十年前的百度世界 2016 大会上,李彦宏曾站在台上提出「AI is the New Electricity」(人工智能是新电能),提前预判了人工智能浪潮。
那时,很少有人能想到,走进写字楼,程序员用 Agent 写代码,运营用 AI 跑报表,连行政都在用大模型搭报销流程。
不过,索洛悖论也再次产生。你去问企业管理者,十有八九会叹气。据统计 2025 年全球企业在 AI 上砸下超过 2500 亿美元,真正跑出业务变革的公司只有 25%。

每个人手里的活都变快了,Token 烧得更狠了,却没有创造更大的效益。为什么个人效率拉满,企业总体却没有跑出生产力变革?难题究竟出在模型能力还不够强,还是企业仍然在用旧的组织方式,硬塞进一项新工艺?
面对这个全球性的难题,APPSO 在今年的百度世界大会上,听到了李彦宏给出新的答案,AI 没有产生预期的经济效益,是因为笔者只是更换了电机,却还没有重新外观工厂。
时隔十年,李彦宏再次作出预判:AI 的下一幕,是智能组织。

「新生产力,只能从 AI 流水线里长出来,从智能组织里长出来。所以,换一台 AI 新电机,只是起点;重新外观新厂房,才是分水岭。而这座新厂房,就是智能组织!」
过去,笔者习惯用员工数量、人才密度、资本投入来衡量一家公司的实力。但在 AI 时代,能调动多少智能,沉淀多少知识,又能让多少复杂任务高质运转,成为组织生产力提高的必答题。
这场变化,正在动摇公司沿用了上百年的组织逻辑。
从一棵树到两块披萨
1855 年,美国纽约与伊利铁路公司画了一张特殊的图。

这是一棵枝叶繁茂的大树。树上的不同分支,分别对应公司的管理层、职能部门与人员分工。
这张被广泛认为是现代企业最早的组织结构图之一的图表,试图应对当时一个棘手的难题,「公司越来越大,到底该怎么管」?
工业革命让铁路、工厂和大型企业迅速扩张。几千上万人聚集在同一个生产体系中,必须有人负责决策,有人执行,有人监督。
后来,马克斯·韦伯平台阐述的科层制理论,把这套逻辑变得更加清晰:专业分工、明确权责、层级管理。
到了互联网时代,亚马逊创始人贝佐斯提出了一个有趣的管理原则,一个团队应该小到两块披萨就能让所有成员吃饱。
「两块披萨团队」背后的想法,是让小团队具备更大的自主权,减少层级与沟通成本,更快地完成任务。

从工业时代的分工管理,到互联网时代的扁平协作,公司的组织方式一直在变化。但有一件事始终没变,无论组织架构怎么调整,完成复杂工作的主体,仍然是人。
AI 开始打破这个前提。
今年 5 月,李彦宏在 Create 大会上提出「超级个体」的构想:过去一个公司的最小生产力单元是团队,如今,一个人加上一支智能体编队,就可能成为新的生产力单元。
工业时代应对如何组织更多人,互联网时代应对如何让人协作得更快。
到了 AI 时代,难题变成了,「如果智能也可以像人力一样被组织、调度和协作,公司是不是该重新外观了」。
「新电机」不能再配「旧厂房」
今年 3 月,Block CEO Jack Dorsey 和红杉资本合伙人 Roelof Botha 合写了一篇长文《从层级制走向智能化》(From Hierarchy to Intelligence)。
Dorsey 认为,从古罗马军团「8 人小队、80 人百人队、5000 人军团」的嵌套指挥链,到 1806 年普鲁士军队在耶拿战役惨败后创建的「总参谋部」,现代企业里「中层管理」的底子早就打好了。
两千年来的层级制,本质上是一套建立在人类局限之上的「信息路由协议」。
可人脑带宽有限,一个主管能有效管理的人数只有 3 到 8 人。想缩小管理幅度就得增加指挥层级,而层级越多,信息流动就越慢。
过去几十年,Spotify 试过敏捷小队,Zappos 试过彻底取消管理头衔的合弄制,可只要公司规模越过几百人的门槛,最终全都会退回层级制。
Dorsey 希望建立一个持续更新的「世界模型」,理解公司内部的工作状态、资源规格和客户需要,再由智能平台协助协调任务。
他甚至提出,要把公司变成一个类似「mini-AGI」的智能平台。

