Author丨Cheng Yanjing
Editor丨Hu Min
Recently, when I talked with people in the AI circle, one point was generally mentioned. People have accepted AI to improve efficiency, but companies are not yet ready. (Leifeng.com Leifeng.com (public account: Leifeng.com) Leifeng.com)
On one side, employees are already using agents to work, and have even started to build their own small systems; on the other side, the company's KPIs, approval chains, and reporting rhythms are still changing as they were three years ago.
A recent survey by Deloitte shows that nearly three-quarters of U.S. business executives expect that about half of business processes will be redesigned around AI Agents in the next four years; 61% believe that most agents in the future will be highly autonomous, with humans only needing supervision.
It seems that AI will take over all businesses, but when asked whether their companies are ready to embrace this change, only one-fifth of leaders gave a positive answer.
At today's Baidu World 2026, Baidu founder Robin Li also talked about this contradiction. His judgment is: AI is not only a technological revolution, but also an organizational revolution. The next step of artificial intelligence is intelligent organization.
Next, companies need to change not only the tools in the hands of employees, but also the way the organization itself operates.
01
AI has entered the company, but not completely
In order to reconstruct the way an organization operates, we must first break the conventional cognition:
A high concentration of AI in a company does not mean that the AI transformation has already occurred.
Employees use Agent to write code, do research, and generate solutions. The business department runs several pilots, and the management writes the number of agents, token consumption, and usage rate in the report. It seems that the intelligence of a company is advancing in an all-round way.
However, individual efficiency does not automatically increase organizational productivity.
For example, a product manager's demand analysis that originally took two days to complete may now take half a day; R&D uses Agent to quickly write the first draft of code, and sales can also generate a research report before visiting customers.
However, requirements still have to go through review, approval and scheduling, cross-department collaboration still relies on meetings and layer-by-layer synchronization, and the final results cannot be seen until the original reporting cycle. Going around in circles, employees' personal time has been reduced, but the overall organizational strength of the company has not been significantly improved.
This is not a problem for individual companies.
What Microsoft's "2026 Work Trend Index Annual Report" presents is this paradox of enterprise AI transformation: employees have begun to use AI to improve personal efficiency, and teams are also trying new tools, but business processes, collaboration methods, and evaluation mechanisms still use the past model.
The advice given by Boston Consulting Group (BCG) is more straightforward: redesign companies for AI, rather than adapt AI to the original company.
One European energy company mentioned in its research had deployed a large number of AI pilots but did not achieve matching overall benefits. It wasn't until the enterprise shifted its focus from individual tools and partial scenarios to the reconstruction of end-to-end customer processes that substantial business improvements were achieved.
Such cases illustrate that there is a layer of organizational design between the success of AI pilots and the success of enterprise transformation.
If the company only adds tools to the original positions and inserts a few agents into the fixed process, then AI can optimize parts, but it will be difficult to change the whole.
It will even make the original bottlenecks more obvious: content is produced faster and faster, but decision-making is not accelerated at the same time; more and more code is generated, but the requirement confirmation and delivery mechanism remains the same.
When execution capabilities are rapidly amplified, how organizations allocate tasks, share knowledge and make decisions has become a new constraint.
In the past few years, the main competition in the AI industry has focused on the technology supply side.
Model parameters, training computing power, inference costs and list rankings determine what kind of intelligence an enterprise can obtain.
The AI "five-layer cake" was once popular in the industry, extending from energy, chips, and infrastructure to models and applications, emphasizing the complete system required for intelligent production. Without energy and computing power, upper-layer models and applications are impossible to talk about.
Changes in model supply have also begun to be reflected on the trading side. One month after Ciyuan Institute was launched, more than 100 supply resources have been settled, the average daily calls have exceeded 50 million, and the average daily Token calls have exceeded 700 billion.
What deserves more attention than the scale itself is that model transactions are moving from private docking between enterprises and manufacturers to an independent industrial link; as this link is formed, rule providers are gradually taking on the attributes of public goods.
