Musheng Consulting proposes the MHCM model: classifying human-machine collaborative team maturity into four levels

Enterprises will not build human-AI collaborative teams overnight; the transition from traditional human teams to a state of human-AI collaboration necessarily goes through an evolutionary process. Based on the theoretical deduction of Moo Consulting and the practical feedback currently collected, we attempt to propose the Moo Human-AI Collaboration Maturity (MHCM) model for human-AI collaborative team organization.

Given our judgment that human-AI collaborative teams will move toward platform-based organizations (or agent-based organizations), their organizational design will naturally move toward the three-platform architecture of platform-based organizations (or agent-based organizations). We have referenced the four-level maturity model for the three-platform architecture (see Figure 1-1). In HACM, we also divide the enterprise's evolution into four levels and match them with the maturity grading of platform-based organizations. This model can be directly used for enterprises' self-assessment and path planning in building human-AI collaborative teams.

Figure 1-1: Platform-based organization maturity model. Source: Moo Consulting

01 Level 1 — Tool Level

The typical manifestation is that employees privately use AI to write emails, revise copy, and make spreadsheets; the enterprise has no AI usage guidelines; and "Shadow AI" emerges. This state naturally leads to various risks, such as data leaks, distorted content, algorithmic bias, and unclear accountability.

For example, in 2023, after Samsung allowed employees in its semiconductor division to use ChatGPT, three leaks occurred in succession within 20 days. Employees pasted core source code into it to debug code, entered complete product data to optimize yield, and even uploaded meeting audio converted to text to organize minutes. Once such information enters a public model, it cannot be retrieved and may be used for subsequent training, which naturally creates risks for Samsung.‌

For another example, in 2025, Deloitte Australia's 237-page welfare benefit review report written for an Australian government department contained fabricated court precedents, nonexistent academic papers, and expert opinions that could not be found in public databases. After the incident was exposed, Deloitte was forced to refund about AUD 440,000 (about RMB 2.06 million) in fees and admitted in a revised version that it had used generative models such as Azure OpenAI.‌ The particularity of this case lies in the fact that Deloitte was at the time loudly announcing the rollout of AI tools to 500,000 employees worldwide, while at the same time a major problem appeared in its deliverables to clients.

This level corresponds to the pyramid organization. Because employees are confined within "small squares," they have ample space to act on their own, and naturally they will also take risks in using AI to save effort, save time, and produce results. In such organizations, without unified and detailed supervision, the above risks are bound to erupt. To cross this stage, enterprises need to establish formal AI usage guidelines and data boundaries, shifting from "default tolerance" to "explicit governance," and at minimum issue an internal "AI Constitution."

Unfortunately, at this stage, the bosses of enterprises have not formed systematic thinking about AI as a productive force. They will call on employees on various occasions: "The future is the era of AI; we must keep up," and "You young people should try AI more." But this is only a verbal call, and the ones being called on are others. They will not make project-level investments for "AI-ization" (at most a token gesture, buying some tokens), and they hold a "whatever happens" attitude toward returns. To put it plainly, at this stage, the bosses of enterprises regard themselves as "outsiders" at heart.

According to our definition of human-AI collaborative teams, such teams clearly do not meet the standard. To be precise, this is traditional human teams "testing the waters in the shallows" before moving toward human-AI collaborative teams. Enterprises and human employees at this stage are beginning to understand the new technology, just as primitive humans obtained fire: there are new possibilities, but also new risks.

02 Level 2 — Assistant Level

The typical manifestation is that all employees are equipped with tools such as Copilot and work buddy; the enterprise has established an AI training system to popularize usage skills and usage guidelines among employees, and even uses post-training certification as an access standard for using AI; but at this stage, "AI-driven efficiency improvement" is still an individual-level phenomenon, and organizational efficiency has not improved in sync.

