# Everyone wants to be AI-pilled. Most Companies Are Still Level 1
**作者**: Ann Miura-Ko
**日期**: 2026-05-01T14:46:13.000Z
**来源**: [https://x.com/annimaniac/status/2050225284277026990](https://x.com/annimaniac/status/2050225284277026990)
---
Over the last few weeks I've expanded our office visits beyond AI-pilled startups to scaled companies — most recently Ramp, a 1,500-person organization. The earlier visits showed me what AI-native looks like at 8 or 50 people. At a tiny startup, it is easy to say the company is AI-native because the founders are. Everyone sits close to the customer. Everyone builds. Everyone experiments. The operating system is mostly the people. At a scaling company, the bar is much higher. AI can no longer be a personality trait of the founding team. It has to become part of the company’s DNA.
Questions around what is truly AI-native reminds me of the debates we used to have about the levels of autonomy in AVs. For years, everyone in AV was chasing Level 5 self-driving. The levels mattered because they forced precision. Cruise control was not autonomy. Lane keeping was not autonomy. Driver assistance was not the same thing as self-driving.
Something similar is happening with AI-pilled organizations.
Right now, “AI-pilled” is being used as though it were binary. You either are or you aren’t. In practice, companies differ both in intensity (how deeply AI is embedded into daily work across the organization) and in technical capability (what AI is actually allowed to see, do, and change).
A company where employees use ChatGPT to summarize meetings is not in the same category as a company where agents can query systems of record, take bounded action, propagate workflows across teams, and improve the way future work gets done. Both may describe themselves as AI-forward. They are not operating at the same level.
So the better question is not: Is this company AI-pilled? The better question is: What level of autonomy has the organization actually achieved? To put a finer point on it:
- What can AI see? Is the work of your company legible to a machine or does it live in someone’s mind, undocumented meetings, and SaaS tools the AI can't read?
- What can AI do? Can it act on systems of record (e.g. open PRs, update CRMs, reconcile invoices) or can it only summarize what humans already wrote down?
- Who can extend the system? Are non-engineers shipping production internal tools, or is every workflow held together by a few power users whose work walks out the door when they leave?
- How has the organization changed? Or are you running 2023's org chart with better autocomplete?
The answers cluster into six levels.
L0: AI as theater
- What can AI see? Nothing structured. Knowledge lives in people's heads, undocumented meetings, and SaaS tools AI can't read.
- What can AI do? Nothing of consequence. Maybe summarize a meeting if a human pastes the transcript.
- Who can extend the system? No one. AI is a personal tool, ungoverned, unintegrated.
- How has the org changed? It hasn't. Same chart, same hiring plan, same handoffs, same dependence on managers as routers.
Hard test: Can AI complete any recurring business process end-to-end?
Common false positive: A CEO who gives an excellent speech about AI transformation while still running the company through the same executive staff meetings, status updates, reporting lines, and headcount plans. Announcements ≠ adoption.
L1 — Personal productivity
- What can AI see? Each individual's personal AI sees only what that person feeds it. Saved prompts, scratch files, private knowledge bases. No org-level visibility.
- What can AI do? Help individuals draft, summarize, brainstorm, code. No action on systems of record.
- Who can extend the system? Each user reinvents independently. Power users are heroes; their workflows leave with them.
- How has the org changed? It hasn't. Same chart. Maybe a "Head of AI" hire that has budgetary influence and has purchased some AI products for the company
Hard test: If your best AI user left tomorrow, would their workflow remain in the company?
Common false positive: "80% of employees use AI weekly!" which is probably true and also meaningless.
L2 — Team workflow
- What can AI see? Teams have shared context like a claude.md per team, shared prompts, function-specific MCP integrations. AI sees within team boundaries.
- What can AI do? Functional workflows. AI for sales prospecting, support tier-1 triage, eng code review. Bounded actions within a team's domain.
- Who can extend the system? Within a team, non-engineers can tap into shared workflows. Across teams, not really. Each function rebuilds the same thing privately.
- How has the org changed? Functional efficiency within roles. A CSM with AI handles 200 accounts vs. 50. Hiring slows but org shape unchanged. Role boundaries intact.
Hard test: Does this workflow cross team boundaries, or is every function building its own private AI stack?
Common false positive: "We have AI workflows in every department." But the workflows don't connect, so the company is a collection of AI-enhanced silos rather than an AI-native organization.
L3 — Organizational infrastructure
- What can AI see? The whole organization is queryable. Cross-functional context accessible. Core systems of record exposed via CLI / MCP / well-defined APIs and integrated into a view on which agents can act and not just observe.
