A New Interpretation of Organizational Operation in the Age of AI
By Geoffrey Chen
This essay is not merely a summary of AI agent development experience, nor is it simply a design note for a specific product. It emerges from the convergence of two lines of work that have occupied me in recent years: on the one hand, an ongoing inquiry into whether the subject remains a necessary foundation in epistemology; on the other, the practical development of AI agents, especially through systems like SmallClaw. As these two lines gradually came together, it became increasingly clear to me that what must be reconsidered in the age of AI is not only the structure of knowledge, but also the structure of organization. The reflections in this essay take shape at precisely that intersection of theoretical inquiry and engineering practice.
For a long time, both organizational theory and software design have shared the same implicit assumption. An organization is presumed to exist first in the human world. People are placed into positions, departments, reporting lines, and approval chains, and software arrives afterward to support communication, coordination, permissions, and record-keeping inside that already established structure. Even the most advanced collaboration systems of the last two decades have mostly dealt not with organization as such, but with its digital representation. Their task has been to help people collaborate, not to create an organization that genuinely runs inside the system itself.
The arrival of AI agents begins to unsettle that assumption at its root. The real question is no longer just whether software can become more intelligent or more autonomous. The deeper question is whether the structural core of organization must still remain outside the system, lodged in the human world, or whether it can now exist internally and be maintained by the system itself. Once that question is taken seriously, many concepts that once seemed stable begin to loosen, including user, role, authority, reporting line, approval path, and perhaps even organization itself.
This essay proposes a model of organizational operation that differs from the logic of traditional collaboration software. It does not begin from the problem of how multiple humans collaborate inside a tool. It begins instead from the idea that roles can become the formal subjects of action inside a system. In this model, roles carry responsibilities, projects define fields of operation, work items serve as the basic units of movement, and approval, routing, and control relations arise dynamically around specific items of work. Human beings still matter, and in many critical moments they continue to bear final responsibility. But they no longer automatically constitute the structural skeleton of the organization. They appear, when needed, as optional shadows attached to role execution. What may look at first like a product design choice is in fact closer to a shift in organizational ontology.
Organization no longer has to begin as a human structure
Traditional organizations have been human-centered not simply because AI did not yet exist, but because many of the core functions of organization could only be maintained by people. Information had to be compressed by managerial layers. Responsibility had to move through stable chains of command. Complexity had to be filtered by human supervisors. Oversight depended on durable reporting structures. Hierarchy became necessary not because it was metaphysically fundamental, but because human capability was limited. Attention is limited. Cognitive bandwidth is limited. Horizontal coordination capacity is limited. The ability to hold large-scale context over time is limited. Because of these limits, organizations had to pre-compress many relations into fixed structures in order to save on judgment, communication, and governance costs.
Seen in this light, hierarchy is first of all an energy-saving arrangement. By fixing reporting, approval, and authority in advance, it avoids the need to recompute control paths for every new situation. Who reports to whom, who approves what, who has final authority over which kind of matter — these relations are compressed into enduring positional order. This gives an organization stability, but it also gives it rigidity.
AI roles alter the underlying condition. As long as a system can maintain project state, task relations, historical context, and governance rules internally, many functions that once required stable human hierarchies no longer need to be carried by permanent chains of command. A role does not need to occupy a fixed place in a managerial ladder before it can meaningfully participate in control or judgment. It can enter a concrete situation and form the necessary approval, confirmation, and routing relations with other roles as the work requires, then allow those relations to dissolve once the matter is finished.
What this suggests is that, in the age of AI agents, organization no longer has to be defined primarily by who permanently outranks whom. It can instead be defined by what specific relations are required for a particular piece of work. This is not the disappearance of organization. It is a different kind of organization. Governance does not vanish. It is released from static hierarchy and returned to the flow of concrete work.
Roles become formal subjects, and work items become the center of relational generation
In such a model, roles are no longer just labels attached to people. They become the formal subjects of action inside the system. Plans are approved by an approval role. Payments are paused by a finance role. External communication is issued by a customer communication role. Projects provide the field in which roles interact, and work items become the units around which action and relation are organized.
This shift matters because it moves the center of organizational logic away from the question of which user has which permission, and toward the question of which role is required to perform which action in relation to which item of work. A permission-centered architecture gives way to a role-centered operational architecture.
