When AI replaces the middle management, who will act as the safety net for AI?
Three months, three companies, three managers.
Charlie Hu, co-founder of OpenMax, downsized the team from 80 to 15 people without affecting the annual revenue;
Li Xuetao led the 60-person R&D team of a software company in Tianjin to complete a comprehensive AI transformation, pushing the profit margin from single digits to 30%;
The company where Elric Liu, Vice President of Product of Moox AI works has only about a dozen people, two major departments, two layers of decision-making, and AI Coding accounts for more than 90%.
These are three typical stories about efficiency.
But looking deeper, the three people encountered the same trouble in different scenarios — AI has thinned out the execution layer and cut off the messenger layer, but management has not become any easier.
Some contradictions that were originally digested by the hierarchy are directly exposed to the top level after the organization becomes flattened.
01
After the middle managers exit
Who has more work to do
All three companies share a common trend, that is the buffer structure of the middle management layer has disappeared.
Neuters
OpenMax took the most thorough measures. Charlie cut the entire BD team, because they served the top leaders of enterprises, and clients would directly respond when they saw BD staff showing up: "Where is your boss?" The BD staff only created groups instead of closing deals, and occasionally even passed on wrong messages.
There was a phenomenon that made him even more intolerable: some BD staff pulled in groups that did not contain real clients at all just to prove their value to their superiors.
After the BD team was cut, Charlie handed over customer development to AI. The Agent analyzes 8,000 contacts on his LinkedIn, and pushes 20 customized outreach suggestions every morning, with customer portraits, business analysis and connection points all sorted out. He spends half an hour reviewing them, and sends them out if there is no problem.
The efficiency has multiplied several times, but there is a price to pay.
After the BD team was cut, all the judgment responsibilities that were scattered among 20 people before, such as whether it is suitable to contact this customer today, whether this paragraph is appropriate to send out, whether the timing is right, etc., are all concentrated on him alone.
AI can provide analysis, but it does not make decisions. The half-hour "review" Charlie does every morning is to make the judgments that AI cannot make.
Li Xuetao's team dismantled four traditional departments and reorganized them into FDE special teams of 3 to 5 people each.
The functional boundaries of requirements, R&D and testing are broken through, and each team faces customers directly. The process is greatly simplified. In the past, it took three passes for requirements to be transmitted from customers to R&D, but now it is done in one step.
However, the decline in communication costs is accompanied by an increase in another type of cost. After the transformation, Li Xuetao regularly crosses project managers every week to talk directly with front-line colleagues one-on-one.
He rarely did communications like "How is the AI working for you? What problems have you encountered recently?" before. With middle managers buffering in the middle, conflicting signals were diluted, filtered and processed. After the middle managers exited, the information transmission path changed from "Customer → Requirement → R&D → Testing" to "Manager → Every individual", and the management radius has actually expanded.
Moox AI has no middle managers from the very beginning.
The dozen or so people are divided into two major sections, working on projects in virtual teams on a daily basis, and product and technology decisions are made by him and the CTO together. The advantage of this structure is extremely fast decision-making: "We only discuss complex things, for simple things, we spend a day or two to build an MVP first to see the effect."
The downside is equally obvious: he and the CTO have become the final outlet for all key judgments. Technical routes, product directions, commercialization logic, every line cannot bypass their judgments.
After the middle managers exit, the manager's to-do list has not become shorter, but the items have changed from "managing processes" to "managing judgments". And judgment cannot be processed in batches.
02
The same AI, two types of people
All three companies encountered a similar internal phenomenon: AI did not receive unanimous welcome.
Two extreme emotions emerged in Li Xuetao's team.
One group of people thinks AI is almighty, they do not even look at the output generated by AI, and "submit it all at once", resulting in a large number of missing requirements and inconsistent details that all emerge intensively in the testing phase.
The other group labels AI as "totally useless", saying "I can write faster than it, why should I use it?"
He had to deal with both types of reactions one by one.
For those who over-trust AI, he repeatedly delineates the capability boundary of AI, clarifying what AI can output directly and what must be manually checked.
For those who do not trust AI, he helps them find scenarios that can truly solve problems, so that they can see the effect first.
The Moox AI team encountered another situation: there were no extreme emotions, but general inertia.
"In the first two months, 80% of the allocated Coding Plan was not used." People are not resistant, but their first reaction when encountering things is always to follow the old ways.
His solution is of typical engineering thinking: he does not push it forcibly through administrative orders, but relies on mechanism design. The company sets up a special bonus every month to reward those who have run through small closed loops with AI in internal business.
At weekly meetings, everyone is invited to share "what things you have done with AI recently that help the business". Those who did it get real money, and those who see it have a reference. "When they see that this thing can bring benefits to themselves, they are willing to use it."
Charlie's solution to this problem is more radical: he directly screens out people who do not have autonomy. "No matter how powerful AI is, if you won't take the initiative to ask it questions, it won't work."
Three managers, facing the same tool, spent three different amounts of energy dealing with the same thing.
The distance between people and AI is not determined by the capability of AI, but by people's perception of AI. And this perception gap is becoming a new added item to the management cost of AI-native organizations.
