From Tencent to Baidu, AI-driven organizational change is moving from isolated experiments to sweeping overhauls.

"We are at a life-or-death moment." After Tencent's internal livestream marking its 28th anniversary on the evening of November 11, Pony Ma's remarks quickly went viral on social media.

Rumors circulated online that he had been "scowling the whole time" and that management had not spoken so sternly in a long while. Talk of "organizational change and personnel adjustments" spread rapidly. He said the goal this time "is not cost reduction, but adapting to the new battlefield of the AI era." The wave has already swallowed hill after hill. "Among the existing troops, who can get on the ship and whether they can fight a sea battle—these things must be prepared in advance. Frankly speaking, it is already a bit late." Those who choose to adapt to the new way of fighting "are not like before, when they still had to be coaxed, persuaded, have conditions negotiated, or first discuss whether they would be given a management position before they were willing to get on board." That night, claims that "Tencent is laying off 50%" flooded social platforms. Tencent later clarified that this was a misreading of the speech.

The next day, Robin Li made organization the theme of his speech at Baidu World. He said AI has not brought the expected economic benefits because companies have only replaced the motor without redesigning the factory. Companies that used to organize work by position and department will in the future operate more around tasks, break down job boundaries, and let people and Agents complete work together. "Positions will not disappear, but they will converge."

Within two days, the founders of two internet companies had both pointed the finger at their own organizations. "The time has come." An internet industry figure judged this in an interview with Digital Intelligence Frontline. In fact, his work over the past two or three years has been accompanied by corporate adjustment turmoil. "You do not wait until everything is ready before acting; you try while doing. There will be successes in this, and there will also be collateral damage."

AI has brought shocks to different companies. A product manager, with seven or eight AI digital employees, developed two systems in two weeks, and the traditional R&D division of labor began to break down; a major internet company split work into small tasks for employees to claim across departments, challenging the original management functions; some software companies began to compress middle-management positions and reassess the value of employees' roles.

The corporate organizational pyramid is beginning to loosen. Some people have gained the productive capacity of an entire former team, while others have found that the value of their original positions is disappearing. Is this an organizational revolution driven by AI, or are companies using AI to accelerate an adjustment that was going to happen anyway?

01 Major companies are taking the knife to their own organizations

This round of organizational adjustment at Tencent first targets management. According to Yicai, citing people familiar with the matter, in the internal speech on November 11, Pony Ma mentioned that some middle managers are unwilling to move, while frontline employees are full of drive. He asked the middle layer to be flatter and more agile, and encouraged frontline employees to charge forward on their own initiative.

But a real question is: if work no longer needs to be assigned layer by layer, what are middle managers supposed to do?

At Baidu, an experiment is also underway. This year, while preparing the communications for a large event, Baidu launched a new workbench. It was built with the company's low-code tools, and its interface was barely beautified. The workbench split the event into modules, and the modules into even smaller actions, openly available for employees to claim, with results submitted and points awarded upon completion. To avoid tasks being too difficult for anyone to claim, each piece of work was broken down simply enough.

"In effect, it turns what used to be systematic work into discrete actions." A Baidu employee told Digital Intelligence Frontline. Such actions "anyone can do," and they do not necessarily have to be done by professional PR staff.

In this employee's view, what this mechanism breaks first is departmental walls. Social media content is light enough to serve as a starting point, first pulling people from various departments onto the same platform. Assignment of work shifted from managers to the platform.

These experiments are not spontaneous employee initiatives. This employee heard that management has convened multiple meetings this year to push organizational transformation, and every business department has been required to explore new ways of working. "The company must have issued a death order that something has to be produced."

On April 28 this year, Baidu adjusted its job-grade system, connecting the professional and management tracks below the middle-management level into a unified digital job-grade system from levels 5 to 12. In this employee's observation, since then one of managers' important tasks has been exploring how organizations should operate in the AI era.

Some employees have observed that many tasks are not yet fully autonomously scheduled by Agents. But the direction of the next step is certainly automatic pushing by AI according to progress.

"You have become a fence or handrail for AI. You tell it that it should go this way, that it should do that. You teach it how this form should run properly."

And in Baidu's product department, other job boundaries have already been broken. At Baidu World on November 12, DODO product manager Chen Yaozong said that in the past he could write a little code but did not dare submit it directly for launch for fear of problems. Now he dares to modify the front end, back end, and even some content in the database, because R&D and testing staff have turned professional standards into AI review rules, and anything that does not meet the requirements is sent back.

