The media loves to spread claims like "machines are replacing humans" and "mass unemployment once AI arrives." Plenty of economists—and even academic research—subscribe to this replacement logic. I want to use that logic as a reference point and introduce you to a different one, one that may serve us better in making sense of the AI era.

The Three Realms of AI Entering the Enterprise

Getting AI into an organization is no easy feat, and the vast majority of companies have not done it yet. Many individuals use AI, but individual use is not the same as AI entering the organization.

A company's embrace of AI can be divided into three realms:

The first is AI adoption. This stays largely at the level of individual users. Employees querying ChatGPT or using AI to write presentation decks—that is AI adoption.

The second is AI deployment. The core here is AI entering business processes, which usually requires integrating AI technology with local workflows and data. The overwhelming majority of AI projects companies tender for today are department-level deployments—process redesign in HR, finance, marketing and so on. Yet a company clearly cannot use processes to replace people, because people are flexible while processes are pre-set.

The third is AI organization. After conversations with many large companies this year, my strongest impression is this: the private deployment companies talked about in previous years was mostly about building company-specific AI on internal data rather than adopting general-purpose large models. A few years on, the money that should have been spent has been spent, and every department reports progress—but the company's financial statements and overall operating efficiency show no fundamental improvement. Where is the problem? There is no holistic view; everyone is still like railway police, each patrolling only their own stretch of track. For AI to truly enter an organization, what changes is not a single process, but the way the entire organization operates.

Whichever realm a company occupies, the pattern is much the same: some organizations start their AI transformation with layoffs, while others dare not start at all for fear of mass layoffs. Both have fallen into the same mental trap—assuming that AI transformation creates value primarily through "AI replacing people."

The Misread AI Layoffs: The Truth Behind the Replacement Logic

A recent US survey produced a surprising finding: 55% of organizations that cut jobs in the name of AI regret it, for two main reasons.

First, many bosses laid people off in a fit of enthusiasm, mistaking AI's marketed promise for deployed reality. The media keeps telling us this AI is "mind-blowing" and that one "surpasses humans." But once it is actually deployed inside a company, it simply does not work. Why such a huge gap? Because testing is run on fixed, standardized task sets, whereas nothing inside a company is fixed or given—everything is dynamic, so expectations and reality fall out of alignment.

Second, when companies cut staff, they also cut away their employees' tacit knowledge and experience. People are carriers of knowledge. Beyond making slides, they carry a great deal of knowledge and experience that can be sensed but not fully put into words. Amazon, for instance, suffered several cloud service outages after its recent large-scale layoffs. Amazon is the progenitor of cloud computing, yet after the cuts its recovery time from outages was far longer than the industry average—startling, to say the least. The reason: the veteran employees who could locate faults within complex server clusters, and who held that tacit knowledge, had been let go.

There is a joke circulating now. How do you achieve "cost reduction and efficiency gains"? Cut people to cut costs; once the cuts are done, the company breaks down, so you have to hire people back—and that is the efficiency gain. Cost reduction and efficiency gains close the loop this way. It is a Chinese joke, but US data bears it out: 50% of laid-off employees are eventually rehired, at lower pay, sometimes as contract workers, because the job market has deteriorated.

Then there is another case: some companies' AI looks dazzling, but behind the scenes they maintain enormous outsourcing teams in India and Vietnam to review AI outputs. That is labor arbitrage, not AI eroding jobs.

Many will say: Silicon Valley layoffs are hard fact. In my view, strategy and organization are two different domains—you have to ask whether those layoffs are strategy-driven or organization-driven. From an organizational standpoint, my core message is: do not assume that just because AI has arrived, people in the organization can be cut at will. Strategy-driven layoffs are a different matter altogether.

Consider another set of data. We used to say internet companies run an asset-light model while telecom operators are asset-heavy, because the latter must build towers and lay cables. But now internet companies' capital expenditure (capex) has surpassed that of telecom operators.