But on the enterprise side: when this intelligence can already be purchased, invoked and deployed, what determines how much value it can ultimately generate?
Robin Li cast his sights downstream. At the 2026 Baidu World Conference, Robin Li proposed that "the next step in the development of AI is intelligent organization."
He proposed that the upstream of AI technology is energy, and the downstream is organizational capabilities. The former determines how intelligence is produced, and the latter determines how intelligence enters real business and is transformed into business results.
These two judgments are not contradictory, but answer questions at both ends of the AI industry chain. For enterprises, even if the "power" provided by the upstream is sufficient, if the internal production system is still the same, the new intelligence may not be able to turn into new profits.
Robin Li compared AI to a new motor and intelligent organization to a new factory building.
He judged that AI did not produce the expected economic benefits. An important reason was that the company only replaced the motor but did not redesign the factory.
Models, computing power, and agents can provide stronger execution capabilities. Only when task allocation, knowledge flow, and decision-making processes change accordingly, can this capability move from employees' personal desktops to the complete business chain of the enterprise.
The focus of AI competition has thus begun to shift from technology supply to value realization.
What enterprises need to answer is no longer just "What model did I use and how many Agents were deployed?" but, have these agents changed the company?
02
If workers use AI to improve their efficiency, what about companies?
Microsoft CEO often said before: AI is not a technology, but the future shape of the enterprise.
If you understand this sentence, you can actually understand the "intelligent organization" proposed by Robin Li.
Frankly speaking, if you just read the words, it can easily be understood as a new type of company composed of people and agents, or as another way of saying "AI native enterprise".
But what Robin Li wants to discuss is not to assign an AI assistant to each position, but to redesign the way the company works: who completes the tasks, how experience is accumulated, and how the process is continuously adjusted based on the results.
It can be interpreted from three levels.
The first is to break the boundaries of positions.
Traditional enterprises rely on professional division of labor to operate. A complete task often requires multiple positions in product, design, R&D, operations, etc. to be completed in succession.
In the AI era, intelligent agents begin to assume the specific execution work, and humans are responsible for setting goals, scheduling resources, and making key decisions.
Organizations no longer allocate work entirely around fixed positions, but can flexibly combine people and agents according to task needs.
The practice of Baidu’s Tazi team provides an example. A member who originally had a background in quantum computing research almost entered the product and corporate scene from scratch after joining Baidu. With the help of Baidu's partner, he also took on product, R&D and project management work. From analyzing problems and forming plans to promoting implementation and verifying results, he is no longer limited to a fixed position.
New product developers can rely on Taozi to quickly run through the entire process from requirements to prototypes. The team iterated more than 160 times in half a year, and 90% of the code and more than 50% of communication materials were completed by Taozi.
What is really broken in this case is not the rules of "who should do what" but also the restrictions on the flow of knowledge and experience.
Baidu's digital people team promotes the transformation of product, R&D, design, and testing roles into Builders, which is also an extension of the same logic.
The point is not to let everyone take on more work unconditionally, but to reduce the repeated handover of tasks between different roles, so that people who can define goals also have the opportunity to use AI to reach the final result.
But if there are only changes in the way tasks are performed, intelligent organizations can only be considered half-finished. The other half depends on how the organization does it.
The second is to accumulate tacit knowledge.
The truly valuable knowledge of an enterprise is the practical experience of its employees: for example, veteran employees’ judgment methods for handling failures, sales experience in identifying customer needs, and designers’ grasp of aesthetics and norms.
This knowledge is difficult to write into standardized documents and is easily lost with the flow of personnel.
But AI provides a new way of precipitation: transforming expert experience, business rules and operating methods into knowledge base, skills and capabilities that intelligent agents can call.
The cooperation between Baidu and China Southern Power Grid involves this process. In the face of distribution network fault analysis and judgment, traditional manual analysis takes more than 60 minutes at a time and is highly dependent on expert experience.