DODO's core team has only a dozen or so employees. Chen Yaozong said that if you count the Agents behind each person, it feels like a team of more than 300 people. DODO's internal user group at peak has two to three thousand pieces of feedback a day. Every night, AI reads the feedback, checks the codebase, and then simulates the review perspectives of R&D, operations, and the boss, questioning the requirements over multiple rounds. By four or five in the morning, the requirements document has already been generated. Chen Yaozong checks it at 7 a.m., offers revision suggestions on the way to work, and by the time he reaches the office at 9 a.m., the content to be updated and launched that day has usually been decided.

At the same time, work that used to belong to different positions is beginning to be completed by fewer people. And the experience of professionals is written into Skills for other employees and Agents to invoke.

Employees' attitudes toward this change are also complicated.

Almost no one feels safe. Younger employees worry that they are too inexperienced; senior employees worry that their salaries are high and their original experience is no longer scarce; middle managers face the question of how much management value formed through coordination and passing information up and down can still be retained. A business leader sighed to Digital Intelligence Frontline that he does not know when he will leave the company and can only prepare mentally in advance.

"The department is still there, and the structure is still there. But some of the work has already bypassed the original management chain and is flowing on the platform."

"Everyone has accepted this." One person said. Some colleagues are proactive, both to respond to the company's requirements and to prove their value in the next round of personnel adjustments. Some take advantage of the company's job-task opportunities to embrace AI, learn to develop products, or try running social media accounts, gaining tens of thousands of followers in a few months, leaving themselves another way out and a bit more confidence for a possible future departure from the major company.

For major companies, this is an adjustment that touches existing interests and management relationships. And in some small companies, founders have already stepped in personally to redesign the entire organization.

02 The mother bug and the embryo

One second before he picked up my call, Wei Jiaxing had just arranged a task thread for AI.

He is the founder of Nanjing Yunfu Intelligence. Founded in 2018, the company uses AI voice agents to provide customer service, telemarketing, and customer follow-up for enterprises. Wei Jiaxing had not written code for many years, and this year he returned to the front line. He told Digital Intelligence Frontline that the new project he is responsible for updates about 1,500 feature points a month; the company's product and R&D team of about 30 people also uses AI programming tools and produces 500 feature points a month.

"That is equivalent to the R&D efficiency of 90 to 100 people." And this was still something he did part-time.

Wei Jiaxing compares his new role to the mother bug in StarCraft. "What does the mother bug do? It produces embryo-level products."

This year, he personally built eight internal company systems, covering everything from demand insight, R&D management, and marketing growth to CRM, human resources, and finance. Each was a rough initial version, which he calls an "embryo version." These were handed to eight executives who understand AI: the R&D management system to the head of product and R&D, the HR system to the head of HR, the finance system to the CFO, for them to continue iterating.

Because they came from one person's hands, the data in the eight systems was connected from the very beginning. Customer requirements enter the system, flow directly into the R&D management system, and are then assigned to engineers' programming Agents; once the code is written, the upgrade log is automatically generated, enters the marketing system, and is published to the company blog.

"In the middle of every pipeline, there is no human collaboration."

In a month and a half, he replaced all of the company's paid SaaS software with these eight systems. In the past, some parts of the company's commission system were unreasonable because the purchased software supported only a few fixed configurations, so the company had to accommodate the system. "I had to design company policy to fit the system." Now he can modify the software according to his own business rules.

Many companies' approach is to add Agents on top of existing systems. Wei Jiaxing feels this is "still a bit low-dimensional," because every transfer burns tokens and is unstable. He wants unnecessary steps to disappear entirely.

For example, in the past, each time Yunfu completed a version, usually containing about 80 requirements, a product manager had to read the requirements documents one by one and write the upgrade log, which took about five hours. Now AI has all the data, and the upgrade log is generated automatically.

"This step has completely disappeared."

In January this year, Yunfu broke the traditional professional division of labor in the R&D department. The previous groups for front end, back end, testing, operations, and UI were reorganized into full-stack group one, full-stack group two, and full-stack group three, with a separate design center and architecture center.

"If an engineer only uses AI to complete the part he is responsible for, with upstream and downstream steps still there, the capacity cannot really be brought out." Wei Jiaxing said. Letting one or two people take on an entire project with AI improves efficiency "not by 20% to 30%, but by 10 to 20 times."