What is strategy? Strategy is the allocation of real money and real resources. The most important items in operating expenditure are R&D and personnel costs. Why are the Silicon Valley giants cutting so many people? Because they need to free up cash to buy GPUs at scale and build AI infrastructure; their existing mature businesses are no longer reliable. Take advertising: in the future, intelligent agents may do the shopping on users' behalf and nobody will watch ads. For them, putting money into large-model infrastructure is the safest bet, and model training also demands massive computing power. This kind of layoff is capital expenditure squeezing operating expenditure—a strategic transformation and resource reallocation that has nothing to do with the organizational logic of "AI replacing people."

Do not follow Silicon Valley's layoffs just because you see them. Their cuts may be strategic, while yours would be merely organizational.

The data also shows something directly at odds with the "all programmers will lose their jobs" narrative: programmer employment in the United States hit an all-time high in 2025–2026. Why? In the past, a traditional company needed at least three people to maintain an IT department, each on a high salary, making the overall budget very large. Now a traditional company needs only one person—at a relatively lower salary—who, working with AI agents, can complete the entire software development cycle. This is the Jevons paradox: programming has become cheaper, so many companies that previously found programmers too expensive can now afford them with AI's help, which in turn increases overall demand for such talent. That is why demand for software engineers in the US market has reached record highs.

One more data point. Statistics from investor Q&A sessions of US-listed companies show that when AI comes up, the ratio of "AI will replace labor" to "AI will augment labor" is 1:8. In other words, the overwhelming majority of listed-company executives believe AI will augment rather than replace labor. A further 55% of the discussion concerns AI business upgrading and AI organizational transformation—how to use AI to upgrade the business, combine traditional products and services with AI, and give users a better experience.

A New Logic: Let Knowledge Flow with the Least Friction

So here is the question: if AI does not replace employees and is not about cost-cutting, then how exactly does it create value for the enterprise?

Let us swap the "replacement logic" for a flux logic—more precisely, the flux logic of knowledge. Under this logic, an AI organization is not an organization with fewer people, but one in which knowledge flows with the least friction; investing in AI and investing in people reinforce each other, enabling large enterprises to break through the bottlenecks of organizational scale and restart their growth engines. The key to benefiting from AI transformation, therefore, lies not in using AI to replace people in performing certain knowledge work, but in using AI to genuinely reduce the transaction costs that arise when knowledge is shared, invoked, translated and combined inside the organization.

In other words, the way AI enters an organization should target knowledge flux at the organizational level, enabling organizational knowledge to reach application faster, with less friction, less gaming, and higher quality.

Let us return to the essence of organizations and firms. According to Coase's transaction cost theory, firms exist because many transactions in the market are simply too costly; move them inside an organization, where "our own people" are easier to deal with, and you reduce the mutual scheming. That is the economists' account. The knowledge-based theory in management, by contrast, holds that firms exist because they can, through a set of internal management mechanisms, integrate knowledge scattered across individuals, teams, departments and external partners, and turn it into the products and services customers need. The firm exists to make knowledge circulate better internally.

Consider also why we need corporate culture at all. Culture lets a group of people within an organization, shaped by the same environment, speak the same language and communicate more smoothly. Tacit knowledge in particular can only be understood through exchange, and market counterparties simply do not have the conditions for such exchange. Only within the same organization do people have long-term opportunities to interact, and only there is knowledge exchange more efficient than in the market.

This is how management scholars understand "why firms exist" from the perspective of the knowledge-based theory. And from this angle, why firms need people becomes clear: because people carry knowledge. When people exchange ideas, they collide into new knowledge, which is then converted into R&D results, products and services, and ultimately into revenue and profit.

So what does "flux" mean? Flux is the efficiency of knowledge flow. Knowledge creates value not through knowledge stock, but through knowledge flux.

Knowledge creates value not through knowledge stock, but through knowledge flux.