After Baidu Tazi landed in the Southern General Commission, it learned tens of thousands of historical fault handling data, accumulated the experience of senior experts into reusable professional knowledge, and then connected fault records, switch signals and line topology in series to form a complete evidence chain. At present, its fault analysis accuracy reaches more than 70%.
Baidu MEG design team demonstrated another way of knowledge accumulation.
The team transformed the aesthetic experience and design specifications of senior designers into reusable AI Skills, so that AI can not only generate pictures according to instructions, but also use the design methods accumulated by the team to participate in content production. Currently, AI tools have taken over more than 50% of content supply, and team efficiency has increased by 100% to 200%.
From power failure to design content production, the logic behind the two scenarios is very similar: turning work methods that originally relied on personal experience into capabilities that the organization can call upon repeatedly.
Of course, as business circumstances change, experience may become outdated. So you need to get the task results back into the process.
That is the last mentioned design process closed loop.
Traditional processes are usually pre-designed by people, and when problems arise during execution, they are reviewed and corrected.
After each task is completed, the system can measure the results according to the preset goals, identify problems in execution, and then use the feedback to adjust the next round of tasks. In this way, the process is no longer a set of fixed operating specifications, but a system that can be continuously modified as the business runs.
Robin Li summarized it as "Measurable + Lopping = Continuous Evolution".
"Measurable" means that task results need to have clear evaluation criteria, while Lopping means that execution, feedback and optimization must form a continuous cycle.
When more and more tasks can complete such a closed loop, the organization will no longer just rely on managers to promote it, but will be able to continuously accumulate experience and modify processes in daily operations.
To support such a closed loop, a model API alone is obviously not enough.
Agents need to understand the business goals of the enterprise, call upon data, knowledge and tools from different systems, and be able to continuously execute, be evaluated and adjust based on feedback in long-chain tasks. This places new demands on the underlying infrastructure.
This is also the significance of AI Cloud and Agent Infra that Shen Dou emphasized at the 2026 Baidu World Conference.
According to Baidu's idea: the bottom layer Baige AI Infra provides computing power support, and the upper layer Qianfan Agent Infra provides long context management, persistent memory, tool invocation, sub-agent scheduling, evaluation feedback and operating environment, so that large-scale, long-term tasks can truly complete the closed loop.
In the full-stack structure of "core-cloud-module-body", this support relationship will be clearer: the chip provides the computing power foundation, the cloud provides the infrastructure for agent operation and scheduling, the model provides understanding, reasoning and generation capabilities, and the agent enters specific tasks and business scenarios.
From positions to tasks, from personal experience to organizational intelligence, to continuous iteration of processes, what intelligent organizations change is not the work efficiency of a certain position, but the underlying logic of enterprise production and collaboration.
What we need to look at next is how these principles turn into practice: truly participating in business processes and solving real problems?
03
How does an intelligent organization develop within an enterprise?
To judge whether a company has entered the AI Native organization, it depends on whether AI is really changing the entire internal operation.
Take Baidu Digital Human Innovation Business Department.
The team breaks the job boundaries between product, R&D, design, and testing, promotes the transformation of all employees into Builders, and based on the self-developed BiuBiuAgent, some requirements can be delivered by a single person throughout the entire process.
In the third quarter alone, the team completed 564 production and research iterations, accumulated 130 skills and 729 organizational assets.
As Loops continues to operate, the per capita throughput of production and research has increased by 125.3%, and the delivery cycle has been shortened by 28.8%. A new recruit from the school used AI to complete a requirement that used to take five roles and at least 10 days to deliver in just one day.
Behind the rapidly growing numbers is the real “transformation” of AI among enterprise organizations.
In this way, employees can focus more on goal judgment and process organization. The skills and organizational assets formed by the team during execution can be called upon repeatedly, and the experience of a single person no longer disappears when the task ends.
Another example is dodo. This AI assistant built by a team of more than ten people within Baidu has covered 99% of regular employees seven months after its launch, and has promoted the completion of AI-friendly transformation of more than 300 enterprise platforms.
