AI Agents and the Demise of Middle Management

Preface: The “Safest” Position is Being Uprooted by AI

Have you noticed that on the layoff lists of 2026, the first to fall are not programmers, but the middle managers who were once considered the “most knowledgeable in business and best at communication”?

This is not alarmist rhetoric. Just this month, a software technology company laid off over 50% of its staff after introducing the AI agent framework OpenClaw. The company’s director said something that sent a chill down the spine of countless middle managers: “Now, we only need to manage five computers to run the company’s operations efficiently.”

Five computers have replaced the coordination work that originally required dozens of middle managers. The management model has shifted from “managing people” to “managing AI”—and decision-making efficiency has actually improved.

We used to say, “management is an art.” Today, AI says, “management can be replaced by algorithms.”

Data from the Tencent Cloud Summit is even more sobering: China’s average daily token calls surged from 100 billion in early 2024 to 140 trillion in March 2026, a more than thousand-fold increase in two years. What does this mean? It means AI is taking over corporate information processing on a massive scale. Every token call could correspond to a coordination, analysis, or decision-making action that once required a human.

And the most endangered group is precisely the one that thrives on “information asymmetry”—middle management.

What is their essential function? They are “information routers.” When the cost of data exchange between AI agents approaches zero, human routers become the most expensive “non-performing liability” on the corporate balance sheet.

Using Coase’s theorem as a scalpel, I will help you dissect a painful truth: the existence of middle management is a “tollbooth” built by companies to reduce transaction costs. When AI allows information to travel directly, the tollbooths should be dismantled.

But don’t despair just yet. I will tell you which middle managers can survive, and how you need to transform.

Chapter 1: The Mid-Career Middle Class, First to Face the Axe in 2026

1.1 The Data Doesn’t Lie: Middle Managers Are Disappearing in Droves

In March 2026, a piece of news sent shockwaves through the professional world.

A knowledge-based service company deeply integrated OpenClaw into its core business, completely restructuring its workflow. The CEO admitted that in the past, managing projects required layer upon layer of coordination across product, technology, and scheduling, leading to information loss. Now, by using five laptops to control 19 AI agents, these “digital employees” are on standby 24/7, handling business analysis, technical architecture design, and even some coding—intellectual work that previously required a medium-sized team.

The layoff rate: over 50%.

The most affected were support departments like logistics and human resources—what are traditionally considered “management-type positions.” A vast number of basic white-collar jobs were replaced by AI.

This is not an isolated case. In February 2026, the American fintech company Block announced a 40% layoff due to efficiency gains from AI tools. Those cut were not frontline employees, but a large number of middle management positions.

If you think this is just an isolated phenomenon in “tech companies,” consider the data from the Tencent Cloud Summit: China’s average daily token calls skyrocketed from 100 billion in early 2024 to 140 trillion in March 2026, a more than thousand-fold increase in two years.

A thousand-fold. What does that signify?

Xia Lixue, co-founder of WENJIAN Core, drew a comparison at the 2026 Zhongguancun Forum: “The last time we saw this kind of growth rate was during the rapid popularization of mobile data in the 3G era.”

Starting from the end of January, her company’s token consumption “doubled every two weeks, and has now increased tenfold.”

Behind every token call, an AI is processing information, coordinating tasks, and executing commands. When these “digital employees” work tirelessly, never make mistakes, and are available on demand, how much value is left for human middle managers?

This is not a gentle adjustment for “cost reduction and efficiency improvement”; it is a structural organizational war. And the first casualties on this battlefield are the middle managers “responsible for coordination.” Your proud management experience may just be a replicable “information processing flow” in the eyes of AI.

1.2 The Middle Manager’s Anxiety: What You Thought Was a Moat is Actually an Information Tollbooth

I’ve spent 30 years in banking risk control and have seen countless middle managers. What is the capability they are most proud of?

“I know the process.””I can align everyone.””I can handle cross-departmental coordination.”

But have you ever considered the essence of these abilities?

They are byproducts of “information asymmetry.”

Imagine your company as a city. What are middle managers? They are the tollbooths.

In the past, for information to get from Department A to Department B, it had to pass through your “tollbooth.” Your justification for collecting the toll (your salary) was that you helped the information find its way—you knew who to talk to, what procedures to follow, and what language to use.

But now, AI has arrived.

AI agents are like an electronic toll collection (ETC) system on a highway. Information can go directly from A to B without passing through any “tollbooth.” When ETC becomes widespread, the tollbooths are either removed or left with a single attendant.