Once that move is made, roles no longer need to stand in fixed hierarchical relations to one another. Sales, finance, legal, and project management may remain structurally parallel. They are not ordered by permanent superiority or inferiority. Instead, they form temporary control relations around particular matters. In one budget revision, finance may hold approval authority. In one contract review, legal may hold veto authority. In one crisis response, project management may hold pause authority. These relations are real, consequential, and binding, but they do not congeal into a permanent hierarchy built into the roles themselves.
This can be described as a new organizational logic in which roles are parallel while relations are dynamic. Roles are not designed as enduring superior and subordinate entities, yet around specific work items the system can generate temporary approval chains, confirmation paths, and routing structures. Hierarchy still appears, but it no longer inheres in the role itself. It inheres in the governance requirements of the work.
This structure may appear flatter than traditional organization, but it is not weaker. It may in fact be closer to the organizational essence of the AI age. In a world of persistent AI role execution, what becomes scarce is no longer the ability of middle management to permanently compress information. What becomes scarce is the ability of the system to generate the most appropriate governance structure for each concrete matter. Organization no longer needs to pre-freeze relations into a fixed ladder merely to conserve human energy. It can allow relations to arise when needed and disappear when no longer necessary.
Humans remain, but they no longer form the structural skeleton
Once roles become the formal subjects of action, the place of human beings inside the system has to be redefined as well. Traditional collaboration software treats human users as the natural first-class entities. Organizational structure is then built out of users, groups, departments, reporting lines, and permissions. If AI agents continue along that path, they easily collapse back into a familiar pattern — multi-user collaboration software with some added automation. That may increase participation, but the soul of organization remains outside the system. Internally, the system is still only a representation.
If organization is to become endogenous to the system, humans cannot continue to occupy the position of sole formal subject. They must be repositioned as participants in role execution rather than as the structural skeleton of the organization itself. More precisely, they should not enter the workflow as primary subjects. They should appear as optional shadows attached to the execution of roles.
The metaphor of a shadow does not imply insignificance. On the contrary, a human may bear final responsibility at critical moments, make decisive judgments, or step in where the consequences of an action require personal accountability. But even then, the formal statement inside the system should remain that the approval role approved the plan, not that a particular person approved it. The latter formulation drags the center of organization back into the human world. The former keeps the role in its place as the formal subject. Information about the human should remain available, but it should live in the layer of execution attribution and audit, not in the structural grammar of the workflow.
This implies that the system must distinguish between two layers of truth. The first is operational truth, which concerns how roles move projects and work items forward. The second is execution attribution, which concerns whether a given role action was performed autonomously by AI or actually carried out by a human through a takeover session. The first layer determines whether organization remains internal to the system. The second preserves responsibility, traceability, and interpretability.
From this follows an important consequence. The necessary subject of workflow should always be the role, not the person who has temporarily taken over that role. Waiting for approval should be represented as waiting for an action by the approval role, not as waiting for a specific person. A plan should not be considered approved by a named human at the workflow level, but by the approval role. Who exactly carried it out, through which channel, and under what takeover state matters greatly, but as part of the explanatory and audit layer rather than as a condition of workflow validity.
Takeover is not a repair function but a governance mechanism
If humans no longer appear as permanent structural actors, then the system needs a way for them to enter the organizational process when necessary. The easiest misunderstanding at this point is to treat takeover as a fallback for AI failure. That interpretation is too narrow.
Takeover is not fundamentally a repair mechanism. It is a governance mechanism by which the executing agency of a role can shift under controlled conditions. AI instability, contradictory output, or insufficient context may certainly trigger takeover, but those are only one class of cases. There are many situations in which a human should carry responsibility even when the AI is capable. These include high-risk approvals, formal commitments, legally sensitive actions, strategic choices, exceptional situations, and decisions whose accountability must ultimately rest with a human. In such cases, takeover occurs not because the AI cannot act, but because the role, in that particular matter, ought to be executed by a human.
Takeover therefore should not be modeled as a user entering the system and intervening. It should be modeled as a change in the execution mode of a role. This must occur at a lower organizational and runtime layer, not at the channel layer. Channels provide access paths. They should not define takeover itself. Feishu, Slack, WhatsApp, and WebChat are merely different surfaces through which human shadows can interact with the system. The logic of takeover belongs to the internal governance of roles.
