The middle management layer is disappearing, specialists are being distilled out, and high-capability super individuals are going solo. How many people do companies actually still need?

Your boss assigns you a task that takes two weeks to complete, but today, with several AI Agents called up, you can finish it in just one hour.

When will you submit the deliverable then?

Li Ning, professor at Tsinghua University School of Economics and Management and author of AI-Native Organizations, once asked his students this question in class. Most people answered: "Two weeks later."

This answer is in fact the answer to another question. Why can corporate organizations hardly perceive efficiency improvement when individuals are advancing rapidly with the help of AI?

Individuals under the water have long been surging in undercurrents, while the organizations on the water surface remain completely calm.

Companies have spent heavily on purchasing tools and consuming Tokens. Everyone seems to work faster, but there is no qualitative leap in corporate output. Instead, a huge amount of duplication and redundancy is generated.

The crux of the problem no longer lies in the technology itself, but in the business foundation that has operated for more than 200 years.

Starting from the "pin-making factory" described by Adam Smith, the essence of the modern corporate system can be summed up in four words: decomposition and integration. Since human capabilities have boundaries, we break a complete goal into pieces, and create job types, departments and job ranks. Then we use long meetings, complex approvals and huge middle management to re-integrate countless "partial individuals" together.

Now, AI is pulling away the foundation of this logic. When a front-end programmer can easily build the back end and the database at the same time, when a business department of 50 people can be quietly replaced by 3 people with Agents, the boundaries of traditional division of labor are rapidly dissolving.

But on the other side of the coin, when "generation" becomes extremely cheap, "verification" has become a new barrier. Hundreds of Agents can produce an astonishing amount of workload in an instant, but cannot share even a second of hesitation, judgment and choice for human beings.

The division of labor logic of organizations is shifting from "splitting by workload" to the more hidden "splitting by cognitive bandwidth".

Deeper games are taking place under the surface. When companies require employees to write their livelihood-dependent experience into Skills and precipitate them into the company's digital assets, it is no different from asking a braised pork chef to hand over his ancestral secret recipe for an ordinary monthly salary. "Anti-distillation", "holding back tricks" and "shadow AI" are beginning to spread in office buildings. This shows that the old employment contract can no longer accommodate the super individuals amplified by AI, but the new benefit distribution mechanism has not yet taken shape. When the execution threshold is completely erased, the last chip human beings have left at the table is only the ability to define problems, and the taste that is hardest to quantify.

Recently, Under the Boil had a long conversation with Li Ning, professor at Tsinghua University School of Economics and Management and author of AI-Native Organizations.

At this moment when everyone is trying to demand certainty from AI, he unpacks the ongoing power transfer between organizations and individuals for us.

01

What else do organizations need to do after one person can handle more work

Under the Boil: Let's first talk about why organizations exist. What is the difference between an organization and a group of people gathering to work together?

Li Ning: This can be traced back to the division of labor discussed by Adam Smith. In The Wealth of Nations, he cited the example of pin making: splitting the work processes so that each person only does a part, and the efficiency will be much higher.

Why do we need division of labor? Because one person cannot learn all skills. In the past, when developing software, some people were in charge of design, some in charge of databases, and some in charge of operation. Our education is the same: we choose a major in college, and finally deeply cultivate in one industry.

But as the division of labor becomes more and more detailed, friction costs arise. The more complex the matter is, the more people need to be coordinated.

Now with AI, after you define the requirements clearly, it can do front-end, back-end, and also build databases. The original division of labor starts to become awkward: I can do your job, and you can also do mine. A front-end developer casually asks how to build the back end, and AI finishes it for him. Organizations need to rethink how these people can work together.

Under the Boil: Then at least in fields like programming, one person with multiple Agents can finish the work. Doesn't that mean organizations may no longer be necessary?

Li Ning: But this is not the same as being omniscient and omnipotent. AI has context limits, and it also needs people to tell it what is worth doing. For example, when we do scientific research, after defining a valuable problem, AI can do a very good job; but if we let it find problems by itself, especially in fields with strong subjectivity, it still cannot do well for now.