At this stage, bosses begin to lead by using AI, enterprises begin to assess token consumption, and they wish every task that can use AI would use AI. They may even launch large-scale layoffs to demonstrate their determination to "transform," but the expected "efficiency explosion" does not arrive as scheduled. Compared with the "Tool Level" stage, individual efficiency improvement at this stage is obviously more pronounced, but under strong assessment pressure, tokens are also consumed frantically. If individual efficiency improvement cannot be transformed into organizational efficiency improvement, then the cost consumed on tokens cannot be recovered.

The Amazon case illustrates the problem well. They once launched an internal AI leaderboard called Kirorank, scoring employees based on token consumption, with the original intention of promoting AI application. But employees quickly found loopholes and began having AI perform large numbers of meaningless tasks (such as repeatedly processing trivial instructions) to inflate token consumption and improve their rankings. This caused the company's computing costs to rise sharply without generating actual business value. Ultimately, Amazon shut down the Kirorank leaderboard, and Senior Vice President Dave Treadwell also warned employees "not to use AI for the sake of using AI."‌

The greater awkwardness of this stage lies in the fact that, although the risks of using AI are reduced to a certain extent and individual efficiency gains are achieved in some areas, the organizational model has not changed, and AI can easily cause "internal dysfunction" within a pyramid organization. Individuals whose efficiency has been improved and those whose efficiency has not will come into collaborative conflict; ultimately, the weakest link determines the height, and organizational efficiency remains low. Moreover, when the use of AI takes individual interests as the primary consideration, conflicts between individuals and between departments will erupt quickly, and large-enterprise diseases will even become more prominent in some areas. Therefore, for enterprises that remain at Level 2, the marginal return on their AI investment will decline rapidly: token consumption will become faster and faster, while returns will become smaller and smaller.

But the advantage of this stage is that, because AI tools are used extensively, a considerable number of human employees engaged in "ineffective running around" begin to be replaced, organizational redundancy is squeezed out, and the organization enters a streamlined state. At the same time, some elite strivers realize that AI tools can bring infinite possibilities, begin to spontaneously optimize work units, promote cross-departmental collaboration, and actively clear work flows, objectively creating incremental customer value. When business owners realize the benefits brought by this spontaneous state, they will adopt a tacitly permissive attitude, provide partial support, and even establish "experimental zones." Compared with the Level 2 state in a platform-type organization, namely the "flexible pyramid organization," AI assistants play the role of an amplifier, making the pyramid more flexible. From the traditional perspective of workforce efficiency, redundant staff being replaced by AI and elites making breakthroughs in "experimental zones" will naturally bring improvements in workforce efficiency data. In addition, because human employees can use AI assistants, the marginal return on human input will show increasing returns. The only thing is that enterprises are unlikely to choose to add people at this stage.

This creates a rather paradoxical scene: AI-empowered bureaucracy coexists with elite-driven human-machine collaboration; traditional workforce efficiency rises while token efficiency falls; traditional workforce efficiency has increasing marginal returns while token efficiency has decreasing marginal returns, producing what Mu Sheng Consulting calls the "marginal scissors gap."

But bosses prefer to see the bright side. It is precisely because elites stand out and workforce efficiency continues to rise that they may be misled into believing that business operations are still good, that there is no need to continue promoting organizational upgrading, and that they only need to wait for the seeds in the "experimental zones" to blossom and bear fruit automatically. But this idea is extremely irrational: the improvement in workforce efficiency comes at the cost of sacrificing token efficiency. It is not that human input has decreased, but that human input has become computing power input. Moreover, if bureaucracy is not restrained, the "internal dysfunction" of the pyramid organization will become increasingly obvious and will gradually eat away at the local efforts of elite strivers, while meaningless token consumption will also cause the enterprise's investment to vanish into the sea. In reality, the vast majority of enterprises will be stuck at this level.

According to our definition of a human-machine collaborative team, such teams still do not meet the standard. AI is only a tool, does not yet possess autonomous awareness, and has not yet taken over certain functional domains. It can be said that this is the "deep-water exploration" of traditional human teams before they move toward human-machine collaborative teams. At this stage, enterprises and human employees begin to gradually master new technologies and integrate them into production.

