AI Is Not Your Bottleneck. Your Organization Is.

AI 不是你的瓶颈，你的组织才是。

AI can now produce more work than many organizations can absorb. The next competitive advantage is not generation. It is throughput. AI has become extraordinarily good at producing work. It can write the memo, generate the code, draft the campaign, summarize the research, design the workflow, propose the experiment, and produce ten alternatives before a human team has finished its first meeting. AI 现在产出的工作量已经超过了许多组织所能消化的程度。下一个竞争优势不再是“生成”，而是“吞吐量”。AI 在产出工作成果方面已经变得极其出色。在人类团队完成第一次会议之前，它就能写好备忘录、生成代码、起草营销活动、总结研究报告、设计工作流程、提出实验方案，并产出十种备选方案。

And yet a strange thing keeps happening inside companies adopting AI: Output rises. Revenue does not rise with it. The usual response is to blame the model. Maybe the prompts need work. Maybe the company needs a stronger model, more context, another agent, or a better orchestration layer. Sometimes that is true. But increasingly, the model is not the bottleneck. The organization is. 然而，在采用 AI 的公司内部，一件奇怪的事情不断发生：产出增加了，但收入却没有随之增长。通常的反应是责怪模型。也许是提示词（prompt）需要改进，也许公司需要更强大的模型、更多的上下文、更多的智能体（agent）或更好的编排层。有时确实如此，但越来越多的时候，瓶颈不在模型，而在组织本身。

Before generative AI, production capacity was scarce. Research took time. Writing took time. Analysis, design, coding, coordination, and revision all consumed expensive human hours. Making any of those activities faster could create obvious value. AI changes that constraint. It makes many forms of production cheap and abundant. But abundance exposes everything downstream. 在生成式 AI 出现之前，生产能力是稀缺的。研究需要时间，写作需要时间，分析、设计、编码、协调和修订都消耗着昂贵的人力工时。让其中任何一项活动提速都能创造明显的价值。AI 改变了这一约束，它使多种形式的生产变得廉价且充裕。但这种充裕也暴露了下游的所有问题。

The draft still needs a decision. The decision still needs an owner. The owner may still need approval. The approved work still needs to be shipped. The shipped work still needs distribution. The distribution still needs measurement. The measurement still needs to change the next action. The operating equation has changed: AI Output × Organizational Throughput = Business Value. If organizational throughput is low, multiplying AI output produces surprisingly little economic value. 草稿仍需决策，决策仍需负责人，负责人可能仍需审批，获批的工作仍需发布，发布的工作仍需分发，分发仍需衡量，衡量结果仍需改变下一次行动。运营公式已经改变：AI 产出 × 组织吞吐量 = 商业价值。如果组织吞吐量很低，那么乘以 AI 的产出所带来的经济价值会少得惊人。

The bottleneck moved downstream. Most AI strategies focus on the left side of the system: Prompt → Model → Output. But businesses make money on the right side: Output → Decision → Execution → Evidence → Revenue. That distinction explains why a team can feel dramatically more productive while the economics barely move. The organization has built a faster factory feeding the same old loading dock. 瓶颈转移到了下游。大多数 AI 策略关注系统的左侧：提示词 → 模型 → 产出。但企业是在右侧赚钱的：产出 → 决策 → 执行 → 证据 → 收入。这种区别解释了为什么一个团队会感觉生产力大幅提升，而经济效益却几乎没有变化。组织建立了一个更快的工厂，却依然在向同一个老旧的装卸码头供货。

Ten reports arrive instead of one. Twenty campaign concepts appear instead of three. Hundreds of leads can be enriched. Dozens of product changes can be proposed. But if one manager must inspect everything, if publishing still requires five manual steps, if nobody owns the next action, or if analytics cannot connect execution to conversion, AI simply creates a larger queue. The faster generation becomes, the more visible that queue becomes. And eventually, the queue becomes the real product problem. 报告从一份变成了十份，营销概念从三个变成了二十个。数百条线索可以被丰富，数十项产品变更可以被提出。但如果必须由一位经理检查所有内容，如果发布仍需五个手动步骤，如果没人负责后续行动，或者分析无法将执行与转化联系起来，AI 只会制造出更长的排队队列。生成速度越快，队列就越明显。最终，队列本身成了真正的产品问题。

Faster AI makes organizational friction more expensive. Consider a slow approval process. When one employee produces one proposal per day, a two-day approval delay is annoying. When an agent system can produce fifty proposals per hour, the same approval process becomes catastrophic. The organization cannot consume what its machines can produce. This is the paradox of AI leverage: The faster generation becomes, the more expensive organizational friction becomes. AI 越快，组织摩擦的代价就越高。考虑一个缓慢的审批流程。当一名员工每天产出一份提案时，两天的审批延迟令人烦恼。当智能体系统每小时能产出五十份提案时，同样的审批流程就成了灾难。组织无法消化机器所能产出的东西。这就是 AI 杠杆的悖论：生成速度越快，组织摩擦的代价就越昂贵。

