转：麦肯锡报告 Agents, robots, and us: Skill partnerships in the age of AI
如何重构人机协作关系：未来的工作，就是你和AI一起，把这项技能发挥到极致
个人理解
如何重构人机协作关系：未来的工作，就是你和AI一起，把这项技能发挥到极致
人类专注判断与情感 -- 过多的依赖工具，人会不会变的更懒了，除了提问不会更多的表达
理论上可以自动化美国经济中约57%的工作时间 -- 中美差距在哪里？中国经济能自动化的比例？也许还没现代化就自动化了
重新设计工作流程，让智能体处理认知任务，让机器人处理物理任务，让人类负责判断和统筹 -- 判断和统筹，对技能要求更高
数字世界里的劳动力，处理非物理工作；物理世界里的劳动力，处理更复杂的物理交互
人类经过数百万年进化的双手和大脑，依然是目前最高效的解决方案；未来的工作不会是机器全面接管，而是一场精细的分工
需要同理心、需要复杂运动控制、以及需要对最终结果负责的关键环节
腾出手来做更有创造性的事 -- 从安步就搬到创造，人类需要适应什么？
纯粹靠单一技能一招鲜吃遍天的时代结束了，技能本身大多不会消失，消失的是旧的使用方式
沟通、管理、运营、解决问题、领导力、注重细节、客户关系、写作
真正发生剧变的，是对AI流利度（AI Fluency）的需求：能够理解AI、使用AI、管理AI，并对其输出结果进行批判性判断的能力
彻底重构工作流（Workflows）-- AI的加入，新工具的引入，过程要素不变，势必引起工作流程的变化
最艰难的决定不是选哪个大模型，而是如何安置被释放出来的人力
未来的世界，知识获取的成本几乎为零 -- 理解、运用、创新？如果只是单纯的用，更深入的发展将如何进行？
麦肯锡报告 Agents, robots, and us: Skill partnerships in the age of AI
来自 百度搜索
AI导读
"自动化将重塑57%的工作时间，但这不是失业危机，而是2.9万亿美元增长的机会。未来属于人机协作：智能体处理44%认知任务，机器人分担13%体力劳动，人类专注判断与情感。掌握AI流利度的员工将成为最稀缺资产，就像指挥家编排人机交响乐团——旧岗位消失的同时，新角色正以更高效的方式涌现。"
现有技术理论上已能自动化美国当前57%的工作时长，但这并非失业的预警，而是通向2.9万亿美元经济增量的新入口。11月25日，麦肯锡全球研究院（MGI）发布了一份足以重塑我们对未来工作认知的报告。Agents, robots, and us: Skill partnerships in the age of AI
这份报告剥离了关于AI会夺走工作的恐慌情绪，将目光聚焦于一个更具建设性的现实：生产力的前沿正在被拓展，工作的本质正在从单纯的人力劳动，转变为人、智能体（Agents）与机器人（Robots）的深度协作。
我们正处在一个十字路口。一边是技术的指数级进化，另一边是人类技能的缓慢迭代。这中间的缝隙，就是未来十年最大的机会所在。
自动化不是终点而是起点
在这个被算法包裹的时代，很多人盯着替代两个字看，却忽略了协作带来的巨大杠杆效应。麦肯锡的研究数据非常直观：以目前已经展示的技术能力，理论上可以自动化美国经济中约57%的工作时间。这个数字听起来惊人，甚至带着一丝寒意。大家会想，如果一半的工作都能被机器干了，人去哪儿？这种线性的思维方式忽略了技术采用的复杂周期。从实验室里的技术可行性，到企业里的规模化应用，中间隔着成本、监管、组织惯性等无数道门槛。电力的普及用了30多年，工业机器人的推广也经历了漫长的数十年。即便是云计算，到了2023年也只有五分之一的公司将大部分应用搬到了云端。这57%的自动化潜力，代表的是一种可能，而非明天的必然。它预示着我们将从繁琐、重复的任务中解脱出来。在这个过程中，有些角色会消失，有些会收缩，但更多的新角色会像雨后春笋般涌现。真正的变革在于，到2030年，如果我们能够重新设计工作流程，让智能体处理认知任务，让机器人处理物理任务，让人类负责判断和统筹，美国每年将能释放出约2.9万亿美元的经济价值。这笔巨额财富的钥匙，不掌握在单纯购买AI软件的公司手里，而掌握在那些懂得如何重构人机协作关系的组织手中。
为了看清未来的工作图景，我们需要厘清两个概念：智能体和机器人。在过去很长一段时间里，机器是死板的。它们只能做你编程让它做的事，一步都不能差。现在的AI改变了规则。通过海量数据的喂养，机器学会了举一反三。
智能体，是数字世界里的劳动力。它们处理非物理工作，比如起草法律文书、分析财务报表、甚至进行复杂的客户沟通。它们不仅能听懂自然语言，还能理解上下文，模拟人类的推理过程。
机器人，是物理世界里的劳动力。它们不再局限于在那儿拧螺丝，而是开始尝试处理更复杂的物理交互。目前美国三分之二的工作时间属于非物理工作。这其中，有三分之一高度依赖人类的情感和社会技能，比如护士的抚慰、教师的引导、领导者的激励，这些是AI目前难以触及的高地。但剩下那部分涉及推理和信息处理的工作，正是智能体大显身手的地方。这部分工作对应了美国工资总额的40%，覆盖了从教育、医疗到商业法律的广泛领域。
物理工作的自动化则相对缓慢。虽然波士顿动力（Boston Dynamics）的视频让人惊叹，但现实是，大部分物理工作需要极高的灵巧度和环境感知力。在这方面，人类经过数百万年进化的双手和大脑，依然是目前最高效的解决方案。所以，未来的工作不会是机器全面接管，而是一场精细的分工。智能体可以拿下44%的工作时间，机器人可以分担13%。剩下的，必须由人来完成。这包括了那些需要同理心、需要复杂运动控制、以及需要对最终结果负责的关键环节。
七种职业原型重塑职场版图
基于这种分工，麦肯锡将工作岗位划分成了七种原型。这不仅仅是分类，更是一张职业发展的藏宝图。
- 最左端是以人为中心的职业。这部分工作约占美国工作岗位的33%，平均年薪71,000美元。医疗保健、建筑维护属于此类。无论AI怎么进化，它很难替代一个护士在病床前的关怀，也很难替代一个维修工在复杂管道井里的灵活操作。这类工作不仅安全，而且因为稀缺性，其价值可能会进一步凸显。
- 最右端是以智能体为中心和以机器人为中心的职业。这部分占到了总岗位的40%。法律助理、行政专员这些高度依赖认知任务的角色，正处于被智能体大规模改造的风暴眼。它们并不是完全消失，而是性质变了。未来的行政专员可能不再是自己在写邮件，而是指挥三个AI助手在处理邮件。还有一小部分是司机、机器操作员，虽然理论上可以被机器人替代，但考虑到成本和现实路况的复杂性，人类依然会长时间留在驾驶座上，只是更多地扮演监督者的角色。
- 最有意思的是中间地带。这里有人-智能体的混合角色，比如教师、工程师、金融专家。AI是他们的外骨骼，帮他们处理数据、批改作业，让他们腾出手来做更有创造性的事。还有人-机器人的混合角色，比如建筑工人和维护人员。机器提供力量和精度，人提供判断和灵巧。以及人-智能体-机器人的全能组合。在现代化的农业、物流和食品服务中，这种组合已经出现。一个农场主，通过iPad指挥着智能体分析气象数据，同时调度自动驾驶拖拉机进行播种，他自己则在田间处理突发状况。
不管你身处哪个象限，纯粹靠单一技能一招鲜吃遍天的时代结束了。很多人担心自己的技能会过时。其实，技能本身大多不会消失，消失的是旧的使用方式。雇主现在要的东西越来越刁钻。十年前，一个岗位平均只需要54项技能，现在涨到了64项。越是高薪的岗位，要求的技能越复杂。数据科学家的招聘简章里，密密麻麻列了90多种技能。但在这眼花缭乱的变化中，有一套硬通货始终坚挺。沟通、管理、运营、解决问题、领导力、注重细节、客户关系、写作。这八大核心技能，像万能插头一样，在各行各业都通电。真正发生剧变的，是对AI流利度（AI Fluency）的需求。在过去两年里，美国招聘信息中对AI流利度的需求暴涨了七倍。这比任何其他技能的增长都要迅猛。这不仅仅是要求你会写两行Python代码，或者是会用ChatGPT生成一段文案。它指的是一种能够理解AI、使用AI、管理AI，并对其输出结果进行批判性判断的能力。目前，这种需求还主要集中在计算机、管理和金融领域。但趋势像水一样，正在向四面八方渗透。流程优化、质量保证、教学指导，这些原本与AI不沾边的技能，现在都开始要求与AI挂钩。反观那些机器已经很擅长的技能，比如基础的研究、常规的写作、简单的数学计算，在招聘信息中的提及率正在下降。这不代表这些技能没用了，而是它们变成了默认配置，或者是被集成到了工具里。
这就引出了麦肯锡提出的技能变化指数（Skill Change Index）。
如果你是做发票处理的，或者只会特定语言的简单编程，你的技能处于高风险区，必须尽快升级。
如果你是做质量保证的，你的技能处于中风险区，你需要学会如何用AI来做质保，而不是自己盯着屏幕找茬。
如果你是做教练、做心理咨询的，你的技能处于低风险区，因为人心的复杂，AI暂时还搞不懂。
到2030年，在技术采用的中性情景下，最抢手的100种技能中，约有四分之一到三分之一的工作时间将被自动化。如果技术跑得再快点，这个比例会更高。这里面有一个关键的认知转变：大多数技能（约72%）是跨界的。它们既用于机器能干的事，也用于机器干不了的事。未来的工作，就是你和AI一起，把这项技能发挥到极致。
不要用AI去做旧流程
有了技术，有了技能，为什么很多企业觉得AI没什么用？因为它们在用牛刀杀鸡。很多公司引进AI，只是把它当成一个更快的打字机，或者更聪明的计算器，用来优化现有的、陈旧的工作流程。这种打补丁式的做法，哪怕投再多钱，收益也微乎其微。真正的爆发力，来自于彻底重构工作流（Workflows）。工作流是企业运转的骨架。它是信息流转、决策制定的一整套链路。麦肯锡分析了美国经济中的190个业务流程，发现60%的潜在收益集中在那些行业特定的核心领域。比如制造业的供应链管理、医疗的临床诊断、金融的风险控制。
让我们看看那些真正玩明白了的企业是怎么做的。
销售：从广撒网到精准狙击
一家全球科技公司，以前的销售模式是人海战术。销售人员每天花大量时间在茫茫多的客户名单里筛选，打电话，被拒绝，再打电话。真正能坐下来跟客户深聊的时间少得可怜。引入AI智能体后，流程被彻底颠覆了。一个优先级智能体负责在后台分析数据，给潜在客户打分排名。一个外联智能体负责发送第一封破冰邮件。一个客户响应智能体负责处理回复，把那些没意向的、甚至还在犹豫的线索自动归类处理。只有当客户表现出明确的兴趣，或者提出了复杂的问题时，这些智能体才会把接力棒交到人类销售专家的手里。这相当于给每个销售配了一支不知疲倦的助理团队。结果是，人类销售不再是推销员，而是变成了谈判专家和关系建立者。他们把时间和精力集中在起草个性化提案、解决客户疑虑上。这直接带来了7%到12%的年收入增长，销售人员从琐事中抢回了30%到50%的时间。
客服：把情绪价值留给人