03 Level 3 — Team Level

In this type of organization, AI agents become "formal members" of the team, with clear responsibilities, permission boundaries, and human owners. Enterprises unleash the power of AI agents through organizational change. As a result, after a certain degree of upgrading, a large number of work units in the organization increasingly rely on AI support, and enterprises begin to connect work units around "human-machine collaboration" and design customer-oriented work flows.

At this stage, because the three-platform architecture has been established, cross-departmental collaboration has become the norm. Moreover, because AI agents have joined cross-departmental collaboration, collaboration can be standardized and connected via APIs, greatly improving efficiency, making contributions to one another clearer, and making incentive feedback more direct. At this time, departments in the organization become increasingly virtual, functioning more like "bases" than the "territory" of department heads. Department heads are more like businesspeople than administrative officials. They begin to focus on how to form more competitive work units and strive to connect them to work flows that create customer value, thereby seeking monetized returns on professional capabilities. Precisely because of this philosophy, stable "1 person + multiple AI agents" structures emerge in scenarios such as customer service, recruitment, and R&D. At the same time, enterprises also establish identity and permission management for AI agents, and the outputs of AI agents are included in quality monitoring. In other words, under the enterprise's "AI Constitution," various forms of legislation are more complete and enforcement is more thorough.

There are two contradictions at this stage:

First, the contradiction between rapidly proliferating AI agents and the level of governance. A 2026 study by Strata Identity shows that only 18% of security leaders believe their identity and access management system permissions can effectively manage AI agent identities, and nearly 80% cannot track agent activity in real time (Strata Identity, 2026). Pandora's box has been opened: various AI agents are rapidly coming online, and the enterprise's governance level has not matched them as expected. "Not knowing how many AI employees you have hired" is the most common reason for failure at the Team Level.

Second, the contradiction between the rapidly erupting productivity of human-machine collaboration and organizational models. As organizations have begun to evolve, human employees can build AI agents to upgrade work units and enter workflows. On the one hand, they see greater possibilities, but organizational upgrading is not yet complete, and the progress of productivity still falls far short of their expectations; organizational flexibility and the level of enterprise empowerment still lag behind. On the other hand, their value-creation capacity has increased substantially, but they have not obtained returns through sufficiently market-based incentive mechanisms.

At the same time, enterprises have begun to popularize the management concepts of "human-machine ratio" and "token efficiency," have begun to develop a new understanding of the "human efficiency" of human-machine collaborative teams (or agent organizations), are gradually abandoning traditional human efficiency indicators (or treating such indicators merely as "auxiliary indicators"), and are forming new human efficiency management concepts and methods. When enterprises pay attention to the "human-machine ratio" and "token efficiency," they will discover that the organizational model is the biggest obstacle to these two indicators. Therefore, enterprises at this stage will not be stuck as assistant-level (Level 2) enterprises are; they will most likely choose to continue promoting organizational change, and their human-machine collaborative teams thus have hope of advancing.

According to our definition of human-machine collaborative teams, such teams have already reached the standard. AI is not only a tool but also possesses autonomous awareness and has begun to take over a large number of functional domains. It can be said that such human-machine collaborative teams are already "treasure hunting in deep waters." At this stage, enterprises and human employees have begun to gradually become familiar with new technologies and have integrated them extensively into production, entering a brand-new era and ushering in broad prospects. Naturally, such "about-to-erupt" enterprises will also become darlings of the capital market, and capital is looking forward to their "visible future."

04 Level 4 — Agentic Organization Level

Human-machine collaboration becomes the mainstream form of the enterprise's mainstream work units. The organization displays the characteristics of "few humans, many AIs; a multi-center dynamic network; and infinite fractaling centered on customers," and the enterprise upgrades into an agentic organization. In terms of organizational structure, the three-platform architecture is fully mature. Work units of "human employees + AI agents" are invoked on demand and connected in a standardized way via APIs, forming smooth workflows that penetrate business scenarios and, for different vertical markets, generate "exponential explosions" of customer value.

