] AI for everything and there was no measurement  around any of it. And so because you couldn't [17:36] measure it, of course nothing looked like it  succeeded. And I I think that will gradually get [17:41] better. We generally talk to clients about three  ways of proving ROI. It's either the unit cost [17:45] of doing something, the throughput, so how many  you can do in a given time and the quality, which [17:50] is not a new way to evaluate processes, right?  That's the same thing. But the problem was that [17:54] the measurement was never in place for those. And  so, um, we always encourage our clients to start [17:59] with things that are measurable, whether it's an  engineering or somewhere else, because otherwise [18:02] you get to the end and it's it's an argument  about whether we if we're still in the middle [18:05] of a short-term versus a long-term project. And so  getting those metrics nailed up front. I think uh [18:10] companies are getting a lot better about because  not only do they now have more processes built [18:15] around their their AI usage, but they also have  more systems in place to measure them and they [18:19] have a little bit of data under when they go down  a path to expect to know what to expect in terms [18:24] of spending. And so the innovation we're seeing  is a transition from um where it used to be when [18:29] you were going to go roll out a project, you would  plan for headcount and you would go through a two [18:33] to three month cycle to hire new people and then  deliver that project. And now tokens are coming [18:37] into that conversation because you can start the  project faster. And so moving on to having ability [18:42] to forecast the tokens it takes to deliver a given  feature or a project is a real discipline that you [18:47] need to be able to deploy inside your enterprise.  And it's going to harp on the same thing unless [18:51] you're measuring what you're doing today. You  can't forecast what's going to cost tomorrow. [18:55] So that's kind of how we build all the blocks  together. Yeah. I think it's also it's it's about [19:00] what what are the kind of the underlying costs you  can kind of put into the denominator. For example, [19:06] you talked about the underlying infrastructure  and IT roll out. How much of that is kind of be be [19:12] uh attributed to specific project and not knowing  what what is the kind of longevity of those pro [19:19] kind of the infrastructure and kind of how how  long that that particular technology is going to [19:23] be relevant going forward. I think there's a lot  of uncertainty uncertainty in the in the air that [19:28] makes it even more it's much harder. Like how do  you deal with that kind of uncertainty? Yeah, it's [19:32] a great question. Um, and there's probably four  main categories of things you need to track as [19:37] you roll these things out. Infrastructure is one  of them. Tokens is the one we talked about. Then [19:41] there's the the time invested by humans. There's  the initial time. And then there's the time humans [19:45] spent rescuing every failed AI workflow. And  you know, anybody in the crowd who's tried to [19:49] deploy agents know that it's a long road from the  first deployment to where you declare success. And [19:55] if you're not tracking the um intervention costs  along the way, you fool yourself in terms of the [20:00] ROI of AI deployments. Uh and then the last one,  which is usually a blind spot, but if we have any [20:07] financial folks in the crowd, you recognize this  like workflows, agents touch outside tools that [20:12] cost a lot of money. And so financial services  is one where it's the the most prevalent. You [20:16] reach out, you do a know your customer call  or an identity call or a credit report call. [20:21] These things cost 20 or $30 per invocation. And  in the old deterministic world, a script runs, [20:26] it happens once and you you're sort of locked in.  You know, in this world, 10 agents might fan out, [20:31] all do the same task, and you get the bill  at the end of the month from five different [20:35] service providers and your AI bill and there's  no connection between the two. So unless you're [20:40] building that instrumentation between all the  third party costs, it just leaves a a big question [20:45] at the end of the month when all those bills  arrive. So um there are ways to solve it but it's [20:50] a blind spot for most companies right now. Right.  