01 定义——AI介入后的“混合编队”
传统管理学语境下的“团队”，默认成员是自然人。团队被定义为“由两个或两个以上相互作用、相互依赖的个体，为实现特定目标而组成的集合”。这个时期，强化人的作业能力的要素被定义为“工具”或“机器”，它们要么强化了人的脑力（运算能力），要么强化了人的体能（物理边界），但无论它们有多强大，始终都是以人为核心的，是人的附属品。
这种对于团队的定义，在AI智能体大规模进入企业后，开始出现了变化。当AI为工具注入了一定范围内的独立思考能力，它们就成为了另一种形式上的“员工”。所以，这种新型的“人机协同团队”成员不再限于自然人，成员之间的“相互依赖”不再只发生于人类之间，还发生在人—AI、AI—AI之间，传统的组织与人力资源管理理论方法，显然应该改变了。
据此，穆胜咨询将人机协同团队（Human-AI Collaboration Team）定义为：以实现特定经营或交付目标为目的，由人类成员与AI智能体（含数字员工、机器人及可调用的外部智能服务）共同构成，通过明确的人机分工、决策权配置与协作协议形成稳定协作关系的作业单元。
这一定义有三个关键限定词。其一，“稳定协作关系”——排除了“人偶尔用一下AI工具”的临时状态。一个人用ChatGPT写一封邮件，不构成人机协同团队；只有当AI智能体被赋予固定职责、进入常态化工作流、并对团队产出承担责任时，人机协同团队才真正存在。其二，“决策权配置”——AI智能体是否拥有决策权，以及在多大范围内拥有决策权，是判断协同深度的核心标尺。其三，“作业单元”——强调这是可以在组织中被设计的最小单位，有明确工作内容和目的，而非抽象概念。
02 五元构成——谁在这个团队里？
人机协同团队的成员可以归为五类角色（如图1-1）。需要说明的是，“五元”是分析性分类，实践中一个团队未必同时具备全部五类。
第一类：人类成员（Human Members）。 穆胜博士在2026年出版的《智能体组织》一书中，按照价值创造的金字塔，把这类组织中的人类员工从高到低分为五种类型，即思想家、设计师、专家、情报员、强化作业者。当下，我们考虑在一个相对成熟的智能体组织里，剔除掉“创世”所需的思想家和设计师，也剔除掉情报员和强化作业者这两类过于基层的一线员工（working-level），人类员工就主要是“专家”，他们的角色主要有三类。
一是方向设定：设定目标和规划路径，即从结果和过程两个维度定义“什么算好”，换句话说，这是建立评估体系，以便作出价值判断。二是监管和接管：既要按照评估体系来监督智能体的运行是否按照既定目标、遵循既定路线，也要处理智能体无法覆盖的长尾、高风险、高情绪场景。三是AI训练与编排者：考虑企业经营目标和商业环境等要素的变化，企业需要适时更新各类智能体，这就需要进行提示（Prompt）设计、知识供给、智能体工作流设计与质量验收。微软将这些角色命名为“智能体的老板”（Agent Boss），并认为，驾驭智能体将不再是高管专属，而是每一位员工的必修课（Microsoft，2025）。事实上，人类管理者也是通过相似的方式来管理人类员工的。
第二类：AI智能体（AI Agents）。 AI智能体是一种以大语言模型为核心大脑的智能系统，具备感知、推理、规划、记忆和执行能力。它的核心特征是目标驱动：你只需给出最终目标，它就能够理解意图、自主拆解任务、规划路径并调用多个工具来完成闭环交付，无需一步步干预。
Gartner对“AI工具—AI助手—AI智能体”作了清晰区分：AI工具“你问它答”，交付一段建议；AI助手“你下指令、它辅助执行”，交付半成品；AI智能体“你交代目标、它自主完成”，交付完整任务结果，但需要清晰的权限与边界（Gartner，2025）。这一区分极其重要：前两者是“人的效率放大器”，只有第三者具备了一定的自主意识，才能真正成为“团队的成员”。事实上，AI智能体的出现，也是“机器”开始被作为一种管理对象的分界线。
第三类：数字员工（Digital Employees）。 数字员工通常由AI智能体、RPA等多项技术组合而成，拥有工号、职责边界和持续记忆，能够像真实同事一样承担特定岗位的持续性工作。它与管理工具的差别在于组织身份：数字员工有“岗位”、有“上级”、有“绩效”，甚至可以被“解雇”。旧金山的实验性零售店Andon Market就由AI Agent Luna负责招聘、排班、请假与员工管理，且Luna曾在一名员工多次迟到后建议终止雇佣关系。这类实践虽属前沿试验，但它标志着数字员工的组织身份正在从“概念”走向“制度”。
