Harvard CIO Says AI Automation Hits Only 18% of Workforce Tasks

哈佛CIO称AI自动化仅影响18%的工作任务 若缺乏人类判断与平台治理，企业AI将难以奏效，这意味着技术将重塑工作角色，而非消除问责机制或永久取代人类劳动力。 Rohit Jain指出，尽管AI模型在众多任务中展现出高水平的智能，但它们缺乏关键的人类判断力、专业责任感和伦理细微考量。他强调只有极少数工作面临完全自动化，并通过历史经验说明技术往往是推动职业角色的演进，而非抹杀劳动力需求。他援引实际案例研究和边缘极端情况，警示工具碎片化和未受监管的智能体部署风险，指明自主系统在脱离人类监督时可能出现的失效。最后，Jain敦促技术领导者将重点放在平台整合、营收增长上，并推行一项结构化的90天策略，将AI作为不带主观偏见的团队成员嵌入业务运营中。 AI模型在跨行业中具备极高的任务执行能力，但缺乏人类判断，因此需要一种由人类主导结果并指引方向的协同智能模式。 (0:08) 仅有18%的可重复性工作能被AI完全自动化，而46%的劳动力几乎不受任何影响。 (2:53) 岗位对AI的高暴露度并不等于责任的转移，因为持牌专业人士、看护人员以及依赖人际关系的岗位无法被自主智能体取代。 (2:53) 企业级AI落地存在巨大差距，92%的应用仍处于理论或试验阶段，尚未动摇核心劳动力结构。 (7:08) 历史研究表明，技术是在重塑工作角色而非永久消除劳动力，自一战以来全球仅有一个职业头衔被彻底淘汰。 (7:08)

Rohit Jain argues that sustainable enterprise AI requires a 'co-intelligent' framework where autonomous tools execute discrete tasks while humans retain ultimate judgment and legal accountability. Rather than pursuing narrow head-count reductions, organizations must bridge the gap between experimental pilots and operational execution through governed platforms and revenue-focused strategies. Judgment and Accountability as Automation Boundaries (0:08) Jain points out that while modern models possess PhD-level intelligence across 1,300 tasks in 44 occupations, they lack human judgment. Referring to OpenAI research, he highlights that only 18% of repeatable workforce tasks are susceptible to full automation, leaving 46% of the workforce essentially unaffected. Consequently, organizational anxiety around widespread replacement misinterprets exposure as a complete handover of responsibility. Accountability, Jain contends, fundamentally relies on professional licensing, legal liability, and human relationships that software agents cannot assume. Regulated professionals such as certified public accountants, attorneys, teachers, and surgeons remain protected from displacement because software cannot take on malpractice risk or build interpersonal trust. Drawing on his MedTech background, Jain notes that while robotic surgery and cardiac monitoring assist practitioners, the human clinician retains ultimate ownership of the outcome. Historical Resilience and the Shift to Skills (7:08) To contextualize fears of technological obsolescence, Jain cites a Harvard Business School study tracking job elimination from World War I through the post-COVID-19 era. Across all continents, the study found only a single occupational title permanently eliminated: the 'knocker-up,' who was hired to knock on windows prior to the invention of alarm clocks. Other historical positions, such as elevator operators or vehicle flag-bearers, evolved into modern safety and logistics roles, demonstrating that technological shifts redefine job descriptions rather than erasing labor entirely. Technological progress also continually creates new job taxonomies. Jain notes that the title 'data analyst' did not officially exist in US labor classifications until 2009, when accessible compute catalyzed demand for data engineering and data science roles. Today, AI is creating positions like 'forward-deployed engineers' and 'forward-deployed CIOs.' Because roles transform rapidly, Jain argues that enterprises increasingly operate in a 'skills commerce' rather than a 'titles commerce.' Operational Pilot Gaps and System Governance Pitfalls (11:53) Despite aggressive enterprise experimentation, Wharton research indicates that 92% of corporate AI adoption remains theoretical or locked in pilot stages, generating minimal frontline operational redesign. Jain references a joint Accenture and Wharton experiment simulating a manufacturing firm operated by 60 agents (35 digital and 25 physical); while utility agents handled routine operational workflows, the system broke down whenever ambiguous, high-level judgment calls were necessary. Navigating real-world edge cases presents similar challenges. Jain cites an autonomous Tesla scenario where an algorithm moving at 70 mph resolved an ethical conflict by choosing passenger safety over an animal crossing the road. Beyond edge cases, Jain cautions that unmanaged tool accumulation creates structural vulnerabilities, pointing to a $500 million publicly traded company using 175 to 185 disjointed AI tools. Fragmented adoption and poor data quality generate permanent organizational 'scar tissue,' distorting executive dashboards and undermining decision-making. Aligning AI Strategy With Revenue and Governance (10:05) Jain advises leaders to redirect AI investments away from defensive cost-cutting and toward top-line revenue expansion. For example, enterprises can pair $1.7 billion in internal productivity gains with $6 billion in revenue growth to preserve and redeploy human capital. While projected sector-specific efficiencies are substantial—reaching roughly 47% in software engineering and 41% in healthcare—realizing these gains requires methodical planning rather than disorganized tool rollouts. For enterprise technology executives, Jain warns against returning from industry conferences to deploy uncoordinated agents across operational units. He advocates for a bounded 90-day implementation framework focused on employee learning, platform consolidation, and direct alignment with profit-and-loss priorities. Jain concludes that enterprise AI success relies on consolidating disjointed tools into governed platforms and treating software as an un-opinionated teammate guided by human judgment. A persistent challenge remains ensuring that organizations maintain data integrity and clear ethical guardrails as autonomous systems handle increasingly complex trade-offs.