Zhang Peng, CEO of Zhipu AI, shared his thoughts on OpenClaw at the Zhongguancun Forum: “Ordinary people can now easily access the capabilities of top-tier models, especially in programming and agent-related tasks. Many ideas that were previously limited because one couldn’t write code or lacked other professional skills can now be brought to life through simple conversation.”

Translated, this means: tasks that once required you to “coordinate” can now be done in minutes by one person with a few AI agents.

So what can these “digital employees” actually do?

A technical guide from the Alibaba Cloud Developer Community provides a clear description: OpenClaw’s multi-agent collaboration system allows multiple AI agents to achieve intelligent division of labor, real-time information synchronization, and flexible role configuration, just like a human team.

A Search Agent focuses on information retrieval.
A Writing Agent is responsible for content generation.
A Code Agent handles development and testing.

Three agents working in concert can complete the entire creative process from raw material to a finished article with a single click.

In the past, this required a product manager to propose a need, an editor to write the copy, a designer to create graphics, and a technician to publish—four departments, coordinated by four middle managers. Now, one person plus three agents gets it done. Your “irreplaceability” in the company is likely just another term for “information asymmetry.” When information can flow directly, the tollbooth should be dismantled.

1.3 A Harsh Analogy: From “Information Router” to “Non-Performing Liability”

Let’s put this more bluntly.

Imagine the router in your home breaks. What do you do? You buy a new one for a few dollars. You don’t feel sentimental about it, because it’s just a “conduit.” The essence of a middle manager is to be the “information router” within a company. They receive information from upstream, process it, and forward it downstream, occasionally adding a bit of “interpretation” and “coordination.” When the cost of data exchange between AI agents approaches zero, human routers are not only slow and expensive but also prone to errors and emotions. For a company, such positions become “non-performing liabilities” on the balance sheet—and must be written off.

Luo Fuli, head of Xiaomi’s MiMo large model, was even more direct in her assessment of OpenClaw: this framework “significantly raises the ceiling for domestic open-source models that haven’t quite reached the level of closed-source models but are already at the forefront of the field.” In the vast majority of scenarios, “the task completion rate is already very close to Claude’s latest model, while also ensuring a very high baseline.”

What does this imply? It means the capability boundaries of AI agents are rapidly approaching those of human experts.

And OpenClaw’s core architecture—”Mission Control (scheduling layer) + Agent Layer + Model Layer + Tool Layer”—is making this capability manageable, reusable, and scalable.

Mission Control is a unified console responsible for agent routing, task distribution, and global monitoring, replacing cumbersome traditional WebUI operations. The Agent Layer consists of “digital employees” with independent identities and capabilities, supporting role definition, personality configuration, and SOP binding, enabling task handoffs.

In plain English: AI agents can not only do the work but can also divide labor and collaborate like a human team to complete tasks in sequence.

This is why that software company could lay off 50% of its staff and use just five computers to manage 19 AI agents. The “communication cost” between AI agents is almost zero, while the “coordination cost” between human middle managers is becoming increasingly high.

1.4 War Game: When AI Starts “Cross-Departmental Communication”

Let’s game out this war more specifically.

Assume you are an operations director at an internet company, managing five operations managers, each with ten specialists under them. What is your value?

It’s “coordination”—allocating metrics, adjusting resources, resolving conflicts, and aligning with other departments.

Now, AI agents enter the picture.

Scenario 1: Requirement Delivery
Past: Product manager raises a requirement -> You hold a meeting to convey it -> Operations managers break it down -> Operations specialists execute -> Feedback is given to the product manager.

Now: Product manager creates a task in Mission Control -> The task is automatically routed to the Operations Agent -> The Operations Agent calls the Data Analysis Agent to get user profiles -> Calls the Content Generation Agent to produce a plan -> Calls the Testing Agent to validate the effect -> The result is automatically fed back to the product manager.

No human is involved in “coordination” throughout this entire chain.

Scenario 2: Cross-Departmental Collaboration
Past: Tech, product, and operations hold a meeting. You are responsible for “translating” requirements and aligning on progress.
Now: Agents from the three departments communicate directly through a protocol layer. The Tech Agent knows what interface the Operations Agent needs, the Product Agent knows the Tech Agent’s schedule, and the Operations Agent knows the Product Agent’s launch plan.
There are no “meeting minutes,” no “alignment sessions,” no “progress syncs”—because all information is synchronized in real-time.

Scenario 3: Exception Handling
This might be the last bastion for human middle managers. When the system throws an error, a user complains, or a crisis erupts, AI can identify the problem, but who makes the decision?

