EARTH LinC / field brief/ 研究简报

Leadership in the AI-native organization

AI 原生组织中的领导力

How many people can one leader truly hold?

一个领导者，究竟能真正托住多少人？

AI has collapsed the cost of knowing. It has not collapsed the cost of trusting. This is a brief on the shape of the organization — span, layers, and the everyday behaviour that keeps a team whole when the work moves faster than the job description.

AI 让「知道」几乎变成免费，却没有让「信任」变便宜。这是一份关于组织形状的简报——管理幅度、层级，以及当工作跑得比职位说明书还快时，那些真正把团队托住的日常行为。

Global desk research 全球案头研究 24 sources, all linked 24 条来源 · 全部附链接 Read time ≈ 9 min 阅读约 9 分钟

“Adaptive change attracts attack. Pace the work, keep your allies, and don’t fight every gate at once.”

「适应性变革必然招来攻击。为工作定速，守住盟友，不要同时去撞每一道闸门。」

Ronald Heifetz & Marty Linsky · A Survival Guide for Leaders · Harvard Business Review, 2002 Ronald Heifetz & Marty Linsky · 领导者生存指南 ·《哈佛商业评论》2002

01

Structure

结构

The shape the best companies actually chose

顶尖公司真正选择的形状

Look past the org-chart slides and the pattern is boringly consistent: small owning teams, few layers, wide context. The famous structures below were not designed for elegance. They were designed so that one person could still know what everyone was doing.

越过组织架构图的华丽包装，规律其实相当朴素：小而自主的团队、更少的层级、更宽的上下文。下面这些著名结构并不是为了好看而设计的，它们是为了让一个人还能真正知道其他人在做什么。

Amazon 6–10 The “two-pizza team”: small enough to be fed by two pizzas, with one single-threaded owner accountable end to end. 「两个披萨团队」：小到两个披萨就能喂饱，并由一位单线负责人对结果端到端负责。 → 04
Spotify (2012) 6–12 Squads of six to twelve; tribes deliberately capped under ~100 people, explicitly citing Dunbar’s limit. 6 至 12 人的小队；部落人数刻意控制在约 100 人以内，明确援引邓巴数。 → 05 · 12
Haier 10–15 Around 4,000 self-governing microenterprises; more than 10,000 middle-manager roles removed in the process. 约 4,000 个自主经营的小微企业；过程中取消了一万多个中层管理岗位。 → 06
Netflix Context 上下文 “Highly aligned, loosely coupled.” Leaders are paid to set context, not to approve decisions. 「高度对齐，松散耦合。」领导者的职责是设定上下文，而不是审批决定。 → 07
Google Safety first 安全第一 Project Oxygen: manager quality moves team outcomes. Project Aristotle: psychological safety was the top predictor. 氧气计划：管理者质量显著影响团队结果。亚里士多德计划：心理安全感是首要预测因子。 → 08 · 09 · 10
Meta, 2023 Fewer layers 更少层级 The “year of efficiency” flattened the org and asked managers to do real individual work again. 「效率之年」压平了组织，并要求管理者重新承担真正的个人产出。 → 24

02

Span & layers

幅度与层级

How many is a good number — and how many is too many?

多少人算刚好，多少人已经太多？

There is no universal number, but there is a solid body of evidence. Neilson and Wulf found the average number of people reporting directly to a CEO roughly doubled — from about five in the mid-1980s to almost ten by the mid-2000s — while Rajan and Wulf documented the disappearance of layers between the chief executive and division heads. Firms did not get simpler; leaders got wider.

世上没有放之四海皆准的数字，但确有扎实的证据。Neilson 与 Wulf 发现，CEO 的直接下属人数大约翻了一倍——从上世纪八十年代中期的约 5 人，增至本世纪初的近 10 人；Rajan 与 Wulf 则记录了 CEO 与事业部负责人之间层级的消失。公司并没有变简单，是领导者变宽了。

The cost of width is attention. Gallup’s long-running data attributes about 70% of the variance in team engagement to the manager alone. A manager with twenty-five direct reports does not have a small amount of attention for each; below a threshold, attention stops being coaching and becomes scheduling.