哈佛商学院校友会首席信息官（CIO）兼数字医疗公司副总裁罗希特·贾因（Rohit Jain）指出，当前人工智能虽能胜任大量专业任务，但本质上欠缺人类判断力与法定问责能力；企业应摆脱单纯追求成本削减的误区，转向以人机协同（co-intelligence）与业务增长为核心的战略重构。 判断力缺失与人机协同的定位 (0:08) 贾因援引OpenAI的研究指出，前沿人工智能模型虽具备博士级别的智力水平，能够处理跨越44个职业的1300多项任务，但其核心缺陷在于缺乏人类判断力。当前企业部署的AI代理大多仅侧重于流程效率，而真正的人机协同必须将人类判断深度嵌入业务流中。在这一框架下，模型训练与结果归属权必须由人类掌控，AI则作为不具个人偏见的协作伙伴辅助执行，而非拥有独立决策权的实体。 埃森哲与沃顿商学院的一项联合模拟实验证实了纯自主决策的局限：由60个AI代理（包含35个数字化代理和25个物理实体代理）运营的制造企业，在应对常规操作时运转良好，但在面临复杂判断时陷入停滞。贾因以特斯拉自动驾驶系统在时速70英里时的决策为例，算法在急刹转向危及车内乘客与撞击横穿道路的动物之间选择保护人类，说明伦理权衡必须经过明确训练。缺乏人类判断的算法若直接接管复杂决策，将给企业带来不可控的系统性风险。 自动化边界与职业问责的不可替代性 (2:53) 针对劳动力市场的技术替代焦虑，贾因引用OpenAI的数据说明，仅有18%的可重复性工作岗位可能被AI完全自动化，多达46%的劳动力几乎不受当前AI工具的实质影响。贾因强调，岗位的高技术暴露度并不等于问责权（accountability）的转移。法定资质认证、行业执照以及深层人际纽带构建了职业问责壁垒，这是自动化代理无法跨越的边界。 例如注册会计师、律师、教师和外科医生等角色，AI可以辅助完成大量计算或分析工作，却无法承担法律诉讼责任或建立信任关系。在医疗领域，即便机器人手术与心脏监测系统普及，临床决策仍由人类医生最终主导。哈佛商学院一项追踪一战至新冠疫情后的长期研究也显示，全球范围内被技术永久消灭的岗位仅有闹钟普及前的“敲窗人”（knocker-up），而电梯员或旗手等工种均演化为新型岗位。历史经验表明，技术演进重塑的是具体职责，而非整体抹除人类劳动。 落地鸿沟、工具过载与劣质数据风险 (7:08) 沃顿商学院的研究显示，当前企业AI的应用与实际业务重构存在巨大脱节，92%的部署仍停留在理论设想或试点阶段，并未改变一线业务形态。与此同时，部分企业陷入了无序采购的陷阱。贾因列举了一家市值5亿美元的上市公司，由于缺乏统一平台，内部堆积了175至185个分散的AI工具，造成治理割裂与决策失误。 数据质量的缺陷进一步放大了算法错误。贾因警告，持续输入低质数据会在企业系统内部沉淀为永久性的“伤疤组织”（scar tissue），导致管理层指标被持续误判。为扭转这种局面，企业必须将AI目标从单向缩减开支转向营收扩张，例如将17亿美元的效率收益与60亿美元的业务增长目标相结合。尽管预估软件开发和医疗健康领域的效率可分别提升约47%与41%，但这些红利必须依托统一平台治理与清晰的商业目标才能真正释放。 技能商业重构与90天执行路线图 (16:11) 技术进步持续催生新的组织角色与分工模式。贾因回顾指出，“数据分析师”（data analyst）直到2009年低成本算力普及后才正式被纳入美国职业分类体系，随后延伸出数据工程师与数据科学家等职能；当下的AI浪潮也正在催生“前沿部署工程师”和“前沿部署CIO”。现代组织的分工逻辑正在由固化的“职位商业”（titles commerce）转向动态的“技能商业”（skills commerce）。 针对CIO、CTO及CAIO等技术决策者，贾因明确告诫切勿在行业峰会后盲目引入数百个分散的AI代理，以免破坏既有的组织业务基座。领导者应当确立务实的90天落地规划：集中力量完成零散工具向统一平台的聚合治理，依托技能导向开展内部持续培训，并将AI技术部署与企业的损益表（P&L）优先目标紧密绑定，以确保技术投入转化为确定性的业务成果。 贾因强调，企业推进AI转型的成败取决于管理者能否在统一平台框架下维持严格的人类问责机制，而非盲目堆砌零散代理或推行防御性裁员。然而，关于软件工程提升47%与医疗领域提升41%等宏观效率预期，仍有待企业跨越92%的试点停滞期后在复杂业务场景中予以长期验证；如何在消除系统内“数据伤疤组织”的同时建立成熟的跨职能技能评估体系，仍是企业落实人机协同战略面临的现实难题。

## Background 该视频由 Ai4 于 2026 年 8 月 20 日发布在 YouTube 平台上。视频内容为哈佛商学院校友会 CIO Rohit Jain 主讲的个人演讲，旨在为技术领袖提供从 AI 实验走向大规模执行的实操指南，并将 AI 重构为嵌入工作流、具备平台级治理与业务价值衡量的企业基础设施。 ## TL;DR Rohit Jain 阐述了企业如何构建“协同智能”（Co-Intelligence）模型，强调 AI 虽然具备博士级智力但缺乏判断力，因此人类必须掌控方向与治理边界。演讲引用研究数据指出 AI 完全替代工作的比例仅为 18%，提醒企业应从单纯的“效率提升”转向“业务增长”。此外，演讲通过虚构 AI 公司实验及特斯拉决策案例剖析了 AI 的道德判断困境，并为 CIO 等技术领袖提出了规避工具碎片化、聚焦损益表（P&L）优先级的 90 天落地行动指南。 ## Atmosphere & style 整个演讲沉稳专业且极具启发性，充满领袖洞察力与务实干货。演讲者以冷静客观的数据和生动的现实案例打破了技术恐慌叙事，将原本抽象的技术治理转化为清晰的高管战略指南。听众能够切身感受到一种从盲目跟风理性回归业务本质的笃定感与使命感。 ## Detailed walkthrough 以下是根据演讲录音整理的详细内容记录与分析： ### [00:08](https://www.youtube.com/watch?v=zoHfM1Bq03E&t=8s) 重新定义 AI 时代：智力与人类判断力的协同 哈佛商学院校友会 CIO、某心脏监测数字医疗公司副总裁 Rohit Jain 明确指出，这场演讲并非纯粹的技术研讨，而是关注 AI 如何深度影响人类与组织。相较于过往关于大语言模型（LLM）等技术细节的讨论，AI 与人类的交互模式呈现出截然不同的动态特征。目前，LLM 虽然拥有博士级别的知识与智力（Intelligence），但严重缺乏“判断力”（Judgment）；人类固然拥有智力与判断力，但 AI 智能体（Agents）目前仅具备执行力而缺乏判断力。这种智力与判断力的分离，正是企业在部署 AI 时面临的核心课题。 在过去三天的会议中，“效率”是最常被提及的词汇，但企业绝不能仅仅停留在利用 ChatGPT 或 AI Agent 自动化执行任务的阶段。AI 智能体犹如没有判断力的团队成员，能够高效处理特定工作，但人类必须承担最终的责任与结果。未来发展的必然趋势是：LLM 将逐步被训练出更精准的判断力，而“协同智能”（Co-intelligence）的核心，正是如何在人类与 AI 之间建立正确的判断力框架与治理约束。 - **核心观点**：AI 拥有博士级智力但缺乏判断力，人类必须掌控最终业务结果并为其设定边界。 - **协同模型**：人类负责设定方向、护栏与评估取舍，AI 智能体在受限范围内高速执行具体任务。 --- ### [00:53](https://www.youtube.com/watch?v=zoHfM1Bq03E&t=53s) 人力资源重塑与“18%自动化”真相 针对企业员工普遍担忧的“AI 会抢走我的工作”这一恐慌情绪，Jain 引用了 OpenAI 针对 44 个职业中的 1,300 多项具体任务展开的研究。研究数据表明，全社会仅有 **18%** 的重复性工作可以被 AI 完全自动化。高达 **46%** 的工作岗位几乎不受影响，其余岗位则处于 12% 至 24% 等不同程度的受影响区间内。这表明世界并非正在走向大规模失业，恐慌叙事与现实相去甚远。 管理者在向团队推广 AI 应用时，必须清楚区分“曝光度/自动化程度”（Exposure）与“法律与业务责任”（Accountability）： ``` +----------------------------------------------------------------------+ | 工作岗位类型与 AI 协同模式 | +----------------------------------------------------------------------+ | [高度责任/牌照类岗位] | CPA 注册会计师、律师 | | | -> AI 提供辅助，但无法替代执业资格与法律责任| +------------------------+---------------------------------------------+ | [情感与关系驱动类岗位]| 教师、教授 | | | -> 疫情证实教学需人际引导，AI 仅作为教学工具 | +------------------------+---------------------------------------------+ | [物理交互与临床类岗位]| 医生、心脏外科医生 | | | -> 达芬奇机器人手术仍需医生操控，AI 无法替代 | +----------------------------------------------------------------------+ ``` - **数据支撑**：OpenAI 官方研究显示，完全可被自动化替代的重复性工作仅占总数的 18%。 - **历史视角**：哈佛团队研究了从一战（1950年代）到疫情结束期间的岗位变迁，发现历史上真正从地球上永久消失的职业仅有“敲窗人”（Knocker-up，过去缺乏闹钟时代人工敲窗催早起的人）。即便经历了工业革命与技术革命，大部分职业如电梯操作员、道路前导旗手等，均是以岗位职责转变的形式继续存在，而非被技术彻底抹杀。 --- ### [07:08](https://www.youtube.com/watch?v=zoHfM1Bq03E&t=428s) 突破效率瓶颈，以“业务增长”为核心驱动 沃顿商学院（Wharton）的研究表明，AI 的理论潜在能力与企业实际落地应用之间存在巨大鸿沟。高达 92% 的应用仍停留在理论与试点阶段，而底层受影响最小的 46% 基础岗位在实际工作中几乎未发生本质变化。AI 的应用浪潮虽然凶猛，但由于缺乏合理的判断力护栏与深度业务整合，目前尚未真正深刻改变企业的人力结构。 企业 CIO 与决策者必须将对话焦点从单纯的“提效（Efficiency）”转向“业务增长（Growth）”。提效只是增长的衍生结果，而非终极目标。 - **业务收益**：领先的企业实践表明，AI 能够带来高达 **60 亿美元** 的年营业收入增长，以及 **17 亿美元** 的生产力提升。 - **行业提效空间**：将数据深度融入业务流后，软件开发效率可提升约 **47%**，医疗健康领域的综合效率可提升约 **41%**。 --- ### [11:53](https://www.youtube.com/watch?v=zoHfM1Bq03E&t=713s) 虚构 AI 公司的实验与道德判断力困境 为了探索 AI 运行的极限，埃森哲（Accenture）与沃顿商学院共同进行了一项实验：他们建立了一家完全由 AI 运行的虚构制造企业。该公司的 CEO 及全体员工均为 AI，共部署了 60 个 AI Agent（其中 35 个数字 Agent，25 个物理 Agent）。实验表明，工具型 Agent 能够完美胜任所有日常例行事务，但一旦涉及需要进行“判断取舍”（Judgment Calls）的复杂决策时，系统便开始陷入混乱。 Jain 引用了现实中的案例来阐述判断力的复杂性：特斯拉在自动驾驶模式下行驶时，面对突发状况（如小狗穿过马路，且周围车道均有车辆高精度通行），系统在毫秒级时间内做出判断——若在时速 70 英里下紧急刹车，会导致车内整家乘客面临毁灭性车祸，最终系统做出了牺牲小狗、保护乘客的决策。这种在极端社会道德与安全约束下的决策逻辑，正是未来企业在训练 AI 模型与构建 Agent 时必须融入的“判断力框架”。 - **数据质量困境**：当前企业面临的最大瓶颈依旧是数据。“垃圾进，垃圾出（Garbage in, garbage out）”，错误的脏数据进入 AI 引擎后，只会在系统中不断放大错误结果。 - **避免工具碎片化**：Jain 举例某家年营收 5 亿美金的上市公司，内部竟然散乱堆叠了 175 至 185 种不同的 AI 工具。这种仅仅是“堆砌工具”而非“构建统一平台”的做法，必然导致 AI 判断力屡屡失效。 ``` +----------------------------------------------------------------------+ | 企业 AI 进化路线：平台化 vs 工具堆砌 | +----------------------------------------------------------------------+ | [堆砌工具依赖] 散乱引入 180+ 孤立 AI 工具 -> 数据割裂 -> 决策判断失效| | | | [统一平台架构] 构建底层数据与 AI 平台 -> 嵌入工作流 -> 可控判断与增长| +----------------------------------------------------------------------+ ``` --- ### [16:11](https://www.youtube.com/watch?v=zoHfM1Bq03E&t=971s) 技能革命与 CIO 90天落地行动指南 技术变革正在迅速颠覆传统的岗位定义。在 2009 年以前，美国官方职业分类中甚至不存在“数据分析师”（Data Analyst）这一岗位，而后随着算力普及与数据爆发，才衍生出数据工程师、数据科学家等一系列岗位。如今在 AI 时代，企业中迅速涌现出“前置部署工程师”（Forward Deployed Engineer）乃至“前置部署 CIO”（Forward Deployed CIO）。这表明，传统的“职位头衔”正在失效，取而代之的是“技能交易”（Skills Commerce）。 对于 CIO、CTO 及 CAIO 等技术领袖，Jain 提出了落地 AI 战略的实用建议： 1. **明确损益表优先级**：AI 项目必须紧密贴合损益表（P&L）的战略重点，拒绝无意义的单纯技术实验。 2. **持续的人才重塑**：通过突破性的领导力（Breakthrough Leadership）与清晰的战略意图，推动团队进行持续学习与技能再造。 3. **避免盲目扩张**：切忌在参加完 AI 峰会后便盲目宣布要在公司内部部署数百个新的 Agent，这会烧毁现有的组织信任与业务桥梁。 演讲最后，Rohit Jain 呼吁各位技术领袖重构框架，以增长为导向，打造兼具信任、治理与可衡量价值的“协同智能企业”，并欢迎与会者通过 LinkedIn 继续深入交流。 ## Notable quotes > [00:08](https://www.youtube.com/watch?v=zoHfM1Bq03E&t=8s) 这场演讲并非纯粹的技术研讨，而是关注 AI 如何深度影响人类与组织。 — Rohit Jain > [00:08](https://www.youtube.com/watch?v=zoHfM1Bq03E&t=8s) LLM 虽然拥有博士级别的知识与智力（Intelligence），但严重缺乏“判断力”（Judgment）；人类固然拥有智力与判断力，但 AI 智能体（Agents）目前仅具备执行力而缺乏判断力。 — Rohit Jain > [00:08](https://www.youtube.com/watch?v=zoHfM1Bq03E&t=8s) AI 智能体犹如没有判断力的团队成员，能够高效处理特定工作，但人类必须承担最终的责任与结果。 — Rohit Jain > [00:08](https://www.youtube.com/watch?v=zoHfM1Bq03E&t=8s) “协同智能”（Co-intelligence）的核心，正是如何在人类与 AI 之间建立正确的判断力框架与治理约束。 — Rohit Jain > [00:53](https://www.youtube.com/watch?v=zoHfM1Bq03E&t=53s) 全社会仅有 18% 的重复性工作可以被 AI 完全自动化。高达 46% 的工作岗位几乎不受影响。 — Rohit Jain > [00:53](https://www.youtube.com/watch?v=zoHfM1Bq03E&t=53s) 历史上真正从地球上永久消失的职业仅有“敲窗人”（Knocker-up）。大部分职业均是以岗位职责转变的形式继续存在，而非被技术彻底抹杀。 — Rohit Jain > [07:08](https://www.youtube.com/watch?v=zoHfM1Bq03E&t=428s) 企业 CIO 与决策者必须将对话焦点从单纯的“提效（Efficiency）”转向“业务增长（Growth）”。提效只是增长的衍生结果，而非终极目标。 — Rohit Jain > [11:53](https://www.youtube.com/watch?v=zoHfM1Bq03E&t=713s) “垃圾进，垃圾出（Garbage in, garbage out）”，错误的脏数据进入 AI 引擎后，只会在系统中不断放大错误结果。 — Rohit Jain > [11:53](https://www.youtube.com/watch?v=zoHfM1Bq03E&t=713s) 这种仅仅是“堆砌工具”而非“构建统一平台”的做法，必然导致 AI 判断力屡屡失效。 — Rohit Jain > [16:11](https://www.youtube.com/watch?v=zoHfM1Bq03E&t=971s) 传统的“职位头衔”正在失效，取而代之的是“技能交易”（Skills Commerce）。 — Rohit Jain > [16:11](https://www.youtube.com/watch?v=zoHfM1Bq03E&t=971s) 切忌在参加完 AI 峰会后便盲目宣布要在公司内部部署数百个新的 Agent，这会烧毁现有的组织信任与业务桥梁。 — Rohit Jain ## Mentioned Rohit Jain, Harvard Business School Alumni Association, Ai4, OpenAI, Wharton School of the University of Pennsylvania, Accenture, Tesla, ChatGPT, LinkedIn ## Video description Rohit Jain - CIO, Harvard Business School Alumni Association This 30-minute solo session gives technology leaders a practical playbook for moving from experimentation to execution. The session reframes AI as enterprise infrastructure: embedded in the flow of work, connected to core systems, governed at the platform level, and measured through business outcomes rather than activity metrics. It also introduces a human-led model of co-intelligence, where people set direction, guardrails, trade-offs and accountability while AI agents execute bounded work at speed. Attendees will learn how to evaluate AI platform readiness, reduce tool fragmentation, govern agentic workflows, prioritize high-value use cases, and redesign work without defaulting to fear-based workforce narratives. The session connects enterprise architecture, responsible AI, workforce transition and ROI into one CIO-ready framework. The key takeaway: CIOs will not win the AI era by buying more intelligence. They will win by designing the system that converts intelligence into trusted, governed and measurable enterprise value. Audience takeaways • A practical architecture for scaling AI beyond pilots • A governance model that embeds control into the platform, not just training • A workforce transition lens for automation, redesign and growth • A 90-day action plan with scorecards for adoption, value, quality, risk and workforce impact [Source](https://www.youtube.com/watch?v=zoHfM1Bq03E)

[0:08] Rohit Jain: 这场演讲并非纯粹的技术研讨，而是关注 AI 如何深度影响人类与组织。 [0:08] Rohit Jain: LLM 虽然拥有博士级别的知识与智力（Intelligence），但严重缺乏“判断力”（Judgment）；人类固然拥有智力与判断力，但 AI 智能体（Agents）目前仅具备执行力而缺乏判断力。 [0:08] Rohit Jain: AI 智能体犹如没有判断力的团队成员，能够高效处理特定工作，但人类必须承担最终的责任与结果。 [0:08] Rohit Jain: “协同智能”（Co-intelligence）的核心，正是如何在人类与 AI 之间建立正确的判断力框架与治理约束。 [0:53] Rohit Jain: 全社会仅有 18% 的重复性工作可以被 AI 完全自动化。高达 46% 的工作岗位几乎不受影响。 [0:53] Rohit Jain: 历史上真正从地球上永久消失的职业仅有“敲窗人”（Knocker-up）。大部分职业均是以岗位职责转变的形式继续存在，而非被技术彻底抹杀。 [7:08] Rohit Jain: 企业 CIO 与决策者必须将对话焦点从单纯的“提效（Efficiency）”转向“业务增长（Growth）”。提效只是增长的衍生结果，而非终极目标。 [11:53] Rohit Jain: “垃圾进，垃圾出（Garbage in, garbage out）”，错误的脏数据进入 AI 引擎后，只会在系统中不断放大错误结果。 [11:53] Rohit Jain: 这种仅仅是“堆砌工具”而非“构建统一平台”的做法，必然导致 AI 判断力屡屡失效。 [16:11] Rohit Jain: 传统的“职位头衔”正在失效，取而代之的是“技能交易”（Skills Commerce）。 [16:11] Rohit Jain: 切忌在参加完 AI 峰会后便盲目宣布要在公司内部部署数百个新的 Agent，这会烧毁现有的组织信任与业务桥梁。

Rohit Jain: CIOs will not win the AI era by buying more intelligence. They will win by designing the system that converts intelligence into trusted, governed and measurable enterprise value. [0:08] Rohit Jain: human-centric discussion rather than a pure technology call [0:08] Rohit Jain: un-opinionated teammate rather than an autonomous decision-maker [2:53] Rohit Jain: high job exposure does not equate to a transfer of accountability [7:08] Rohit Jain: Technology transforms job descriptions rather than permanently erasing human labor. [16:11] Rohit Jain: skills commerce rather than a titles commerce