The answer may surprise you: AI is also learning to make decisions.

Luo Fuli mentioned at the Zhongguancun Forum: “When a model starts executing longer-term tasks, you find that it can learn and evolve on its own. If this self-evolution mechanism can operate continuously, its potential will be immense, like a top scientist exploring things that didn’t exist in the world before.”

She predicts that the time window for this to happen may only be one to two years.

In other words, by 2028, AI agents may well have the ability to handle “exceptions.”

When AI can independently complete cross-departmental communication, the value of human middle managers shifts from “manager” to “managed.” If your job is merely to pass information, AI is a thousand times cheaper.

1.5 n8n and Coze: Making Agent Orchestration as Simple as Building Blocks

OpenClaw isn’t the only player. In the AI agent ecosystem, two other important tools are changing the game: n8n and Coze.

n8n: The “Command Center” for Agent Collaboration
If you find AI agents too complex, n8n is the tool that simplifies everything.

At Mattermost, engineers used n8n to build an “automated task processing system.” When a Jira ticket is marked as “ready for development,” n8n automatically adds it to a queue, triggering a Cursor Automations agent to write code, run tests, and open a pull request, finally notifying the team for review in a Mattermost channel.

No human “coordination” is needed for the entire process.

n8n plays the role of the “orchestration layer”—it knows which agent should do what, when, and what the next step is after completion. Its core value is liberating the agent scheduling logic from code and turning it into a visual, configurable workflow.

A technical blog post from Vonage even demonstrated how to use n8n to build a five-node AI WhatsApp receptionist that automatically answers customer questions, handles emergencies, and logs conversation history.

Coze: Making Agent Operations a New Profession
Coze is a game-changer from another angle.

A guide from the Alibaba Cloud Developer Community points out that the engineering value of the Coze API is that it evolves agents from being “prompt-driven” to “workflow-driven.” By using conditional branches, code nodes, and parallel plugin calls, it can handle tasks with engineering-grade stability.

What does this mean?

It means an AI agent is no longer just a “chatbot” but a “production system” that can be embedded in business systems, continuously optimized, and maintained over the long term.

This has given rise to a new position: the AI Agent Operations Engineer.

The skill model for this role includes:

Understanding of system architecture: The ability to design a rational agent structure.
Tuning of operational states: Continuously correcting agent behavior through logs, feedback, and data metrics.
Business integration capability: Ensuring the agent is integrated into existing systems, not just an add-on tool.

The future workplace isn’t a competition between humans and AI, but between “people who can manage AI” and “people who can’t.” The former is a new species; the latter is an endangered one.

1.6 The Shifting Anchor of Middle Management’s Value

By now, you should be feeling the chill.

The first layer of logic for the demise of middle management is the replacement of the “information routing” function. But if you think this is merely an “efficiency problem,” you are thinking too superficially.

The fatal blow is this: when AI agents can autonomously handle cross-departmental coordination, the very organizational form of the enterprise will fundamentally change—from a “pyramid” to a “protocol stack.”

In the past, the organization was connected by “people” linking various departments.
Now, the organization is connected by “protocols” linking various agents.

In this new world, the core value of middle managers—coordination, communication, alignment—is being replaced by algorithms and protocols.

Does this mean all middle managers will disappear?

Not necessarily.

But those who survive will no longer be the ones who “know the process best,” but those who can “handle exceptions”; not those with “strong coordination skills,” but those who can “define problems accurately”; not those who “manage many people,” but those who are “adept at managing AI.”

Next, I will take you to the economic bedrock—Coase’s theorem. You will discover that AI agents are not just optimizing the firm; they are rewriting the very reason for its existence. The era of the information router is over. The era of the problem solver has begun. Do you want to be the dismantled tollbooth, or the roadside assistance team still needed in the age of ETC?

Chapter 2: Coase’s Theorem Strikes Back—When Coordination Costs Hit Zero, Why Does the Firm No Longer Need You?

2.1 An Experiment by Zuckerberg: Bypassing All Middle Managers with AI

In March 2026, a piece of news rocked the management world.

Meta CEO Mark Zuckerberg is building a personal “CEO agent” for himself. According to The Wall Street Journal, this AI tool can retrieve information for him directly, bypassing the layers of reporting previously required to get answers.

In plain English: Zuckerberg is using AI to circumvent his own management.

This isn’t a tech news story; it’s a signal. When the CEO can use AI to bypass middle management, what reason is left for middle management to exist?