一家大型公用事业公司，每年要接700万个求助电话。以前的自动语音系统（IVR）笨得要命，只能解决10%的问题，剩下的全靠人工硬扛。客户体验差，员工压力大。新的代理式AI系统上线后，情况变了。身份验证智能体先确认你是谁；意图识别智能体搞清楚你要干嘛；调度智能体帮你约时间；自助服务智能体直接连通后台帮你改套餐、查账单。这一套组合拳下来，40%的电话完全不需要人介入就能解决。更重要的是，解决率超过了80%。那些真正需要人接听的电话，往往是复杂的、情绪化的，甚至是客户正在气头上的。这时候，人类客服代表接手，看到的不再是一个冷冰冰的号码，而是屏幕上已经整理好的客户背景、问题描述和情绪预警。他们可以专注于安抚客户、解决疑难杂症。平均每次通话成本降低了一半，客户满意度反而提升了6个百分点。这就是把机器的效率和人的温度结合的典范。
医药：让新药跑赢死神
在制药行业，时间就是生命。研发一款新药，光是写临床研究报告，就要耗费大量医学专家的时间。他们要整理成吨的数据，起草冗长的文档，还要经过无数轮的合规审查。一家全球生物制药公司决定用AI来啃这块硬骨头。他们开发了一个AI平台，专门负责写初稿。这个智能体能从结构化和非结构化的数据中提取信息，几分钟内生成一份格式规范、数据详实的报告草稿。它甚至能自己检查错误。医学专家的角色变了。他们不再是码字工，而是变成了主编和审稿人。他们利用自己的临床判断来验证AI写的内容对不对，逻辑顺不顺。效果立竿见影：初稿的人工接触时间减少了近60%，错误率下降了50%。这意味着新药上市的流程可以缩短好几周。对于那些等待救命药的患者来说，这几周可能就是生与死的距离。
IT现代化：旧代码的重生
银行的IT系统往往是活化石。有些核心系统用的代码语言，可能比现在的程序员岁数都大。要更新这些系统，以前需要几十个工程师没日没夜地干好几个月，还容易出错。一家区域性银行尝试让AI智能体来干这活。评估智能体先去扫描那些古老的代码库，理清里面的依赖关系；功能智能体设计新的架构；编码智能体负责把旧代码翻译成新语言，并自动运行测试。在这个过程中，人类开发者变成了指挥官。他们每个人指挥着15到20个智能体，审核它们写出的代码，确保架构的完整性。代码准确率达到了70%，人工工时减少了一半。这不仅是效率的提升，更是让银行从沉重的技术债务中解脱出来，能够更快地推出新的金融产品。
这些案例都在传递一个信号：AI转型，不是买个软件那么简单。它是一场彻头彻尾的组织变革。如果你是管理者，你现在的任务不是盯着员工有没有在摸鱼，因为那些能摸鱼干完的活儿，AI大概率都能干。你的任务变成了编排（Orchestration）。你要像指挥交响乐团一样，指挥由人、智能体和机器人组成的混合团队。你要决定哪个环节交给机器，哪个环节必须留给人。你要建立信任机制，确保机器不出乱子，确保数据不带偏见。你还要建立一种容错和实验的文化。在早期，AI肯定会犯傻，智能体肯定会出错。如果一出错就叫停，那永远也迈不出第一步。对于企业来说，最艰难的决定不是选哪个大模型，而是如何安置被释放出来的人力。是单纯地裁员以降低成本？还是把这些人投入到更高价值的创新工作中去？聪明的企业会选择后者。因为AI可以复制，但那些懂业务、懂客户、懂情感、并且掌握了AI流利度的员工，才是无法被复制的核心资产。
教育体系也面临着大考。我们的学校还在教孩子死记硬背吗？未来的世界，知识获取的成本几乎为零。真正稀缺的是批判性思维，是提出好问题的能力，是能够一眼看穿AI一本正经胡说八道的能力。从小学开始，我们就应该培养孩子的AI流利度。这不只是技术课，更是逻辑课、伦理课。
社会机构、政府、工会，都需要行动起来。我们需要建立更灵活的技能认证体系，让人们在不同的职业赛道之间切换变得更容易。我们需要社会安全网，托住那些在转型期暂时掉队的人。
未来不是人与机器的战争，而是人利用机器去探索更高维度的创造力。在这个新时代，最危险的不是AI，而是抱着旧地图找不到新大陆的我们。
参考资料：
https://www.mckinsey.com/mgi/our-research/agents-robots-and-us-skill-partnerships-in-the-age-of-ai#/
AI is expanding the productivity frontier. Realizing its benefits requires new skills and rethinking how people work together with intelligent machines.
Agents, robots, and us: Skill partnerships in the age of AI
Full Report (60 pages)Appendix (PDF-3 MB)
At a glance
- Work in the future will be a partnership between people, agents, and robots—all powered by AI. Today’s technologies could theoretically automate more than half of current US work hours. This reflects how profoundly work may change, but it is not a forecast of job losses. Adoption will take time. As it unfolds, some roles will shrink, others grow or shift, while new ones emerge—with work increasingly centered on collaboration between humans and intelligent machines.
- Most human skills will endure, though they will be applied differently. More than 70 percent of the skills sought by employers today are used in both automatable and non-automatable work. This overlap means most skills remain relevant, but how and where they are used will evolve.
- Our new Skill Change Index shows which skills will be most and least exposed to automation in the next five years. Digital and information-processing skills could be most affected; those related to assisting and caring are likely to change the least.
- Demand for AI fluency—the ability to use and manage AI tools—has grown sevenfold in two years, faster than for any other skill in US job postings. The surge is visible across industries and likely marks the beginning of much bigger changes ahead.
- By 2030, about $2.9 trillion of economic value could be unlocked in the United States—if organizations prepare their people and redesign workflows, rather than individual tasks, around people, agents, and robots working together.
Introduction
Work in the future will be a partnership between people, agents, and robots—all powered by artificial intelligence. While much of the current public debate revolves around whether AI will lead to sweeping job losses, our focus is on how it will change the very building blocks of work—the skills that underpin productivity and growth. Our research suggests that although people may be shifted out of some work activities, many of their skills will remain essential. They will also be central in guiding and collaborating with AI, a change that is already redefining many roles across the economy.
In this research, we use “agents” and “robots” as broad, practical terms to describe all machines that can automate nonphysical and physical work, respectively. Many different technologies perform these functions, some based on AI and others not, with the boundaries between them fluid and changing. Using the terms in this expansive way lets us analyze how automation reshapes work overall.1
This report builds on McKinsey’s long-running research on automation and the future of work. Earlier studies examined individual activities, while this analysis also looks at how AI will transform entire workflows and what this means for skills. New forms of collaboration are emerging, creating skill partnerships between people and AI that raise demand for complementary human capabilities.