[0:08] This is not going to be a technology call. [0:11] Let's be clear on that. This is more about how  does it impact humans. What Barry was talking [0:17] about was all technology and this is something  very different because it interacts with humans. [0:22] So, what My name is Rohit Jain. I am the CIO  for Harvard Business School Alumni Association. [0:29] In addition to that role, I'm a VP at a digital  health company that does cardiac monitoring. [0:35] I also sit on the board for the Alumni  Angels, the Harvard Alumni Angels, [0:40] and I'm an investor there as well. And I am so  happy that you all are spending time here just as [0:48] the conference is getting ready to close. And you  know, everyone's been talking about this forever [0:55] that the LLMs, they have PhD level knowledge,  intelligence, but do they have that judgment? [1:04] That's the question. What Barry was talking about  was all the intelligence part, but do humans have [1:10] that intelligence? They do. Do humans have the  judgment? They do. But do agents have? They don't. [1:17] And that's the whole dynamics here that we  are dealing with. That when you think of it, [1:23] a chat GPT model can do tremendous amount of  work. And there was all this research done [1:30] over 1,300 tasks spanning 44  occupations. And they could do well. [1:38] But they could do it without judgment.  And that is the whole thesis of this talk. [1:46] What's been happening till now? We've got  agents that keep running. They execute a [1:51] task. We heard about efficiency. We've been  hearing about efficiency for the past 3 days. [1:57] But where does the judgment part sit in?  The humans have to train the models to [2:03] have the right judgment. And over a period of  time, what is really going to happen is this. [2:11] LLMs will start having judgment as well.  And that's when the real things kick in. [2:18] It's not only about efficiency  that we need to be thinking about. [2:22] We need to be thinking about how the  judgment works. And that's core intelligence. [2:28] We've got companies now that are thinking about  agents in that HR organization. You've got people, [2:35] you've got agents. When you look at an HR  organization, you see them sitting there. But, [2:41] it's a teammate without a judgment. Humans own the  outcomes. Agents do the work. And that's the goal. [2:53] Look at these numbers. [2:57] And I'll let that sit in for a second. Who  created the study? OpenAI. And what did they say? [3:05] They said 18% of the jobs that are  repeatable can be automated using AI. [3:13] It's not that the world is falling apart. It  is not. Here's a company that's doing this [3:19] that we believe are the biggest  disruptors, and that's what they quote. [3:25] So, when you think of the human layer working  with the agents, which jobs get automated first, [3:33] which are repeatable? It's only 18%. You've  got this big huge chunk of 46% of the workforce [3:41] that doesn't get impacted. And in between,  you've got the range between 24%, 12%, and so on. [3:48] I've provided a link there. Have have at that. Uh  when you all get a chance, I don't know if these [3:54] things are going to be online or not. So, but  this is one thing that I really want you all to [3:59] take back to your workplaces. When you work with  people and you say, "I am deploying an AI agent." [4:08] The first thing is, "Am I going to  lose my job?" And the answer is no. [4:12] It's not true. That's not what really  happens in a regular workplace. [4:19] And why is that? [4:24] When a job is automated, you've got exposure.  We looked at that 18% number, right? [4:31] So, if that 18% of the job and workforce  is automated, is there an accountability? [4:40] The exposure is not equal to accountability. And  that's what we want to be thinking of as leaders [4:46] or as technologists or people who  are working within the AI ecosystem. [4:51] That's the key thing that we need to be thinking  of. Because when you think of accountability, [4:59] the you need some licenses. For example,  you've got CPAs and so on. You need licenses. [5:05] Can an agent help a CPA? Sure, an agent can help  a CPA. Can an agent help an attorney? Of course, [5:11] it can. But it cannot replace an attorney,  at least in the foreseeable future. [5:17] In the middle, you've got  the