The changes within Meta are more radical than what is seen externally. The tech giant, with its 78,000 employees, is implementing a comprehensive AI work system. Employees are already using a personal agent tool called “My Claw”—these AIs can access an employee’s chat history and work files, and can even communicate on the employee’s behalf with colleagues—or their colleagues’ own personal agents.

You read that right: AI agents can converse directly with each other to accomplish cross-departmental communication.

Meta has even created an internal group where employees’ personal AI agents can interact. It sounds like science fiction, but it is the reality of March 2026.

Zuckerberg was even more blunt on the January earnings call: “We are investing in AI-native tools to enable Meta’s employees to get more done. We are elevating the status of individual contributors and flattening our team structures. Projects that once required large teams can now be done by one very talented individual.”

This statement is incredibly dense with meaning. Zuckerberg is saying that the organizations of the future will no longer need so many “managers,” because AI allows one person to do the work of an entire team. When the CEO starts using AI to bypass the management layer, the “necessity” of that layer is no longer an academic question, but a layoff question.

2.2 Revisiting Coase’s Theorem: Why Do Firms Exist?

To understand the underlying logic of this transformation, we need to go back to 1937.

In that year, a young economist named Ronald Coase asked a deceptively simple question: Why do firms exist in the world? Why can’t the market coordinate everything?

His answer, which later won a Nobel Prize, was: “transaction costs.”

Market transactions have costs—you need to find suppliers, negotiate, sign contracts, supervise execution, and handle disputes. When these costs are too high, the firm emerges. The firm turns “external transactions” into “internal coordination,” replacing market negotiations with a management hierarchy, which is cheaper.

Imagine you need to cook every day. You could eat at a different restaurant for every meal—this is the “market model.” But it’s cumbersome: you have to choose a restaurant, wait for a table, order, and pay each time.

Alternatively, you could hire a full-time chef—this is the “firm model.” You pay a fixed salary, the chef handles all the meals, and you don’t have to worry about “where to eat today.” Although you pay a salary, you save the hassle of “choosing a restaurant every day.”

This is the essence of the firm: a machine for reducing coordination costs.

And within this machine, what are middle managers? They are the “coordinators”—they are responsible for translating the directives of senior leadership into actions for the execution layer, and for summarizing information from the execution layer into reports that senior leadership can understand.

The reason you have this job is not because your boss likes you, but because, on the matter of “internal coordination,” you are cheaper than the market.

2.3 How AI Agents Demolish the “Coordination Cost” Moat

Now, let’s recalculate Coase’s theorem in the context of 2026.

Coase stated that the boundaries of a firm are determined by coordination costs. When internal coordination costs are lower than market transaction costs, the firm expands. When internal coordination costs are higher than market transaction costs, the firm shrinks.

What AI agents are doing is reducing both types of costs simultaneously, but in completely different ways.

First, AI is driving internal coordination costs toward zero.
In the past, internal coordination required meetings, emails, group chats, and alignment sessions. With each additional department, coordination complexity increased exponentially. A 5-person team has 10 lines of communication; a 10-person team has 45—this is O(n²) growth.

Now, AI agents can communicate directly. A protocol layer provides unified scheduling, and each agent only needs to connect to the protocol layer once for tasks to flow automatically. The coordination cost is reduced from O(n²) to O(n).

Research teams have termed this new type of enterprise the “Protocol-Coordinated Firm.” Its core feature is a protocol layer separating the intent layer from the execution layer.

In simpler terms:

Top layer: You state “what I want” (intent).
Middle layer: The protocol layer translates, validates, and schedules (done automatically by AI).
Bottom layer: Countless specialized AI agents execute the tasks.

In this structure, the middle management position disappears. The function of “coordination” has been replaced by the protocol layer and AI agents.

Second, AI is also drastically reducing “market transaction costs.”
In the past, outsourcing a project required finding people, negotiating, signing contracts, supervising, and accepting deliverables. Now, a company can directly call upon external specialized AI agents via the protocol layer, paying on a per-use basis.

YiVal’s “Wanzhi 2.5 Enterprise Multi-Agent” has already demonstrated this possibility. A user simply inputs a simple prompt, and a “Marketing Director Agent” comes online, instantly breaking down the task and assembling a team—visual designers, marketing managers, content managers, media experts, and other sub-agents, each fulfilling their roles and synchronizing their professional knowledge and process progress in real-time.

What does this mean? It means a company can break down its execution layer into countless tiny, specialized, and replaceable agents without being overwhelmed by complexity.