More precisely, role execution should not be reduced to a binary between pure AI and pure human control. A more realistic model allows multiple modes of execution, such as AI-primary with human supervision, or human-primary with AI assistance. This prevents human intervention from severing the continuity of context, and it prevents AI from disappearing completely whenever a human appears. The role remains stable as the formal subject while its execution mode changes according to the needs of the work.
A single permanent human supervisor and optional human shadows
If humans are no longer to serve as the structural center of the system, then who occupies the position of final control? There must be one exception. A system that aspires to organizational integrity requires a single permanent human supervisor. This is not an ordinary user. It is the only stable human axis of control inside the system. It holds global visibility, final authority to grant or revoke access, and full audit access.
Apart from this supervisor, other humans are not permanent constituents of the system. They exist only as optional shadows attached to particular roles when needed. Roles do not require one-to-one human correspondence. In normal operation, most roles should not depend on a human being present. Only for specific projects, specific work items, and specific actions does the supervisor authorize an external identity to serve as the shadow executor of a role. What is granted is not general system usage by a person, but a restricted right for a given external identity to access a role through a designated channel, within a specified scope, at a specified level, and for a specified set of actions.
This transforms traditional user management into a narrower and more disciplined logic of shadow authorization. A shadow human does not need to log into the system core. They do not need to understand the overall internal structure. They should not be granted full system identity. They interact only through an authorized channel, receive only the matters related to the role they are permitted to shadow, and can perform only those actions explicitly allowed. The supervisor uses the system itself. Shadows use only channels. This separation is essential if the introduction of multiple humans is not to pull the model back toward ordinary multi-user collaboration software.
From static hierarchy to work-generated relations
If the argument so far is accepted, then the flatness of this organizational model should not be regarded as a temporary limitation. It should be recognized as a more fundamental characteristic of organization in the age of AI. Parallel roles do not imply the absence of governance. The absence of fixed hierarchy does not imply the absence of control. What changes is that control no longer has to sediment into permanent rank relations between roles. It can instead be generated dynamically around projects and work items.
This marks an important theoretical shift. Traditional organization attaches hierarchy in advance to roles and positions in order to compress human governance costs. AI-era organizational systems can leave roles parallel and allow approval, confirmation, veto, and routing relations to be generated around concrete matters of work. Hierarchy no longer serves first to determine who is permanently above whom. It serves to determine what governance structure is required by this particular item of work at this particular time. Humans rely on static hierarchy to conserve energy. AI relies on dynamic relations to operate. That may be the deepest distinction between the two organizational forms.
What follows from this is that the truly important problem is no longer how to map conventional organizational structures into AI systems. The deeper problem is how to allow systems to generate governance relations that are sustainable, auditable, retractable, and reconfigurable from within. Organization is no longer an external human structure supported by software. The system itself becomes the environment within which organization runs. The significance of a system like SmallClaw is therefore not merely that it is a product. Its significance lies in the fact that it constitutes an implemented model of AI-native organizational operation.
Conclusion
The central claim of this essay is not that we need a new kind of software feature. It is that, once AI agents can persistently execute roles, organization no longer has to treat human users and fixed hierarchy as its unquestioned foundation. Organization can exist first inside the system. Roles can become formal subjects. Humans can recede into the status of optional shadow executors. Global control can be concentrated in a single permanent supervisor. Approval and control relations can be generated dynamically around specific projects and work items. Traditional hierarchy is not the essence of organization. It is an energy-saving compression structure shaped by human limitation. In the age of AI, there is no reason to assume that this static structure must continue to serve as the default condition of organization.
The organizational model of the AI era should therefore not be understood as conventional collaboration software with stronger automation. Nor should it be understood as a more intelligent digital mirror of human institutions. A more radical possibility is that the structural core of organization itself becomes endogenous to software, and that software ceases merely to support organization and begins to serve as the place where organization actually exists and runs. That change is still unfolding, but it is already sufficient to define a new problem for technical theory and to force a reconsideration of organization, role, responsibility, hierarchy, and the place of the human.