What many large and mid-sized enterprises face is not a stock problem but a flux problem: they nominally possess enormous amounts of knowledge, yet when it is actually needed it is hard to discover, understand, invoke, combine and convert it quickly, and hard to update it in time against fast-changing external conditions. Conversely, the higher an organization's knowledge flux, the faster it can turn dispersed knowledge into customer solutions, business judgments and organizational action—and the more competitive it becomes. If knowledge sits unused inside experts' heads, it produces no value; if experts do not talk to the product, technology and marketing departments, their knowledge remains mere stock.

Three Obstacles Blocking the Flow of Knowledge

Knowledge that will not circulate is the root of "big-company disease." A company sets out to build a culture of open exchange and simply cannot do it. The organization runs more and more slowly because knowledge will not flow.

Whether it is a large enterprise or a small one that has contracted large-company disease, there are three core reasons knowledge struggles to flow, corresponding to three kinds of obstacles—three kinds of transaction cost.

Knowledge friction: the objective obstacle, rooted in bounded rationality

Knowledge itself has two characteristics. First, different trades are as distant from each other as mountains: different professions, functions, business units and technical communities differ enormously in language, experience, mental frameworks and problem awareness. Second, professional knowledge is inherently complex—much of it cannot be nailed down on paper, must be applied in concrete contexts, requires antecedents and consequences, and involves a great deal of scattered information. On top of that, no one can know everything and everyone's energy is limited. So when different departments understand the same issue differently, it is not necessarily parochialism at work; it may simply be that within a short time it is genuinely hard to grasp another department's professional knowledge and the positions that flow from it. This is why organizations need mechanisms such as documentation, processes, training and job rotation, as well as cross-departmental meetings, to overcome knowledge friction.

Knowledge gaming: the subjective obstacle, rooted in opportunism

Many people describe organizations as warm communities where everyone exchanges freely across boundaries. Ask yourself honestly: is that really true inside your organization? Probably not—especially in large companies. Why? Because knowledge is not only the capacity to solve problems; it is also the foundation of organizational power. Holding key knowledge usually means control over customer relationships, the authority to define problems, bargaining power over resources, and organizational status. Knowledge holders therefore do not necessarily want to share; they may share selectively, delay sharing, or deliberately keep information opaque. Why can "departmental walls" never be torn down? Because people lock knowledge away in their own interest. This gaming is textbook opportunistic behavior, aimed at personal rather than organizational benefit. To mitigate it, organizations need carefully designed KPI incentives, peer evaluation, internal reputation mechanisms and culture building—but these generally only reduce the intensity of opportunism; they cannot dissolve the structural relationship between knowledge and power. This obstacle is a large one.

Knowledge integration: the structural obstacle and the biggest one, rooted in external uncertainty

The division of labor in a formal organization is usually relatively stable, but external problems do not arrive neatly parceled out along organizational lines. Once an organization is established, how should it invoke knowledge to meet customer needs? That depends on the needs themselves. Why did companies end up as matrix organizations, with both product lines and sales lines? Because satisfying a single customer need requires both the product department's knowledge of technical performance and the regional sales team's knowledge of the customer's industry or company specifics. Only when the two kinds of knowledge come together does a solution emerge. That is knowledge integration.

Formal arrangements like the matrix exist precisely because most organizations must integrate knowledge, because customer needs are not fixed, and because predicting customers' true individualized needs is hard. Hence the process committees, project teams, steering groups, project management offices and virtual teams. These extra organizational arrangements add enormous management cost—and yet, to meet uncertain demand, organizations have no choice but to keep adjusting themselves, layering on informal structures to integrate internal knowledge and satisfy external demand.

Three Core Changes as AI Reshapes the Organization

How does AI solve these problems? The problems themselves have not changed; the means of solving them have.

• Knowledge invocation shifts from "finding people" to "finding knowledge," resolving knowledge friction;

• Knowledge capture shifts from "archiving results" to "tracing process," cracking the knowledge gaming problem;

• Knowledge flow shifts from "departmental collaboration" to "task connection," lowering the cost of knowledge integration.