AI is not just optimizing the firm; it is rewriting the very reason for the firm’s existence. When coordination costs approach zero, “internal coordination” is no longer a service worth paying for.

2.4 From “Process Management” to “Protocol Coordination”: A Silent Organizational Revolution

To understand the depth of this transformation, we need to look at the evolution of the AI tool ecosystem.

The Division of Labor Among AI Tools: Who Does What?
If you’re bewildered by the array of AI tools, let me clarify:

RPA (Robotic Process Automation): Like a tireless mechanical arm, faithfully repeating the operational steps you teach it. Suitable for fixed-process, highly repetitive work. But it’s “brittle”—a website redesign can break it.
Coze: An AI application-building platform from ByteDance, allowing even novices to build a chatbot via drag-and-drop. Lowest barrier to entry, but has a capability ceiling.
n8n: An open-source workflow automation tool that can connect different systems, software, and databases. For example: “customer places order -> automatically syncs to ERP -> sends confirmation email -> pushes notification to Lark group.” More powerful, but requires some technical foundation.
OpenClaw: An open-source project that exploded in popularity in late 2025, with over 230,000 stars on GitHub. It’s essentially an AI assistant that runs on your computer, but unlike a normal chatbot, it has “hands.” It can read and write files, open a browser, run code, send emails, and even control WeChat to send messages for you.
Skills: OpenClaw’s plugin system. The skills you install for the AI determine what it can do. The ClawHub currently has over five thousand skills, from code development to smart home control.
Agent: A broad concept, encompassing everything from the simplest prompt-based bots to the most complex fully autonomous digital employees.

The emergence of these tools is reshaping the “labor structure” of corporations.

An independent developer built an operations agent system using OpenClaw with staggering results: weekly content output surged from 2-3 articles to 15-20, daily operations time was compressed from 3-4 hours to 20 minutes of review, and content conversion rates rose from 1.2% to 3.8%.

He concluded: “The scarcest resource for an independent developer isn’t money; it’s time and energy. A good operations agent system is like hiring an operations team that works 24/7, never complains, and continuously improves.”

The firm of the future has no “departmental walls,” only “protocol stacks.” Your proud cross-departmental communication skills are worthless in the face of a protocol.

2.5 Why “Flattening” Isn’t a Slogan, But an AI-Era Inevitability

You’ve probably heard many companies tout “flat management,” but most of the time, it’s just a slogan. Because the reality is, the human brain cannot process that much information.

A manager can directly supervise 7-10 people at most. Any more than that requires an intermediate layer. This is an inevitable result of organizational entropy.

But AI changes everything.

An AI tool called “Second Brain” has already appeared within Meta, built by an employee based on Claude, which can index and query project documentation. Its creator says the AI tool is “designed to be an AI chief of staff.”

At the same time, Meta is building a brand-new applied AI engineering organization. The structure of this organization is highly unusual: as many as 50 individual contributors report to a single manager.

A 50-to-1 ratio. This is unthinkable in traditional management.

But Zuckerberg believes it is entirely achievable—because AI has taken on the function of “management.” The manager no longer needs to do “process management,” only “results management” and “exception handling.”

This is the organizational form of the AI era: the pyramid has been flattened, not because managers have become stronger, but because AI has taken over the information processing function of the middle layer.

The pyramid organization exists because the human brain cannot process so much information. When AI can, the pyramid is destined to fall.

2.6 Language Friction Reduced to Zero: The Ultimate Reversal by Coase’s Theorem

Let’s pull the perspective back to the highest level.

Human civilization is built on language. We use language to communicate, collaborate, form organizations, and develop commerce. Language is humanity’s greatest invention, and also its greatest bottleneck.

The existence of most jobs is, in essence, due to “language friction”—what A says is understood differently by B; the needs of Department A must be translated before Department B can execute them. A large part of a middle manager’s job is to perform this “translation” and “lubrication.”

Then came large models. What they do is “automate expression and understanding.”

When large models become the high-bandwidth, low-friction intermediary layer, the true nature of many people’s jobs is exposed: they are merely “friction tax collectors.”

At this moment, Coase’s theorem recalculates the books. The reason for a firm’s existence is to reduce transaction costs. When AI reduces language friction to near zero, the act of “helping others understand” is no longer valuable. The most expensive cost in human society is not ignorance, but “understanding.” When large models crush the “marginal cost of understanding” to near zero, the act of “helping others understand” faces devaluation.