He calls the new organization a "bubble factory." ByteDance and Tencent can run a bunch of business units at the same time and make countless apps, eventually producing winners; now a company of a hundred people can do the same. Each "bubble" consists of one to several core employees plus a group of Agents, independently advancing a project. If it achieves agreed performance, it gets 5% of the project's lifetime dividend. "The team then has a feeling of internal entrepreneurship, and without much risk."

However, Wei Jiaxing admits that this also creates some problems. After the middle platform is dismantled, code standards and reuse and accumulation across projects weaken. What worries him more is architecture. If a project is not well designed at the beginning, the Agent will keep writing along the wrong context, getting messier and messier.

"An engineering team without business thinking, if allowed to take independent responsibility for a project, could be disastrous." So he does not dare hand new projects to young people lacking architecture experience. He first writes the "embryo" himself, then hands it over.

The logic of hiring has also changed. Yunfu has hired very few people this year, "not because performance is declining, but because we really do not need to." In the past, if you passed the written test, you could get in and start with simple work. "Now there is no simple work for you to do; AI finishes the simple work." He recently hired two master's graduates from 985 universities and had them partner on a project, one understanding programming and the other understanding aesthetics, and both also needing to understand the business quickly.

The threshold for talent has risen. What about older employees? Wei Jiaxing admits that so far this year Yunfu has only let two people go, and the rest have mainly been absorbed through reassignment, even if their original salary was higher than that of the new position. He acknowledges that some people are no longer suited to their original work, but these employees have followed the company for many years and have not made mistakes.

"In the old era, they were fine. They were professional engineers."

What puzzles Wei Jiaxing is a set of numbers. "If capacity is expanded dozens of times, what if revenue only grows 20% to 30%?" The number of business opportunities cannot suddenly grow dozens of times.

The extra capacity can only be channeled into new products and extended into adjacent areas. But every new product has to start from zero—and it also diverts attention from the main battlefield.

"Very philosophical." He said he had been thinking recently about how to optimize, "maybe after another quarter, I'll have a clearer picture."

Now, this CEO-turned-programmer again assigns 20 tasks to programming Agents at 11:30 every night, then goes to sleep. When he wakes up the next day, the first thing he does is check the results.

03 Who's gaining, who's losing

In March this year, Tuya's operations team spotted an abnormal bill: one employee had burned through several thousand dollars' worth of tokens in a single day. Operations and HR suspected someone was using company computing power for private work and asked the security department to investigate.

It turned out to be a product manager. He was running seven or eight digital avatars simultaneously, working for him around the clock, and had built two systems in two weeks. "If it had been handed to internal R&D, it might have taken two or three months." Tuya CISO Liu Longwei told Digital Intelligence Frontline.

In Liu Longwei's view, the people in the company who use AI best are product managers who understand a bit of R&D. In the past, product managers controlled requirements but had to fight for people and scheduling across R&D departments; now, "raising a bunch of AIs" is enough.

Tuya has also begun reorganizing by business line. Departments originally divided by function—cloud, app and front end, hardware—have become business lines and project teams, with a small number of core people leading Agents. In March this year, the company also set a rule: each existing application should have no more than two maintainers.

In many companies, the people who can define requirements are the first to obtain the productivity amplified by AI.

A Baidu employee observed that the company is also bringing in a group of young AI talents from outside. Some call them "genius kids": young, able to produce results independently with a few people and computing power, and paid well. At the same time, the execution layer is becoming cheaper and cheaper. Almost all companies see that the next few years will be a "talent dividend period," with a large base of fresh graduates, allowing companies to recruit excellent students at a not-so-high cost.

Talent is being pulled apart at both ends, while the middle layer is squeezed.

The IT department is also being redefined.

Yang Senlin, vice president of the paper products company C&S, said the company's IT department no longer accepts any BI report development requests. In the past, IT spent more than 4 million yuan a year on outsourcing and manpower to write various reports for business departments; now business people use AI to do their own analysis. "The knowledge that used to be IT expertise has now been lowered in threshold. So what does IT do? It builds a stable foundation and provides technical help." President Gao Bo said that the IT department's job now is "not to help them generate an analysis or generate a model," but to act as a manager of tools and solve technical problems between tools and the business.