Yeah. Speaking of uncertainties, I think that [20:56] um the regulatory environment also puts kind of u  another layer of burden to healthcare industry and [21:03] how do you take this this into into consideration  and kind of the how how the landscape is changing [21:08] to affect your kind of RARI calculation? That's an  excellent question. I mean regulatory environment [21:15] really impacts any deployment that's clinical  and indeed I would dare to say that you know [21:21] the technology was way ahead of regulation which  is probably not a surprise but sometimes you may [21:27] one may underestimate the impact it has had cuz  as we are shifting from narrow to more general [21:32] applications which really are more clear value  in healthcare if you're doing diagnostics the [21:38] regulatory environment wasn't built for that  it's like approved one by one and that has [21:42] actually been a significant barrier to bringing  some of those solutions. There also other things [21:48] like in imaging where we are quite advanced, you  know, taking an image and producing a draft text. [21:54] I personally think that's going to be the next  step change and there are the examples of that. [21:58] The regular framework doesn't exist. Now, we have  to be careful. We need regulation because we want [22:04] to have tools that are safe. At the same time it  has to be balanced especially when overseen by an [22:10] expert to also allow that. The good news is we  have seen some breakthrough device designations [22:16] this year that are facilitating that but that's  always a consideration and it also puts in another [22:23] layer for healthcare that may not be present in  other industries. So as you want to let's say [22:28] you want to generate internally right you want to  use the foundation models so have agents sure for [22:34] purely administrative things you can do that as  long as it's still HIPA compliant so there's still [22:39] that level but when it comes to diagnostics that  gets very tricky and actually you know it's not [22:47] fully worked out so I don't know if strategically  we can do those things internally that let's say [22:54] another industry can do that so that may continue  to rely on third party vendors because I would [23:00] imagine most healthc care organizations don't want  to become vendors of record and get FDA clearance [23:05] and all of that. So it's certainly an important  impact. Steve, I mean I think was kind of reading [23:13] uh what happened at what you did at Blue Diamond.  I was surprised that like Blue Diamond is such an [23:18] AI company doing all the I mean it's as a kind of  I'm a I love uh Blue Diamond. um almonds and then [23:27] I didn't realize that they're doing like using  AI for logistics, marketing, customer service [23:33] and crop forecasting and also taste trends. I  mean it's just such a wide variety of things that [23:40] you're applying AI technology for like what was  your experience like in rolling out the technology [23:45] in such a diverse kind of um the functions  within the organization? what was there any [23:51] kind of similarities and differences that you had  to take into consideration and kind of different [23:56] ROI measures that you you took into consideration  but make make those investment decisions. No like [24:04] like so I think we took that approach of let's  not focus initially on the short-term ROI. Let's [24:08] really get the strategy. Let's see where there's  going to be benefit to the company. Um couple [24:14] examples like crop forecasting because of mother  nature it was really difficult to kind of predict [24:20] the future of what that crop was going to look  like. So it moved the needle a little bit but [24:24] not enough. Demand sensing was another one. Trying  to really predict what the customers were going [24:28] to order. Same thing with demand planning didn't  move things so much. But again we just tried it [24:34] to kind of understand what it was. Um we did have  success predominately in our manufacturing side of [24:40] things with whole predictive maintenance, yield  management, um quality and and uh um the quality [24:48] side of it from from that. So we could actually  say for this variety of almonds, what's what's [24:55] that um yield going to be? How's it going to tie  into the the product that we're producing as well? [25:02] So really saw you know seven figure savings  because of better um dashboarding metrics etc [25:09] to to really see what what was going on there. The  other part of that was the workforce planning side [25:15] of it where we could really gauge again byproduct  by line line manufacturing you know compliance [25:23] measures as well as overtime measures. So really  seeing where we could reduce cost from that. So, [25:27] so great benefits from the manufacturing side.  And then on the marketing side, the same thing, [25:32] you know, we're being able to get a product  to market faster by by using things of, hey, [25:37] what's going to be the next popular flavor trend  in Asia or Canada, you know, it helped us define [25:43] our our ketchup flavored almond for Canada as  an example. Um, so so getting better insights [25:49] there to go to market faster and then leveraging,  you know, AI to generate some of the content and [25:55] the graphics and things like that. So saving  money, getting to market faster, you know, [25:59] some of those those cases. Um, so it really  just went back around to where where is it [26:05] moving the needle needle, where are we seeing  the the value to the business. Um, but again, [26:11] we had to experiment across multiple disciplines  to really get that some key components, you