At this stage, the middle and back offices provide services via APIs and dispatch combinations of "human BPs + AI agents (such as procurement agents and performance management agents)" into the front office. The front office's core human employees connect these combinations to form operating units and steer customer-oriented workflows. In form, the front office consists of small human-machine collaboration teams. At the same time, front-office operating units make co-investments, claim performance, and enter into profit-sharing bets. The human employees participating in operating units obtain profit sharing through co-investment claims, while the middle and back offices that dispatch AI agents obtain standardized fees or profit sharing through service agreements.

Such enterprises have already formed purely market-based incentive mechanisms and have fully popularized the new-generation human efficiency management model of "human-machine ratio" and "token efficiency," with measurement of such human efficiency penetrating every workflow and work unit. The enterprise becomes an investment platform that, based on the logic of human-machine ratio and token efficiency, invests resources to support each function in building and operating AI agents and directs them toward the workflows most capable of generating economic returns, ensuring maximum efficiency, maximum operating returns, and maximum profit. Moreover, the enterprise's governance level is also more mature. Because the organizational model simplifies governance, they only need to focus on basic clauses and monitor core indicators. They are not managing in greater detail; they are managing more cleverly.

At this point, the enterprise enters a virtuous cycle of exponential growth and is no longer troubled by large-enterprise diseases. In other words, in this organizational model, both humans and machines automatically increase efficiency. They begin to aim at market feedback: humans strive to be strivers, while machines, driven by humans, work hard for their own survival. For machines, if there is no performance, or if efficiency falls below a threshold, they will be abandoned because they cannot obtain token investment.

According to our definition of human-machine collaborative teams, such teams have fully reached the standard. AI not only possesses autonomous awareness but has also begun to become the core pillar of various functional domains. It can be said that human-machine collaborative teams at this time are comprehensively carrying out "marine resource development." At this stage, enterprises and human employees are thoroughly at home with new technologies and act freely in a brand-new era. Of course, no enterprise is currently recognized as having fully reached Level 4. The value of this level lies not in "whether it has been achieved" but in the clear "North Star" it provides for enterprises.

05 The trajectory of human-machine collaborative team maturity

Summarizing the above four stages, we can clearly identify the evolution process of human-machine collaborative teams:

Figure 1-2: Mu Sheng Human-Machine Collaborative Team Organizational Maturity Model (MHCM). Source: Mu Sheng Consulting

In this process, enterprises have three growth trajectories:

First, the depth of AI technology application and the matching level of governance. From treating AI as a tool, to treating it as an assistant, to forming AI agents and granting them employee status, AI is exercising greater and greater autonomy and increasingly fulfilling the responsibilities and powers of an "employee." At the same time, governance logic has also moved from complete absence, to full coverage, to grasping the major and letting go of the minor, and the direction from "managing AI" to "managing the platform" is very clear.

Second is the depth of organizational change, including the three-tier architecture, incentive mechanisms, and enablement mechanisms. Along this trajectory, the shift from pyramid organizations to platform organizations remains the same, fully consistent with the platform organization maturity model we described. For AI to unleash its power, enterprises have no choice but to build the most open and flexible organizations. We also emphasize that this trajectory is the core trajectory of the evolution of human-machine collaborative teams.

Third is the depth of understanding of effectiveness. Initially, enterprises treated AI as a tool, so investment was naturally limited, and of course there was no need to focus on the return on investment in this area; later, AI became an employee, investment gradually increased, and human input gradually became computing power input, and focusing on the "human-machine ratio" and "token efficiency" became mainstream. Only by shifting the perspective on effectiveness, focusing on the return on investment in AI, and rationally allocating investment chips can AI unleash its maximum productivity. In this regard, Silicon Valley has responded most swiftly. Starting in 2025, a large number of financially sound Silicon Valley companies launched a number of layoffs under the name of "AI replacement," the most direct reason being to please Wall Street (broadly referring to the capital markets). Because capital is not optimistic about those tech companies that hoard large numbers of employees rather than GPUs.