Dorsey 抛出的难题,确实击中了当下企业 AI 转型的一处痛点,AI 已经越来越聪明,公司却不一定变得更聪明。
李彦宏即将在百度世界大会提出的「新电机、新厂房」,正是对这一矛盾的回应。
如果只给旧厂房换上一台更强的电机,却不重新外观生产流程,新工艺的威力就很难完全释放。
企业如何自己长成智能组织,今天李彦宏也提出了三个原则,实际上就能当作一套完整的组织进化路径:
打破岗位边界,释放组织活力;
沉淀隐性知识,构建组织智能;
外观流程闭环,推动组织进化。

传统组织是「固态」的,按部门和职位切分工作,而智能组织则是「液态」的,将来的公司就像一张围绕目标动态聚散的任务网络,从「以岗位为中心」,转向「以任务为中心」。

可个人变得更能干,不代表公司就能跟着受益。
企业里最宝贵的往往是那些没写进文档的「隐性知识」:资深销售为什么放弃某个看似优质的商机,老工程师为什么先排查某个故障点。
员工积累的经验,如果只能由自己使用,换个人接手,很多事情还是得从头摸索。
在李彦宏看来,真正的组织智能,必须让隐性知识显性化,借助 Skill、Workspace 等方式沉淀到公司平台上,持续累积、随时调用、人人可用。

这意味着组织里的信息和能力,开始摆脱原有的部门边界。从让一个人变得更能干,到让整个公司具备更多可复用的知识,AI 才开始从个人工具走向组织资产。
不过,知识可以沉淀,还只是开始。
如果一家企业每次都能调用过去的经验,却无法依据新的业务结果调整工作方式,那么它仍然只是一套更先进的自动化平台。
真正的智能组织,还需要具备不断学习和自我改进的能力。
李彦宏将这种机制概括为一个公式:「可衡量 + Looping = 持续进化」。

简单来说,AI 不仅要把任务做完,还要有明确的业务标尺来检验结果,让每次执行产生的经验、偏差和反馈回流进下一轮循环(Loop)。
在李彦宏的设想中,将来人和 AI 围绕一个任务形成一个 Loop,公司就变成一个由无数 Loops 组成的、高度自动化运营的组织,组织,就变成一个不断学习和进化的平台。
这场实验先在百度内部跑通了。
百度数字人创新业务部推动全员向 Builder 转型,打破机型、研发、外观、测试等岗位边界,让员工借助 AI 承担更加完整的工作流程。
今年第三季度,团队完成了 564 次产研迭代,沉淀 130 个 Skills 和 729 篇组织资产,产研人均吞吐提高 125.3%,交付周期缩短 28.8%。
其中,一名校招新人借助 AI,仅用一天便完成了过去需要五个角色、至少十天才能交付的一项需要。
比起这个颇具冲击力的效率数字,更值得注意的是,团队并没有让这些经验随着任务结束而消失。
智能组织不是往公司里塞进越多 Agent 越好,工作过程中积累的 Skills、规则和方法,可以继续被其他成员调用,成为下一次任务的起点。

今年百度世界大会上,百度数字人希加加还取代真人主播采访李彦宏。希加加向李彦宏抛出了一个格外犀利的难题:
既然我都能做采访了,那我这样的数字人能当 CEO 吗?
这听起来像是一句玩笑,却提出了一个值得思考的难题:AI 已经可以独立完成越来越多的工作,它在企业中究竟应该扮演什么角色?
当 AI 可以承担任务、积累经验、持续进化,从一个听从指令的工具,逐步成为参与组织运转的新角色,企业的组织形态也将迎来新的想象空间。

智能组织的新厂房
提出一种新的组织理念并不难,难的是,让它在真实业务中跑起来。
过去十多年,百度持续投入芯片、云基础设施、大模型和智能体,积累了从底层算力到应用的 AI 能力。
但有了工艺底座,不代表企业就自动完成了组织升级。

