AI News

Li Yanhong Proposes Three Principles for Intelligent Organizations at Baidu World Conference 2026

TMTPost · AGIUpdated Nov 13 16:01

AI summary

At Baidu World Conference 2026, Li Yanhong proposed three principles for intelligent organizations: breaking down job boundaries, accumulating tacit knowledge, and designing closed-loop processes, and presented the formula "Measurable + Looping = Continuous Evolution." Baidu disclosed that over the past year, internal daily active agents increased from 5,000 to 37,000, the number of effective tasks completed per week rose from 84,000 to over 1 million, and organizational Skills accumulated more than 20,000 in total. After all members of Baidu's Digital Human Innovation Business Department transitioned to Builders, they completed 564 product R&D iterations in Q3, with per-capita throughput increasing by 125.3% and delivery cycles shortened by 28.8%.

Body · AI translation

Employees are getting faster and faster, but companies remain the same, and this gap has become a challenge faced by enterprises worldwide.

On November 12, Baidu World 2026 opened at the China National Convention Center Phase II in Beijing. In his speech, Baidu founder Robin Li offered a direct explanation: "AI has not produced the expected economic benefits because we have only replaced the motor, but have not yet redesigned the factory."

Compared with the releases of models, chips, and applications, this topic seems rather plain. How jobs are divided, how knowledge is shared, and how workflows are adjusted sound like a management class.

But for Baidu and all AI companies, this is an AI story that must be explained with extreme clarity: with massive AI investment, is what comes back a vague "ghost story" about the company's future, or a good story in which the company is certain to win in the AI era?

Baidu is a typical example. It continues to invest in chips, cloud, models, and agents. What customers ultimately buy cannot be only new technology. After AI enters a company, whether products can be delivered faster, services can be improved, and costs can be reduced determines how much companies are willing to keep investing.

Individuals efficient, organizations inefficient

A misjudgment easily arises in enterprises' AI transformation: companies encourage employees to use AI well, token consumption keeps hitting new highs, and they conclude that the company's AI transformation is in great shape. But the more these results are concentrated in individuals and single roles, the more likely they are to conceal the fact that processes have not changed.

The truly difficult adjustments involve authority, responsibility, and interests between departments, and are hard for employees to complete on their own. If the success of local pilots makes managers stop reforming, the result may be that everyone is busier and faster, while the company shows no obvious change.

Microsoft's 2026 Work Trend Index Annual Report describes the same phenomenon and calls it the enterprise AI transformation paradox: employees are already using AI to improve personal efficiency, but business processes and evaluation mechanisms remain stuck in the past, and the organization has not changed much.

Similarly, BCG published an article with a very blunt title, "Redesign the Company for AI, Not AI for the Company." In its research, a European energy company deployed a large number of AI pilots without achieving overall returns, and only after shifting to rebuilding end-to-end customer processes did it achieve substantive improvement.

The root of the gap lies in the company's factory settings. The organizational form of the modern company was designed for a purely human workforce.

The world's first organizational chart came from a railway company. In 1855, the New York and Erie Railroad drew a tree-shaped chart, with branches labeling functional departments and personnel division of labor. Responsibility division plus hierarchical management supported the scale of the industrial age. Weber later summarized this system into bureaucratic theory. It managed scale, at the cost of rigidity.

In the internet era, Bezos proposed the two-pizza team, small enough to be fed by two pizzas, with decision-making delegated and layers compressed. But no matter how company layers are compressed, for more than a century the organization has always had only one kind of participant: people.

For the first time, organizations have participants who can work without being on the payroll. Processes and role settings designed for a purely human workforce have become the old factory building that has not moved amid the roar of the new motor. Improvements in individual efficiency occur in the cracks of old processes, and efficiency in the cracks does not automatically accumulate into organizational productivity.

Baidu chose this moment to discuss the intelligent organization, precisely pushing the question to this level: employees have already gained new work capabilities, so how does the company plan to use these capabilities?