这里需要重点区分数字员工和AI智能体。尽管数字员工以AI智能体为核心，但后者和前者并不是一种进阶关系。这里，有必要进行区分：
第四类：机器人（Robots）。 具身智能形态的物理执行者。它把协同的边界从“信息流”推进到“物理流”。在制造业与供应链场景中，机器人是AI智能体的“手”。在人机协同的语境下，机器人也包括RPA（机器人流程自动化），它是与人类、AI智能体共同构成协同团队的核心执行角色之一。
第五类：外部生态伙伴（External Partners）。 包括外部API服务、专业智能体、外包“数字劳动力”。Gartner在供应链研究中发现，76%的供应链领导者已实施或计划实施相关功能，以自动化地与“机器客户”（即能独立谈判和采购的非人类经济体）互动（Gartner，2025）。当交易对手也可能是AI时，“团队边界”就必须重新定义。
03 能力矩阵——人机能力如何互补？
人机协同的底层逻辑是“能力互补”，而非“能力替代”。当下，包括穆胜咨询、麦肯锡与微软等主流研究机构发布的内容，都指向一个共同的结论：AI在“广度、速度、一致性、规模”上具有压倒性优势；人类在“判断、共情、价值取舍、非常规创新”上仍不可替代。穆胜博士对此有一个更具文学性的表述：无论技术如何演进，高级创意、复杂判断与共情仍属于人类，构成AI难以进入的“上帝禁区（好比上帝在造物时就设置好的禁区）”（穆胜，2025）。
据此可以绘制人机能力矩阵（以四个维度描述）：
穆胜博士在《智能体组织》一书中，分析了这类组织里少量人类员工的价值，我们可以基于这些价值推演出上述矩阵。其他研究机构的结论也证实了这种人机协作的必要性。微软Work Trend Index的数据显示：员工今年转向（更愿）与AI而非同事合作的第一原因是“24×7可用性”（42%），其次是“机器速度与质量”（30%）、“无限的创意想法”（28%）——恰是“人类无法提供”的三项（Microsoft，2025）。同时，领导者认为AI时代最重要的人类技能是“质量控制”（50%）与“批判性思维”（46%）（Microsoft，2026）。这两组数据的对照，最清楚地说明了人机分工的分界线：机器负责“做得多”，人类负责“判断得对不对”。
04 人机比&token效——效能的组织度量
“人机比”（人类成员数 : AI智能体数）正在成为一个新的组织度量。它是一个人效（HR Efficiency）的中介变量，有极其明显的“杠杆效应”，人类员工越是能够驾驭更多的智能体，人效就会越高。但是，这个指标并不容易衡量。穆胜咨询提示，有两个计量方向需要注意：
一是AI智能体的“当量”不统一。一个负责全公司考勤巡检的代理，与一个负责单个客户报价的代理，工作量与价值完全不同。穆胜咨询的建议是，把对“数量”的计量转化为对“工时”或“任务量”的计量。事实上，“数量比”远不如“工时比”或“任务量比”有意义。
二是AI智能体的成本结构不同。AI智能体的边际人工成本低，但存在算力、模型调用、数据与运维的持续投入。
这正是穆胜咨询提出的“人效口径重写”的现实基础：分母应当是“人机总投入”——人工成本加上AI相关的算力、模型、数据与运维投入；人效的定义应走向“token计效”（穆胜咨询，2026）。 穆胜咨询认为，“人机比”只有与“人机总投入产出比”结合，才是有管理意义的指标，企业应该致力于提升“人机比”和“token效”，两者的结合构成了智能体组织的“人效”，精确地说，智能体组织的人效=人机比*token效。
穆胜咨询研究院预测：到2028年前后，“人机比”与“token效”将成为规模化企业的常规经营指标，并在部分先进企业进入董事会与经营分析会的常规议题；但“人机比”不会出现统一的行业标杆值，因为其合理区间高度依赖业务场景的标准化程度，更应该是企业的定制指标。
05 治理结构——新型组织契约的核心
人机协同团队最容易被忽视、却最关键的静态构成要素，是治理结构。
原因很直接：当团队里出现一个能够自主行动的非人类成员，“谁批准了它”就成了一切治理问题的起点。HCLTech¹ 在研究中描述了一个典型的困境场景：凌晨2点03分，一个关键企业工作流突然启动，没有工程师部署变更，没有分析师发起自动化，但一系列决策在财务、客户系统与云基础设施之间精准移动；到天亮时，高管们只问一个问题——“谁批准的？”答案是：“没有人，是系统自己做的。”（HCLTech，2026）