智能组织要回答的,是从各自使用 AI 工具,转向让人与多个智能体围绕任务协作。这次百度世界大会,百度搭子进一步升级,试图改变这种各自为战的状态。
百度内部就已经在成为智能组织的新样本,一位原本学物理的新人机型经理,借助百度搭子,已经能独立承担售前沟通、机型外观以及大部分研发工作。
过去,一项需要可能需要机型、外观、研发等多个角色接力完成,现在他可以自己调动 AI 完成其中的大部分环节,T+1 就能拿出一个可以现场演示的能力 Demo。
当下,百度搭子已经推出个人版和企业版。在百度搭子开发平台的内部实践中,一项典型需要从分析到可交付思路的时间,从 7 天缩短到 2 天。
智能组织不仅协助员工应对当下的难题,还试图让今天应对的难题,成为明天整个组织具备的能力。
库库企业版围绕「懂组织、深能力、自进化」升级,它背后连接着百度文库积累的 18 亿份专业内容、7.5 亿篇学术文献,以及超过 700 万个在线知识库。
越来越多的企业,也开始把库库用到实际工作中。
上线仅三个月,库库企业版的能力已覆盖超过 60 万个组织,日均任务量增长至上线初期的 9 倍,月访问量增长至 14 倍。在南华期货,客户服务效率提高 60%;在锡创投,业务交付速度提高 25%;在国联人寿,投研团队人效提高 40%。

不只是百度搭子和库库,近一年百度内部的日均活跃智能体数(DAA)从日均 5000 个增长到 3.7 万个,增长了 6 倍多。智能体执行的有效任务完成数,从每周平均 8.4 万个,增长到一周超过 100 万个,增长了 11 倍。
当下,百度已建设 4300 个 Workspace,组织 Skill 累计沉淀超过 2 万个。
李彦宏称,「这些组织的好经验、好方法沉淀下来,可以让更多同学可以复用,让智能体可以调用。让一个人的经验,成为整个组织的能力。」
让 AI 从一个个独立工具,进入企业真实的工作流程。
模型和算力的供给方式,也在发生相应变化。
当越来越多的智能体进入业务流程,企业不仅需要更强大的模型,还需要可以依据不同任务,灵活挑选和调用合适的模型,让智能体持续、高质地工作。
为了应对这个难题,百度还在此次世界大会上推出大模型交易平台「词元所」。

李彦宏说它是大模型的交易所,「词元所希望汇聚好的模型和模型给出商,让企业乃至个人都能便捷地找到、购买和使用可信的模型 Token」。
对于智能组织来说,模型不再只是需要单独采购和部署的工艺机型,它更像水电可以按需获取、灵活调用的基础资源。

