AI is replacing middle managers

Jensen Huang has about 60 executives who directly report to him. He doesn’t appoint a second-in-command, and he doesn’t hold one-on-one meetings. He says, “I never hold one-on-one meetings with them. If there’s a problem, put it on the table—everyone tackles it together.”

According to traditional management theory, the upper limit for how many people a leader can manage effectively is 3 to 8; beyond that, things easily get out of control. But Jensen Huang has used this extremely flat structure to reach the top three in global market value.

When Musk took over Twitter in 2022, within a few weeks he slashed 7,500 employees down to around 1,500. Outside observers widely predicted the platform would collapse. Instead, Twitter (now X) not only survived—the product iterations continued as well.

These two cases reveal the unspoken question inside modern companies: What role do the middle layers in a pyramid structure actually play?

Middle layer: A relay station for information

The truth may be a bit brutal: we’ve kept a hierarchical management system for over two thousand years, and it has never been the best solution for organizational evolution—it’s just a reluctant fallback dictated by human biological limitations.

Whether it’s Qin’s commandery-and-county system, the nested command structure of the Roman army, or the staff officers specifically set up when the Prussian army was rebuilt in the 18th century, the hierarchical design of every large-scale organization in human history essentially solves the same problem: how information can flow accurately across large populations.

The core function of a middle-layer manager boils down to only one thing: translate the decisions above into instructions that can be executed below, and aggregate the situations below into information that those above can understand. They’re not tools of power; they’re information protocols.

When Ren Zhengfei expanded Huawei in its early days, he spent a great deal of effort building a business unit system and establishing collaboration processes. The underlying logic was the same—how to enable tens of thousands of people to coordinate, and how to prevent information from being distorted during transmission.

Now looking back at Twitter’s layoffs, what Musk did is logically clear: he was verifying which roles were truly creating the product, and which roles were merely handling information transmission and internal coordination.

When Musk bypassed these information relay stations and had engineers face the code directly, the machine kept running. But Twitter’s approach was brutal. After large-scale layoffs, many coordination functions that were genuinely useful disappeared as well, and some of the platform’s problems were exposed shortly afterward.

This shows: simply cutting people won’t solve the root problem—you need a replacement mechanism.

Three experiments by Chinese tech giants

Before a replacement mechanism appears, intense competition among Chinese internet companies had already made them aware of the losses occurring during information flow, and they tried to push human coordination efficiency to its limits using extreme systems.

ByteDance tried to break down information barriers through extremely high transparency. Through internal collaboration tools like Feishu, they gave employees enough background context to reduce reliance on command and approval and reduce dependence on middle-layer command. But as the company grew, ByteDance ultimately moved back toward a business-line-based divisional structure—hierarchy quietly returned. Pure transparency wasn’t enough.

PDD (Pinduoduo) streamlined its hierarchy to the extreme and imposed strong result orientation, allowing it to achieve e-commerce efficiency numbers no one else could match. But this path is nearly impossible to replicate. It relies on extreme talent density and cultural pressure—most companies can’t afford that cost.

Zhang Ruimin of Haier, in 2005, split 80,000 employees into more than 4,000 micro start-up units. Each unit was directly responsible to users, and it eliminated a large number of middle management positions. But over the next 10 years, more than 20,000 employees left because they couldn’t adapt to such a highly self-driven model. The larger the scale, the higher the coordination cost; pure restructuring by system eventually hits a ceiling.

These three paths confirm a reality: in purely human organizations, pursuing extreme flattening often faces a dilemma between losing control or applying high pressure. Remove the middle layer, and the information flow is likely to break. Using systems and culture to improve coordination efficiency—what human mechanisms can do has basically already topped out.

Manus: 78 people, $125 million

A replacement mechanism that can truly break through this ceiling is emerging.

In 2025, an AI company called Manus, which was founded in China and later relocated to Singapore, provided the ultimate answer. With only a 78-person team, it achieved in nine months roughly $125 million in annualized revenue. By the end of 2025, Meta acquired it for more than $2 billion.

At Manus, there’s no traditional approval process, and no project managers coordinating in the middle. Their system works through multiple AI programs operating in cooperation—one responsible for breaking down tasks, one responsible for execution, and one for verifying results. The AI autonomously completes all coordination, while these 78 people focus on handling the parts the system can’t temporarily reach: instinctive judgment, value boundaries, and the first step in entirely new scenarios.

This isn’t a lucky start-up. It’s proof of how far efficiency can go once a loophole is plugged. A small team unlocking immense commercial value—that is the efficiency transformation enabled by AI taking over coordination authority.

AI: The infrastructure that takes over the coordination layer

Manus isn’t an exception. This is happening on a much larger scale.

Jack Dorsey, founder of the U.S. fintech company Block (which owns the well-known payment apps Square and Cash App), made a seismic architecture adjustment earlier this year. The company cut headcount from over 10,000 to below 6,000, mainly by eliminating middle-management roles, product manager roles, and analyst roles.

In a letter to shareholders, the executive wrote: this is to rebuild a new way of working by using intelligent tools plus smaller, flatter teams. The market reacted positively, and the stock price jumped significantly on the day.

But more worth watching than this round of layoffs is what they’re building. To let AI truly replace the middle layer, you need two things.

One is a real-time view of what’s happening inside the company. Block is a remote-work company; decisions, code, design—everything leaves digital traces. By reading these data, the system can grasp the company’s internal state in real time: who is doing what, where things are stuck, and where resources are located. In the past, this required managers to align through meetings; now, AI automatically aligns internal status.

The other is a customer truth model. The core responsibility of traditional middle layers and product managers is to understand user needs and translate them into product directions. But interviews and questionnaires naturally introduce bias—what people say and what they do often don’t match. Block turns the movement of funds into behavioral facts by processing millions of real transaction records. Changes in merchants’ cash flow, user migration, and shifts in consumption patterns can all be captured and combined into inputs for decision-making.

With these two layers of capability stacked together, the information integration and decision-coordination functions originally carried by the middle layer are essentially replaced as a whole. The organization begins to shrink into three types of roles:

Deep specialists who work directly within the complete context provided by the system; domain owners who receive a specific problem and, within a limited time window, have the authority to mobilize all resources needed—after the problem is solved, the role is dissolved; people-development specialists who both do hands-on work and train others around them.

A large number of progress-alignment meetings disappear, because the system has already synchronized information at the underlying layer. The long-term form of middle-layer management is starting to be rewritten.

But for all of this to work, the prerequisite is: does the company have data with sufficiently high density and sufficiently real digitization, so the system can construct these two “reality maps”? Without that foundation, AI is only a more efficient tool—not the organization itself.

The middle layer doesn’t disappear because they aren’t trying

Jensen Huang uses 60 direct reports to bypass hierarchy, and at Nvidia with 30,000 people the company keeps the decision-making speed of a start-up; after Twitter cut 80% of its workforce, the platform continues operating as usual; Block hands half of the coordination work to the system; Manus achieves the output of a mid-sized company with 78 people.

These aren’t coincidences, and they aren’t special cases of individual geniuses in isolation. They’re manifestations of the same trend showing up in different places.

For over two thousand years, there was only one reason hierarchy existed: humans were then the only available information relay mechanism. That premise has changed.

The disappearance of middle-layer managers isn’t because they aren’t working hard—it’s because what they rely on to survive—the collection, aggregation, transmission, and translation of information—has already been done faster, more accurately, and more cheaply by systems.

After two thousand years of making do, it ends here.