know, [26:17] really making sure you have the business in  place to work with you on that. you know, [26:21] being a a a real ad advocate for it was one um  to Jason's point of what can we measure really [26:30] making sure we had that initial measurement  so we could see the movement from that side [26:33] of things. And then the other component was the  data to make sure you had at least some initial [26:39] data points. Um we were lucky to have a tool  that actually created some of the automated [26:44] um ontology that we needed. So by having disperate  data systems, it kind of pulled things together [26:50] for us so we we could have a better sense of you  know what what data matches there were. Um so [26:56] those key components you know those three three  pieces but really try it see what the results [27:02] were but then move on with the high value items.  I really like that approach and I think it's u I [27:08] mean I think it's that through that iterative  process you probably see what are the kind of [27:13] signals and data points that you that help us make  decision about like double down on certain project [27:18] and kind of move away from it. What are the key  kind of data points that that you kind of you [27:23] look for? Um yeah again in our our case it's like  what improves our on time and full delivery to our [27:31] customers. You know that's at one measure. um the  other so that yield, manufacturing, logistics, all [27:39] those components lead to that to make sure you've  got good insights from that side of it. Um and [27:44] then we're focused naturally on that bottom line  numbers again to maximize the return. Um so so [27:52] where where are their cost savings opportunities  like like I mentioned with the manufacturing side [27:56] of it where we saw seven digit savings from the  work we put in on that side of the equation. So [28:02] um so I guess do you have like a specific target  number in terms of kind of time to market time to [28:07] delivery and like where is a gap in like your um  yeah I mean I think before we started leveraging [28:14] some of these tools time to market could be  eight months or so I think we're getting that [28:18] down now to three to four months to kind of get  to that point. So a lot of work behind that to get [28:23] there. Um so so again just seeing where there's  some metrics that we can can show reductions in [28:29] times and and reductions in cost as well. That's  great you feel. Um so I think congratulations on [28:36] your your role at Cargill. That's really that's  amazing. Um and I think you have worked you have [28:41] extensive work at at 3M across multiple geography  and I think you're you have a unique experience of [28:47] rolling out this tools across different different  countries and different jurisdiction and can you [28:52] tell us about like uh what the experience was  like what you have learned uh by applying these [28:58] tools in in different different culture in  different different market yeah I mean I [29:03] think the it's a completely different playing  field um when when you're trying to deploy AI [29:11] um or data products in let's say one region  versus trying to do that in multiple regions right [29:17] um and that's because of multiple things that  change there's u regulatory environments that's [29:22] changed the the nature of the business is  different um and the the the product and the [29:28] tools and the infrastructure that's available  is different as well right um so ultimately [29:32] it comes down to kind of a a very simple thing  right the first thing is there's no AI strategy [29:39] Like I I think that's the first part you need to  push out of the picture. There's only a business [29:42] strategy, right? So you have to AI is just an  enabler to to compete more effectively, right? And [29:49] and I think otherwise it's like a hammer searching  for a nail, right? Um so so I think the first part [29:55] is getting that out of the way. So usually when  you're doing global deployments, the first part [29:59] usually when I sit and think about it as a chief  AI officer, you're looking at it in terms of what [30:04] are the common areas of overlap that you have  across regions where you have similar areas of [30:11] competitive differentiation, right? Whether that's  your sales process, whether that's your demand [30:15] management or any of those pieces, right? And once  you hone in on that, then it's about innovating [30:21] with a global perspective, right? um there's  there's there's experimentation you do now within [30:26] that scope because part of what you want to do in  order to drive that ROI on a global perspective is [30:32] going to be about reusability. So you want to  build once deploy many times right um so when [30:37] you take that picture you're able to get it done  but there's always again the second part of the [30:42] complexity that comes in which is infrastructure  which is now you also need partners to work with [30:46] because let's say in all the companies that I  worked in you always