真实数据印证了治理缺口：Strata Identity² 在2026年研究显示，仅18%的安全负责人认为其身份与访问管理系统能有效管理AI智能体身份，而对这个议题表示“中等信心”“轻微信心”“没有信心或不确定”的样本分别为35%、29%和18%。近80%的企业无法实时追踪AI智能体的活动，即它们正在做什么、由谁负责。仅21%的企业拥有准确地运营中的AI智能体清单（Strata Identity，2026），而32%的企业依赖非实时记录，约32%的企业计划在一年内实施优化，8%的企业则完全没有清单。换言之，绝大多数企业其实并不清楚自己“雇了多少AI员工”，也不清楚它们到底在以什么方式运作，企业的人机团队管理处于混沌状态。
据此，AI智能体的治理应遵循四条原则：
其一，独立身份。 每一个AI智能体都必须拥有第一类身份（first-class identity），不得借用人类凭证或共享服务账号，且必须绑定一个可问责的人类所有者（HCLTech，2026；Cisco Agentic Identity方案）。
其二，任务级最小权限。 权限应依据“这个代理要做什么任务”而非“哪个部门拥有它”来授予。一个审批小额采购的采购代理，可以访问采购流程、供应商记录与采购政策，但不应对企业财务系统拥有无限制访问权（Crucial Logics，2026）。
其三，区分“读取数据”与“采取行动”。 一个代理可能需要广泛读取客户信息以分析投诉，但这不意味着它可以自动向客户发送回复。高风险动作（退款超过一定金额、变更生产权限、导出受监管数据、对外发布、执行资金划转）必须设置人类审批闸门（AI Cyber Check，2026）。
其四，全生命周期审计。 包括代理的注册、认证、授权、凭证轮换、会话监控、审计日志与注销。代理退役时，其凭证、令牌与权限必须被同步清除——“停掉工作流”并不等于“回收了权限”。
06 人机团队——全面升级的纸面实力
根据人机协同团队的特点和上述治理原则，我们可以总结出它们与人类团队的如下区别：
这张表中最值得注意的是最后一行。穆胜咨询在谈到现代企业的赋能机制（让员工“有能力干”的机制，主要通过知识管理来实现）时明确指出：“高度依赖超级英雄的企业绝非真正的优质企业”，企业应搭建“知识萃取→知识沉淀→知识分享”的闭环知识体系，让平凡人做非凡事，让非凡人做超凡事（穆胜咨询，2025）。AI智能体的加入，第一次让“下限托底”具备了工程化的可能——这可能是人机协同团队对组织理论最重要的贡献。
注释：
¹ HCLTech是一家总部位于印度的全球性科技公司，也是印度顶级的IT服务与咨询巨头之一。它区别于普通外包商的核心标签是“工程与研发（E&R&D）服务领域的全球领导者”，近年来正大力向AI驱动的技术服务转型。
² Strata Identity 是一家专注于身份编排（Identity Orchestration） 的技术公司，核心是解决企业在混合云和多云环境下，碎片化身份系统的集成、现代化改造与安全管理问题。这家企业于2019年成立于美国科罗拉多州博尔德，2026年6月已被数据安全公司 Rubrik 收购，目前是 Rubrik 旗下公司。其CEO Eric Olden是 SAML 标准的共同作者之一。公司还主导创建了身份查询语言 IDQL 和开源项目 Hexa，以推动多云身份管理的标准化。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
01 Definition — “Mixed formation” after AI intervention
“Teams” in the traditional management context, where the tacit members are natural persons. Teams are defined as “composed of two or more individuals who interact and are interdependent to achieve specific objectives”. During this period, elements of enhanced human operational capability are defined as “tools” or “mechanicals”, which either strengthen human brain (calculation) or human physical energy (physical boundaries), but however powerful they may be, they remain human-centred and subsidiary.