Baidu's answer

AI chips, large models, and agents can all be bought with money, but intelligent organizational capability must grow within the enterprise itself. The famous science fiction writer William Gibson has a widely circulated quote: "The future is already here—it's just not evenly distributed." This is exactly a portrait of the present.

The same model can be sold to different companies in the same industry, but after it enters an enterprise, what data it can access, what rules it uses to make judgments, and who is responsible for the results differ from company to company. Whether the experience a company has accumulated over the years can be used by AI requires more long-term cultivation than which model was chosen at procurement.

This is also the significance of discussing the intelligent organization.

Back to Baidu World 2016, the theme of that edition was "AI is the New Electricity." At that time, Robin Li already proposed that AI is the new electricity and judged that the next act after the mobile internet would be artificial intelligence.

At Baidu World 2026, Robin Li proposed that the intelligent organization includes three principles: break job boundaries to unleash organizational vitality; precipitate tacit knowledge to build organizational intelligence; and design closed-loop processes to drive organizational evolution. The three concern how work is allocated, how experience is preserved, and how results are improved.

First, work allocation. Traditional organizations are centered on roles, while organizations in the AI era will become centered on tasks, meaning staffing must be rearranged around delivery results. Professionals still need to gatekeep, but which work must be handed to specialized roles and which can be completed by business staff with the help of agents need not fully follow past divisions.

Next, knowledge. Baidu proposes that true organizational intelligence is making knowledge explicit and precipitating it onto the company's platform, where it can continuously accumulate, be called upon at any time, and be available to everyone.

"Precipitating tacit knowledge" must deal with much experience that is difficult to transfer. By organizing the basis for judgment, the handling process, and actual results, agents can call upon it, and other employees can reuse it. What enterprises accumulate should include the methods for getting something done, not just documents from past work.

Finally, an intelligent organization should be an intelligent system capable of autonomous evolution. How can autonomous evolution be achieved? Li Yanhong gave a formula—Measurable + Looping = Continuous Evolution.

After a task is executed, the results need to be checked, the method corrected based on feedback, and then used for the next round of work. The hardest step here is often determining what counts as doing it well.

In the future, a company will become a highly automated operating organization composed of countless Loops, and ultimately, an organization will become a system that continuously learns and evolves.

Judging from these three principles, Baidu is attempting to push enterprises' use of AI into daily management. Complementing this is its layout of chips, cloud, models, and agents: models handle understanding and reasoning, the cloud platform supports task execution, agents connect to tools and business systems, and the organization provides specific goals, knowledge, and permissions.

Baidu's four layers—chips, cloud, models, and agents—are all self-developed. Its own chips run its own cloud, its own models run on the cloud, and the models grow their own agent applications. These layers reinforce one another and optimize cyclically; the entire architecture itself is a Looping.

The first batch of explorations of intelligent organizations

One of the most complete samples of an intelligent organization at present is Baidu itself, and it already demonstrates both scale of use and depth.

At the conference, Baidu disclosed that over the past year, the number of internal daily active agents increased from 5,000 to 37,000; the number of effective tasks completed per week rose from 84,000 to more than 1 million; organizational Skills accumulated more than 20,000 in total, and 4,300 Workspaces were built.

Take a specific example: dodo is an AI assistant built by a small team of about a dozen people inside Baidu. In seven months since launch, it covered 99% of regular employees, drove more than 300 enterprise platforms to complete AI-friendly transformation, and this year won Baidu's highest award, a million-dollar prize.

Look next at Baidu's Digital Human Innovation Business Department: the fixed division of labor among product, R&D, design, and testing was broken down entirely, and all staff shifted toward becoming Builders. In Q3 alone, it completed 564 product-R&D iterations, accumulated 130 Skills and 729 organizational assets, increased per-capita throughput by 125.3%, and shortened the delivery cycle by 28.8%.

Why did this change happen first inside Baidu? Besides awareness, the full-stack layout of chips, cloud, models, and agents is an important reason.