从角色等级到工作生成关系——AI时代组织运行模型的一种新解释
本文并不只是一次 AI Agent 开发经验的总结，也不只是一个具体产品的设计说明。它来自我近来两方面工作的交汇——一方面是对认识论中主体必要性的持续思考，另一方面是对 AI Agent 尤其是 SmallClaw 的实际开发。在这两条线索逐渐靠近的过程中，我开始意识到，AI 时代需要被重新思考的，不只是知识的结构，也包括组织的结构。本文中的这些判断，正是在这种理论思考与工程实践的交汇中形成的。
By Geoffrey Chen
长期以来，无论是管理学、组织理论，还是软件系统设计，对“组织”这一对象的理解都默认建立在一个前提上，也就是组织首先存在于真人世界之中。人先被配置在职位、部门和汇报链条里，之后软件才进入这个已经成形的结构，承担信息传递、流程记录、权限控制和协作支持的功能。即使是过去二十年中最先进的协作软件，真正处理的也大多不是组织本身，而是组织在系统中的映射。系统的任务，是帮助人协作，而不是在系统内部形成一种真正独立的组织运行结构。
AI agent 的出现，使这个前提第一次发生了根本动摇。问题不再只是软件是否更聪明，自动化是否更强，而是组织的主骨架是否仍然必须外置于真人世界之中。换句话说，AI时代真正值得重新提出的问题不是“AI如何帮助组织”，而是“组织是否可以首先存在于系统内部，并由系统自身维持运行”。一旦这个问题被认真对待，很多传统上被视为理所当然的概念都会失去稳定性，包括用户、岗位、权限、上下级、审批链，甚至组织本身的定义。
本文试图提出一种不同于传统协作软件逻辑的组织运行模型。它不从“多个人如何在系统中协作”出发，而从“角色如何作为系统内部正式主体运行”出发。其基本判断是，在AI时代，组织的第一性单位不再必须是人，而可以是系统内部的角色。角色承担职责，项目提供工作域，work item 构成推进单元，审批、流转和控制关系则围绕具体事项动态生成。真人仍然重要，甚至在许多关键节点上仍然承担最终责任，但真人不再天然构成组织主骨架，而只是作为角色执行的可选影子被系统吸纳进来。这种变化看起来像是产品设计上的差异，实际上更接近一种组织论上的转向。
组织不再首先是人的结构，而首先是系统中的运行结构
传统组织之所以必须以人为中心，不只是因为过去没有 AI，而是因为组织的很多基本功能只能由人来维持。信息要靠中间层压缩，责任要靠职位链条传递，复杂决策要靠管理者逐层过滤，监督也要通过长期稳定的等级关系来完成。层级之所以存在，并不主要因为它具有某种形而上的必然性，而是因为真人有限。人的注意力有限，认知带宽有限，横向协调能力有限，持续掌握大规模上下文的能力也有限。正因为如此，组织才必须把许多关系预先固化成稳定层级，以节省判断、沟通和管理的能耗。
从这个意义上说，传统 hierarchy 首先是一种节能结构。它通过长期固定的上下级关系，避免每一件事项都重新计算控制路径、审批路径和责任路径。谁向谁汇报，谁批准谁，谁拥有某类事项的终审权，这些关系被预先压缩进岗位和职位秩序之中。组织由此获得稳定性，但也同时获得了僵硬性。
AI角色的出现改变了这个基础条件。只要系统内部能够持续保存项目状态、任务关系、历史上下文和控制规则，很多原本依靠真人层级承担的压缩功能，就不再必须通过长期稳定的等级结构来完成。某个角色不需要先在一个固定的上下级链条中占位，才有能力参与控制和判断。它可以在具体事项到来时，围绕该事项与其他角色形成临时的审批关系、确认关系和流转关系，并在事项结束后让这种关系自然消散。
这意味着，AI时代的组织不必首先表现为“谁长期高于谁”，而可以首先表现为“当前这项工作需要哪些角色之间生成何种控制关系”。这不是无组织，而是另一种组织。它不是取消治理，而是将治理从静态等级结构中释放出来，使其回到具体工作事项之中。
角色成为正式主体，工作项成为关系生成中心
在这种模型中，角色不再只是某个人的系统标签，而是系统内部的正式行为主体。批准是由审批角色完成的，暂停是由项目管理角色完成的，对外回复是由客户沟通角色完成的。项目提供了角色协同的语境，work item 则成为真正触发关系生成的最小单位。
这一步非常关键，因为它把组织的中心从“人拥有权限”转移到了“角色承担动作”。在传统软件里，系统首先问的是哪个用户拥有什么权限；在这种新模型里，系统首先问的是哪一个角色面对哪一个事项需要完成什么动作。用户中心的权限逻辑，开始让位于角色中心的运行逻辑。