This definition of team began to change after AZAI agent entered the enterprise on a large scale. When AI injects a range of independent thinking into the tool, they become another form of “staff”. Thus, this new type of “human teamwork” is no longer limited to natural persons, and “interdependence” between members is no longer confined to human beings, but also between human beings—AI, AI—AI—AI—AI—AI, the traditional organization and the theoretical approach to human resources management should clearly be changed.
Accordingly, the Musheng Advisory defines the Hulman-AI Coordination Team as the operational unit for a stable collaborative relationship between members of the human race and AIAI agent (comprising digital workers, robots and accessible external intelligence services) for the purpose of achieving specific business or delivery objectives, through a clear division of labour, decision-making power and collaboration agreements.
This definition has three key qualifiers. First, “stabilizing collaborative relationships” - excludes the temporary status of “persons occasionally using an AI tool”. An individual who writes an e-mail with ChatGPT does not constitute a human team; it only exists when AZAI agent is assigned a fixed role, enters normal workflow and takes responsibility for team output. Second, the “allocation of decision-making power” - whether and to what extent AZAI agent has decision-making power is the core rule for determining the depth of synergy. Third, the “operational module” — emphasises that it is the smallest unit that can be designed in an organization with a clear job content and purpose, not an abstract concept.
02 Five-dollar composition -- who's on this team?
Members of the HRCC team can be classified as five categories of players (see figures 1-1). It should be noted that the “five dollar” is an analytical classification and that in practice a team may not have all five categories simultaneously.
Category I: Human Humanity. In a book published in 2026, " AI agent Organization ", Dr. Musheng divided the human workforce of such organizations into five categories, from high to low, according to the pyramid created by value, namely thinkers, designers, specialists, intelligence agents and enhanced operators. At this point, we are thinking of removing the thinkers and designers needed for the creation of the World in a relatively mature AI agent organization, as well as the two categories of over-stretching-level employees (working-level) who are mostly “experts” and whose roles are mainly three.
The first is direction setting: setting targets and planning paths, that is, defining what is good from the two dimensions of the outcome and process, in other words, establishing an assessment system to make value judgements. The second is supervision and takeover: both monitoring the operation of AI agent in accordance with the assessment system and following established objectives and routes, and addressing long tails, high-risk, high-emotional scenarios beyond the reach of AI agent. Third is AI training and organizer: taking into account changes in factors such as business objectives and business environment, enterprises need to update the various types of AI agent in due course, which requires Prompt design, knowledge supply, AI agentworkflow design and quality acceptance check. Microsoft named these roles “Agent Boss” and believed that managing AI agent would no longer be the exclusive preserve of executives, but a compulsory course for every employee (MSC, 2025). Indeed, human managers manage human employees in similar ways.
Category II: AIAI agent (AI Acts). AZAI agent is an intelligent system with a large-language model of the brain with sensory, reasoning, planning, memory and implementation capabilities. Its core feature is goal-driven: by giving the ultimate goal, it can understand intent, break down tasks, chart paths and use multiple tools to deliver in a closed circle, without a step-by-step intervention.
Gartner drew a clear distinction between “AI Tool-AI Assistant-AIAI agent”: AI Tool “You answer”, delivering a proposal; AI Assistant “You give the order, it supports the execution”, delivering the semi-finished product; AIAI agent “You answer the target, it does it autonomously”, delivering the full mission results, but clear permission and the boundary (Gartner, 2025). This distinction is extremely important: the first two are “human efficiency amplifiers” and only a certain sense of autonomy on the part of third parties can truly become “group members”. In fact, the emergence of AZAI agent is also the line that the “mechanical” began to be used as a management object.