- What can AI do? Agents act across systems. They update CRMs, open PRs, route tickets, run analyses, draft customer communications, reconcile invoices. Cross-functional but still bounded.
- Who can extend the system? Non-engineers don't just consume shared skills — they author them. Sales rep packages call analysis as a shareable skill. CX engineer packages a ticket investigation pattern. Skills move horizontally across functions.
- How has the org changed? The org chart looks materially different from a 2023 equivalent. Specific shape varies — zero-PM teams, PM-as-agent-orchestrator and product curator, or pure role convergence into "builders". The unifying signal: the company has made an explicit structural choice about how AI changes who does what, and the choice is visible. Token-maxing over headcount-maxing running uncomfortable API bills.
Hard test: Can an agent answer, across systems: what shipped last sprint, who asked for it, what broke after launch, what customers said, and what the company should do next — without convening a cross-functional meeting?
Common false positive: A landfill of meeting transcripts and dashboards with no synthesis. Capture is not legibility. An inert archive is not an operating system.
L4 — Compounding operating system
- What can AI see? Not just what happens but the relationships between what happens. The system maintains its own context so that agents update agents, skills marketplaces propagate wins and removes duplicate efforts, the system learns what to surface. Capture + synthesis + query are continuous.
- What can AI do? Agents have policy-driven decision authority within scoped domains. Security agents detect then validate then fix then open PR with human review at the merge step. Custom internal harnesses purpose-built for the work the company does most. Active removal of software blockers that prevent agents from being useful.
- Who can extend the system? Non-engineers ship production internal tools. A finance person builds an automated contract reviewer. An AE shipped a sales tool in under an hour. None of them are engineers. They didn't file a ticket. They found their own pain, prototyped a fix, and pulled engineering in only when it was time to go to production.
- How has the org changed? Hierarchy collapses toward "channel managers" of agent workflows. New archetypes emerge. Compensation/promotion explicitly tied to AI proficiency. Customer signal-to-ship measured in hours.
Hard test: Show me a workflow that got better because the system learned from prior runs, not because one heroic person manually improved it. Plus: show me three production tools shipped by non-engineers in the last quarter.
Common false positive: Agent sprawl. A hundred brittle automations don't equal a compounding operating system. L4 requires managed compounding (lifecycle, observability, evaluation), not chaotic proliferation. Without compaction discipline, the factory clogs.
L5 — Virtually self-driving organization
A clean operational definition (with the caveat that I realize L5 does not exist yet so I’m describing what I think it might look like): an L5 organization is one where the core operating loops can sense reality, diagnose issues, initiate work, execute within delegated authority, update shared memory, and improve future behavior — with humans governing strategy, taste, risk, values, and exceptions rather than running the loops themselves.
The six L5 markers:
1. The system notices something important without being asked.
2. The system synthesizes across multiple sources of context.
3. The system decides whether action is warranted.
4. The system acts within delegated authority.
5. The system escalates when uncertainty or consequence exceeds its authority.
6. The system updates shared memory so future behavior improves.
Through the four-question lens:
- What can AI see? Generative — the system asks its own questions, identifies gaps in its own knowledge, proposes investigations and runs them.
- What can AI do? Delegated authority for novel decisions, not just configured policies. The L4→L5 leap: at L4, the system improves because humans direct it to. At L5, because it notices it should.
- Who can extend the system? Non-engineers contribute directly to the customer-facing product itself, or the product is reshaped so anyone can extend it without writing code. The boundary between "internal tool" and "product feature" dissolves.
- How has the org changed? Truly fluid. Agents are organizational members with meaningful delegated authority. The org self-modifies or proposes role changes, team boundary shifts. Onboarding becomes system-driven. Institutional knowledge survives transitions perfectly because it lives in the system, not in any individual.
Hard test: What important thing did the company notice, decide, act on, and learn from recently without a human initiating the process? Not a threshold alert; not a configured automation; not an agent summarizing what people surfaced. Something the system synthesized that humans hadn't framed as a question yet.
Common false positive: The "fake autonomy" pattern. The company claims self-driving behavior, but the system is only executing preconfigured rules or surfacing threshold-based alerts. Humans are still doing all the noticing. Distinguishing real generative behavior from glorified observability is the open challenge at this level.
Interestingly, a company rarely answers all four questions at the same level but the asymmetry tells you where the next intervention should focus. Sometimes AI sees a lot but can't do much. AI might do a lot but only engineers can really extend it. The org chart could have changed but the substrate is thin.