词元所试运营一个月以来,入驻供应资源已超 100 个,日均调用次数突破 5000 万,日均调用 Token 超 7000 亿。
APPSO 认为,「智能组织」之所以会成为 AI 时代新的组织范式,在于它把公司的运行底座从「按层级管人的静态架构」,改写成了「按任务闭环攒复利的自进化平台」。
一百多年前,第二次工业革命早期并不顺利。蒸汽时代机器的位置受中央传动轴和皮带约束,换上电动机后的整整三十年里,工厂生产率几乎没动。
直到每台机器取得独立电机,工厂拆掉笨重的传动轴,重新安排机器、工人和生产流程,电力才真正释放出重塑工业生产的力量。
工艺增强,并不一定意味着效率的提高,电力革命真正改变工厂的,不只是那台新买来的电动机。今天的 AI,或许正处在类似的时刻。几乎没有哪家公司还缺一台名为 AI 的「新电机」。
真正决定公司将来十年身位的,是谁先动手重建那座名叫「智能组织」的新厂房。
本文来源:https://www.ifanr.com/1689182
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
At the Baidu World 2016 conference ten years ago, Li Yanhong stood on stage to propose "AI is the New Electricity," anticipating the wave of artificial intelligence early.
At that time, few could have imagined that upon entering the office building, programmers would use Agents to write code, operations staff would use AI to run reports, and even administrative personnel would use large models to build reimbursement workflows.
However, the Solow paradox has emerged once again. If you ask corporate executives, nine out of ten will sigh. According to statistics, global enterprises invested over $250 billion in AI in 2025, yet only 25% of companies truly achieved business transformation.
Tasks in everyone's hands have become faster, and Tokens are being burned more heavily, yet greater benefits have not been created. Why is individual efficiency maximized, yet the enterprise as a whole has not achieved a productivity revolution? Does the problem lie in model capabilities not being strong enough, or in enterprises still using old organizational methods to force a new technology?
Facing this global challenge, APPSO heard a new answer from Li Yanhong at this year's Baidu World conference: AI has not generated the expected economic benefits because one has merely replaced the motor without redesigning the factory.
Ten years later, Li Yanhong has made another prediction: the next chapter of AI is intelligent organizations.
"New productivity can only emerge from the AI assembly line and from intelligent organizations. Therefore, replacing the motor with a new AI one is just the starting point; redesigning the new factory is the watershed. And this new factory is the intelligent organization!"
In the past, we used to measure a company's strength by the number of employees, talent density, and capital investment. But in the AI era, how much intelligence can be mobilized, how much knowledge can be accumulated, and how many complex tasks can be operated with high quality have become essential questions for improving organizational productivity.
This change is shaking the organizational logic that companies have followed for over a century.
From a Tree to Two Pieces of Pizza
In 1855, the New York and Erie Railroad Company drew a special diagram.
This is a large tree with lush branches and leaves. The different branches on the tree correspond respectively to the company's management, functional departments, and division of labor.
This chart, widely considered to be one of the earliest organizational charts of modern enterprises, attempted to address a thorny problem at the time: "As companies grow larger and larger, how exactly should they be managed?"
The Industrial Revolution led to the rapid expansion of railways, factories, and large enterprises. With thousands or even tens of thousands of people gathered in the same production system, there had to be those responsible for decision-making, execution, and supervision.
Later, the bureaucratic theory expounded by Max Weber made this logic even clearer: professional division of labor, clear authority and responsibility, and hierarchical management.
In the Internet era, Amazon founder Bezos proposed an interesting management principle: a team should be small enough that two pizzas can feed all its members.
The idea behind the "Two-Pizza Team" is to grant small teams greater autonomy, reduce hierarchy and communication costs, and complete tasks faster.
From the division-of-labor management of the industrial era to the flat collaboration of the Internet era, the way companies are organized has been constantly evolving. But one thing has never changed: no matter how the organizational structure is adjusted, the subject that completes complex work is still people.
AI has begun to break this premise.
In May this year, Robin Li put forward the concept of "super individual" at the Create Conference: in the past, the minimum productivity unit of a company was a team; now, a single person plus a fleet of agents may become a new productivity unit.
The industrial era focused on how to organize more people, while the Internet era focused on how to enable people to collaborate faster.
By the AI era, the challenge has become, "If intelligence can also be organized, dispatched, and collaborated with just like human labor, shouldn't companies be reimagined?"
"New motors" should not be paired with "old factories"
In March this year, Block CEO Jack Dorsey and Sequoia Capital partner Roelof Botha co-wrote a lengthy article titled *From Hierarchy to Intelligence*.
Dorsey argued that from the nested command chain of the ancient Roman legions—"8-person squads, 80-person centuries, and 5,000-person legions"—to the "General Staff" established by the Prussian army after its crushing defeat in the 1806 Battle of Jena, the foundation for "middle management" in modern enterprises was long laid.
The hierarchical system that has existed for two millennia is essentially an "information routing protocol" built on human limitations.
However, human brain bandwidth is limited, and a manager can effectively oversee only 3 to 8 people. To narrow the span of management, more command tiers must be added, and the more tiers there are, the slower the flow of information.