变宽的代价是注意力。盖洛普长期数据显示，团队敬业度差异中约有七成仅由管理者一人决定。一个带 25 名直接下属的管理者，并不是「每人分到一点点关注」——低于某个阈值后，关注就不再是辅导，而只剩排程。

One manager · the people they hold

一位管理者 · 他所托住的人

418
8 Direct reports 直接下属
5 Layers to reach 10,000 people 覆盖一万人所需层级
14% Share of headcount who manage 全员中管理者占比

Arithmetic, not destiny: a perfectly uniform tree of 10,000 people. Narrow spans buy attention and pay for it in layers; wide spans buy speed and pay for it in loneliness. Most large organizations that work well land between four and seven layers in total.

这是算术，不是命运：假设一棵一万人的均匀树。窄幅度买来关注，代价是层级；宽幅度买来速度，代价是孤独。多数运转良好的大型组织，总层级落在四到七层之间。

Standardised, repeatable work 标准化、可重复的工作 15–25 The process carries the knowledge; the manager removes blockers and holds quality. 知识由流程承载，管理者负责扫除障碍、守住质量。
Stable operations with judgement 需要判断的稳定运营 10–15 Exceptions are frequent enough that people need a person, not a playbook. 例外足够频繁，人们需要的是一个人，而不是一本手册。
Mixed knowledge work 混合型知识工作 8–12 Weekly one-to-ones stay real at this size; beyond it, they quietly become status updates. 在这个规模上，每周一对一还是真的沟通；再多，它就悄悄变成了进度汇报。
Novel, AI-assisted, adaptive work 前所未有的 AI 协作与适应性工作 5–8 Verification, taste and re-scoping are constant. This is coaching load, and it is heavier than it looks. 校验、品味与不断重划边界是常态。这是辅导负荷，比它看起来重得多。
Any unit that must stay one culture 必须保持同一种文化的单元 ≤150 Dunbar’s limit on stable relationships — the reason Spotify capped tribes and W.L. Gore split plants. 邓巴给出的稳定关系上限——这正是 Spotify 限制部落规模、W.L. Gore 拆分工厂的原因。

Ranges synthesised from sources 02, 03, 05, 12–17. Treat them as a starting hypothesis for your own org, then measure whether one-to-ones still change anything.

区间综合自来源 02、03、05、12–17。请把它当作对你自己组织的初始假设，然后去检验：一对一是否还真的改变了什么。

03

The shift

位移

What AI actually changes in a people manager’s week

AI 究竟改变了管理者的哪一周

Most of what a manager used to hoard — knowledge, translation, access, the shape of the role — has become cheap and fast. What has become expensive is judgement, verification, and the slow accumulation of trust.

管理者过去所囤积的大部分东西——知识、翻译、信息通道、岗位的形状——如今都变得又便宜又快。变贵的，是判断力、校验，以及信任那缓慢的累积。

Knowledge transfer

知识传递

months of shadowing a senior person to learn the domain.

跟着资深同事影随几个月，才摸到门道。

A competent first draft of almost anything, in seconds. The manager’s value moves from knowing to judging.

几秒钟就能得到一份像样的初稿。管理者的价值从「知道」移向「判断」。

Copilot: −55.8% task time → 22

Information asymmetry

信息差

the quiet basis of managerial authority.

管理权威悄悄依赖的地基。

Near zero. Authority has to be re-earned through clarity of context and quality of decisions.

几乎归零。权威必须靠上下文的清晰与决策的质量重新赢得。

Novices gain most: +34% → 21

Language & distance

语言与距离

translation as a filter, and a bottleneck, on global teams.

在全球团队里，翻译既是过滤器也是瓶颈。

Effectively free. Fewer intermediaries, more direct contact — and far less excuse for vague context.

几乎免费。中间人更少，直接接触更多——含糊的上下文再无借口。

Writing tasks: −40% time, +18% quality → 20 写作任务：耗时 −40%，质量 +18% → 20

The job description

职位说明书

rewritten once a year, if ever.

一年重写一次，甚至从不重写。

Moving faster than HR can publish it. Hire and staff for direction and adjacent skill, not for the box.

变化快过 HR 的发布速度。按方向与相邻能力去招人和配人，而不是按那个格子。

39% of core skills change by 2030 → 18 到 2030 年 39% 的核心技能将改变 → 18

Open study

开放学习

a training budget, a curriculum, a queue.

一份培训预算、一套课程、一条排队的名单。

Anyone can run their own study group. The manager becomes curator of what matters and protector of the hours to do it.

任何人都能自己发起学习小组。管理者成为「什么重要」的策展人，以及那段时间的守护者。

Output & risk

产出与风险

measured in volume produced.