Category III: Digital Employes. The digital workforce is usually made up of a combination of AZAI agent, RPA etc., which has the number of jobs, the boundaries of duties and the memory of continuity, and is able to perform the same continuous work as a true colleague in a given position. It differs from management tools in terms of organizational identity: digital workers have “jobs”, “superior”, “performance” and can even be “dismissed”. The experimental retail store in San Francisco, Andon Market, was managed by AI Agent Luna for recruitment, scheduling, leave and employee management, and Luna recommended termination of the employment relationship after a staff member had been late on several occasions. Such practices, although at the forefront of the experiment, mark the movement of the organizational identity of the digital workforce from “concept” to “system”.
There is a need to distinguish between digital employees and AZAI agent. Although AZAI agent is the core of the digital workforce, the latter is not an evolutionary relationship with the former. Here, it is necessary to distinguish:
Category IV: Robots. Physical implementer in intellectual form. It moves the synergistic boundary from “information flow” to “physical flow”. In manufacturing and supply chain scenarios, robots are AZAI agent's hands. In the context of human synergy, robots also include the RPA (automation of robotic processes), which is one of the core executive roles of the team with human beings, AZAI agent.
Category V: External Partners. Includes external API services, professional AI agent, outsourcing of the “digital workforce”. Gartner found in the supply chain study that 76 per cent of supply chain leaders had implemented or planned to implement functions to interact automatically with “mechanical customers” (i.e. non-human economies capable of independent negotiation and procurement) (Gartner, 2025). When the counterparty may also be AI, the “team boundary” must be redefined.
03 Capacity matrix — How does human capacity complement each other?
The bottom logic of human synergy is “complementarity of capabilities” rather than “capability substitution”. At the moment, the contents of publish, the mainstream research institutions, including Mussung, McKinsey and Microsoft, point to a common conclusion: AI has an overwhelming advantage in “widening, speed, consistency, scale”; and humans remain irreplaceable in “judgment, commonality, value trade-offs, non-conventional innovation”. Dr. Musheng has a more literary expression of this: no matter how technology evolves, high-level creativity, complex judgement and mutuality remain human beings and constitute a “no-go-go-go zone” that is difficult for AI to access (like a no-go-go-go-go-go-go zone) (Musheng, 2025).
A matrix of human capacity (described in four dimensions):
In a book entitled " AI agent ", Dr. Musheng analysed the value of a small number of human employees in such organizations, on the basis of which we can use the matrix. The conclusions of other research institutions confirm the need for such human collaboration. Microsoft World Trend Index's data: the first reason for employees to turn (and prefer) to work with AI rather than their colleagues this year is “24x7 usability” (42%), followed by “machine speed and quality” (30%), “unlimited creative ideas” (28%)— – It's just three things that “humans cannot provide” (MSC, 2025). At the same time, leaders see “quality control” (50 per cent) and “critical thinking” (46 per cent) as the most important human skills of the AI era (MS, 2026). The contrast between these two sets of data is the clearest illustration of the dividing line between human beings: machines are responsible for “doing much” and humans are responsible for “judging right”.
04 Human power ratio &token - Organizational measure of effectiveness
The “human ratio” (human members: AZAI agent) is becoming a new organizational measure. It is an intermediate variable for HR Efficacy, with a very obvious “leveraging effect” and the more human employees are able to handle more AI agent, the better it will be. However, this indicator is not easy to measure. In the Musheng advisory note, two measurement directions require attention:
One is that AZAI agent's “equivalent” is not uniform. An agent responsible for company-wide attendance and inspection, as well as an agent responsible for individual client offers, has a completely different workload and value. It is recommended that the measurement of “quantity” be translated into the measurement of “working hours” or “task volume”. In fact, “quantity” is far less meaningful than “hours of work” or “task ratio”.
The second is that AZAI agent has a different cost structure. The marginal labour costs of AIAI agent are low, but there are ongoing inputs for computing, modeling, data and transport.
This is the real basis for the “human-calibre rewrite” approach proposed by the Musheng Advisory: the denominator should be “total human-machine input” — labour costs plus AI-related calculus, models, data and transport inputs; the definition of human-effects should move towards “tokeen measurement” (Musheng Consulting, 2026). According to the Mu Sheng, the “human ratio” is a management indicator only when combined with the “total human input output ratio” and enterprises should work to improve the “human ratio” and “token effect”, which constitute the “human effect” of the AI agent organization, which, to be precise, is the human impact of the AI agent organization = human impact of *token.