Although the analysis focuses on the United States, many of the patterns it reveals—and their implications for employers, workers, and leaders—apply broadly to other advanced economies.
We find that currently demonstrated technologies could, in theory, automate activities accounting for about 57 percent of US work hours today.2 This estimate reflects the technical potential for change in what people do, not a forecast of job losses. As these technologies take on more complex sequences of tasks, people will remain vital to make them work effectively and do what machines cannot. Our assessment reflects today’s capabilities, which will continue to evolve, and adoption may take decades.
AI will not make most human skills obsolete, but it will change how they are used. We estimate that more than 70 percent of today’s skills can be applied in both automatable and non-automatable work. With AI handling more common tasks, people will apply their skills in new contexts. Workers will spend less time preparing documents and doing basic research, for example, and more time framing questions and interpreting results. Employers may increasingly prize skills that add value to AI.
To measure how skills could evolve, we developed a Skill Change Index (SCI), a time-weighted measure of automation’s potential impact on each skill used in today’s workforce. Nearly every occupation will experience skill shifts by 2030. Highly specialized, automatable skills such as accounting and coding could face the greatest disruption, while interpersonal skills like negotiation and coaching may change the least. Most others, including widely applicable skills such as problem-solving and communication, may evolve as part of a growing partnership with agents and robots.
Employers are already adjusting. Demand for AI fluency—the ability to use and manage AI tools—has jumped nearly sevenfold in two years. The need for technical AI skills employed to develop and govern AI systems is also growing, though at a slower pace. About eight million people in the United States work in occupations where job postings already call for at least one AI-related skill—a fraction of what may be needed in the years ahead. Demand is also rising for complementary skills such as quality assurance, process optimization, and teaching, as well as for some physical skills such as nursing and electrical work. In contrast, job post mentions are declining for routine writing and research, both areas where AI already performs well, although these skills remain essential for much of the workforce.
In our midpoint scenario of automation adoption by 2030, AI-powered agents and robots could generate about $2.9 trillion in US economic value per year.3 Capturing this may depend less on new technological breakthroughs than on how organizations redesign workflows—especially complex, high-value ones that rely on unstructured data—and how quickly human skills adapt. Integrating AI will not be a simple technology rollout but a reimagining of work itself—redesigning processes, roles, skills, culture, and metrics so people, agents, and robots create more value together.
Leaders will play a central role in shaping this partnership. The most effective will engage directly with AI rather than delegating, invest in the human skills that matter most, and balance gains with responsibility, safety, and trust. The outcomes for firms, workers, and communities will ultimately depend on how organizations and institutions work together to prepare people for the jobs of the future.
Chapter 1
The workforce of the future will be a partnership of people, agents, and robots
AI is redefining the boundaries of work and unlocking new potential for productivity.4 Work will be reconfigured as a partnership between people, agents, and robots.5
AI has made agents and robots more autonomous and capable
For much of the past century, machines have been built to follow rules. Robots executed physical routines like assembling parts while software automated predictable clerical and analytical tasks. Both types of machines operated in a predetermined way; they did what they were programmed to do, and little more. The rise of AI has begun to change that and to broaden the scope of what automation can do. (See sidebar “How technology is advancing.”)
AI agents and robots—machines that perform cognitive and physical work, respectively—are becoming more capable as they learn from vast data sets. This enables them to simulate reasoning and to respond to a wider range of inputs, including natural language, and to adapt to different contexts instead of simply following preset rules.
Share
Sidebar
How technology is advancing
We estimate that today’s technology could, in theory, automate about 57 percent of current US work hours. This figure compares the capabilities of existing technologies, including those demonstrated in a lab, with the level of human proficiency required for different work tasks.6 As technology advances, the picture will continue to evolve and should be updated regularly.
Actual adoption depends on more than technical capability. Factors including policy choices, labor costs, implementation expenses, and development time all influence when and where automation is deployed. Electricity took more than 30 years to spread, and industrial robotics followed a similar multidecade path. As recently as 2023, only about one in five companies ran most of their applications in the cloud, despite the technology being widely available since the mid-2000s.7 (See the technical appendix for details.)
In this chapter, we focus on technical automation potential—mapping the frontier of what today’s technologies can do and identifying the types of work that could be most affected in the years ahead.
AI can have an impact on all types of work
We distinguish between physical and nonphysical work. Robots are needed to automate the former, agents the latter. Not all automation requires agents or robots in the narrow technical sense of those terms, but we use them broadly to capture the full range of technologies that automate work.
Nonphysical work accounts for about two-thirds of US work hours. Roughly one-third of those hours draw on social and emotional skills that mostly remain beyond AI’s reach, while the rest involve tasks—such as reasoning and information processing—that are better suited to automation. These more automatable activities represent about 40 percent of total US wages and span roles in fields from education and healthcare to business and legal (Exhibit 1).
Exhibit 1
Image description: An animated GIF presents a series of horizontal butterfly charts showing the distribution of work hours in the United States by occupation group for 2024. The first frame shows a chart with two main sections—physical work (bars extending left from the center) and nonphysical work (bars extending right)—representing the capabilities required in each occupation. A narrow column at the far right shows each occupation’s share of the workforce. Physical work dominates in building and grounds cleaning and maintenance; construction and extraction; and transportation and material moving. Nonphysical work dominates in legal; business and financial operations; computer and mathematical; office and administrative support; and educational instruction and library. The rightmost column shows employment sizes, with office and administrative support the largest group (about twelve percent), followed by transportation and material moving, sales, management, and healthcare practitioners and technical. The second frame follows the same format, with a dashed-line rectangular overlay inside the bars to indicate the share of work hours that require social and emotional capabilities. Overlays appear within both physical and nonphysical portions for each occupation. The overlays show a greater share of required social and emotional capabilities in people-facing roles such as management; healthcare practitioners and technical; and educational instruction and library, and a smaller share in highly physical roles such as farming, construction, and production. End of image description.
The near-term influence of automation on physical work may be narrower. Activities that require physical as well as cognitive capabilities account for about 35 percent of current US work hours. Robots have made major progress, but most physical work still demands fine motor skills, dexterity, and situational awareness that technology cannot yet replicate reliably (see sidebar “Robots in the workplace”).
Share
Sidebar
Robots in the workplace
Even so, the effects could be significant for some workers. Physical tasks make up more than half of working hours for about 40 percent of the US workforce, including drivers, construction workers, cooks, and healthcare aides. Advances in robotics are expected to change occupations in areas like production and food preparation, including some lower-wage roles. Robots may also continue to perform work that is hazardous or otherwise unfeasible for people, such as underwater tasks, search and rescue, and inspections of dangerous environments.
AI-powered automation will change work, but people remain indispensable
At current levels of capability, agents could perform tasks that occupy 44 percent of US work hours today, and robots 13 percent (Exhibit 2).8
Extending automation further would require technologies that can match a range of human capabilities currently unmatched. Agents would need to interpret intention and emotion. Robots would need to master fine motor control, such as grasping delicate objects or manipulating instruments in surgery.
Tasks occupying more than half of current work hours could potentially be automated, primarily by agents. Yet, that does not mean half of all jobs would disappear; many would change as specific tasks are automated, shifting what people do rather than eliminating the work itself.
In addition, work that draws heavily on social and emotional skills remains largely beyond the reach of automation even under a full-adoption scenario. This is because many tasks require real-time awareness, such as a teacher reading a student’s expression or a salesperson sensing when a client is losing interest. People also provide oversight, quality control, and the human presence that customers, students, and patients often prefer.
Exhibit 2
Image description: A 2×2 matrix chart shows the distribution of US work hours in 2024 by technical automation potential. The chart is divided into four quadrants and uses different shades to indicate work hours covered by people, agents, and robots. The top left quadrant represents work that is not automatable and requires nonphysical capabilities, accounting for 21 percent of total hours covered by people. The top right quadrant represents work that is not automatable and requires physical capabilities, accounting for 22 percent of total hours covered by people. The bottom left quadrant represents work that is automatable and requires nonphysical capabilities, accounting for 44 percent of total hours covered by agents. The bottom right quadrant represents work that is automatable and requires physical capabilities, accounting for 13 percent of total hours covered by robots. Most hours are technically automatable, with the largest block being automatable work requiring nonphysical capabilities, while automatable physical work is smaller. End of image description.
Image description: A continuation of the previous exhibit following the same format with an additional layer indicating the share of total hours that require social and emotional capabilities, represented by a dashed outline placed on top of the previously mentioned values. The top left quadrant, non-automatable work requiring nonphysical capabilities, includes 15 percent of hours requiring social and emotional capabilities. The top right quadrant, non-automatable work requiring physical capabilities, includes 8 percent requiring social and emotional capabilities. The bottom left quadrant, automatable work requiring nonphysical capabilities, includes 8 percent requiring social and emotional capabilities. The bottom right quadrant, automatable work requiring physical capabilities, includes 1 percent requiring social and emotional capabilities. End of image description.