relationship workforce. [5:22] All of those pieces, they  are relationships. Teachers. [5:27] When the pandemic hit, what was the first  thing that really turned on its head? I come [5:34] from an education background in the education  industry. Since the 15th century, 16th century, [5:41] you had a teacher {slash} professor standing  teaching a classroom. And the pandemic hit, [5:49] everything turned on its head. Whether it was an  elementary school or a grad school, everything was [5:54] online. But still, you needed a professor or  a teacher to walk the classroom through that. [6:02] That's that physical layer. Can agents help the  students and the teachers? Of course they can. [6:07] And then you've got the physical  jobs, where you need doctors, nurses. [6:15] The agents help them. Agents don't replace them.  Like I said, I'm the VP at a MedTech company. [6:22] We make cardiac monitors, but we don't replace  cardiac surgeons. You have robotic surgery, [6:29] but that robotic surgery is still being  driven by a human. When you think of agents [6:34] and co-intelligence, that's the layer I want  everyone to think about when you go out from here. [6:39] Don't focus only on that 18% which was there on  the previous slide. Please think across the board. [6:54] There's a gap, really, when you think of it. And  I you know, the text that's coming on is really [7:00] taking away the links, but all of these things are  backed by research. This research is from Wharton. [7:08] There's a gap between what AI can really do [7:12] and what it is actually doing. This is one slide  again. Just please take it back to your teams. [7:21] And put this as an overarching model when you all  are working with your teams, with your customers, [7:28] because if you are a service provider, you're  going and talking to a customer, and that's the [7:32] biggest pushback that you get. What happens  to my workforce? And when you really see of [7:38] look at this, organizations that have actually  adopted they still have room for growth. [7:46] There's 92% of it is just theoretical  AI. That's the topmost row, right? [7:54] And on the bottommost which goes back to that  46% anchor that I had there's hardly any change. [8:03] Everyone really needs to think of this because  AI adoption it's tremendous. It's happening. [8:11] It's real. It's an avalanche. It is coming. But it  is not impacting the workforce. It's not impacting [8:19] humans. And it wouldn't because we need to get  into this mode where we have the judgment. We [8:24] have the guardrails for AI. I want to digress  a bit from my presentation. After the pandemic [8:33] we did a study at the school and we said from the  First World War in 1950 to the end of the pandemic [8:42] how many jobs were eliminated as in that  job doesn't exist in the world anymore [8:49] as a job description. Nothing  exists anywhere in any continent. [8:53] Any guesses from this audience? How  many jobs were eliminated forever? [9:01] One. Just one job. And that job is called  upper knocker. So before the First World War, [9:11] people didn't have alarm clocks and so on.  They paid people to come in and knock on their [9:16] window to be woken up at sunrise at 6:00 a.m. or  whatever. That's the only job that is eliminated. [9:24] Imagine that. We've had the industrial  revolution. We've had the technology revolution. [9:30] Everything else exists. Elevator  operators, they don't exist here, [9:34] but they exist in other parts of the world.  When the cars were on the main roads, [9:39] you had a person walk with a flag in front of  the car telling that there's a car coming behind. [9:44] You still have that job description when  you go to the Home Depots of the world, [9:48] you've got someone walking around with a flag  saying there's a truck coming behind. The point is [9:55] technology comes, jobs get transformed.  