Go further: can one person plus AI handle everything? There is still the limit of human cognitive resources. I spend ten minutes explaining an idea, AI may work for a whole day, but it will come back to me for decision-making in the middle. While waiting for it, I can start another project, but when I have five or six projects running at the same time, many people will say: "I have no more ideas, and I can't let AI work for me anymore."

So division of labor may still be needed, but the content being divided has changed. I can run a product from the front end to the database, which does not mean I can think clearly about every detail. The page looks good, but is it the optimal solution? How do users feel when using it? AI generates a version, I look at it for ten minutes, then think for half an hour, which still consumes my cognitive resources.

One project can let hundreds of Agents run together, but they solve the workload problem, not necessarily the problem of human judgment and attention.

Under the Boil: Even in model companies, for such a complex task as model training, one person cannot master everything. Do we still need a leader, with other people supplementing his bandwidth to form a smaller pyramid?

Li Ning: Yes, smaller. It is indeed the case for a period of time, but the gap between people is not that big. In the past, interns and CEOs could hardly sit together to discuss strategies, because their cognition did not align. Now the knowledge people master and the background for understanding problems may be closer, and the communication methods will also change.

Under the Boil: I am also thinking that besides division of labor, organizations have other advantages such as supply chain and negotiation capabilities. These do not seem to be completely replaced by AI?

Li Ning: Yes, what AI mainly processes now is still the digital world. There are also design and information flow in agriculture and manufacturing, but there are also things like land and production. AI can optimize them, which does not mean replacing them. So not all organizations will become the same in the end, there will be hybrid types.

We also divide tasks into two dimensions: generation and verification. In the past, many people were responsible for working, and finally experts accepted the results, the bottleneck was mainly on generation; now a lot of content comes out at once, and verification cannot keep up instead.

Of course, there are also some tasks that AI rarely sees, so generation itself is difficult; some judgments are highly subjective and have no clear standards. These areas are less affected for now.

Under the Boil: That means organizational transformation has to make people better at verification, right?

Li Ning: Yes. Why do mathematics and code rank at the forefront? Because they can be verified: give it a clear reward function, it knows whether it is doing better or worse, and can continue to improve. But when it comes to what makes a good article, besides language and grammar, there is taste, which is not that simple.

However, this part is also changing. For example, I sort out the comments I gave to students in the past into a Skill, which contains my experience in judging research. We also fine-tuned a small model dedicated to evaluating research ideas to see if it meets the taste of top journals. The effect even exceeds that of experts and front-end large models, because it has learned the implicit signals in the domain data.

It can't do other things, but it can do this one thing well.

In the past, consulting companies needed many junior employees to help search and write reports, and finally experts talked to customers. In the future, reports can be written out quickly, but do experts have time to read them? Now we can publish one paper a day, and we can't read all of them at all.

The bottleneck that organizations need to solve is different from before.

02

The impact of super individuals

Under the Boil: Some time ago, people discussed token maxing: everyone used a lot of Tokens and produced a lot of output, but the organization did not change, and finally still could not pass the verification. Everyone makes their own Skills, which may also lead to repeated labor. How to solve this problem?

Li Ning: I often ask a question in class: your boss gives you a task, which takes two weeks and only excellent employees can finish it. You use Agent to finish it in one hour, when will you submit it? Many people say two weeks later.

Individual efficiency is improved, but the organization cannot see it. There are also things that originally do not need ten people to do, but the posts remain unchanged. Ten people each bring an Agent, and the output they produce is naturally similar.

So we need to clarify the requirements first. For example, we have ten new products to make, and then let everyone do it. If it is still this one product and these few customers, and we just ask everyone to use more Tokens, where is the exit? It is not a complete waste in the initial stage, at least everyone learns how to use it, but in the long run we can't just stay at this stage.

The same goes for Skills. Employees have work badges, permissions and training after joining the company, but Skills are very "wild": you make one that you think is good, and share it with me. Are there any risks in it, who certifies and manages it? None of these have kept up.

Under the Boil: You have investigated many companies. For current AI-native organizations, what changes have taken place?

Li Ning: One type is companies established after the emergence of generative AI, with a small number of people, and when encountering problems, they first think about whether AI can do it. The other type is traditional enterprises, very few of which have fully transformed, and most of them start to change from one department or one business point.