As technology advances, the work requiring people will also change as some roles shrink, others expand or shift focus, and new ones are created. Radiology illustrates this dynamic. Between 2017 and 2024, radiologist employment grew by about 3 percent per year despite rapid advances in AI, and it is expected to continue growing.9 AI augmented radiologists’ work, improving accuracy and efficiency while enabling doctors to focus on complex decision-making and patient care.10 The Mayo Clinic, for example, has expanded its radiology staff by more than 50 percent since 2016 while deploying hundreds of AI models to support image analysis.11
AI is also creating new types of work and roles. Software engineers are creating and refining agents while designers and creators are using generative tools to produce new content.
Share
Sidebar
Framing the jobs debate as AI reshapes work
Overall US demand for labor has remained strong through multiple waves of automation, with new activities having been created faster than technology has replaced existing ones.12 Yet AI’s broad reach raises concern that this time may be different. The outcome will depend on whether new demand, industries, and roles emerge to absorb displaced workers—a question beyond the scope of this research. If history is a guide, employment is likely to evolve rather than contract, although there is no certainty that AI will follow the same pattern (see sidebar “Framing the jobs debate as AI reshapes work”).
The mix of people, agents, and robots varies across a spectrum of seven archetypes
The overall level of employment and mix of occupations in the economy depend on how industries evolve. Within occupations, the configuration of work differs markedly based on their reliance on physical, cognitive, and social and emotional capabilities.
To understand the variation, we analyzed roughly 800 occupations and grouped them according to their physical and nonphysical automation potential.13 This exercise yields seven archetypes that show how people, agents, and robots could collaborate.
Occupations with the lowest automation potential were classified as people-centric, while those with high shares of automatable tasks were labeled agent-centric or robot-centric. Roles with a more even balance were grouped into mixed or hybrid archetypes that combine substantial shares of two or all three (Exhibit 3).
Exhibit 3
Image description: An animated GIF follows the same layout format as the previous exhibit, showing seven 2x2 matrix square charts that classify US occupations into seven archetypes based on who covers work hours—people, agents, or robots—in 2024. In the first frame, each chart is divided into sections representing work hours covered by people, agents, and robots, with varying sizes indicating the share percentage. A two-headed horizontal arrow across the full figure runs from less automatable (left panels) to more automatable (right panels). The seven archetypes are: People-centric, People–agent, Agent-centric, People–robot, Robot-centric, People–agent–robot, and Agent–robot–centric, with a balanced mix of all three. An additional shaped line outlines corresponding archetypes within the quadrants. Example roles appear beneath each panel along with a donut chart to indicate the current workforce in each archetype. In the second frame, following the same format, dashed-line squares appear overlayed inside the shaded squares to indicate the share of total hours that require social and emotional capabilities. Overlays appear in both nonphysical and physical portions, with a greater share visible in people-centric and people–agent archetypes and a smaller share in robot-centric and agent–robot-centric archetypes. End of image description.
This framework applies across labor markets and can help leaders see where change may come first and how workforce transitions could unfold, highlighting roles that may evolve into human–agent–robot coworker models and those likely to be largely automated by agents or robots under human supervision. For workers, it offers a view of how their own roles might change.
At one end of the spectrum are roles that remain largely human. These people-centric occupations—found, for example, in healthcare and in building and maintenance—make up about one-third of US jobs and pay an average of $71,000 a year. Physical activity that current technologies cannot replicate accounts for about half of the work hours in these occupations.14
At the other end of the spectrum are roles with the highest potential for automation by agents or robots. These occupations make up about 40 percent of total jobs. With an average pay of $70,000, most are agent-centric roles in legal and administrative services. They involve large shares of cognitive tasks—such as drafting documents—that could technically be handled by AI systems. Some of this work may end up being fully automated, but people will still be needed to guide, supervise, and verify.
A smaller subset of these highly automatable jobs involves physical work. These robot-centric roles—such as drivers and machine operators—are physically demanding, sometimes hazardous, and typically pay about $42,000 a year. In theory, they could be almost fully automated, but cost and other real-world constraints may keep people in the loop.
Agent–robot roles form an even smaller category, accounting for only about 2 percent of workers. They pay roughly $49,000, and physical tasks occupy 53 percent of work time. These jobs appear mainly in production settings where software intelligence directs physical systems, such as automated manufacturing or logistics operations.
Between the extremes lies a diverse set of occupations that combine humans, agents, and robots. These hybrid roles employ about one-third of the workforce and differ significantly in pay, physical intensity, and automation potential—yet people remain essential in every setting. As automation is adopted, productivity rises, and people’s roles shift from performing tasks to directing how machines perform them. Hybrid roles break down as follows:
- People–agent roles, which include teachers, engineers, and financial specialists whose work could be enhanced by digital and AI tools. These pay an average of $74,000 per year and account for about one in five US workers.
- People–robot roles, found in maintenance and construction, involve machines that add strength and precision to human efforts. About 81 percent of these work hours involve physical tasks, and annual pay averages $54,000. Fewer than one in a hundred US workers hold these jobs.
- People–agent–robot roles, found in transportation, agriculture, and food service, combine all three forms of labor in roughly equal measure. About 43 percent of the work hours involve physical tasks, and annual pay averages $60,000. Roughly 5 percent of US workers are employed in these roles.
This analysis reflects the current US task mix and what is technically possible with today’s technologies rather than a forecast of what will happen.
The mix of activities will evolve as technology advances and companies adapt their workflows. The distribution of roles across work archetypes also differs by economy and industry. For example, in regions where manufacturing is more prevalent, people–robot roles may be more common than in economies that rely more heavily on services.
Regardless of where one sits, collaboration between people and intelligent machines is likely to deepen. The illustrations below offer examples of how this might work in practice (Exhibit 4).
Exhibit 4
Image description: A conceptual illustration showing a solar-farm workflow with humans, AI agents, and robots working together for facility inspection and repair. Humans oversee and make decisions while AI agents coordinate continuous monitoring, and robots execute inspection, cleaning, and repair tasks. The illustration features various AI-powered technologies, including a drone, rover, and agents, represented as two tablets in enclosed circles, alongside a field technician. These features are in a different shade to stand out. An AI-powered drone conducts visual and thermal inspections to spot faults and shares findings with field technicians. An AI-powered rover cleans panels, removes vegetation, and performs minor repairs while reporting progress. In one circle, a tablet with a solar panel powered by an AI agent monitors system health, predicts failures, and displays prioritized repair tasks with estimated times. In the second circle, a graph shows total energy output over time, where AI agents monitor system health to predict component failures and optimize energy generation and grid interaction in real time. The field technician oversees these operations, validating diagnostics and handling complex repairs. End of image description.
Image description: A conceptual illustration showing an order-fulfillment workflow in a building materials store, where customer interactions remain human-led while AI systems and robots streamline recommendations, inventory, and material movement. It features AI-powered agents represented as computer screens enclosed in two circles, a store manager, a customer, a humanoid robot, and a wheeled mobile manipulator robot with a claw hand. These features are in a different shade to stand out. In the first circle, a bar graph on a computer screen represents inventory data, with the AI-powered agent managing inventory data, coordinating with suppliers, and scheduling logistics. In the second circle, a screen shows recommended materials, where another AI agent generates personalized recommendations based on design specifications and budget constraints. The AI-powered humanoid retrieves smaller items for customers using a basket, while an AI-powered wheeled mobile manipulator transports heavy materials with its claw hand to pickup areas. The store manager engages with customers, providing advice on material selection with AI support. End of image description.
Chapter 2
Human skills will evolve, not disappear, as people work closely with AI
Employers hire workers for their skills. The skills they need evolve as technology and ways of working change. AI accelerates this shift.
To understand how AI could reshape demand for human skills, we analyzed job postings, which offer the most up-to-date view of what employers are seeking.15 Lightcast data, widely used by labor economists, provide a detailed and consistent record of the language employers use to describe roles and skills. While postings reflect hiring intentions rather than the actual work people do, they offer the most comprehensive picture of skill demand.
From this source, we identified roughly 6,800 skills cited frequently in more than 11 million job postings, providing a representative snapshot of the US labor market.16 We then examined how employer requirements differ across occupations.17
Our analysis shows that nearly all occupations have at least one highly disrupted skill—defined as being in the top quartile of change by 2030—and that a third of occupations will see more than 10 percent of their skills highly changed.
We also find that employers now expect a broader and more specialized mix of skills across nearly all occupations. A core set of eight high-prevalence skills—communication, management, operations, problem-solving, leadership, detail orientation, customer relations, and writing—remains essential across industries. Demand for AI fluency, the ability to use and manage AI, is rising faster than demand for any other set of skills.
Skill requirements have become more specific and specialized over time
The number of distinct skills associated with each occupation has risen on average to 64 from 54 a decade ago, reflecting greater specificity in how employers describe roles.18 Higher-wage fields tend to require more skills and greater specialization. Job postings for data scientists and economists, for example, list more than 90 unique skills, compared with fewer than ten for motor-vehicle operators.
Higher-wage jobs that require more skills tend to place greater emphasis on management, information, and digital skills. Lower-wage roles focus on hands-on work, operating equipment, and providing care and assistance (Exhibit 5).