They don't get eliminated forever. [10:05] And this is the main thing that we want to think  about. Growth is what everyone's looking for, [10:11] not efficiency. Efficiency is a factor of growth.  Most of the conversations that we have with AI is [10:19] it is efficiency. No, we need to be  thinking about growth at this stage. [10:24] How do we grow our companies? How do we grow our  enterprises? 6 billion in annual revenue growth [10:31] plus 1.7 billion in productivity gains.  Okay. That's good. But how do you make it [10:40] it without impacting humans? That's  the thing that we have to think about. [10:44] It's not about more output.  It's not only about efficiency. [10:54] We looked across industries.  Let me be more specific, right? [10:59] Working across multiple industries and  specifically this one's focused on MedTech and [11:03] the healthcare industry. If anyone from the life  sciences world here, this one is for you because [11:07] that's my focus. That's where I invest as well.  Software efficiency can go up by around 47%. [11:17] Healthcare uh efficiency can go up around 41% when  you marry all the data that I've been presenting [11:24] all till now. That's the end state that we are  looking at. You are still left with a big chunk [11:30] where humans have still got to play a role. And  this is why planning and AI rollout matters. [11:38] It kind of goes back to the previous  presentation and the one before that as well. [11:42] That it's not a question of just rolling out AI, [11:45] it's the business impact that you're  trying to get and what's your end state. [11:53] So, let's get practical. Accenture and  Wharton, they created a fictitious company [12:02] using AI. The CEO is AI. It's got a complete  staff of AI. It's a manufacturing organization. [12:11] Does business and they let it run. And it had  around 60 AI agents. 35 digital, 25 physical. [12:23] And how did that organization run? Utility  agents, they're able to handle routine work. [12:29] We all know that. But when the question  came, they had to make a judgment call. [12:36] That's where the issues started. Actually, just  3 weeks ago, uh test and I don't know how many [12:43] of you read this in the news. Tesla was uh on  a self-driving mode with a family in the car. [12:53] There was one dog crossing  the [clears throat] street. [12:57] And the car had to make a judgment call.  Kill the dog or because of the speed, [13:04] something would happen to the passengers.  It did the right judgment. It killed the dog [13:10] because if it would have braked at 70 mph, there  would have been problems for the passengers. [13:17] That's the judgment that we have to think of  for AI. Is it sad for the dog? Of course, it is. [13:24] But then, as a human, that's what we would  have done. Because you can't change your lanes, [13:29] there's traffic all across, and there's a  dog right in the front. What would you do? [13:35] And that's where AI needs to start moving into.  And that's how we train our models moving forward. [13:46] I'll skip this slide. It's a lot more Medicare,  Medicaid, and the healthcare industry. But, [13:52] think of the numbers that we're  spending in here. $570 billion is [13:58] Is that something that we can change? Of course,  we can. But, why is there resistance to change? [14:05] That's what we, as people dealing in the  AI world, need to think about and go solve. [14:12] And yes, this was touched, I think, in almost  every presentation that I have sat through today. [14:17] Data is the problem. Cannot repeat that  anymore. I think we've all got scar tissues [14:25] uh with bad data for producing bad results, and  you push it into AI, it's producing bad results [14:31] forever. You can't go clean that up. So, that's  something that we need to