2.7 The Shifting Anchor of Middle Management’s Value

You should now grasp the underlying logic of this transformation.

Coase’s theorem tells us that firms exist to reduce coordination costs. AI agents are completely upending this rationale.

When AI agents can autonomously handle cross-departmental communication, the organizational form of the enterprise will fundamentally change—from a “pyramid” to a “protocol stack.” The core value of middle managers—coordination, communication, alignment—is being replaced by algorithms and protocols.

But this does not mean all middle managers will disappear.

Those who survive will no longer be the ones who “know the process best,” but those who can “handle exceptions”; not those with “strong coordination skills,” but those who can “define problems accurately”; not those who “manage many people,” but those who are “adept at managing AI.”

Zuckerberg’s experiment is not over. Meta is planning a large-scale layoff, potentially affecting 20% or more of its total workforce. The reason is simple: to offset the enormous costs of AI infrastructure investment and to prepare for the higher efficiency brought by AI-assisted employees in the future.

This is the harshest workplace truth of 2026: AI is not here to eliminate you; it is here to force you to upgrade. But the window of opportunity for that upgrade may only be one to two years.

The era of the information router is over. The era of the problem solver has begun. Zuckerberg is using AI to bypass management—not because he dislikes managers, but because AI is faster, more accurate, and cheaper.

Chapter 3: Survivor’s Bias—Which Middle Managers Make It to the Next Episode?

3.1 A Question That Chills Middle Managers to the Bone

In March 2026, at the Zhongguancun Forum, an audience member posed a soul-searching question to the AI experts on stage:

“I am a middle manager with 15 years of experience, managing a team of over 20 people. My company is now introducing OpenClaw and n8n, and I hear that AI can automate many of the things I do daily. I want to know, in three years, what will be left for me to do?”

The expert on stage paused for two seconds before replying: “What will be left for you to do are the things in your current job that AI cannot do. The question is, are you sure you know what those are?”

This sentence sent a shiver down the spines of many in the audience.

Because most middle managers are not clear on which parts of their job are “replaceable by AI” and which are “irreplaceable by AI.” They haven’t even thought about this question—until the moment the layoff list is announced.

But if we analyze it calmly, the answer is quite clear: AI can replace “information routing,” but it cannot replace “meaning-making”; AI can replace “process management,” but it cannot replace “exception handling”; AI can replace “coordination and communication,” but it cannot replace “trust endorsement.”

Now, let’s dissect these three things.

AI can calculate the optimal solution, but only a human dares to sign off on it. The middle manager of the future is not a “manager,” but a “hub of trust.”

3.2 Redefining “Management”: From “Managing Processes” to “Managing Exceptions”

I’ve worked in banking risk control for 30 years and have seen over ten thousand companies. I’ve discovered a pattern: good managers are not those who “follow the process most smoothly,” but those who “handle exceptions the best.”

Imagine a company as a highway. There are two types of middle managers:

The first type: Traffic controllers—directing how vehicles should move, which lane is faster, and where to turn. This is “process management.”
The second type: Roadside assistance teams—handling accidents, landslides, traffic jams, and extreme weather. This is “exception management.”

AI can be a traffic controller. It knows the optimal routes, performs real-time dispatching, and makes dynamic adjustments. OpenClaw’s Mission Control is already doing this—automatically routing tasks, assigning agents, and monitoring execution.

But AI cannot be a roadside assistance team. Why? Because every “exception” is different.

The rescue plan for a car accident depends on the condition of the injured, the extent of vehicle damage, the weather, and traffic flow.
The solution for a landslide depends on soil quality, available equipment, and the impact on the surrounding environment.
Handling a customer complaint depends on the customer’s emotions, their history with the company, corporate policy, and legal risks.

These “exceptions” require human judgment, experience, and even “intuition.” More importantly, they require accountability—whoever makes the call is responsible.

AI can manage the “normal,” but it cannot manage the “abnormal.” The future of middle management lies not in being a “guardian of the process,” but in being a “terminator of exceptions.”

3.3 Survivor Profile One: The “Problem Definer”

Let’s return to a fundamental question: what is AI actually good at?

AI excels at getting “from A to B”—given a clear objective, it can find the optimal path. But AI is not good at “defining A”—that is, answering the question, “What is the real problem that needs to be solved?”

OpenAI’s Harness Engineering experiment revealed a key insight: when AI is responsible for execution, the human’s job becomes “defining the problem, setting the boundaries, and judging the outcome.”