Exhibit 5
Image description: A set of 22 horizontal stacked bar charts showing the distribution of skills across U.S. occupation groups, expressed as percentages. A separate column on the right displays the average wage for each occupation group, represented as circles of varying sizes with dollar amounts in thousands, ranging from 38 to 155 thousand dollars. Each bar represents an occupation group and is divided into eight skill categories: communication, collaboration and creativity, management skills, information skills, digital skills, assisting and caring, handling and moving, working with machinery and specialized equipment, and constructing. Higher-wage groups such as legal, management, and computer and mathematical have a larger distribution of communication, management, information, and digital skills. In contrast, physical categories—handling and moving, machinery, and constructing—dominate trades such as construction, production, and transportation, which have lower average wages. Healthcare roles show large shares of assisting and caring. End of image description.
Even within a single field—software development, for example—the skills required for similar-sounding jobs can differ sharply. Python developers, AI engineers, and C++ developers share fewer than half of their required skills, reflecting how technology drives specialization.
Because skills are becoming increasingly specific and work is evolving rapidly—with some roles disappearing, others changing, and new ones emerging—adaptability and ongoing learning are essential.
The speed of technological change raises the importance of transferable skills, including eight high-prevalence ones
Each wave of technology has changed what workers do. The difference today is speed. Until 2023, the need for AI-related skills grew at roughly the same pace as for cloud computing, cybersecurity, and other digital skills. After the rise of generative AI, it accelerated sharply: Nearly 600 new skills appeared in job postings over the past two years—about one-third of the total added in the past decade—many of which are tied to AI and its enabling technologies.
This rapid churn heightens the value of transferable skills. Despite growing specialization, a core set of eight high-prevalence skills—among them communication, customer relations, writing, problem-solving, and leadership—has stayed relevant across industries and wage levels.
These skills form the connective tissue of the labor market and are key to workforce development. Building them makes workers more adaptable and better prepared for change. Their application is likely to evolve as people work more closely with AI-powered agents and robots, a theme we explore below.
Many other skills are also transferable across occupations. For example, more than half of the skills required for account executives also appear in 175 other occupations. These range from similar sales positions to roles in marketing and human resources. The overlap allows companies to widen their talent pipelines by drawing from adjacent roles or redeploying employees with similar skills.19 For workers, it opens pathways to new—and often more people-centric—positions that build on existing strengths (Exhibit 6).
Exhibit 6
Image description: Bubble scatter plot comparing skill overlap with an account executive and the technical automation potential of various occupations in the US. The horizontal axis shows skill overlap from about 50 to 90 percent; the vertical axis shows automation potential ranging from 0 to 100 percent. A dotted horizontal reference line marks 52 percent—the automation potential for an account executive. Each bubble represents a different occupation, with the size of the bubble indicating the number of full-time equivalent workers, categorized into three sizes: less than 10, between 10 and 100, and more than 100 thousand workers. Different shades distinguish 11 occupation groupings such as management, STEM, business, finance, arts, legal, and others. Key occupations labeled include insurance sales agent, sales consultant, and advertising sales representative near the high-automation, high-overlap area. Sales representative and travel/tour guide lie above the reference line, sharing substantial skills with account executives—often with higher automation potential—while city/town manager, marketing consultant, director of sales, and business development manager are below the reference line at high overlap but lower automation potential. End of image description.
Demand for AI fluency is growing faster than any other skill
As AI technology matures, demand for related skills is spreading beyond development roles. Demand for AI fluency jumped nearly sevenfold in the two years through mid-2025. It is now a job requirement in occupations employing about seven million workers. Demand for technical AI skills—building and deploying AI systems—has also grown, albeit at a slower pace (Exhibit 7).20
Exhibit 7
Image description: A set of three vertical stacked bar charts illustrates the increase in demand for AI fluency and technical AI skills in US occupations from 2023 to 2025, divided into STEM and non-STEM occupation types, with STEM occupations showing a larger increase. The charts show the number of employees in occupations where an AI-related skill was listed in at least 5 percent of job postings, measured in millions. The first set on the left shows employees in occupations seeking AI fluency, rising from 1.0 million in 2023 to 7.0 million in 2025, indicating a 6.8 times increase. The second set, middle chart, shows employees in occupations seeking technical AI skills, rising from 2.1 million to 3.3 million, a 1.6 times increase. The third set on the right shows employees in occupations seeking any AI-related skills, growing from 2.2 million to 7.5 million, a 3.5 times increase. End of image description.
So far, however, most AI skill demand today is concentrated in a few fields. Three-quarters of all AI skill demand in the United States is found in three occupational groups: computing and mathematics, management, and business and finance (Exhibit 8). The rest comes from ten other groups in which the technology is starting to become more prominent, including architecture and engineering; installation, maintenance, and repair; and education. Demand for AI-related skills remains limited in nine other occupational groups, such as construction, transportation, and food service, which together account for about 40 percent of the workforce and fall below the median income.
Exhibit 8
Image description: A table-style horizontal bar chart with 13 bars showing occupation groups in millions of US workers whose jobs require AI-related skills, segmented into STEM and non-STEM occupation types. Each bar represents the total number of full-time equivalent workers in millions, with a segment indicating the share and count of workers whose jobs require AI skills. Seventy-five percent of the demand for AI skills comes from three groups: Computer and mathematical shows the largest count at about 2.6 million, followed by management at about 2.3 million, and business and financial operations at about 0.8 million. The remaining 25 percent of demand is spread across ten smaller bars, each adding less than about 0.3 million. Additionally, there is a bar representing nine other groups with no AI skill demand. At the bottom, a totals row notes about 161 million full-time-equivalent workers overall, with roughly five percent having AI-skill demand. End of image description.
While the core demand is still concentrated, AI’s influence is beginning to ripple outward. Employers are increasingly seeking more AI-adjacent capabilities such as process optimization, quality assurance, and teaching—skills employed to redesign work with AI, supervise and verify AI systems, or train people to use them.
Meanwhile, the number of mentions in job listings is falling for skills that machines already perform well or significantly enhance—research, writing, and simple mathematics—though these skills remain essential for much of the workforce (Exhibit 9).
Exhibit 9
Image description: Two side-by-side horizontal bar charts, each with eight bars, show changes in the number of US occupations with job postings that mention each skill subcategory from 2023 to 2025. The vertical axis lists skill subcategories; the horizontal axis shows the change in number of occupations. The left column shows the greatest decreases, with general science and research down 140 occupations and writing and editing down 134. The right column shows the greatest increases, with artificial intelligence and machine learning up 185 occupations and people management up 138. Overall, AI-related and several business skills show the largest increases in occupational demand, while traditional science, writing, and basic technical categories decline. Numbers of occupations in 2023 appear in columns at the far left and far right of the chart for each skill. End of image description.
Most human skills will remain relevant, but AI will change how they are used
Our analysis finds that roughly 72 percent of skills are required both for work that could be done by AI and for work that must be done by people (Exhibit 10). For details, see sidebar “How we assess skill exposure to automation.”
Exhibit 10
Image description: A two-part exhibit summarizes where skills are used across activities with different automation potential. At the top, a horizontal stacked bar shows the distribution of skills by technical automation potential: eleven percent people-led, seventy-two percent done by a combination of people and AI, and seventeen percent AI-led. Below, each bar is subdivided by type of skills required for work into five vertical stacked bar charts showing the distribution of work hours by eight skill categories such as communication and collaboration, management, information skills, digital skills, assisting and caring, handling and moving, working with machinery and specialized equipment, and constructing, for "work hours by people," "by people with agents," "by people with robots," "by agents," and "by robots." The mix of categories shifts across bars—human-centric categories such as communication and management dominate people-led work, while physical and machinery-related categories feature more prominently in agent- and robot-led work. Most skills are applied in activities that are partly automatable, with only a minority of work hours concentrated at the extremes of purely people-led or AI-led tasks. End of image description.
Share
Sidebar
How we assess skill exposure to automation
A small set of skills is likely to remain uniquely human. These are rooted in social and emotional intelligence such as interpersonal conflict resolution and design thinking, which depend on empathy, creativity, and contextual understanding and will be challenging for machines to replicate.
At the other end of the spectrum are skills likely to become largely AI-led, including data entry, financial processing, and equipment control. In these areas, people will step back from hands-on work to focus on design, validation of results, and exception handling—making sure AI agents and robots run properly as they operate mostly on their own.
Between these poles lies a broad middle ground where people and AI work side by side. Here, a skills partnership is emerging: Machines handle routine tasks while people frame problems, provide guidance to AI agents and robots, interpret results, and make decisions. The work blends collaboration and oversight, as humans bring judgment and contextual understanding that machines still lack.
The eight high-prevalence skills described earlier fall largely within this middle ground. They remain relevant but will evolve as people, agents, and robots take on different aspects of the same work (Exhibit 11).
Exhibit 11
Image description: A table with eight high-prevalence skills and three columns describing how work shifts among people, AI agents, and robots. Relevance across occupations is shown with small horizontal bars, with high relevance in communication, 99 percent, followed by management, 94 percent, and operations, 84 percent. For each skill, the table describes how people, agents, and robots will collaborate, what agents and robots will do, and what people will do. Some collaboration tasks include drafting and interpreting information, planning projects, and forecasting demand. Agents and robots take on content generation, automating schedules, executing routine tasks, identifying patterns, and more. People focus on refinement and storytelling, coaching hybrid teams, strategy and process design, judgment, and interpretation. End of image description.