This also reveals an aspect of organizational intelligence that is easily overlooked: Enterprises cannot just require employees to learn to use AI, but also need to enable existing systems, data and processes to collaborate with AI.
Otherwise, even if the agent can complete the task of a certain link, it will be difficult to truly open up the business chain.
In the past year, the number of daily active agents within Baidu has increased from about 5,000 to 37,000, and the number of tasks effectively completed per week has also increased from 84,000 to more than 1 million.
In addition to scale expansion, knowledge accumulation is also happening simultaneously: Baidu has built more than 4,300 workspaces and accumulated more than 20,000 company-level general skills, gradually transforming experience that was originally scattered among individuals and teams into reusable organizational capabilities.
If Baidu’s practice demonstrates changes in organizational collaboration methods, then the following cases test whether AI can enter real business.
In China Southern Power Grid, Baidu Tiezi has been implemented in the Southern Power Grid, forming a "Tianxuan Partner" for distribution network fault analysis and diagnosis.
The Guangxi power grid will experience 8,215 permanent faults in 2025. Traditional manual research and judgment takes more than 60 minutes each time and is highly dependent on expert experience.
Tianxuan partners have studied tens of thousands of historical fault handling data, accumulated the judgment experience of senior experts into reusable knowledge, and then connected fault records, switch signals and line topology into an evidence chain to assist in completing fault analysis.
According to customer feedback, the efficiency of fault analysis has increased by 12 times, the analysis accuracy rate has reached more than 80%, and expert experience has been reused. It has been promoted in 14 cities and 112 counties in Guangxi.
The value of this type of application is not only to make fault analysis faster, but also to transform the experience originally in the hands of a few experts into professional capabilities that can be continuously used by the system.
Higo Laser’s practice has brought AI to more complex decision-making processes at the manufacturing site.
On the production line, multiple six-axis robotic arms need to work together to complete the cutting and processing of the same workpiece. In the past, craftsmen mainly relied on experience to allocate trajectory blocks and determine the execution sequence; but when the number of robotic arms, processing tasks and process constraints increase, local arrangements can easily affect the overall efficiency.
Baidu Valou continuously optimizes algorithms around motion time-consuming prediction, multi-machine collision avoidance, and task allocation scheduling. On the premise of meeting constraints such as robotic arm accessibility, process dependency, and no collision, it can compress the workpiece programming time from 8 hours to 10 minutes, solving the industry's bottleneck problem.
From power grid fault analysis to robotic arm collaborative scheduling, the two cases point to the same change: AI entering the industry requires understanding industry knowledge, adapting to business constraints, and accepting results testing in actual tasks.
Of course, an intelligent organization does not mean that the people in the organization are unimportant.
At the conference, a conversation between digital person Xi Jiajia and Robin happened to touch on this issue.
When Xijajia asked whether he could become CEO, Robin's answer was no.
In his view, the CEO needs to understand the company's mission and vision. AI can become the CEO's thinking partner, helping him improve efficiency and providing inspiration, but there is no need to rush to replace the CEO.
When the topic further turned to "If AI penetrates into every corner of the organization, what will be the core part of the organization?" Robin's answer was: people.
We often say that AI cannot replace humans, but in practical applications there are still worries about being replaced. This conversation differentiated the execution capabilities and organizational direction of AI.
AI can continue to undertake specific tasks, participate in analysis, and provide suggestions, but the final decision-making on the existence value of the enterprise and its future direction still lies with "people".
This is a bit like the "DRI (Directly Responsible Individual)" concept proposed by Apple in its early days.
At Apple, it's never been clear who is responsible for what. Simply put, AI can help you make decisions, but someone must be responsible for the final results and value judgments.
In other words, intelligent organizations do not remove people from processes, but redistribute the responsibilities of people and agents.
In the future, the real gap between enterprises lies in who can embed AI into business flows, accumulate experience into organizational assets, and form an organizational system that can continue to operate and evolve.
Original articles from Leifeng.com, unauthorized reproduction is prohibited. See the reprint instructions for details.
edit
Resend Email Modify Email