have operations for example [30:51] in China and the landscape in China is very  different from anything else in the world right [30:57] um so you have to work with partners at that point  as well um to be able to understand how you will [31:03] drive this reusability across languages across  um you know kind of uh regional differences. Um [31:11] but the key part of it there I think step one  no AI strategy it's a business strategy find [31:16] overlapping areas of competitive advantage deploy  experiment within the scope of that competitive [31:22] advantage work with partners to ensure that you  have reusability um across the different regions. [31:28] If you're able to do that effectively you have a  winning strategy. Yeah, I really like that. Uh so [31:33] there's no AI strategy but there's only business  strategy. Um what what are some of the things that [31:39] uh surprise you like in in in deploying the  technology in across the region and what you [31:46] what worked much better than you expected and  what didn't work although you had a very high [31:52] expectations and and and I mean I think it's and  why was the case. Got it. Maybe let me rephrase it [31:58] a little bit this way. Okay. Let me uh a lot  of battle scars from having failed for sure, [32:04] right? Um and have had failures. None of those  failures have been because of technology. All of [32:12] those failures have either been because of a lack  of readiness from a people standpoint or from a [32:18] process standpoint. It has never been because of  the technology, right? Um and so so so some of the [32:26] some of the painful pieces has been um you know as  we start looking at it from a process standpoint [32:32] for example we deploy AI technology and expect  for example that the customer is going to like [32:37] it. I don't know about you guys but I hate being  on the receiving end of a AI customer service bot [32:43] when I just want to get to a human being right.  Um and so we've had failures of that sort where [32:50] um you know you're kind of in that recursive  loop with an AI chatbot when you're just like [32:54] at some point it's like get it to a human being  the customer is more important than you know [32:59] uh you know a few dimes that you would have saved  in the process. So, so I think some of the painful [33:04] learnings have been have come from that aspect of  you just cannot replace the human in the loop and [33:10] if the customer experience is the most paramount  thing that you want um you know then you have to [33:16] make sure um that that you're thinking through  the process implications. What does that business [33:21] process look like? What is the impact of that  process on your business? That's the first part. [33:26] The second part is a people aspect of it which  is uh when you think about the workforce in [33:32] the company and you start thinking about enabling  them uh there's not a lot we put into the efforts [33:38] around literacy uh and so when I say literacy  most companies started off with AI literacy from a [33:44] riskmanagement standpoint how do you safely use AI  and what do you not put into a chatbot or or an AI [33:50] bot so to speak right but there's I think there's  growing realization and And it was a painful [33:56] realization for me as I was driving the program  was you actually have to teach people to integrate [34:01] AI into the flow of their work. That is when you  see value, right? Um driving AI into a business [34:08] process is not just about workforce reduction.  It is about employee expansion. And when you flip [34:14] that lens to more of a growth mindset, you start  seeing the importance of getting employees to [34:20] integrate AI into their day-to-day work, helping  them see the opportunities. So to me those were [34:24] kind of the like some of the painful lessons but  none of the failures were because of technology. [34:30] It was because of lack of thought into readiness  from a people and process perspective. That's a [34:36] very interesting example about the customer  service. Um I mean there is a notion about [34:40] uncanny valley right I mean so I'm curious whether  that the lesson was whether that we should have a [34:49] human and customer support or we should not have  like premature technology and customer support [34:55] like what what would you think um I I think I  would look at it this way which is you have to [35:01] focus on the customer experience right you have  to look at it from the perspective of the customer [35:07] it's exciting to have techn technology in place  but I don't think AI technology is at a point [35:12] where it can serve every need of the customer and  again this depends from company to company right [35:16] if you're offering very transactional service like  you're putting a banking app and the focus is to [35:21] prevent the or to help the customer to not go to  the bank yeah sure you can use a lot of chat bots [35:27] because it's very deterministic transactions  for example at 3M we built out and released [35:32] this product called AS3M it is a non-deterministic  product that deeps that deals with deep material [35:37] science these questions which is just is extremely  difficult to build an AI solution around. So you [35:42] have to have the human at the loop. That's that's  fantastic. I'm going to go back to Jason uh about [35:49] this kind of human question. So you mentioned that  the best performing team and least performing team [35:54] had like 20x difference in productivity and and  and um an RI like where does that difference come [36:02] from? like uh how do you assess that uh kind  of difference from without without asking for [36:09] external help as a as a as a measurement? Yeah.  