The Skill Change Index shows widespread shifts in skills by 2030
Among the 100 most in-demand skills, the effects of AI will differ widely. People-focused skills such as coaching face the least exposure to automation, while manual and routine skills like invoicing face the most. Skills such as quality assurance fall near the middle of the distribution—areas where AI is changing how people use skills rather than replacing them outright.
To gauge the extent of these shifts, we developed the Skill Change Index (SCI), a time-weighted measure of each skill’s potential exposure to automation in different adoption scenarios. The SCI shows where the most significant shifts in skills are likely to occur (Exhibit 12).
In the midpoint scenario, roughly one-quarter to one-third of work hours tied to the 100 most in-demand skills could be automated by 2030. For instance, about 28 percent of the work associated with quality assurance could be carried out by machines.
In a faster-adoption scenario, exposure rises sharply. Under this trajectory, the most affected skills among the top 100 could reach 60 percent, while about half of the work hours associated with quality assurance could be automated.
Exhibit 12
Image description: A scatter-and-line chart shows the Skill Change Index for about 6,800 skills under a midpoint scenario of automation adoption. Circles represent index values of the top 100 skills, with different shades indicating skill categories. The vertical axis displays the Skill Change Index from 0 to 70, where a higher index indicates more exposure to automation and a lower index indicates less exposure. The horizontal axis orders skills by percentile. The scatter-line, labeled "Midpoint scenario of skill change," increases steadily in an upward curve across skills, with most skills concentrated in the low-to-mid 30s on the index. Quartile markers indicate index values of about 23 percent at the 25th percentile, 28 percent at the 50th percentile, and 33 percent at the 75th percentile. Labeled examples along the rising curve include lower-exposure skills such as good driving record, leadership, coaching, and negotiation; mid-range skills such as communication, customer relations, management, writing, quality assurance, and detail orientation; and higher-exposure skills such as inventory management, invoicing, and SQL. End of image description.
Image description: A continuation of the previous exhibit follows the same format with an additional scatter-and-line chart showing a second scenario labeled "Early scenario" above the midpoint scenario scatter-line. The early-adoption scenario sits well above it across the distribution, with higher quartile markers of about 43 percent, 51 percent, and 59 percent. The same labeled examples appear along the curves—from lower-exposure leadership and coaching to higher-exposure inventory management, invoicing, and SQL—with different shades denoting categories. Faster adoption would substantially raise automation exposure for most skills—roughly doubling index values versus the midpoint scenario—while preserving the relative ordering of skills. End of image description.
Across the broader set of 7,000 skills, exposure remains uneven. Digital and information-processing skills rank highest on the SCI, reflecting AI’s growing proficiency in data handling and analysis. By contrast, assisting and caring skills are likely to change the least (Exhibit 13).
Exhibit 13
Image description: Eight horizontal bar chart showing the distribution of eight skill categories in the US by position on the Skill Change Index, with automation exposure levels indicated to experience the most change by 2030. The eight skill categories are Digital skills, Information skills, Working with machinery and equipment, Constructing, Communication, collaboration, and creativity, Management skills, Handling and moving, and Assisting and caring. The horizontal bars for each category are divided into three segments representing low, moderate, and high automation exposure. Digital and information skills are expected to experience the most change, with high automation exposure, 42 percent and 29 percent, while assisting and caring skills see the least change, with low automation exposure of 10 percent. On the far right side of each skill category, examples are provided and divided into lower and higher automation exposure levels. End of image description.
The SCI reveals three broad paths for how skills may evolve.
Highly exposed skills—those in the top quartile of the index—are more likely to decline in demand. These are often specialized skills, such as accounting processes and programming in specific languages, that AI can already perform well.
Skills in the middle quartiles are more likely to evolve, changing in nature and application rather than simply rising or falling in demand. These are often transferable skills that combine human judgment with digital tools; AI fluency itself is one of these. As workers collaborate with AI, they apply skills like writing and research in new ways rather than being made obsolete.
Finally, low-exposure skills—those in the bottom quartile—are likely to endure. These are often grounded in human connection and care, such as leadership and healthcare skills.
Over time, the overall demand for skills will depend on how the mix of jobs in the economy evolves and on how rapidly organizations adopt AI and other technologies. As adoption accelerates, some skills that are only partially automatable today may become more exposed, while entirely new forms of work and skills may emerge.
Chapter 3
Entire workflows can be reimagined around people, agents, and robots
Share
Sidebar
How we estimate the economic value of AI
AI-powered automation could unlock $2.9 trillion of economic value in the United States by 2030, according to our midpoint adoption scenario.21 Realizing these gains requires more than automating individual tasks. It will mean redesigning entire workflows so that people, agents, and robots can work together effectively. (See sidebar “How we estimate the economic value of AI.”)
Reimagining workflows is key to capturing the economic potential of AI
Workflows—multistep processes involving collaboration, information exchange, and decision-making—form the backbone of how organizations operate. Most were designed for a pre-AI world, so applying AI to individual tasks within these legacy processes is unlikely to deliver the productivity gains now possible.
This may explain why relatively few businesses report tangible benefits from AI so far. Nearly 90 percent of companies say they have invested in the technology, but fewer than 40 percent report measurable gains.22 The gap may reflect the fact that many projects are still in pilot or trial phases or that organizations are applying AI to discrete tasks rather than redesigning entire workflows. In banking, for example, this would be the difference between offering employees access to a chatbot for ad hoc use and deploying custom agents alongside people in a reimagined process to approve, process, and manage loans more efficiently and deliver better customer service. Unlocking larger productivity gains from AI will require reimagining workflows along the lines of the latter, rather than taking a task-based approach.
We analyzed 190 business processes across the US economy to identify where the greatest opportunities may lie. About 60 percent of potential productivity gains are concentrated in workflows related to sector-specific domains—activities at the core of each industry. In manufacturing, these include supply chain management; in healthcare, clinical diagnosis and patient care; and in finance, regulatory compliance and risk management. Additional gains come from cross-cutting functions such as IT, finance, and administrative services that support every sector (Exhibit 14).
Exhibit 14
We strive to provide individuals with disabilities equal access to our website. If you would like information about this content we will be happy to work with you. Please email us at: McKinsey_Website_Accessibility@mckinsey.com
Share
Sidebar
An early view of workflows across the US economy
In finance and insurance, for example, there are seven key workflows within the IT function (Exhibit 15). Every sector–function combination has its own set of workflows, which represent the critical unit for realizing gains from human–AI collaboration. (See sidebar “An early view of workflows across the US economy” for more examples.)
Exhibit 15
Image description: A three-column flowchart illustrating functions and workflows in the finance and insurance sector. The chart is divided into three columns: Sectors, Functions, and Workflows. The Sectors column lists various economic sectors, highlighting Finance and insurance. The Functions column details sector-specific functions such as Finance, Sales, Planning and management, and others. In this column, Information Technology is highlighted as it stems out from the Finance and insurance sector. The Workflows column lists specific tasks like Infrastructure and network management, IT operations, and service management, among others from the Information Technology function and Finance and insurance sector that was highlighted. Arrows and captions indicate that sectors are broad segments of activity, functions are distinct operational areas, and workflows are structured sequences of tasks. End of image description.
From a utility to a bank, early movers are experimenting with AI-embedded workflows
Some organizations are redesigning workflows around AI, offering early evidence of how these transformations look in practice. We identified 80 implementation cases—from pharmaceuticals to banking and sales—and looked closely at several to glean insights from their approaches.
Managers and specialists are increasingly acting as orchestrators and validators rather than executors, while domain experts such as data analysts, underwriters, and engineers partner with agents that perform initial analysis or generate draft outputs. As a result, the most valuable human skills are shifting toward AI fluency, adaptability, and critical evaluation of outputs, enabling people to focus on higher-value work.
We present four cases that illustrate how these changes are unfolding. A technology firm uses AI agents to prioritize sales leads and manage outreach, freeing specialists to spend more time negotiating and building relationships. A pharmaceutical company applies AI to draft clinical reports, reducing errors and accelerating regulatory submissions. In customer service, agents now resolve most routine inquiries, while a regional bank uses them to speed up software modernization.
These deployments illustrate how increasingly specialized agents could reshape entire business processes. They also show that people remain at the center of work because AI still depends on human guidance, interpretation, and quality control.
Sales case: AI-powered agents enabled specialists to redirect time from routine tasks to selling activities
A global technology company sought to expand its reach and deepen customer relationships while navigating growing complexity and customer volume. In its traditional model, human sales teams used inconsistent prioritization methods and had limited capacity to tailor outreach to thousands of smaller accounts. As a result, only top prospects received customized attention.
To overcome these limits, the company introduced several AI agents to automate the early stages of the sales process (Exhibit 16). A prioritization agent scores and ranks accounts based on public and proprietary data. An outreach agent contacts customers, while a customer response agent manages follow-ups and categorizes leads as interested, not interested, or uncertain. A scheduling agent sets up calls and reminders for high-potential leads. When a case requires human judgment, a handoff agent transfers the file to a specialist.