think this as we start [14:38] walking through our journeys with AI. How do you  build judgment? How do you get clean data? And [14:46] we talked of agents in the previous call, uh  previous session. And I'm sitting there and [14:51] thinking that, okay, if a agent gave me the  right answer right at that time, I'm going to [14:57] interpret it the way I want. However, if I've  got another teammate who got the same answer, [15:02] but is dealing with a different set of parameters,  how would they interpret that answer? How does [15:08] that judgment layer of an output coming out from  an agent with incorrect data, because there was [15:15] a relationship to the dashboard, how does it work  out? It will not work out, and that's why getting [15:21] these building blocks changing your organizations  the right way is very very critical for success. [15:30] Are you building a platform where  your organizations are going to change [15:35] or are you accumulating tools? I know of a  company it's a mid-size company around 500 million [15:44] publicly traded. They've got 175 or  185 AI tools. And I'm sitting there [15:53] and saying, "Seriously, you're not building  a platform. You're not collecting agents. [15:57] You're just accumulating tools." And  therefore the AI judgments start failing. [16:11] I skipped a couple of slides  in the interest of time. [16:16] When you think of a data analyst role that did  not exist in the US job market definition at all [16:24] till 2009. That was when it was created for the  first time. Why? Because compute became easier, [16:32] data was readily available, a data analyst  role got created, data engineer and data [16:38] scientist and so on and so forth. Come AI,  you suddenly see forward deployed engineer. [16:45] That's a new role that suddenly has gotten  created. And now with a forward deployed [16:49] engineer, you've got a forward deployed CIO  as well. I've seen those roles coming along. [16:54] It's the not the titles that are mattering, it's  actually the skills that are more important now. [17:00] And that's what we want to think. It's the  skills commerce, it's not the titles commerce. [17:06] As new tools and new technologies roll out  those are the things that will matter most. [17:17] Talent reinvention is a big huge deal and you  really want to encourage your teams and the [17:25] people you work with, whether you're customers,  your teams, whoever you lead or you work with. [17:32] You have to have breakthrough leadership  and that comes with clarity. What are [17:36] you trying to achieve? It's a little bit of a  theoretical piece, but it's very very important [17:42] for everyone to think about what are the most  critical factors for you to retain your talent. [17:53] My last slide and then I can pause. I still have  further slides, but I'll just pause at this one. [17:59] If there are CIOs, CTOs, CAIOs in  this room, this one is for you all. [18:05] How do you put it all together? [18:09] Jobs are not going. They are going [18:11] to get transformed depending on  where you are in the hierarchy. [18:17] Set X plus it P&L priorities. That is  very very critical. And most importantly, [18:26] evaluate your talent and your team throughout. And  how do you do that? You have to embrace continuous [18:32] learning. That is very very very critical for  every organization and every team to grow. [18:41] Here's a plan for you all to use when you  go back. Don't just go and say, "All right, [18:47] I came back from AI 4 conference. We  are going to have 600 new agents." No. [18:54] You'll really burn every known bridge. And just  take these things when you leave from here. [19:02] And with that, I'll pause. I've got a minute for  questions. I really ran through these slides. [19:08] And if you want more information,  connect with me on LinkedIn.