Specifically:

AI can write 100 lines of code, but it cannot write “what problem this feature should solve.”
AI can generate 10 solutions, but it cannot judge “which solution best aligns with the company’s strategy.”
AI can work 24/7, but it cannot decide “what is worth doing and what should not be done.”

Let’s understand this through the n8n case:
At Mattermost, engineers used n8n to build an automated system: when a Jira ticket is marked “ready for development,” n8n automatically triggers the Cursor Automations agent to write code, run tests, and open a pull request.

But one step requires a human: “Should this ticket really be marked as ready for development?”

This question seems simple, but it involves business priorities, technical feasibility, resource allocation, and strategic alignment. AI can offer suggestions, but the final decision must be made by a person.

Because there is something here that AI cannot replace: shared risk.

In the AI era, the scarcest ability is not “how to do things right,” but “what are the right things to do.” The former is delegated to AI; the latter is reserved for humans.

3.4 Survivor Profile Two: The “Trust Endorser”

Against the backdrop of rapid AI technological iteration, an easily overlooked trend is that the value of interpersonal trust is actually increasing.

Why?

Because AI can process information, but it cannot process “trust.” Why does a client give an order to you instead of to an AI? Because of trust. Why does a team follow you instead of an algorithm? Because of trust. Why does a boss entrust an important project to you instead of letting an AI execute it automatically? Because of trust.

What is the essence of trust? It is shared risk.

AI will not “take the fall” for you. When a project fails, an AI will not step up and say, “This is my responsibility.” When a decision is wrong, an AI will not say, “I was wrong, I will bear the consequences.” When a customer is dissatisfied, an AI will not say, “I understand how you feel, and I will do my best to help you solve the problem.”

These abilities to “take the blame,” “bear responsibility,” and “empathize” are things AI can never replace, because they are rooted in the ethical and emotional systems of human society.

This is corroborated by the practice of Coze:
A guide from the Alibaba Cloud Developer Community points out that the Coze API allows agents to evolve from “prompt-driven” to “workflow-driven,” capable of handling tasks with engineering-grade stability. But even so, companies still need to create the position of “AI Agent Operations Engineer”—not because these AIs aren’t smart enough, but because the company needs someone to be “responsible” for the AI’s output.

This person is the “hub of trust.”

AI can do the work for you, but it cannot carry the burden for you. The middle manager of the future is not a “manager,” but a “hub of trust”—others are willing to entrust their backs to you not because of your ability, but because of your accountability.

3.5 Survivor Profile Three: The “AI Legion Commander”

If the first two types of survivors are those who “don’t compete with AI,” then the third type is those who “make AI work for them.”

OpenAI’s experiments have already shown that a 3-person team can command AI agents to produce a million lines of code in five months. The “super-individual” of the future is not a lone wolf, but a combination of a “human + an AI legion.”

From “Leading a Team” to “Leading an AI Legion”:Past “management ability” was about “how to manage people”—motivation, assessment, coordination.
Future “management ability” will be about “how to manage AI”—what instructions to give, what boundaries to set, how to verify the output.

How to manage AI? Let’s break it down using the practice of OpenClaw:The core of OpenClaw is the “Skills” system—a plugin marketplace with over five thousand skills currently available. The skills you install for the AI determine what it can do.

A manager who understands AI will do three things:

Select the right skills: Know which agent to use for which task, and which plugin for which scenario.
Set boundaries: Give the AI clear instructions, telling it what it can and cannot do, and to what extent.
Verify the output: The AI’s output is not always perfect and needs human review, correction, and feedback.

This is the ability to “command an AI legion.” It doesn’t require you to know how to code, but it does require you to know “how to communicate with AI.”

The high earners of the future will not be those who “manage the most people,” but those who are “most adept at managing AI.” Your subordinates may not be human, but your salary could be higher than it is now.

3.6 A Brutal Conclusion: The Collapse and Bifurcation of the Middle Class

Let’s briefly summarize this analysis using the “Four Forces Model.”

Force of Expansion: AI is rapidly penetrating all knowledge work domains. From RPA to Coze to n8n to OpenClaw, the tools are becoming more powerful and their applications more widespread. This wave will not stop.

Force of Contraction: A large number of traditional middle management positions are disappearing. Meta plans to lay off 20%, the software company laid off 50%, Block laid off 40%—these are not isolated cases, but a trend. Every company is asking the same question: “Which positions can be done by AI?”

Force of Equilibrium: There may be policy interventions, such as “transitional placement for AI-displaced jobs” and “vocational skills retraining.” But policy cannot stop a trend; it can only delay the pain, not reverse the direction.