Exhibit 16
Image description: A flowchart illustrating people-agent collaboration at a global B2B tech company, focusing on redesigning commercial workflows to reallocate time from routine tasks to selling activities. The chart is divided into four numbered workflow stages: 1) Target and enable, 2) Personalize outreach, 3) Qualify and convert, and 4) Serve and grow. Each stage lists agents' activities and people's activities, along with relevant skills in two columns. In stage 1, a prioritization agent reviews and ranks accounts while a business-development specialist reviews agent outputs. In stage 2, an outreach agent creates tailored outreach for each account. In stage 3, several agents handle response management, scheduling, coaching, and handoff, while people update CRM software and complete sales based on agent feedback. In stage 4, an admin agent takes on administrative tasks, and customer service representatives manage requests and plan account strategy. A legend at the top right uses symbols to mark skills with higher (triangle) versus lower (square) change on a Skill Change Index. End of image description.
This process expanded outreach and improved conversion rates, delivering a projected annual revenue increase of 7 to 12 percent from new sales, cross-selling, and increased retention. Across sales roles, time saved ranged from 30 to 50 percent. Business development specialists were able to spend more time on strategic engagement—drafting proposals, negotiating partnerships, and building relationships.
Looking forward, this model could be extended by introducing additional agents to support sales. A coaching agent could provide real-time feedback to sales teams, while an admin agent could act as an assistant, handling routine administrative tasks.
Customer operations case: AI agents improved customer experience and reduced cost per call
A large utility company handles more than seven million support calls each year, even with multiple self-service options available on its app and website. Its interactive voice response system had previously resolved only about 10 percent of inquiries, leaving the rest to human customer-service representatives.
To improve efficiency and customer experience, the company deployed agentic conversational AI across its entire customer base (Exhibit 17). The system includes several agents: an inbound call agent that authenticates customers, an intent identification agent that determines the purpose of the call, a call scheduling agent that manages appointments, and a self-service agent that integrates with back-end systems. Together, these now handle roughly 40 percent of all calls, resolving more than 80 percent without human involvement. When escalation is needed, customers are transferred with verified account details and conversation history, ensuring a seamless handoff.
Exhibit 17
Image description: A flowchart illustrating people-agent collaboration at a leading utilities firm, focusing on reimagining service workflows to improve customer experience in issue resolution. This exhibit follows a similar format to the previous one but features three workflow stages instead of four. The stages are: 1) Call initiation and routing, 2) Issue identification, and 3) Issue resolution, along with a list of agents' activities and people's activities and the legend symbol described in the previous exhibit. In stage 1, an inbound call agent handles calls with recognized customer information, an intent identification agent flags call intents, and a call scheduling agent schedules calls. A customer service representative builds trust with customers using pre-populated information. In stage 2, a self-service agent handles simpler calls, and a customer issue identification agent transcribes and updates systems with new information. People answer more complex questions or take calls when requested by customers. In stage 3, a coaching agent assists representatives with real-time data to suggest solutions. End of image description.
The new process has cut the average cost per call by about 50 percent and increased customer satisfaction scores by six percentage points, driven by shorter waiting times, more consistent handling, and faster resolution. Human representatives now manage more complex, emotionally sensitive, and high-value issues, improving both the quality and the impact of service.
Future applications could go further. A customer issue identification agent could monitor systems to detect service interruptions and contact customers proactively, while a coaching agent could provide real-time guidance to human representatives during live calls. In such models, AI would handle most routine inquiries while people concentrate on complex or relationship-based issues, supported by continuous insights and automated follow-up. Advanced AI agents could eventually handle 80 to 90 percent of customer inquiries, documenting each interaction and initiating follow-up to ensure continuity and consistency.
Medical writing case: Gen AI platform accelerated report drafting and improved accuracy
A global biopharmaceutical company sought to improve its process for drafting clinical study reports, which document safety and efficacy data for new drugs. In the traditional model, medical writers manually compiled study data, drafted lengthy reports, and coordinated multiple review cycles. Limited capacity and long turnaround times constrained the ability to meet growing submission demands.
To improve the speed and quality of clinical study reports, the company developed an AI platform that reconfigures workflows for report writing (Exhibit 18). This AI companion synthesizes structured and unstructured study data, generates comprehensive drafts in minutes, applies company style and compliance templates, and self-reviews for errors. These tools shift the medical writers’ role from manual drafting to collaborating with AI systems and applying clinical judgment. Writers can regenerate and edit sections of text, review potential issues, and validate data against source materials to ensure accuracy and regulatory compliance.
Exhibit 18
Image description: A flowchart illustrating people-agent collaboration at a global pharmaceutical company, focusing on streamlining clinical study reporting workflows to enhance collaboration between people and agents. This exhibit follows a similar format to the previous ones, featuring four workflow stages, a list of agents' activities and people's activities, and the legend symbol described in the previous exhibit. The stages are: 1) Initiation and setup, 2) Drafting, 3) Reviewing, and 4) Quality control and submission. In stage 1, a clinical study planning agent automates planning tasks, while a researcher or medical writer uses the agent's planning to communicate with regulators. In stage 2, a data mapping agent synthesizes study data, a report-drafting agent creates a first draft, and a validation agent ensures data accuracy. People validate drafts and add clinical judgment. In stage 3, a reviewing agent regenerates drafts based on feedback, and coauthors review sections to ensure a cohesive narrative. In stage 4, a submission draft agent prepares a submission-ready draft, and a publishing team finalizes and submits the report. End of image description.
Early data indicate substantial efficiency gains. Touch time for first human-reviewed drafts dropped by nearly 60 percent and errors declined by roughly 50 percent. Go-to-market efforts accelerated by weeks when combined with other related processes and technology changes, and further improvements are expected as writers build AI skills and additional agents are introduced. The company reports that scaling these efforts can be challenging, and a combination of technology and people skills, including resilient data engineering, prompt engineering upskilling, and bold organizational leadership, is key.
Looking ahead for life science companies, agents could be leveraged to support key stages of clinical research, from study planning through to submission. A clinical study planning agent could help assemble trial protocols, a data mapping agent could analyze and synthesize data, and a report drafting agent could produce full drafts. A validation agent could then check for compliance, and a reviewing agent could scan for errors. Finally, a submission draft agent could help generate regulator-ready documents. Applied across the research cycle, these tools could shorten timelines by several months.
IT modernization case: AI agents streamlined code migration and shifted human roles to orchestration
A regional lender used AI agents to modernize its banking application for small and medium-size enterprises. The aim was to update various programming languages to speed up internal development. The project would previously have required months of work, large budgets, and extensive engineering capacity for manual documentation, code refactoring, and testing of millions of lines of code.
To accelerate the process, the bank launched a pilot using AI agents for multiple modernization tasks (Exhibit 19). An assessment agent scans legacy code bases identifying dependencies, while a functionality agent generates the target-state architecture. A coding agent migrates code to new frameworks and performs automated tests. Developers collaborated with 15 to 20 agents each, verifying and refining outputs to ensure architectural integrity, compliance, and functional accuracy. The modernization also shifted applications from desktop to mobile, on-premises to cloud, and monolithic to microservice architectures.
Exhibit 19
Image description: A flowchart illustrating people-agent collaboration at a regional bank, focusing on automating IT modernization workflows to elevate related roles to focus more on orchestration. This exhibit follows a similar format to the previous ones, featuring three workflow stages, a list of agents' activities and people's activities, and the legend symbol described in the previous exhibit. The stages are: 1) Planning, 2) Module modernization, and 3) Testing and validation. In stage 1, a modernization planning agent develops team structures and timelines, while a program manager provides inputs for planning sprints and resources. In stage 2, an assessment agent evaluates legacy applications, a functionality agent creates drafts of functionality, and a coding agent writes code drafts. An application architect and product manager cross-checks outputs and works with users, while a software developer directs agents and reviews code drafts. In stage 3, quality assurance and testing agents test functionality, and a QA tester and business team create testing tasks and validate with users before launch. End of image description.
As AI agents took on most of the repetitive execution, the focus of human work shifted toward planning, orchestration, and testing. Early results show up to 70 percent code accuracy.
Following the pilot module, the bank now plans to extend the use of agents to the entire modernization effort. It estimates that this could reduce required human hours by up to 50 percent. A modernization planning agent could coordinate the process, supported by quality assurance agents and testing agents.
AI is reshaping managerial work and skills
Our case studies show that as AI takes on more analytical and decision-support tasks, the nature of managerial work is shifting from supervising people to orchestrating systems in which people, AI agents, and robots collaborate. This change allows managers to redirect time to higher-value work involving skills such as influencing and mentorship, while also demanding greater technical fluency (Exhibit 20). For example, a sales manager might spend more time coaching teams to use AI-driven insights and strengthen relationships, while a customer service manager might oversee a hybrid workforce of people and AI agents, training both AI systems and staff to improve service.
Exhibit 20
Image description: A table showing example shifts in leadership and management skills in the US among people, agents, and robots. The table is organized with rows listing skills from top to bottom: prioritization, decision-making, planning, coordinating, budgeting, accountability, innovation, coaching, influencing, and mentorship. Columns describe what "people, agents, and robots will collaborate on," what "agents and robots will" do, and what "people will" do. Each skill is associated with a level of change indicated by three shades ranging from high, medium and low, under the "Position on Skill Change Index, quartile" column. Higher change is at the top rows and lower change in the bottom rows. The table details how each skill will be adapted in collaboration with AI, with specific tasks and responsibilities outlined for people, agents, and robots. End of image description.