That means improving the model can actually make a poorly designed operating system feel worse. More intelligence enters the company, but the pathways that convert intelligence into action remain fixed. The result is not leverage. It is congestion. This is why adding another agent often disappoints. The company does not necessarily need another intelligence source. It may need a shorter path from intelligence to reality. 这意味着改进模型实际上可能让设计糟糕的操作系统感觉更糟。更多的智能进入了公司，但将智能转化为行动的路径却依然固定。结果不是杠杆效应，而是拥堵。这就是为什么增加另一个智能体往往令人失望。公司不一定需要另一个智能源，它可能需要的是从智能到现实的更短路径。

Throughput is the missing AI metric. Companies measure AI adoption with convenient numbers: hours saved; prompts run; assets generated; agents deployed; tasks automated. Those numbers describe production capacity. They do not necessarily describe business throughput. A more useful question is: How quickly can a useful machine-generated signal become a verified real-world result? 吞吐量是缺失的 AI 指标。公司用一些方便的数字来衡量 AI 的采用情况：节省的时间、运行的提示词、生成的资产、部署的智能体、自动化的任务。这些数字描述的是生产能力，并不一定描述商业吞吐量。一个更有用的问题是：一个有用的机器生成信号，能以多快的速度转化为经过验证的现实世界结果？

That journey might look like this: Signal → Analysis → Decision → Action → Evidence → Revenue. Every handoff introduces latency. Every unclear owner introduces waiting. Every unnecessary approval introduces friction. Every manual copy-and-paste step creates dependency. Every missing measurement point makes the organization less capable of learning from what it shipped. Throughput therefore depends on more than model speed. It depends on whether the organization can decide, execute, verify, and learn at approximately the same speed that AI can generate. Most cannot yet. 这个过程可能是这样的：信号 → 分析 → 决策 → 行动 → 证据 → 收入。每一次交接都会引入延迟，每一个不明确的负责人都会引入等待，每一个不必要的审批都会引入摩擦，每一个手动复制粘贴的步骤都会产生依赖，每一个缺失的衡量点都会使组织更难从已发布的内容中学习。因此，吞吐量不仅取决于模型速度，还取决于组织能否以与 AI 生成速度大致相当的速度进行决策、执行、验证和学习。目前大多数组织还做不到。

The unit of automation should be the closed loop. This changes what companies should automate. A weak automation looks like this: Request → Generate → Done. The system produced something, so the automation is considered successful. But nothing necessarily changed in the world. A stronger automation looks like this: Goal → Generate → Decide → Execute → Verify → Measure → Improve. The output is not the endpoint. It is an intermediate state. The workflow is complete only when the work reaches reality and the result can influence the next action. 自动化的单位应该是闭环。这改变了公司应该自动化的内容。一种薄弱的自动化是这样的：请求 → 生成 → 完成。系统产出了东西，所以自动化被认为是成功的。但现实世界中不一定发生了任何改变。一种更强大的自动化是这样的：目标 → 生成 → 决策 → 执行 → 验证 → 衡量 → 改进。产出不是终点，它是一个中间状态。只有当工作触及现实，且结果能够影响下一次行动时，工作流才算完成。

This is why the useful unit of AI automation is not the task. It is the closed loop. At minimum, an important AI workflow needs: A measurable objective. Not “use AI,” but increase qualified leads, reduce resolution time, improve checkout completion, or raise retention. A decision rule. What happens when the system produces an answer? Who or what decides GO, ITERATE, REDIRECT, or KILL? An execution path. The output must be able to change reality: publish, send, deploy, update, contact, deliver, or transact. Evidence. The action should leave proof: a public URL, commit, delivery receipt, event, lead, conversion, or payment. A feedback loop. Measurement must alter the next run. Otherwise automation is merely repetition. Without these pieces, an AI agent is often just a very fast worker placing documents on somebody else’s desk. 这就是为什么 AI 自动化的有用单位不是“任务”，而是“闭环”。至少，一个重要的 AI 工作流需要：一个可衡量的目标（不是“使用 AI”，而是增加合格线索、减少解决时间、提高结账完成率或提高留存率）；一个决策规则（当系统产生答案时会发生什么？谁或什么来决定执行、迭代、重定向或终止？）；一个执行路径（产出必须能够改变现实：发布、发送、部署、更新、联系、交付或交易）；证据（行动应留下证明：公共 URL、提交记录、交付收据、事件、线索、转化或付款）；一个反馈循环（衡量必须改变下一次运行，否则自动化仅仅是重复）。没有这些要素，AI 智能体往往只是一个把文件放在别人桌上的超快员工。

Human involvement should be intentional, not inherited. None of this means removing humans from every workflow. Judgment, accountability, taste, relationships, ambiguity, and irreversible risk can justify human gates. In many cases… 人类的参与应该是刻意的，而不是继承性的。这并不意味着要将人类从每一个工作流中移除。判断力、责任感、品味、人际关系、模糊性和不可逆转的风险都可以成为人类把关的理由。在许多情况下……