So, usually um it comes down to well the most [36:18] the most effective way to change behaviors with  gentle nudges, not strong guard rails. That's [36:23] what we found. And so organizations that put that  have put none of those in place that just roll out [36:28] tools like claude and codeex with no guidance  on what to use um end up having a small portion [36:33] of engineers who naturally gravitate towards the  right decisions. They'll make they'll make model [36:37] routing decisions that make sense. But the the  short answer to your question is it's a choice of [36:43] which models engineers are using to complete tasks  or employees if you take it out wider. And so for [36:48] knowledge workers, this is outside of engineering.  A lot of, you know, bringing that efficiency [36:52] down is just making slight adjustments to the the  tuning that most of these tools give you. Choosing [36:57] default models, only allowing certain groups  in the company access to the top tier models [37:02] because most most people don't need them for a lot  of their work. Um, and so that's stuff that is not [37:07] really hard to do, but has to be done to just  control exposure. And then let somebody ask you [37:11] for access to something like Fable versus give it  to everybody and then see what happens. And then [37:16] and then be permissive. I mean it's generally  what we say if somebody asks it and they have [37:19] a reason give it to them. Um so that's a starting  point. And then um inside of engineering it's uh [37:26] the worst cases are people who are experimenting  with no mind toward cost and and you want to want [37:31] to be careful about how you control that because  you want people to experiment. And so in general [37:35] what we recommend companies do to control that  is to give um this group of employees, however [37:41] you define them, a flagship model budget that is  fixed. say whatever your number is, $1,000 a week, [37:45] $1,000 a day, that's your budget to use the  highest tier anthropic or or codeex models, [37:52] and after that's gone, you keep working because  there's nothing more expensive than an engineer [37:56] who can't work. I mean, that's more expensive than  all the tokens. Um, but you have to go back to the [38:00] standard models, which are still very capable. And  so that is an example of a gentle nudge that every [38:05] time a developer is about to touch their keyboard,  they have to think, do I use my fixed flagship [38:10] budget or do is this general work? And that little  nudge in the right direction is generally hugely [38:15] impactful from the organization's overall bill.  And then as we look at the difference between [38:19] those groups, usually that's how we bring the  top group down more towards the middle. What [38:23] about the training and upskilling of the talent?  How did it work? And and kind of how where do [38:29] what where's approach that to see the best most  success? I think everybody's looking for the the [38:35] magic solution to training and skilling people  on AI. Um but the reality is you you just have [38:40] to dive in. I mean that's really what we've seen  is um the the good news is at this point there is [38:46] someone in every group or every function who has  harnessed it and is like working magic with it [38:50] from the perspective of their peers and bringing  those people together to show what they're doing [38:55] with it and how it's impacting their workflow with  a team of you know everybody else who should have [38:59] the same workflow but doesn't that's the most  effective way we've seen to roll out knowledge [39:03] across groups so um things like AI happy hours  which are not in a bar they're at your desk not [39:08] that fun um but they're really effective and we  still do one inside of our company and all we do [39:14] all day is AI and every week we learn from them.  