The Mushen Consulting Institute predicts that, around 2028, the “human ratio” and the “token effect” will become the standard operating indicators for larger enterprises, and that some of the advanced enterprises will enter the regular board and business analysis sessions; however, the “human power ratio” will not have uniform industry markers, as its rational areas are highly dependent on the standardization of business scenes, and should be more customized for enterprises.
05 Governance structure — core of the new organizational compact
The most vulnerable, but critical, static component of the human rights team is the governance structure.
The reason is straightforward: when a non-human member of a team emerges who can act on its own, “who approves” becomes the starting point for all governance issues. HCLTech1 describes in the study a typical dilemma scenario: at 2:03 a.m., a key enterprise, workflow, was suddenly launched, there was no change in the deployment of engineers, no automation by analysts, but a series of decisions moved precisely between finance, client systems and cloud infrastructure; at dawn, the executives asked only one question — “Who approved?” The answer is, "No one, the system does it."
The real data confirm the governance gap: Strata Identity2 in 2026 showed that only 18 per cent of security managers believed that their identity and access management systems could effectively manage AZAI agent identity, while sample, which expressed “medium confidence” on the subject “without confidence or uncertainty”, was 35 per cent, 29 per cent and 18 per cent, respectively. Almost 80 per cent of businesses are unable to track AZAI agent activities in real time, i.e., what they are doing and who is responsible for them. Only 21 per cent of enterprises have AZAI agent lists (Strata Identity, 2026) that are accurately operating, while 32 per cent rely on non-real-time records, about 32 per cent plan to optimize within one year and 8 per cent have no list at all. In other words, the vast majority of businesses do not know how many AI employees they “hire” and how they actually operate, and the management of their crews is in a state of confusion.
Accordingly, the governance of AIAI agent should be guided by four principles:
First, independence. Each AIAI agent must have a first-class identity (first-class identity), may not borrow a human voucher or shared service account, and must bind an accountable human owner (HCLTech, 2026; Cisco Agency Initiative).
Second, the minimum mission level permission. permission should be granted on the basis of “what task this agent will do” rather than “which department has it”. A procurement agent who approves small purchases has access to the procurement process, vendor records and procurement policies, but should not have unrestricted access to the enterprise ' s financial system (Crucial Logics, 2026).
Thirdly, a distinction is made between “access to data” and “action”. An agent may need extensive access to client information to analyse complaints, but this does not mean that it can automatically send a response to the client. High-risk actions (refunds exceeding a certain amount, changes in production of permission, export of regulated data, external publish, execution transfer of funds) require human approval gates (AI CyberCheck, 2026).
Fourth, full life cycle audits. Includes the registration, certification, delegation of authority, rotation of vouchers, monitoring of sessions, audit logs and write-off of agents. Upon discharge from service, the documents, tokens and permission must be synchronised - “Closed workflow” does not amount to “recovered permission”.
06 Human crew — fully upgraded paper power
Depending on their characteristics and the principles of governance mentioned above, we can sum up the difference between them and the human team as follows:
The most noteworthy line in this table is the last line. When it comes to the empowerment mechanisms of modern enterprises (the “capable” mechanisms for employees, mainly through knowledge management), it is made clear that “businesses that are highly dependent on superheroes are by no means truly high-quality enterprises” and that enterprises should put together a closed-ring knowledge system of “knowledge extraction and knowledge-sharing”, in which ordinary people do extraordinary things and non-humans do extraordinary things (Musheng Consulting, 2025). The addition of AIAI agent, for the first time, gave the “bottom line” the possibility of engineering - which may be the most important contribution of the human team to organizational theory.
Comment:
1 HCLTech is a global science and technology company based in India and one of India ' s top IT service and consultancy giants. It is distinguished from the core label of an ordinary contractor, “Global Leadership in Engineering and Research and Development (E&R&D) Services”, which has been undergoing a major shift towards AI-driven technology services in recent years.
2 Strata Identity, a technical company focused on identity organization, focuses on the integration, modernization and security management of debrisized identity systems in mixed and cloudy environments. Founded in Bordeaux, Colorado, United States, in 2019, the enterprise was acquired in June 2026 by Rubrik, a data security company, and is currently a Rubrik company. Its CEO Eric Olden is one of the co-authors of the SAML standard. The company also led the creation of IDQL and open source projects to promote standardization of cloudy identity management.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.