由此产生的直接结果是，角色之间并不需要天然形成固定等级。销售、财务、法务、项目管理，可以在结构上保持平行。它们不是通过稳定的上下级关系彼此约束，而是通过 project 和 work item 的推进，在具体事项上临时生成不同的控制关系。某一次预算审批中，财务角色拥有批准权；某一次合同修订中，法务角色拥有否决权；某一次外部危机处理中，项目管理角色拥有暂停权。这些关系都真实存在，也都具有治理效力，但它们并不沉淀为角色本体上的永久等级秩序。
这可以被概括为一种新的组织逻辑，也就是角色平行而关系动态。角色本身并不被预设为稳定的上下级实体，但围绕具体工作事项，系统会即时生成所需的审批链、确认链和流转链。等级依然存在，但不再附着于角色本身，而附着于具体工作的治理要求。传统 hierarchy 是角色层面的静态秩序，这里出现的是 work item 层面的动态秩序。
这种结构看起来比传统组织扁平，但它并不比传统组织弱。相反，它可能更接近AI时代组织的本质。因为在 AI agent 体系中，真正稀缺的已不再是中间层对信息的长期压缩，而是如何让系统在具体事项上生成最精确的关系结构。组织不再需要通过预先固化所有关系来节省真人能耗，而可以通过系统内部对 project、role 和 work item 的精确维护，让关系在需要时出现，在不需要时消失。
真人仍然存在，但不再构成组织的主骨架
一旦角色成为正式主体，真人在系统中的地位也必须被重新定义。传统协作软件天然把真人用户当作一等对象，组织结构由真人用户、群组、部门、汇报链和权限关系构成。AI agent 如果沿着这条路继续推进，很容易退化成“多人协作软件外加一些 AI 功能”。这样做或许可以快速扩展用户参与，但组织的灵魂仍然留在系统之外，系统内部始终只是一层映射。
如果组织要内生于系统，就不能让真人继续作为唯一正式主体。真人必须被重新定位为角色执行的参与者，而不是组织骨架本身。更准确地说，真人不再直接作为工作流主语存在，而只是作为某个角色的实际执行影子被吸纳进来。
“影子”这个说法并不意味着真人不重要。恰恰相反，真人可以在某些关键节点上承担必须由人承担的责任，比如付款批准、对外正式承诺、法律敏感决策、策略方向修正等。但即便如此，系统里的正式表达仍应是“审批角色批准了计划”，而不是“某个具体的人批准了计划”。后者会把组织重心重新拖回真人世界，前者则保持了角色作为正式主体的地位。真人的信息应被保留，但应留在执行归因与审计层，而不应进入工作流主骨架。
由此，系统内部便自然分成两层真相。第一层是运行真相，也就是角色如何推动项目和工作项前进。第二层是执行归因，也就是这次角色行为究竟是由 AI 自动完成，还是由某个真人通过接管会话实际执行。前者决定组织结构是否仍然内生于系统，后者保证责任、可追溯性和解释性。
这种分层带来一个重要推论，也就是工作流的必要信息始终应是角色，而不是接管者。等待审批，不应写成等待某个人，而应写成等待审批角色动作。某次计划获批，不应从流程上表达为某位真人批准，而应表达为审批角色完成批准。接管者是谁，在哪个 channel 上完成，是否处于接管状态，这些都应成为审计层和执行记录中的重要信息，但不应反过来成为流程成立的条件。
接管不是故障补救，而是组织治理机制
一旦真人不再作为系统骨架中的常驻主体出现，如何让真人在必要时进入组织运行之中，就变成一个新的问题。这里最容易出现的误解，是把接管理解为 AI 出错后的人工兜底。事实上，这种理解过于狭窄。
接管的本质不是故障处理，而是角色执行权在不同执行者之间受控切换的一种治理机制。AI 输出不稳定、上下文理解不足、判断冲突严重，当然会触发接管，但这只是其中一种情况。还有许多节点，即便 AI 具备相当强的处理能力，也依然应当由真人承担责任，比如高风险审批、策略性选择、例外情况裁决、需要明确责任归属的控制动作等。也就是说，接管不是因为 AI 不会，而是因为某个角色在这个具体事项上此刻应当由真人来承担责任或做出判断。