For example, supply chain experts, in the past, had to hand over requirements to product managers and then find IT developers; now they may build the system by themselves. The ones doing better are letting people who understand the business finish the whole thing.

But the difficulties often lie inside. A department of 50 people led by a founding veteran can be done by 3 people plus Agents. How can this department be adjusted? Who has the motivation to adjust? The business cannot stop, and no one knows whether the adjusted result can reach the original effect.

We have seen some enterprises set up a small separate team, while the old team continues to work, and let two or three AI-native employees test in parallel. It's like you need to renovate the house but you still live in it, so you prepare a new set outside first, and move in after it's done.

The early cost may even rise instead.

Under the Boil: Besides fewer people and working around projects, what else is different?

Li Ning: The responsibilities of people will expand. In the past, many things were assigned to different posts, now one person can do more work in both front and back ends, and the collaboration method has also changed.

This is different from the past digital transformation. In the past, the information flow was moved to the cloud, and the process was roughly unchanged; now the process itself needs to change, and it often starts from one department or one point.

In the past, what tested the organization was whether it could organize a thousand people to move towards one goal. Now what is more important is whether it can find super individuals and motivate them well. The way to manage a large army and the way to manage special forces are definitely different.

Under the Boil: Looking further ahead, where will AI-native companies go?

Li Ning: I think they must become innovative organizations. In the past, high efficiency and low cost meant competitive advantage. Now we look at many startup teams, their working methods are already very similar: large screens, voice input, one computer hanging five or six projects, and human cognitive bandwidth is almost fully used.

The difference lies in whether they can find new demands. It does not have to be a very large demand. Solving user pain points well in a small field can also bring opportunities.

Under the Boil: But sometimes the demand is very clear, for example, I just want to sell the product. Large factories use data and A/B testing to make judgments, and AI may be better at statistics and prediction. Will human judgment also be replaced?

Li Ning: What you said is the visible scope. How to place website buttons, how to price e-commerce products, how to combine promotions, these incremental improvements can be done through testing.

But there are unknown spaces, those demands that have never been seen before. This is like what March said about "exploration and exploitation". The exploration part still needs people to give directions for now.

Under the Boil: Then in the future, it may be that several people with genius ideas form a small team to do disruptive innovation?

Li Ning: I think it's not just geniuses.

Large factories have very high talent density, why haven't they made it? There is a problem of organizational orientation here.

We have contacted many large factories, they talk about ROI every day. But innovative things have never been seen before, how can you use historical data to calculate how long it will take to get the return? If everything has to answer this question first, many attempts cannot be made.

Under the Boil: Current organizations are more inclined to maintain operation, right?

Li Ning: Yes. When an employee has an idea, he needs to demonstrate it, hold a review meeting, get the green light and resources before he can do it. It made sense in the past, because development required a lot of people and time to invest.

Now the cost structure has changed, and holding a demonstration meeting may be the most expensive link. With the time for a meeting, you can give the task to AI and the demo will come out.

Some small teams require to come up with several ideas every week, and then select from them.

It's not just empty talk, there must be a visible prototype.

03

How far can a person's ability expand

Under the Boil: Let's talk about super individuals again.

Everyone's experience and skills are biased. I think it's still very difficult to make judgments in all fields. If 99% of the people in an organization do not have this ability, will we only organize that 1% of people in the future?

Li Ning: I don't think only 1% of people have potential. There are several problems that have not been solved now: how to identify such people, who can be trained, and how to connect the long-term talent echelon.

For example, junior posts are reduced, and everyone needs to judge experts. Where do experts come from? In the past, people started "learning by doing" as interns, and slowly became experts after seven or eight years or more than ten years. But this is only one path under the past division of labor, not necessarily the only path.

With AI, the way of learning may change. I have seen some people who did not seem to be super individuals at first, but after contacting Agents, it was like opening a door, they dived very deep, and may become very proficient in a field in just two or three weeks.

This is not only the difference in IQ, but also the direction, method and goal of learning. In the past, we spent a lot of time practicing operations and learning skills, but we may not have thought clearly about what the field is all about. Now the actual