Shen Dou, Executive Vice President of Baidu Group and head of Baidu AI Cloud Business Group, mentioned that the new AI cloud will penetrate enterprises' underlying resource systems, IT systems, and even operating systems as never before. Because this infrastructure no longer serves only any one object among people, tools, and processes, but serves the individual tasks jointly constituted by these units.

At the same time, agents internalize enterprise tasks and drive the evolution of intelligent organizations. AI no longer exists independently as an optional tool, but is restructured into business flows and task loops, truly improving enterprise operating efficiency. AI cloud infrastructure, supporting large-scale agent workloads for agents, requires new infrastructure. At the bottom, Baige AI Infra provides more powerful and more cost-effective computing services; at the top, Qianfan Agent Infra provides an environment supporting large-scale agent operation and ultimate model services. These two layers enable large-scale, long-cycle tasks to truly complete a closed loop.

"The future of intelligent organizations will not be completed by any one company alone. We hope to build the infrastructure well so that intelligent organizations across all industries can take root and grow on the cloud," said Shen Dou.

In addition, Baidu launched the large-model trading platform Ciyuansuo, which is also a very interesting move.

What Ciyuansuo solves is not "where can I buy a model," but "is the model I bought really the model it claims to be." As model supply increases at a rate of several per day, the cost of choice is shifted to the buyer—what enterprises truly lack is reusable access rules and continuous monitoring mechanisms, not yet another model catalog.

The relationship with intelligent organizations lies in this: once AI enters daily business, enterprises need to continuously judge service quality, rather than relying only on a one-time test at procurement. When model identity, response speed, and availability are verifiable, it becomes easier to find the cause when a task goes wrong. Ciyuansuo provides the basis for judgment on the model supply side; how much revenue these investments ultimately generate still has to be accounted for in the enterprise's own business results.

The platform focuses on two major capabilities, "authenticity verification" and "credit reporting." It conducts resource verification through related technologies such as model fingerprint comparison, capability testing, and performance and availability testing, while also adopting methods such as regular spot checks and dial testing and full-link real-time monitoring of calls to establish credit ratings for model providers. It provides customers with neutral, credible, open, and transparent model trading services. Since Ciyuansuo began trial operation one month ago, more than 100 supply resources have onboarded, daily average calls have exceeded 50 million, and daily average tokens called have exceeded 700 billion.

Through this complete technology stack of chips, cloud, models, and agents, the form of intelligent organizations is also spreading beyond Baidu. China Southern Power Grid is a sufficiently solid sample: Baidu Dazi was deployed in the China Southern Power Grid General Dispatch Center, becoming a Tianxuan partner, serving the power supply guarantee for more than 200 million people across five southern provinces.

Take distribution network fault assessment as an example. In 2025, Guangxi Power Grid had 8,215 permanent faults. Traditional manual assessment took more than 60 minutes per case, relied heavily on expert experience, and was highly prone to misjudgment.

The China Southern General Dispatch Center and Baidu built Tianxuan Partner based on cooperation on Baidu Dazi. It learned tens of thousands of historical fault handling records, precipitated senior expert experience into reusable knowledge, and linked fault records, switch signals, and line topology into a complete evidence chain, realizing closed-loop assessment from fault to analysis, with an analysis accuracy rate of more than 70%, a 12-fold improvement in fault analysis efficiency, and full promotion in 14 prefecture-level cities and 112 counties in Guangxi.

Baidu Dazi has more than 160 ecosystem partners, has integrated more than 1,800 ecosystem capabilities, and covers more than 30 professional fields. For more enterprises, the threshold for connecting to an intelligent organization is much lower than building a set of capabilities on their own.

The replacement of organizational forms has always been a slow variable. It took half a century for bureaucracy to go from Weber's theory to the default setting of global companies. How fast intelligent organizations will transform corporate forms may well be the standard answer to the timeline for AI to deliver value. (This article was first published on TMTPost APP)

Source

TMTPost · AGI