因此，接管不应建模为“某个用户进入系统干预”，而应建模为“某个角色的当前执行方式发生变化”。这件事必须发生在更底层的组织与角色运行层，而不是发生在 channel 层。channel 只提供接入路径，接管本身则是内部组织能力。Feishu、Slack、WhatsApp、WebChat 只是把真人影子带进系统的不同入口，它们不应定义接管逻辑本身。
在更精确的设计上，角色并不只有两种状态，也就是纯 AI 与纯真人。更合理的情况是允许多种执行模式并存，例如 AI 主执行、真人监督，或者真人主执行、AI 辅助。这样一来，真人的进入不会切断 AI 对上下文的持续维护，AI 也不会因为真人出现而完全退出。角色作为正式主体保持稳定，执行者则根据工作需要发生变化。
唯一常驻真人 supervisor 与可选影子真人
如果真人不应普遍进入系统，那么谁来承担全局控制中心？这里必须有一个例外。一个组织运行系统若要保持统一控制与最终可撤销性，就需要存在一个唯一常驻的真人 supervisor。这个位置不是普通用户，而是系统内唯一稳定的人类控制轴心。它拥有全局视野、最终授权权、最终撤销权以及完整审计访问权。
除这个 supervisor 外，其他真人都不是系统的常驻组成部分。他们只是可选影子，按需附着在某个角色上。角色默认不需要与真人一一对应，绝大多数时间里也不应依赖真人在场。只有在特定项目、特定工作项、特定动作上，supervisor 才会把某个外部身份授权为某个角色的影子执行者。被授权的对象不是“这个人能不能使用系统”，而是“这个外部身份能否在特定范围内、通过特定 channel、以特定级别接入某个角色，并执行哪些动作”。
由此，传统用户管理逻辑被转换成一种更克制的影子授权逻辑。影子真人不需要登录系统本体，不需要理解系统全局结构，也不应被赋予系统级身份。他只通过被指定的 channel 与系统互动，只接收与自己被授权角色相关的事项，只能执行被允许的有限动作。Supervisor 使用 SmallClaw，影子只使用 channel。这种分工非常关键，因为它确保多真人的引入不会把系统拖回传统多人协作软件的逻辑。
从静态等级到工作生成关系
如果以上判断成立，那么目前这种看起来扁平的组织模型，就不应再被视为一种未完成的阶段，而应被理解为 AI时代组织的一个更本质的特征。角色平行，并不意味着没有治理；没有固定 hierarchy，也不意味着没有控制关系。真正发生变化的是，控制关系不再必须沉淀为角色之间的长期等级秩序，而可以围绕具体 project 与 work item 动态生成。
这可以被视为一种重要的理论转向。传统组织把等级预先附着于角色与职位之上，借此压缩真人治理成本。AI时代的系统组织则可以让角色保持平行，把审批、确认、否决、流转等关系交给具体事项即时生成。等级不再首先服务于“谁长期高于谁”，而服务于“这项工作此刻需要什么样的治理结构”。真人靠静态等级节能，AI 靠动态关系运行。这也许正是两种时代的组织形式之间最深的差别。
如果这个判断成立，那么未来真正值得研究的问题就不再是如何把传统组织结构映射进 AI 系统，而是如何让系统在内部形成可持续、可审计、可收回、可重组的动态治理关系。组织不再是外部人事结构加上一套软件支持，而是系统本身作为组织运行环境的事实。SmallClaw 一类系统的意义，也不再只是产品层面的创新，而是提供了一个已经落地的、可以被分析和推广的 AI 时代组织运行模型。
结语
这篇文章真正想提出的，不是一个新软件如何设计，而是一个更基础的问题——在 AI agent 可以持续承担角色执行的条件下，组织是否仍然必须以真人用户和稳定层级作为第一前提。本文给出的答案是否定的。组织可以首先存在于系统内部，角色可以成为正式主体，真人可以退居为影子执行者，监督中心可以收束为唯一常驻的 supervisor，审批和控制关系则围绕具体项目与工作项动态生成。传统 hierarchy 并非组织的本质，而是真人局限下的一种节能压缩结构。AI时代的组织，未必需要继续背负这种静态结构作为默认前提。
因此，AI时代的组织运行模型，不应被理解为传统多人协作软件加上更强的自动化，也不应被理解为把人类组织原样搬进一个更聪明的系统之中。更可能的方向是，组织的主骨架本身开始内生于软件系统，而软件第一次不仅支持组织，而且成为组织本身的运行地点。这样的变化还远未结束，但它已经足以构成一个新的技术理论问题，并要求我们重新思考组织、角色、责任、层级和人的位置。