Force of Evolution: Some middle managers will evolve into “super-individuals”—capable of defining problems, providing trust, and commanding AI. Others will be eliminated. There is no middle ground.

A Question Through the Humanity Prism:
Under the same macro trend, why do different people make different choices?

Some choose to resist: “AI is not as good as me,” “My experience is irreplaceable.” The result is elimination. Some choose to lie flat: “I can’t learn it anyway, I’ll just wait for the severance package.” The result is also elimination. Some choose to evolve: “Since the trend is here, I will ride it.” The result is survival, and perhaps even a better life.

This is not a gap in ability, but a gap in cognition, a gap in mindset, a choice of “value rationality.”

The middle class of the AI era will not disappear, but it will polarize—you either become someone who “masters AI” or someone who is “mastered by AI.” The choice is in your hands, but the window of opportunity is only one to two years.

3.7 You Have Time, But Not Much

Reading this, you may already feel a sense of urgency.

The surviving middle managers are not lucky; they have done three things right: defining problems, providing trust, and commanding AI.

But don’t rush to pigeonhole yourself—these three abilities are not innate; they can be learned.

“Defining problems” can be practiced: Every day, ask yourself, “What fundamental problem is the work I’m doing solving?”
“Providing trust” can be accumulated: Step up at critical moments and take risks for your team.
“Commanding AI” can be learned: Start by building a simple agent with Coze, then move to automation with n8n, and eventually master OpenClaw.

You have time, but not much.

Luo Fuli’s prediction at the Zhongguancun Forum is that AI’s self-evolutionary capabilities will mature within one to two years. By then, AI may already have the ability to handle “exceptions.” By then, the bar for survival will be much higher.

So, start now.

AI won’t make everyone unemployed, but it will make everyone retake the exam. Those who pass will be upgraded; those who fail will be out.

Conclusion: AI Isn’t Here to Eliminate You, But to Force Your Evolution

At this point, you might be asking: So what should I do?

My answer is this: delete the words “management experience” from your resume. Replace them with “problem definition,” “trust building,” and “AI command.”

Let me give you an actionable “transition checklist”:

First, re-examine your work.

Before you leave work each day, ask yourself three questions:

Of the things I did today, which ones could be replaced by AI?
Which ones can AI not do, or not dare to do?
If my position were replaced by AI tomorrow, where would I be reassigned?

If the answer to the first question is “a lot,” you are in danger. If the answer to the second question is “very little,” you are in danger. If you can’t answer the third question, you are in even greater danger.

Second, deliberately practice “exception management.”

Don’t be a “process expert”; be a “problem solver.” When the team encounters difficulties, don’t just say, “I’ll coordinate”; say, “I’ll solve it.” When a project encounters an anomaly, don’t just think, “follow the process”; think, “how do we break this deadlock?”

Remember: AI can handle 80% of normal situations, but it cannot handle the 20% of exceptions. That 20% is your anchor of value.

Third, learn AI, but don’t just learn the tools.

Don’t just learn how to use a specific piece of software; learn “how to command AI.” This includes:

How to write clear instructions for AI.
How to set the working boundaries for AI.
How to verify the output of AI.
How to embed AI into business processes.

Tools will change, but the ability to command AI will not become obsolete. Today you might use Coze, tomorrow OpenClaw, and the day after something else. But as long as you know how to “command,” you can keep up.

Fourth, build your “trust account.”
Make your team, your clients, and your boss believe that you are reliable in critical moments. How do you build trust?

Do what you say: Always follow through on your commitments.
Dare to take responsibility: When problems arise, don’t pass the buck or make excuses.
Communicate with empathy: Understand the emotions and needs of others, not just transmit information.

Trust is something AI cannot replace. It is your final moat.

There is a saying in the Tao Te Ching: “Thirty spokes share one hub. It is the center hole that makes it useful.”

The utility of a wheel comes not from its many spokes, but from the empty space at its center—the hub that gives the spokes purpose. The spokes can be changed, added, or removed, but the hub is the center of the wheel, the point where everything converges.

AI agents are those “spokes”—they are tools, executors, and assistants. And you? You must become the “hub”—the center of the wheel. AI can replace any of the spokes, but it cannot replace the center.

AI won’t make everyone unemployed, but it will make everyone retake the exam. Those who pass will be upgraded; those who fail will be out. Don’t wait until the layoff list is announced to ask yourself: am I the tollbooth destined for demolition, or the roadside assistance team still needed in the age of ETC?

【End of Text】