以产出的数量来衡量。

Measured in what survives verification. Inside the frontier AI lifts quality sharply; just outside it, confident work gets confidently wrong.

以能通过校验的部分来衡量。在能力边界之内，AI 大幅提升质量；一旦越出边界，自信的产出会自信地出错。

+12.2% tasks, +40% quality — but −19pp outside → 19 任务量 +12.2%，质量 +40%——边界之外 −19 个百分点 → 19

04

Trust

信任

Trust is a behaviour, repeated. Not a speech, delivered.

信任是被重复的行为，而不是被发表的讲话

Frei and Morriss describe trust as standing on three legs — authenticity, logic and empathy — and note that when trust wobbles, it is almost always one specific leg, not all three. Edmondson’s two decades of research point the same way: people take interpersonal risk when the evidence of everyday behaviour says it is safe to. Neither can be produced by talking about trust more.

Frei 与 Morriss 把信任比作三条腿的凳子——真实、逻辑与同理——并指出信任动摇时，几乎总是某一条腿出了问题，而不是三条都塌。Edmondson 二十年的研究指向同一处：只有当日常行为的证据表明「这样做是安全的」，人们才会去冒人际风险。这两者，都不会因为多谈几次信任而产生。

In an AI-native team the practice gets more literal, because the work itself is now partly a machine’s draft. Showing how you decide is the whole job.

在 AI 原生的团队里，这件事变得更具体——因为工作本身已经部分是机器的初稿。把「你如何做决定」摊开来，就是这份工作的全部。

  1. Say the constraint out loud within 24 hours of learning it. 知道约束的 24 小时内，把它说出口。 Withheld context is read as either contempt or chaos. Both cost more than the bad news. 被扣住的上下文，会被读成轻视或混乱，两者的代价都高于坏消息本身。
  2. Show your own AI drafts — including the ones you threw away. 展示你自己的 AI 草稿——包括你扔掉的那些。 It is the fastest way to make verification normal instead of shameful. 这是让「校验」变成常态而非羞耻的最快方式。
  3. Name the decision-maker, the deadline and the evidence, in writing. 用文字写明：谁决定、什么时候、依据什么。 Ambiguity about who decides is the most expensive form of politeness. 对「谁来决定」含糊其辞，是最昂贵的一种礼貌。
  4. Give credit with a name and a specific verb. 给认可时，带上名字和一个具体的动词。 “Great work, team” is noise. “Mei rewrote the eval set” is a signal others can copy. 「大家辛苦了」是噪音；「小美重写了评测集」才是别人能模仿的信号。
  5. Protect two uninterrupted hours a day that you never break into. 每天守住两小时不被打断的时间，你自己也不闯进去。 Attention is the manager’s only non-renewable budget — and the team’s too. 注意力是管理者唯一不可再生的预算——对团队也一样。
  6. Repair fast: name the miss, state what changes, then actually change it. 快速修复：说出失误，说明会改什么，然后真的去改。 Trust is rebuilt by the third step, not the first two. 重建信任的是第三步，而不是前两步。
  7. Re-scope roles quarterly, out loud, with the person in the room. 每季度当面、公开地重划一次角色边界。 If the boundary is going to move anyway, move it with consent rather than by surprise. 边界反正会移动，那就带着同意去移动，而不是靠突袭。

05

Pacing

定速

Change this big has to be paced, or it eats its own leaders

这么大的变革必须定速，否则它会吞掉发起它的人

Heifetz and Linsky’s warning is twenty years old and has never been more useful: adaptive change asks people to give up something they value, and so it attracts attack — usually aimed at the person, not the idea. Flattening layers, redrawing roles and putting AI inside daily work is exactly that kind of change.

Heifetz 与 Linsky 的警告已有二十年，却从未如此适用：适应性变革要求人们放弃他们珍视的东西，因此必然招来攻击——而攻击通常指向人，而不是想法。压平层级、重划角色、把 AI 放进日常工作，正是这样一种变革。

So: get on the balcony before you get on the floor. Give the work back to the people who own it. Open one gate at a time and let the organisation digest it. Keep your allies close, and keep the dissenters where you can hear them — they are usually holding the part of the truth you skipped.