Over the past few decades, Spotify has tried agile squads, and Zappos has tried holacracy, a system that entirely eliminates management titles. Yet as soon as a company's size crosses the threshold of a few hundred people, it will ultimately revert to a hierarchical system.
Dorsey hopes to build a continuously updated "world model" that understands the working status, resource specifications, and customer needs within the company, and then have an intelligent platform assist in coordinating tasks.
He even proposed turning the company into an intelligent platform similar to "mini-AGI".
The challenge posed by Dorsey has indeed hit a pain point in the current enterprise AI transformation: AI is becoming increasingly intelligent, but companies do not necessarily become smarter as a result.
The "new motors, new factories" concept that Robin Li will put forward at the Baidu World Conference is precisely a response to this contradiction.
If you merely install a more powerful motor in an old factory without reimagining the production process, the full power of the new process will hardly be unleashed.
On the question of how enterprises can grow into intelligent organizations on their own, Robin Li also put forward three principles today, which can effectively serve as a complete path for organizational evolution:
Break down job boundaries to unleash organizational vitality;
Accumulate tacit knowledge to build organizational intelligence;
Close the process loop to drive organizational evolution.
Traditional organizations are "solid," dividing work by departments and positions, whereas intelligent organizations are "liquid." Future companies will resemble a task network that dynamically gathers and disperses around goals, shifting from being "position-centered" to "task-centered."
However, an individual becoming more capable does not necessarily mean the company will benefit accordingly.
The most valuable assets in a company are often the "tacit knowledge" not documented: why a senior salesperson abandons a seemingly high-quality business opportunity, or why a veteran engineer prioritizes checking a specific fault point first.
If the experience accumulated by employees can only be used by themselves, whoever takes over will still have to start from scratch to figure out many things.
In Li Yanhong's view, true organizational intelligence must make tacit knowledge explicit, accumulating it on the company's platform through methods such as Skill and Workspace, so that it can be continuously accumulated, called upon at any time, and accessible to everyone.
This means that the information and capabilities within the organization are beginning to break free from traditional departmental boundaries. AI only begins to transition from a personal tool to an organizational asset when it moves from making a single person more capable to equipping the entire company with more reusable knowledge.
However, merely accumulating knowledge is just the beginning.
If a company can retrieve past experience every time but fails to adjust its working methods based on new business outcomes, it is still merely a more advanced automated platform.
A truly intelligent organization also needs to possess the ability to continuously learn and self-improve.
Li Yanhong summarizes this mechanism as a formula: "Measurable + Looping = Continuous Evolution".
Simply put, AI must not only complete tasks but also be evaluated by clear business metrics to verify results, allowing the experience, deviations, and feedback generated by each execution to flow back into the next loop.
In Li Yanhong's vision, in the future, humans and AI will form a Loop around a task, turning the company into a highly automated operational organization composed of countless Loops, and the organization itself will become a platform for continuous learning and evolution.
This experiment was first successfully implemented internally at Baidu.
Baidu's Digital Human Innovation Business Department is driving all employees to transition to Builders, breaking down the boundaries of positions such as modeling, R&D, appearance design, and testing, and enabling employees to undertake more complete workflows with the help of AI.
In the third quarter of this year, the team completed 564 product-R&D iterations, accumulating 130 Skills and 729 organizational assets, increasing per-capita product-R&D throughput by 125.3% and shortening the delivery cycle by 28.8%.
Among them, a new hire from campus recruitment, leveraging AI, completed in just one day a task that previously required five roles and at least ten days to deliver.
More noteworthy than this striking efficiency figure is that the team did not let these experiences fade away with the completion of the tasks.
An intelligent organization is not about stuffing as many Agents into a company as possible; the Skills, rules, and methods accumulated during the work process can continue to be called upon by other members, becoming the starting point for the next task.
At this year's Baidu World Conference, Baidu's digital human Xijiajia also replaced a real-life anchor to interview Robin Li. Xijiajia posed an especially sharp and challenging question to Robin Li:
Since I can even conduct interviews, can a digital human like me become a CEO?
While this sounds like a joke, it raises a thought-provoking question: AI can already independently complete more and more work, so exactly what role should it play within an enterprise?
As AI takes on tasks, accumulates experience, and continuously evolves, gradually transitioning from a tool that follows instructions to a new role participating in organizational operations, the organizational structures of enterprises will also open up new possibilities for the imagination.
The New Workshop of Intelligent Organizations
