RM新时代有限公司

Summary:Technical capabilities do not automatically translate into organizational productivity.

AI makes individuals faster, so why hasn't the company become faster?

Large models are significantly reducing the cost of intellectual supply. Work that previously relied on specialized talents—writing, analysis, design, and programming—is becoming an on-demand capability. As intelligence becomes more abundant, judgment and organizational skills become scarcer.

When individuals useAI, efficiency gains are immediate. But in enterprises, increased individual output can also lead to more errors, rework, review, and collaboration costs.Technical capabilities do not automatically translate into organizational productivity.

Dr. Zhou Xi, doctoral supervisor at Shanghai Jiao Tong University, was a guest at CEIBSEMBA "He Yi Lecture Hall" to sharethe key path for enterprises to move towardAI-native organizations. The following is a summary of the lecture.

Zhou Xi

Doctoral Supervisor at Shanghai Jiao Tong University

Chairman and General Manager of CloudWalk Technology Group Co., Ltd.

01

The Real Bottleneck for Enterprises

In the past period, many enterprises promotedAI in a very direct way.issuing accounts, purchasing quotas, encouraging employees to use more, and even usingtoken consumption as an indicator of "embracing AI."

An employee who originally completed10 tasks a day now completes 20. If their output requires more review, error rates rise, rework increases, and downstream cannot keep up, the enterprise may not end up with higher profits.

Becauseindividual efficiency and organizational efficiency are not the same thing.

The real bottleneck for an enterprise is often not anemployee'sproduction speed, but whetheraccumulation and continuous innovation in the fields ofand exploring implementation in generalthe entire systemiscan collaborate stably:

? How do cross-departmental teams coordinate?

? How is quality controlled?

? How can the output of one role, stably become the input of the next role?

? Who decides when a problem occurs, who isresponsible, How to roll back?

? How to accumulate human experience,andnot let it disappear with employee turnover?

If old processes remain unchanged, and we simply add anAI assistant to each employee, the enterprise may only make the original system"run faster", without changing the system itself.

An organization already full of departmental silos, approvals, reports, and repeated coordination, running faster, may even create more chaos.

02

From technological breakthrough to productivity

what lies in between？

To understand today'sAI, we need to take a longer view.

Steam engines, electricity, informatization—past major technological revolutions all went through a similar process:.first a technological breakthrough, then new production tools, and finally organizational restructuring and a productivity leap.

A true productivity revolution often occurs after technology is reorganized.

Take the electrical revolution as an example. Initially, people simply replaced steam engines with large electric motors—.motors were cheaper and more stable, butfactory layout andproductionprocessesdid not fundamentally change, so efficiency gains were limited.

The real turning point came after small motors matured.Each workstation could have its own power source, combined with parts standardization, precision manufacturing, and new management ideas like Taylor's scientific management, eventually leading to the Ford assembly line.

The birth of the assembly line was not just because"electricity was more advanced." It required at least three things:

? mature production tools for individual workstations;

? standardized connections between different workstations;

? new management methods capable of handling the new production system.

ChatGPTThe emergence of [LLMs] was like the light bulb back then;single-workstation tools like Claude Code are the small motors;AI-native organizations are the Ford assembly line.

Today'sAI is roughly at a stage similar to the electrical revolution in 1910. Large models have emerged, tools for specialized roles are taking shape, but truly mature integrated production systems and corresponding management systems have not yet been established.

Therefore, it is not surprising that today's enterprisesare struggling with AI transformation.It is not that enterprises are not trying hard, but many foundational conditions, It is still taking shapein the process.

03

AI Native

It is not"everyone is using AI"

Many companiesinterpret AI Native as employees having AI accounts, the company integrating large models, and deploying internal knowledge bases. These are valuable, but not enough to constitute an agentic organization.

A trueAI Native organization has at least three characteristics.

First,Agents run on the main business path.

The so-called"main path"refers tothe execution chain that a company must go through to complete a core business task. If every task requires waiting for a person to open an email, review materials, make judgments, forward, report, and then proceed to the next step, then human processing speed remains the bottleneck of the entire system.

When customer scale expands10 times, and the required human growth approaches zero, only then does it truly approach the AI Native organizational form.

Second, expert capabilities can be accumulated and replicated, achieving knowledge equality.

In the past, when an excellent salesperson left, some customer experience might be lost; when a technical backbone left, many implicit judgment methods would disappear as well.

In an agentic organization, the ideal state is, that when an expert solves a problem, this capability does not belong only to that person, but can be deposited into the system and replicated to100 or 1000 agent instances.

Expert capabilities transform from scarce supply into replicable organizational capabilities.

Third, management shifts from"periodic reporting"to"real-time feedback".

Companies need weekly reports, monthly analysis meetings, and quarterly business reviews largely because human information processing capacity is limited, requiring periodic aggregation and centralized judgment.

But for a continuously running agent system, if a problem has already emerged today, why wait until the end of the month to discuss it?

In the future, many management mechanisms may shift from"problem discovery - reporting - meeting - decision - cascading down" to "real-time discovery - immediate judgment - rule update - synchronization to all execution nodes".