Across industries, companies are finding that the biggest gains come from redesigning entire workflows rather than automating individual tasks. Doing so requires new operating models, data foundations, and skill pathways for people as their collaboration with agents and robots deepens. In the next chapter, we examine how leadership could evolve to guide this transformation.
Chapter 4
Leadership is crucial as agents and robots reshape work and the economy
AI adoption is reshaping how organizations operate, creating new ways of working built around the strengths of people, agents, and robots. Managing this transition will require business leaders to make deliberate choices about its pace and purpose, and to work with other institutions to ensure that workers are well prepared.
Key questions for business leaders
For businesses, embedding AI successfully depends on recognizing the enduring importance of people. This is as much a practical concern as an ethical one. As technology takes on more tasks, the judgment and oversight people provide will be even more vital to keeping organizations on course. Work will be organized differently: Employees will need retraining as workflows are reshaped around what people and intelligent machines do best, and performance measures will need to reflect contributions from both. The questions below highlight some of the choices and trade-offs leaders face in implementing AI.
Are you reimagining your business for future value?
Early AI efforts often aim to improve existing workflows rather than rethink them. Larger gains come from redesigning processes entirely. Building for future value means looking several years ahead and working backward to identify which roles, skills, and structures may need to change in relation to AI. Leaders must choose where they invest in major redesigns now versus refining current models for nearer-term gains.
Are you leading AI as a core business transformation?
AI will touch nearly every function. Leaders can approach it as either a technology project or a broader business transformation. Delegating responsibility to the IT department may speed implementation, but lasting change and real strategic advantage will depend on visible commitment from senior leadership and sustained attention to how AI affects people and business across the organization.23
Are you building a culture of experimentation and learning?
Implementing AI involves uncertainty, especially at the start. Organizations that test and adapt quickly tend to learn fastest. This depends on a culture that supports curiosity, risk-taking, learning from setbacks, and collaboration. Changing culture is difficult but essential for transformation on the scale AI is likely to require.24
Are you building trust and ensuring safety?
AI changes how businesses stay accountable and maintain oversight. The focus is shifting from checking individual outputs to setting clear policies, validating AI logic, dealing with exceptions, and determining when human involvement is most needed. The challenge is to keep the right balance, maintaining enough oversight to manage risk and ensure safety without limiting efficiency and innovation.25
Are you equipping your managers to lead teams of people, agents, and robots?
AI is redefining what it means to manage. Routine supervision may be automated, freeing managers to focus on coaching, influencing, and orchestrating hybrid teams of people, agents, and robots. They will also play a key role in testing for bias, validating performance, and upholding integrity. As automation reduces direct control, staying accountable for outcomes may become more challenging. New performance metrics and feedback systems will be needed to assess human and machine contributions and how they interact.
Are you preparing your workers for new skills and roles?
Companies will need to decide how to use capacity freed up by AI—whether to reinvest it in developing people and higher-value work or to focus on greater efficiency and cost reduction. Most will do some of both. Managing this shift means identifying which roles can evolve and giving employees clear, skill-based pathways to grow into them.
AI makes continuous learning and training even more important to organizational strength. As jobs change and skill needs evolve faster, helping workers understand how their skills transfer to new types of work will make both people and businesses more resilient. AI fluency will need to extend across all levels of the organization. Companies can use digital tools, hands-on projects, and coaching to build these skills, while partnerships with other organizations and institutions can expand access to learning and open new opportunities.
Key questions for institutions
Periods of economic disruption often force societies to strengthen the systems that help people adapt. Since the Industrial Revolution, nations have expanded education, training, and social safety nets. In the United States, the New Deal and the GI Bill built modern social infrastructure, while the digital revolution extended inclusion through online learning and telehealth.26 The coordinated response to the COVID-19 pandemic showed how quickly institutions can mobilize when livelihoods are at stake.
The rise of AI may call for similar renewals. Public, private, and civic institutions can lead by example in retraining people and expanding opportunity. The questions that follow invite leaders to rethink how education and job mobility can evolve in the age of AI.
How can education and training keep pace?
Education will play a pivotal role as skill needs evolve. Foundations of AI fluency—competencies such as critical thinking, questioning results, challenging assumptions, and recognizing bias or error—should be developed from primary school onward so people learn to use and guide these technologies effectively.
Curricula could be redesigned to combine technical knowledge with transferable human skills such as adaptability, analytical thinking, and collaboration. This approach could help prepare workers for a more fluid job market. Universities might integrate AI across disciplines, while vocational and community colleges expand training in skilled trades.
AI could also support more personalized and continuous learning. As demand for reskilling grows, investments in lifelong learning will have to be made. Education systems and employers may need to work more closely together, using shared programs, flexible models, earn-as-you-learn apprenticeships, and faster credentialing to help people move across jobs and industries.
What systems are needed to ensure that transferable skills lead to new opportunities?
As AI transforms work, many people will need to move into entirely new occupations. Transferable skills will be essential to make those shifts, but they will matter only if the labor market can recognize and reward them. Clear definitions of skills, trusted ways of demonstrating ability—through testing or verified credentials—and better matching platforms could make this possible. Building links between employers, schools, and credentialing institutions could expand access to work and opportunity.
How can local economies and communities respond?
The impact of AI will vary widely across industries and regions. Understanding those differences through data is the first step toward effective action. With a clear picture of where change is happening, industry groups, educators, workforce agencies, and unions can work together on training and job-transition strategies that fit local needs.
Share
Sidebar
Acknowledgments
The partnership between people, agents, and robots is already taking shape as businesses embed the technologies in their workflows, changing skill profiles for jobs in many industries.
Today’s technologies offer vast opportunities to increase productivity and enhance human skills and will continue to advance. How work evolves depends on choices made now. Investing in workers and their skills—not just in technology—will be decisive in expanding human potential and ensuring that the benefits of AI are widely shared.
Glossary of terms
Adoption. The deployment of AI and automation technology into real work activities and workflows within an organization or labor-force context, determining how much of the automation potential is captured, how fast, and how broadly.
Agents. Machines that perform work activities in the digital world, augmenting or substituting a person’s nonphysical capabilities (e.g., natural language generation, social and emotional reasoning, and creativity).
AI-powered agents. Agents with AI embedded, allowing them to act more autonomously and orchestrate workflows; also known as agentic AI.
AI-powered robots. Robots with AI embedded, allowing them to act more autonomously and orchestrate workflows.
Artificial intelligence (AI). The ability of software to perform tasks that traditionally require human intelligence, potentially augmenting or substituting people’s capabilities.
Capabilities. Physical or nonphysical abilities that support the application of skills, assessed based on human levels of performance required to perform work activities. Nonphysical capabilities include cognitive (e.g., natural language, logical reasoning, creativity, and navigation) and social and emotional capabilities.
Generative AI. Applications of AI that take unstructured data as inputs and generate unstructured data through foundation models (i.e., large artificial neural networks trained on vast amounts of varied data).
Nonphysical work. Work that involves cognitive or social/emotional capabilities rather than physical movement, such as problem-solving, information processing, creating, or collaborating with others.
Occupations. A set of jobs that share similar tasks or work activities that can be described in terms of their skills, work contexts, and other qualifications. In the United States, occupations are formally classified using the Standard Occupational Classification system, maintained by the Bureau of Labor Statistics.
Physical work. Work that involves direct interaction with the physical world, requiring motion-based capabilities such as gross motor skills, fine motor skills, and mobility. These tasks typically include operating or moving objects, tools, or machinery; assembling or positioning materials; and performing actions that depend on human strength or dexterity.
Robots. Machines that perform work activities in the physical world, augmenting or substituting a person’s physical capabilities (i.e., gross motor skills, fine motor skills, or mobility).
Skills. Knowledge, competencies, and attributes that people deploy to perform work activities, often acquired through formal education, training, or work experience. Lightcast and ESCO provide a market-driven classification system for skills.
Technical automation potential. The share of work hours that theoretically could be automated with certain levels of technical capabilities. We assessed the technical automation potential across the US economy through an analysis of the detailed work activities of each occupation. We used databases published by the US Bureau of Labor Statistics and O*NET to break down about 800 occupations into approximately 2,000 activities, and we determined the capabilities needed for each activity based on how humans currently perform them at work.
Work activities. Observable work behavior that represents what people do to accomplish the objectives of an occupation. In the United States, activities are formally classified by O*NET into detailed work activities (DWAs).
Workflows. A structured sequence of work activities that collectively advance work toward a defined goal, guided by processes (e.g., rules, dependencies, information flows) and involving people and technologies.
小龙虾开发者社区是 CSDN 旗下专注 OpenClaw 生态的官方阵地，聚焦技能开发、插件实践与部署教程，为开发者提供可直接落地的方案、工具与交流平台，助力高效构建与落地 AI 应用
更多推荐
0
0- 0
已为社区贡献4条内容
所有评论(0)