So, group specific meetups that are um informal, [39:20] moderated by someone, but usually a peer, not  a manager, are a super effective way to upskill [39:25] people. And then making sure for like something  we can do centrally as as leaders is measure [39:31] the skills that are most used across all of your  teams. And most of the most of the LLM providers [39:35] have telemetry that will tell you which of the  skills are most used. Find the novel use cases [39:40] and make sure other people are aware of those  and then just extending those out to your teams. [39:45] That's great. Like going back to the like a  measurement RARI measure. I think it's a it's not [39:51] just a budget question. It's it seems like kind of  hardly strategy. It's a kind of process question, [39:57] right? I mean like what is there what's a  the best practice that you have seen across [40:02] different companies? How how do you treat  that number and how can can how can we make [40:06] it continue to be relevant with with like  all the changing environment? What's really [40:12] interesting about the budget question is we we  have a continuum and we we try kind of we try to [40:17] put someone on this when we first meet them and  I've heard it already in some of the questions. [40:21] Part of the continuum is just driving adoption  like we're we want everybody to use it as much [40:25] as possible and then somewhere in the middle  there's this transition point. It's like oh [40:28] my gosh not that much. And so um we're trying to  help people drive adoption by identifying groups [40:35] of people that aren't using it and uplifting  them through things like we just talked about. [40:39] Um, but actually what's really surprising, and I'm  not sure if it's the case for people in the room, [40:43] is that when most companies we talk to say they  still have a relatively blank check from the board [40:50] to use AI because they're afraid of curtailing  innovation. Um, that's still the majority of [40:55] conversations we have. I see you nodding your  head. Yeah. However, I will say that because we [41:00] see the continuum, everybody gets past that point.  That is a point in time because what happens is [41:04] eventually employees become so capable at using  it. It's what you wanted, but it becomes a place [41:09] where you have to start asking questions about  what did we get? This person spent $10,000 this [41:14] week. Is that good or bad? And if you you have  to be able to answer that eventually. And so, [41:19] um, going back to measuring what they're doing  and just making sure as a it it comes down to [41:23] driving accountability down into managers just  like it always has being able to defend their [41:28] token use budget. But in order for managers to do  that effectively, there has to be some measurement [41:32] system in place so that uh they don't have to  manufacture an answer every time the CFO sends [41:36] a question down the chain. That's great. Yeah,  we have time for a couple of questions from the [41:40] floor. Do you have any questions? I mean, I'll  repeat the question just in case everybody didn't [41:46] hear it, but when we rolled out the AI happy hour,  how did it expand and how do people accept it? Um, [41:51] I think we're lucky. We're a startup, so we  have a lot of natural curiosity. However, [41:57] um, everybody showed up. It was surprising. It  started as an engineering thing. Uh sales and [42:01] marketing showed up, the CEO showed up and you  know, so um I usually ahead of time would ask [42:07] one or two people to have something to share. Um  and as it turned out that wasn't very hard because [42:12] every week somebody is experiencing something new  that feels like magic to them. Um so I always try [42:18] to curate each call with a couple of stories to  make sure and usually momentum carries it from [42:22] there. We always do them on Fridays, you know,  just because that's when people are the most [42:26] laid-back. We did the midm morning so that Europe  can pick it up in their afternoon. Um, and we can [42:31] do it in the morning. So, some people bring  beer, some people bring coffee. Um, but yeah, [42:36] that's how we do it. I think you'd be surprised.  People want to share the cool things they're doing [42:41] and if you give them a forum to do it, it kind of  takes off. That's our experience at least. Let me [42:46] just jump in on that because I mean it's it's  I'm at the other end of the spectrum, you know, [42:50] um 100 billion plus like, you know, 100,000  plus employee base, right? Um and and it's the [42:56] transformation is about the culture, right? And  so events like this. So there's there's different [43:01] pieces we put in which is we have guided events  where we actually are showing them things they [43:07] can do. We have cross-pollination events which is  inviting you know like white papers and speakers [43:13] on topics. And then we what we really look  for is organic acceleration. So we want to see [43:19] employees selforganize around you know lunches  and whatever things they want to organize and [43:24] and we facilitate that but ultimately the metric  or indicators that I look for is the amount of [43:30] organic activity that's happening. So everything  that happens otherwise when you're trying to do [43:35] it at scale is to just give enough momentum  that the organic activity starts. Once that [43:39] organic activity starts you know you've done  your job. Yeah. That's great. Thank you so [43:44] much for sharing your insight this morning.  Please give him a round of applause to the