所以：先上阳台，再下舞池。把工作交还给真正拥有它的人。一次只打开一道闸门，让组织有时间消化。守住盟友，也把异见者留在你听得见的地方——他们手里往往正握着你跳过的那部分真相。

The org chart is a hypothesis about attention. AI changed how fast we can know. It did not change how slowly we come to trust.

组织架构图，是一个关于注意力的假设。AI 改变了我们知道得有多快，却没有改变我们信任得有多慢。

Sources & further reading — every claim above, linked

来源与延伸阅读 —— 以上每一项主张，均附链接

Read the originals

去读原文

  1. A Survival Guide for Leaders Heifetz & Linsky · HBR · 2002 The pacing doctrine this brief is built on. 本简报据以成立的「定速」原则。
  2. How Many Direct Reports? Neilson & Wulf · HBR · 2012 CEO spans roughly doubled, ~5 → ~10. CEO 管理幅度大致翻倍：约 5 → 约 10。
  3. The Flattening Firm Rajan & Wulf · NBER WP 9633 · 2003 Layers between CEO and division heads disappearing. CEO 与事业部之间的层级正在消失。
  4. Amazon’s Two-Pizza Teams AWS Executive Insights Small teams, single-threaded ownership. 小团队，单线负责。
  5. Scaling Agile @ Spotify (PDF) Kniberg & Ivarsson · 2012 Squads, tribes, and the ~100-person cap. 小队、部落，以及约 100 人的上限。
  6. The End of Bureaucracy Hamel & Zanini · HBR · 2018 Haier’s microenterprises, in detail. 海尔小微模式的详细拆解。
  7. Netflix Culture Netflix · current Context, not control. 用上下文，而非控制。
  8. Google re:Work Google · Project Oxygen & Aristotle The manager-behaviour and team-effectiveness research. 管理者行为与团队效能的原始研究。
  9. What Google Learned About the Perfect Team Duhigg · NYT Magazine · 2016 Psychological safety, told as a story. 用故事讲清楚的心理安全感。
  10. Psychological Safety and Learning Behavior in Work Teams Edmondson · ASQ · 1999 The founding paper of the field. 该领域的奠基论文。
  11. Begin with Trust Frei & Morriss · HBR · 2020 Authenticity, logic, empathy — and which leg wobbles. 真实、逻辑、同理——以及哪条腿在晃。
  12. Neocortex size as a constraint on group size Dunbar · J. Human Evolution · 1992 Where the number 150 comes from. 数字 150 的出处。
  13. Why Great Managers Are So Rare Gallup ≈70% of engagement variance sits with the manager. 敬业度差异约七成取决于管理者。
  14. State of the Global Workplace Gallup · annual Engagement and manager wellbeing, worldwide. 全球敬业度与管理者状态的年度数据。
  15. The State of Organizations 2023 McKinsey Delayering, spans, and the squeezed middle manager. 去层级、管理幅度，与被挤压的中层。
  16. The State of AI McKinsey · annual Adoption rates and where value actually lands. 采用率，以及价值真正落在何处。
  17. Global Human Capital Trends Deloitte · annual Skills-based organizations and boundaryless roles. 以技能为基础的组织与无边界岗位。
  18. Future of Jobs Report 2025 World Economic Forum 39% of core skills expected to change by 2030. 预计到 2030 年，39% 的核心技能将改变。
  19. Navigating the Jagged Technological Frontier Dell’Acqua et al. · HBS / BCG · 2023 +12.2% tasks and +40% quality inside the frontier; sharply worse outside it. 边界之内任务量 +12.2%、质量 +40%；越界则明显更差。
  20. Experimental evidence on generative AI and productivity Noy & Zhang · Science · 2023 Writing tasks: 40% faster, 18% higher quality. 写作任务：快 40%，质量高 18%。
  21. Generative AI at Work Brynjolfsson, Li & Raymond · NBER · 2023 +14% overall for support agents; +34% for novices. 客服整体 +14%，新人 +34%。
  22. The Impact of AI on Developer Productivity Peng et al. · arXiv · 2023 GitHub Copilot: 55.8% faster on the benchmark task. GitHub Copilot：基准任务快 55.8%。
  23. Building the AI-Powered Organization Fountaine, McCarthy & Saleh · HBR · 2019 Why the org design, not the model, is the bottleneck. 瓶颈在组织设计，而不在模型。
  24. Update on Meta’s Year of Efficiency Meta · 2023 A large-scale flattening, described by the company itself. 一次大规模压平，由公司自己讲述。