From "finding people" to "finding knowledge"

In a traditional organization, to draw on an expert's knowledge you first had to know who held the relevant experience. Department names were the address of knowledge; the demand side had to negotiate with a specific department and initiate cross-departmental collaboration to invoke it, and the department manager was the gatekeeper of knowledge. This arrangement was rational enough: people are the carriers of knowledge, their time and energy are limited, the use of a scarce resource had to be controlled, and so the power to set priorities fell to department managers. Experts therefore sat behind departmental walls, and invoking them meant crossing those walls—an arrangement rational under the premise of "scarce expert attention."

In the AI era, materials, databases and AI itself can fully retain an expert's entire body of knowledge. In the future, a department's external interface may well be a digital expert—a knowledge interface—that proactively retrieves relevant departmental knowledge in response to invocation requests. Even when the expert is unavailable, the knowledge can still be retrieved: knowledge is separated from the human body that holds it. On one hand, such invocation is no longer constrained by human experts' time and energy, so controlling invocation priority is no longer necessary and departments lose that portion of administrative power. On the other, AI digital experts possess near-unbounded rationality and can understand cross-departmental, cross-disciplinary knowledge well—precisely overcoming the knowledge friction caused by bounded rationality. Once knowledge is decoupled from people, knowledge supply expands dramatically and the efficiency of knowledge consumption rises sharply, ultimately driving growth in knowledge flux.

Some will ask: does that mean experts face unemployment, since AI can "distill" their knowledge and replace them? Not really. A digital expert performs only the knowledge-circulation function; the richness and accuracy of knowledge still need human experts to guarantee, and human experts continue to play an irreplaceable role. The real change lies in the department's organizational positioning: power will come more purely from knowledge authority and responsibility rather than from administrative authority and monopoly. What experts must do is continuously update the knowledge base behind their digital colleagues, ensuring knowledge stays accurate and fresh—that is the expert's new core value.

From "archiving results" to "tracing process"

Knowledge management has never worked well. Neither consulting firms nor law firms have found a way to get employees to actively contribute all their knowledge, and the root cause is knowledge gaming: as the carrier of knowledge, handing your knowledge over means handing over your core competitiveness, and you will naturally lose out in promotion contests. There was no fundamental solution in the past—knowledge lives in people's heads, and you cannot force anyone to share it.

AI now brings a turning point. In the future, when knowledge workers do all of their work—reading documents, writing emails, revising proposals, building decks—if the entire process takes place within an integrated working environment (IDE), then meeting conversations, proposal iterations and decisions will be captured in full and in structured form as they happen; every action leaves an automatic trace. Take the simplest example: two experts in the strategy department are competing for a director position. One spends three hours finding a paper critical to the company's strategy. He would never voluntarily share it with his competitor. But within an integrated working environment, the fact that he browsed that paper is automatically logged in the system, and when the other expert searches the same topic, the article automatically ranks near the top of the results.

Process tracing helps increase the observability of knowledge transactions and suppress knowledge gaming: contributions become easier to identify, invocation becomes easier to track, and the room for withholding and selective sharing is greatly compressed. Knowledge thus turns from an individual instrument of power into an organizational asset, resolving the knowledge gaming problem at its root.

From "departmental collaboration" to "task connection"

In traditional organizations, the more complex the requirement, the more units must be coordinated and the higher the management cost. Cross-departmental collaboration used to require departmental approval and leadership allocation: the department first decided whether to join and then dispatched participants, who in turn had to shuttle information back and forth between the project and the department—a very expensive way to manage.

In the AI era, the power structure of department leaders will be deconstructed. Everyone will have their own digital assistant or digital twin serving as an interface for external collaboration. These agents are embedded within business units while also crossing their boundaries, connecting directly to customer needs, expert experience, historical cases and process knowledge. When a requirement appears, the lead department's digital twin can engage other digital twins directly, invoke the relevant knowledge modules and generate a preliminary proposal—largely eliminating cross-departmental knowledge friction and dynamically recombining knowledge around the task at the base layer.