Tuya's R&D is doing something similar: transforming all existing code and business into a unified platform that AI can read. "Before, it was collaboration between people." In the future, interaction between AI and AI will be established through knowledge bases.

This force will in turn hit software companies. "Some IT departments certainly have the incentive to write their own systems to replace some paid software. Otherwise the IT department will be laid off." One person said.

In addition, those hit hardest in this round will be middle managers. Kai-Fu Lee recently offered an analysis: a company of 10,000 people may need about 3,000 managers reporting layer by layer. But if most work is handed to AI, AI does not need layer-by-layer reporting, "no politics, no information barriers." He said bluntly that if a manager only holds meetings, passes information, and manages people, there will be no place for him in the company of the future.

Software service companies are already feeling the change. A senior figure at iSoftStone told Digital Intelligence Frontline that now, from a requirement to a demo, "it is very fast; two or three people, or even one person, can get it done." A senior engineer can be used as 10 people, and front end, testing, and architecture can all be handed to AI, "as long as the requirements can be thought through clearly." Junior staff are decreasing, and the middle layer has also begun to consolidate.

In his eyes, the shape of the organization is slowly changing from a pyramid to a pine tree, with fewer and fewer people at the bottom, and the middle management layer will also be compressed. "Probably no company boss likes middle management." If AI can make every task crystal clear, "what do I need a director for?"

However, quite a few people believe that this round of organizational change is not all brought about by AI.

Zuo Chun, Honorary Chairman of SinoSoft, believes that hierarchy will naturally expand, and there also exists a mechanism for periodic adjustment. Taking Silicon Valley as an example, employees have to be promoted one level every few years, "without layers, how can you get promoted?" With more layers, people grow in the middle who only relay messages up and down. So Silicon Valley lays off a round about every ten years; old employees' salaries keep rising, but the work they do is similar to that of young people.

"AI has become an excuse, the best excuse."

Even "one person matching a whole team" does not feel new to him. "In a 10-person development team, originally only two people are the real core; AI just widens the gap even more."

Every company is changing, but at different speeds. Wei Jiaxing observed that bosses at traditional large enterprises are generally anxious, but the anxiety "cannot be transmitted to the frontline"; the frontline will find ways to push AI back again. Small companies are much simpler; there are only about ten middle managers in total, "you raise everyone's pay, and everyone immediately 'surrenders'." Among some industry clients served by iSoftStone, organization and processes have barely changed; what all parties are really fighting over is who should lead the AI platform, because whoever controls AI gets the budget and power.

But large companies do not necessarily stay slow forever. Wei Jiaxing has seen a major client that was initially very conservative about AI, until one of its distributors aggressively embraced AI and grew rapidly; the major client immediately turned around. "When a big company truly faces a crisis of survival, it will be more proactive than anyone."

04 The final signature is still from a human

At Tuya, the R&D team is being reorganized, but the security department is very stable. Because someone always has to be responsible for the code written by AI.

Liu Longwei said that the security team is also changing. In the past, they checked item by item according to a checklist; now they develop security "digital humans" for the product and research process to check automatically, then have humans confirm, and also set rules and define permissions for digital employees.

"Previously, IT security managed people; now it manages AI."

Zuo Chun has seen similar problems. AI has made code generation much faster, but the speed of confirmation has not kept up. "It's like a huge number of people writing articles, and there are not enough reviewers." AI can also review, but it is "irresponsible"; the final decision still has to be made by a human.

Kai-Fu Lee has a similar judgment: no matter how powerful AI is, it cannot bear responsibility. "If a CEO does a bad job, the board can dismiss him and cut his pay, but with AI, at most you can only stop using it." Therefore, he proposes that future enterprises will need more DRIs, that is, directly responsible individuals. In the past, managers managed people; future managers will manage AI and be responsible for the results. Enterprises will no longer need so many people who relay messages and assign work, but will need more people who can define requirements, set rules, judge results, and step forward when problems arise.

Li Yanhong also gave a similar answer at the Baidu World Conference: jobs will converge into two categories. Builders are responsible for creating new things, and Sellers are responsible for communicating with people and building trust. Enterprises will no longer need so many people who relay messages and assign work, but will need more people who can define requirements, set rules, judge results, and step forward when problems arise.

Ma Huateng asked, who can get on the boat. Those who get on the boat ultimately still have to be responsible for the direction of the voyage.