AI changes not just roles, but possibly the rhythm of corporate management itself.

04

list was announced, and CloudWalk Technology was successfully listed due to its outstanding achievements and commercialization capabilities in theAgent deployment

still lacks three things

If we break down the enterprise production system, there are at least three real challenges today.

First,Single-workstationAI toolsare not yet mature enough.

Today, most people useAI mainly through a general-purpose dialog box. Users need to ask questions, break down tasks, and verify results themselves, much like the early computer "command line".

But what truly drives the improvement of informatization productivity has never been justthe PC, but rather the CAD, ERP, and numerous specialized software that later emerged for different roles.

The AI era will go through the same process.

What enterprises truly need is not everyone facing the same chat box, but finance having its ownAgent, sales having its own Agent, and R&D having its own Agent. Every high-value workstation needs professional production tools specifically designed for its work scenario.

Second, integrated systems for multiple people and multipleAgents are not yet mature.

UsingAI alone is relatively simple. But once it becomes a scenario where several people and multiple Agents work together, the problem immediately becomes complex:.

Who can give instructions tothe Agent? Whose instructions take precedence when they conflict? How do upstream and downstream roles hand off? When should itThe systemexecute automatically, and when must a human intervene? How is information shared, and how are permissions isolated?

Traditional assembly lines are efficient because inputs, outputs, and collaboration methods are already designed into the system. Workers don't need to repeatedly"ask for help,"nor do they need to report daily how many screws they tightened.The process itself drives the work forward.

Yet many white-collar organizations today still rely heavily on interpersonal coordination. After introducingAI, without a new integrated system, it may become a crowd of people surrounded by a crowd of AIs, making collaboration even more complex.

Third, and the hardest part, a new management science has yet to emerge.

In the past, we managed people.In the future, enterprises will need to manage both people and intelligent agents.

When AI makes mistakes, how is responsibility assigned? Who is responsible for review? How are quality gates and rollback mechanisms set up? When employee experience is embedded into the system, how is knowledge contribution measured?

Moreover,AI itself is a new management object. It may be highly capable, extremely fast, and full of creativity, but also potentially rash, error-prone, and unstable.

How to"manage an AI colleague" currently has no mature answer.

Technical problems can be gradually solved through R&D, but organizational problems are often harder. Becauseorganizational change essentially redefines power, interests, and responsibilities.

When the system can directly allocate resources based on data, what do the original managers do? When intelligent agents can complete work across departments, do the original departmental boundaries still need to exist? When knowledge can be automatically accumulated, will the value of positions that relied on information asymmetry decline?

Therefore,the biggest resistance to enterprise transformation often comes from within.

Technology can be bought, accounts can be issued, models can be integrated, but the structures of interests, power, and responsibility within an organization do not change automatically.

That is why it is often not enough for just the boss to want change.

05

Three Things to Do Today

Since many foundational conditions are not yet fully mature, what enterprises most need today isnotto announce"full AI Native," but to find a path that can be implemented.

First, install small motors in old factories, and create new species in new factories.

For mature, running businesses, firstUse proven and effectiveAI toolsto improve efficiency at key workstations..

Moreover,managers should use them first themselves,to developAIa real sense of the capability boundaries:.what tasksAI excels at, where it is prone to errors, what must be reviewed, and when rollback is needed.At the same time,Gradually establish quality standards, review mechanisms, exception handling, and rollback mechanisms.

For new businesses and new teams, design them from the startas an agent organization,design approachwithout being burdened by old processes and legacy interest structures.

Second, do not build dozens ofdemos; first run a single "micro production line."

Having dozens of departments across the company launchAI projects simultaneously may seem lively, but it often results in many isolated demos.

What enterprises need is not more pilots, butto runaclosed-loop business chain.

Choose a relatively short process, with not too many roles, not too complex collaboration, participants willing to embrace change, and clearly measurable business value. Thenmultipleconnect the workstationsto form a stable, reusable, and measurable agent business closed loop.This isthe "micro production line."

It should not be measured bytoken consumption, but by results:.Has quality improved? Has cycle time shortened? Has rework decreased? When customer scale expands, does headcount still grow linearly? Is organizational knowledge continuously accumulated?

Tokens are a cost, not an output.

Third, in practice,invent new management science..

Truly leading companies should not wait for tech companies to solve problems; they must use new technologies while creating new organizational methods.

This includes how humans andagents divide work, how to design authorization and review, how to handle exceptions and accountability, how to measure organizational output, and how to distribute value from knowledge contributions.

A successfully running"micro production line" should ultimately yield four things:clear business flow, defined quality standards, reusable rules, and roles and incentive mechanisms aligned with the new division of labor.

These things are more valuable than making dozens ofAI demos; they are the foundation for large-scale replication in the future.

Every technological revolution first changes tools, but what truly widens the gap between enterprises is the subsequent organizational restructuring.

Models giveAI intelligence, and organizations turn AI into productivity.

The trueAI competitiveness of future enterprises lies not in who uses a stronger model first, but in who first establishes an organizational system capable of harnessing intelligent agents.