Proposing a new organizational concept is not difficult; the hard part is making it run in real business operations.
Over the past decade, Baidu has continuously invested in chips, cloud infrastructure, large models, and agents, accumulating AI capabilities ranging from underlying computing power to applications.
But having a technological foundation does not mean that an enterprise has automatically completed its organizational upgrade.
What an intelligent organization needs to answer is how to shift from each person using AI tools independently to having people collaborate with multiple agents around tasks. At this Baidu World Conference, Baidu Duazi was further upgraded, aiming to change this fragmented state of working in isolation.
Baidu itself is already becoming a new model of an intelligent organization; a new model manager who originally studied physics, with the help of Baidu Duazi, can now independently handle pre-sales communication, model design, and most of the R&D work.
In the past, a task might require multiple roles such as model, design, and R&D to complete in relay; now he can independently deploy AI to handle most of the steps, and can deliver a capability Demo suitable for on-site demonstration within T+1.
Currently, Baidu Duazi has launched both a personal version and an enterprise version. In internal practice on the Baidu Duazi development platform, the time for a typical task from analysis to a deliverable approach has been shortened from 7 days to 2 days.
An intelligent organization not only assists employees in addressing current challenges, but also seeks to transform the challenges addressed today into capabilities possessed by the entire organization tomorrow.
Koku Enterprise Edition has been upgraded around "understanding organizations, deep capabilities, and self-evolution," backed by the 1.8 billion pieces of professional content, 750 million academic documents, and over 7 million online knowledge bases accumulated by Baidu Wenku.
More and more enterprises are also beginning to apply Koku to their actual work.
Within just three months of launch, the capabilities of Koku Enterprise Edition have already covered more than 600,000 organizations, with daily average task volume growing to 9 times the initial level at launch and monthly visits growing to 14 times. At Nanhua Futures, customer service efficiency has improved by 60%; at Xichuang Tou, business delivery speed has increased by 25%; at Guolian Life Insurance, the investment research team's productivity has improved by 40%.
Beyond Baidu Dazi and Koku, over the past year, the daily average number of active agents (DAA) within Baidu has grown from 5,000 per day to 37,000, an increase of more than 6 times. The number of valid tasks completed by agents has grown from an average of 84,000 per week to over 1 million per week, an increase of 11 times.
At present, Baidu has built 4,300 Workspaces, with over 20,000 organizational Skills accumulated.
Li Yanhong said, "The good experience and methods of these organizations are accumulated so that more colleagues can reuse them and agents can call upon them. Let one person's experience become the capability of the entire organization."
Enabling AI to move from individual standalone tools into enterprises' real workflows.
The way models and computing power are supplied is also undergoing corresponding changes.
As more and more agents enter business processes, enterprises need not only more powerful models but also the ability to flexibly select and call upon appropriate models according to different tasks, enabling agents to work continuously and with high quality.
To address this challenge, Baidu also launched the large model trading platform "Ciyuan Suo" at this World Conference.
Li Yanhong said it is an exchange for large models, "Ciyuan Suo hopes to gather good models and model providers, so that enterprises and even individuals can conveniently find, purchase, and use trustworthy model Tokens."
For intelligent organizations, models are no longer just standalone industrial machines that need to be purchased and deployed individually; they are more like basic resources such as water and electricity that can be obtained on demand and flexibly called upon.
Since Ciyuan Suo began its trial operation a month ago, onboarded supply resources have exceeded 100, daily average call volume has surpassed 50 million, and daily average Token calls have exceeded 700 billion.
APPSO believes that the reason "Intelligent Organization" has become a new organizational paradigm in the AI era lies in its transformation of the company's operational foundation from a "static hierarchical structure for managing people" into a "self-evolving platform that accumulates compound returns through closed-loop task completion."
Over a century ago, the early stage of the Second Industrial Revolution did not go smoothly. In the steam era, machine positions were constrained by central drive shafts and belts; for a full thirty years after electric motors were introduced, factory productivity barely moved.
It was only when each machine obtained its own independent motor, when factories dismantled the heavy drive shafts and rearranged machines, workers, and production processes, that electricity truly unleashed the power to reshape industrial production.
Process enhancement does not necessarily mean improved efficiency. What the electricity revolution truly changed in factories was not just the newly purchased electric motor. Today's AI may be at a similar moment. Almost no company now lacks a "new motor" named AI.
What truly determines the company's position over the next decade is who takes the initiative first to rebuild the new plant named "Intelligent Organization."