At the same time, departments used to be inherently conservative about cross-departmental collaboration because human experts' time bandwidth is scarce. Now digital twins have near-infinite time bandwidth and can respond to collaboration requests by default. Humans need only to set rules and permissions, and to review and revise preliminary proposals; only extreme exceptions require human intervention to veto and restart. The path of knowledge flow is no longer "Department A → Department B → customer," but a recombination of humans, digital twins and knowledge systems pulled by the task itself. In the future, even customers and suppliers may dispatch their own digital twins to join in, enabling cross-boundary knowledge flow and dramatically reducing the management cost of knowledge integration.

The Three-Layer Architecture of the AI-Native Organization

Building on a diagnosis of where knowledge meets resistance, AI organizational transformation can unfold around a three-layer operating architecture.

The first layer: consolidate the underlying knowledge infrastructure. Companies must bring online, structure and semanticize the knowledge scattered across documents, meetings, projects, customer interactions, process systems and expert experience, so that AI can recognize, understand, invoke and update it. The point here is not simply to build a bigger knowledge base, but to turn organizational knowledge from static archiving into dynamic infrastructure. The human role at this stage is knowledge producer, quality validator and feedback provider. The key discipline is not to rush into building intelligence, but to build flowability first—letting knowledge prove its value through circulation.

The second layer: cultivate a middle layer of digital agents (a shadow organization). This includes department-level digital agent interfaces and a digital twin assistant for every employee. Two points are critical. First, permissions must be set carefully: any member may question any digital agent (initiating a knowledge invocation request), but the knowledge and data that agent can access must be controlled. Second, mobilize employees to train their own digital assistants, and assign AI deployment specialists to each department—but more important still is inspiring employees to actively help these assistants grow and improve. To be clear: digital agents are nodes of knowledge flow, not nodes of responsibility; they are assistants, not replacements.

The third layer: build the top layer of human organization. Once the first two layers take shape, companies need to redefine the roles of human employees, managers and departments. Departments should no longer be primarily units of knowledge monopoly, but units of knowledge production, validation and accountability. Managers should no longer coordinate mainly through information asymmetry and administrative authority, but take greater responsibility for rule-setting, permission governance, exception judgment and the management of external uncertainty. Employees should no longer be mere executors of processes, but producers, trainers and updaters of organizational knowledge. This process may squeeze out some employees who cannot adapt to AI—but that is an outcome, not a precondition.

Humanity's Moat Lies in Managing Uncertainty

What AI changes about organizations essentially overturns two default premises on which organizations have always run: knowledge friction arising from bounded rationality, and knowledge gaming arising from opportunism.

If we use AI to lower knowledge transaction costs and unclog management bottlenecks, and then release human capacity—or even hire more people—to unlock the growth potential of existing resources, we can restart the growth engine entirely. But an AI transformation that begins with layoffs goes wrong from the very start: the trust crisis caused by layoffs makes it very hard for employees to genuinely participate in opening up knowledge and rebuilding processes, AI investment never converts into returns, and the company enters a negative cycle of contraction, losing its growth prospects.

Management work that exists only to police internal processes, relay information and play internal politics—"management work oriented toward internal complexity"—does face real replacement risk. But I want to emphasize one conclusion that matters enormously: do not look for the reasons we are irreplaceable in human attributes such as emotion or creativity. That is anthropocentric talk. The real moat comes from demand that does not change. The uncertainty of this world will not disappear because AI has arrived; coping with external uncertainty is, and always will be, the core value of human beings and of firms.

Do not fear machines taking over operational work. Let machines handle internal complexity, and free people to do what machines cannot.

One last thing. Companies should generously release their human resources: let machines handle complexity inside the organization, and liberate people to do what machines cannot do—exploring the frontier of the organization, coping with uncertainty, finding new sources of growth, building new business models, and forging consensus among stakeholders. That is the part of an enterprise truly worth relying on and taking pride in.