03
The people who stay
The work gets lighter, the responsibility gets heavier
In AI-native organizations, the work of "getting things done" is getting lighter and lighter.
AI Coding accounts for more than 90%, sales copy is generated with one click, customer portraits are automatically tagged, and weekly reports are automatically sorted out by Agents by capturing group chat information. But on the other side, the work of "making judgments" is getting heavier and heavier.
Li Xuetao has an accurate schedule for this.
Before the transformation, most of his energy was spent on tracking progress, checking codes, and going through approval processes. After the transformation, AI took over these tasks, and the time he freed up did not become free time, but was filled with new affairs.
He talks to those who over-trust AI about boundaries, talks to those who do not trust AI about methods, and talks to those who resist transformation about mindsets. He says, when there are fewer people, you will pay more attention to each individual.
Elric described a similar scenario.
In the past, when a new requirement came in, the salesperson would understand it first, then pass it on to the product manager, who would then translate it to the technical team. In the three-layer transmission process, information is lost every time.
Now salespeople can directly use Vibe Coding to build a Demo for customers to see, the intermediate translation chain is broken through. But the pressure of final decision-making has also come back: what exactly the customer wants, whether this requirement is worth doing, these judgments can no longer be partially undertaken by sales, product and technical teams respectively, and they are all concentrated back to Elric and the CTO.
Charlie took this logic further. He says, humans are the periphery of AI. Humans reach the external world through offline activities and customer contacts, obtain information and feedback, and then hand them over to AI.
The core task of these 15 people remaining in the company is not execution, but collection and judgment. Every customer meeting, every observation of competitors, every market fluctuation, does not end after being "done". It has to become corpus, input, and the basis for the next round of decision-making.
This transformation seems to liberate productivity, but in fact it widens the boundary of responsibilities that one person bears.
In the past, one person was only responsible for the output of his own position, but now he may be responsible for the understanding of requirements, the judgment of technology, the empathy for customers, and the review of AI output at the same time.
The work is lighter, but behind this "lightness", one person undertakes the decision-making volume of three or four people.
Charlie says, AI is an amplifier of capability. If a person has no capability, even after being amplified by AI, he still has no capability.
04
Who will take ultimate responsibility for uncertainties
During the three interviews, there was a word that never came out of anyone's mouth, but it almost lingered at the bottom of every conversation, that is, taking the ultimate responsibility for uncertain results.
The BD team was cut, and Charlie spends half an hour every morning reviewing 20 customer messages — he is taking the ultimate responsibility;
The four departments were dismantled, and Li Xuetao crosses project managers every week to talk to everyone individually about their AI usage — he is taking the ultimate responsibility;
Tokens are allocated but no one uses them, and Elric designs incentive mechanisms, sets benchmarks, holds sharing sessions, and pushes everyone forward one by one — he is taking the ultimate responsibility.
There is no standard KPI that can define these things.
They are not processes, not technologies, not things that can be clearly written in job descriptions. They are the part of responsibilities that AI cannot take on, which managers take over one by one and bear in their own way.
The three people have different practices, but share the same core: AI can achieve 99% of the work, and someone has to make the remaining 1% of judgments.
Li Xuetao has been in the company for more than ten years, and most of the team members were brought up by him personally. He has high prestige, but he still encountered all kinds of resistance when promoting the transformation.
"If you are not the top leader of this team, or you just joined not long ago with a complicated interpersonal relationship, the resistance to transformation will be much greater."
The implication behind this judgment is that AI-native organizations have extremely high requirements for managers' trust reserves. In the past, trust could be scattered in different hierarchies and different people, but now it has to be concentrated in a small number of people.
When Elric was asked "how would you design the organization if you started all over again", he said, let everyone use it first, feel the benefits, and then push it forward.
There is no standard answer, no perfect path.
This is not modesty, it is their real feeling. When facing people, facing the change of perception, facing the reshaping of habits, AI cannot give any help, and managers have to tackle them one by one.
Charlie's expression on this issue is the sharpest.
He says, it depends on whether AI is treated as a second-class citizen or a first-class citizen. Some people treat AI as a workhorse: I pay you, so you have to work. AI-native organizations are the opposite: humans serve AI.
Behind this judgment is the choice of where the manager positions himself. Are you the one who manages AI, or the one who takes ultimate responsibility for AI? The daily workload of these two positions is vastly different.
Up to now, the most unexpected gain in the direction of AI-native may have nothing to do with algorithms.
It unexpectedly illuminates the essence of management: the organization can become smaller, the process can become faster, the tools can become more powerful, but in the end, someone has to step up and take responsibility for the remaining uncertainties.
AI can write code for you, write plans for you, analyze customers for you, generate all texts and numbers for you, but it will never say a sentence for you: "I will take responsibility if something goes wrong."
The middle management layer is removed by AI, the execution layer is thinned by AI, there are no more messengers, and all redundancies are burned out completely.
The remaining layer of managers are doing work that is closer to the original definition of management than ever before: making decisions in ambiguity, taking ultimate responsibility in uncertainty, and moving forward on your own in the gray area where no one can make judgments for you.
This may be the most underestimated cost of AI-native organizations so far.
This article is from the WeChat Official Account "Neuters" (ID: Neuters), author: Alex, published with authorization from 36Kr.