Faced with such evolutionary laws, how will Chinese enterprises respond? The Mu Sheng Consulting Research Institute predicts that large Chinese enterprises will generally complete the leap from Level 1 to Level 2 around 2027; between 2028 and 2030, a batch of enterprises entering Level 3 may emerge; but before 2030, enterprises entering Level 4 will still be a minority, except that their demonstration effect will accelerate the overall evolution of the industry. It should be particularly noted that crossing levels does not depend on the amount of AI investment, but on the depth of organizational change—this is the core proposition repeatedly argued in the first part of this report.

Regarding maturity models related to human-machine collaborative teams, other enterprises and research institutions have also made related statements, such as the five-level classification jointly proposed by Lenovo and the Renmin University of China Business School; Alibaba, ByteDance, Baidu, and Haier have each proposed their own three-level classifications; Meituan has proposed a four-level classification. However, we insist on using the above classification and believe that this model is better able to guide enterprises toward correct practice.

Among these classifications, the most influential should be the classification system of Lenovo and Renmin University of China¹. In September 2026, the two institutions jointly released the "AI-Native Organization Transformation Guide" and proposed an L1 to L5 five-level maturity classification: L1 ("+AI" local attempts) → L2 ("+AI" widespread application) → L3 (hybrid AI foundation) → L4 ("AI+" operations reconstruction) → L5 (AI-native maturity). This model points out that "moving from '+AI' into 'AI+'" corresponds to the "perilous leap" between L3 and L4.

In Lenovo's own context, its agents have indeed reached the L3 level of "limited autonomy," capable of autonomously planning, decomposing tasks, and reversely influencing processes. And Lenovo has indeed done a great deal of work on the "AI foundation": self-developed "Qingtian 5.0" hybrid AI foundation, the Agent OS unified operating system, and pushing token costs to extremely low levels. From this perspective, it possesses the technical infrastructure characteristics of L3 and is a pioneering representative among Chinese enterprises. But a deeper look at its governance mechanisms reveals that the "autonomy" of AI agents is still strictly confined to the "execution" layer, and "decision-making power" has not been substantively transferred. If classified according to the above five levels, technology application has reached L3, while organizational governance remains at L2.

In fact, this contradiction between appearance and substance is precisely why we insist on using MHCM. A large number of classification models on the market share a distinct orientation²: they are essentially "technology application maturity" models, not "organizational change maturity" models. Using the degree of technology application as a yardstick will mistake "using tools well" for "deep organizational transformation." In fact, we fully understand why enterprises choose such classification models—because organizational change is too difficult to quantify, too painful, and too touching on interests. Or to look at it this way: saying "we have deployed 100 agents" is much easier than saying "we have handed a certain decision-making power to AI and borne the corresponding risks." To put it more plainly, the technology yardstick allows enterprises to claim they are "undergoing transformation" without touching the power and responsibility structure.

MHCM breaks away from technological fixation and returns to the origin of organizational evolution. It insists that a stage classification with real reference value must take factors such as the transfer of decision-making power, responsibility traceability mechanisms, organizational structure reshaping, incentive system reconstruction, and upgrading of effectiveness concepts as core criteria, rather than merely the number of agents and their level of autonomy.

¹ This "Guide" was led by the World-Class Enterprise Research Institute of Renmin University of China and the Business School of Renmin University of China, and jointly completed with Lenovo Group. This is a team with deep industry-academia-research integration; Lenovo provided the practical sample of its three-year transformation, while the Renmin University side was responsible for refining and constructing the academic framework. The academic leader of this research team is Professor Zhang Jin, Vice Dean of the Business School of Renmin University of China and Executive Dean of the World-Class Enterprise Research Institute.