Steve Blank once said that a startup is not a small version of a large company. Similarly an AI-pilled company is not simply an AI-assisted version of an old company. They are organizations rebuilt around a new operating model. We are still learning what this looks like but those who are curious and engaged will have a compounding advantage as they make their way quickly up the stack and see real world impact in their business operations and hopefully margins.
If you’re learning to build a company in this way I would love to speak with you!
## 相关链接
- [Ann Miura-Ko](https://x.com/annimaniac)
- [@annimaniac](https://x.com/annimaniac)
- [166K](https://x.com/annimaniac/status/2050225284277026990/analytics)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [10:46 PM · May 1, 2026](https://x.com/annimaniac/status/2050225284277026990)
- [166.5K Views](https://x.com/annimaniac/status/2050225284277026990/analytics)
- [View quotes](https://x.com/annimaniac/status/2050225284277026990/quotes)
---
*导出时间: 2026/5/2 12:01:15*
---
## 中文翻译
# 人人都想被“AI点化”，但大多数公司仍处于第1级
**作者**: Ann Miura-Ko
**日期**: 2026-05-01T14:46:13.000Z
**来源**: [https://x.com/annimaniac/status/2050225284277026990](https://x.com/annimaniac/status/2050225284277026990)
---
在过去的几周里，我拓宽了我们办公室拜访的范围，从那些已经“被AI点化（AI-pilled）”的初创公司扩展到了规模化企业——最近一次是拜访了拥有 1500 名员工的 Ramp。之前的拜访让我看到了 8 人或 50 人规模时，“AI原生”是什么样子的。在一个微型的初创公司，很容易说这家公司是 AI 原生的，因为创始人们就是。每个人都紧贴着客户。每个人都参与构建。每个人都在做实验。那时的“操作系统”基本上就是人。但在一家规模化的公司，门槛要高得多。AI 不能再仅仅是创始团队的个性特质，它必须成为公司 DNA 的一部分。
关于什么才是真正的“AI原生”的问题，让我想起了过去我们关于自动驾驶（AV）自动化级别的辩论。多年来，自动驾驶领域的每个人都在追逐 L5 级自动驾驶。这些级别很重要，因为它们强制要求精确。巡航控制不是自动驾驶。车道保持不是自动驾驶。驾驶员辅助与自动驾驶不是一回事。
类似的事情也正在发生在那些“被AI点化”的组织身上。
目前，“被AI点化”这个词的使用似乎是非黑即白的。你要么是，要么不是。但在实践中，公司在强度（AI 在组织日常工作中嵌入的深度）和技术能力（AI 实际上被允许看什么、做什么、改变什么）两方面都存在差异。
一家员工只是用 ChatGPT 来做会议纪要的公司，与一家智能体可以查询记录系统、采取有界行动、跨团队传递工作流并改进未来工作方式的公司，完全不在一个类别。这两者都可能自称是“AI 领先”的。但它们的运作水平并不相同。
所以，更好的问题不是：这家公司是否被“AI点化”了？更好的问题是：该组织实际上实现了什么级别的自主权？具体来说：
- AI 能看到什么？公司的工作对机器来说是清晰可读的，还是存在于某人的脑海中、未记录的会议中，以及 AI 无法读取的 SaaS 工具中？
- AI 能做什么？它能否作用于记录系统（例如打开 PR、更新 CRM、对账发票），还是只能总结人类已经写下来的内容？
- 谁可以扩展系统？是非工程师在发布生产级内部工具，还是每个工作流都由少数几个“高级用户”维系，而一旦他们离职，这些工作也就随之带走了？
- 组织发生了怎样的变化？还是说你们只是拿着 2023 年的组织结构图，用上了更好的自动补全功能？
这些答案汇聚为六个级别。
**L0：AI 作秀**
- **AI 能看到什么？** 没有任何结构化内容。知识存在于人们的头脑中、未记录的会议中，以及 AI 无法读取的 SaaS 工具中。
- **AI 能做什么？** 做不了任何有实质意义的事。也许如果人工粘贴了文字记录，它能总结一下会议。
- **谁可以扩展系统？** 没人。AI 是一种个人工具，无治理、未集成。
- **组织发生了怎样的变化？** 没有变化。同样的架构图，同样的招聘计划，同样的交接流程，同样依赖管理者作为路由器。
**严苛测试：** AI 能否端到端完成任何周期性的业务流程？
**常见的假阳性：** 一位 CEO 发表了关于 AI 转型的精彩演讲，但仍然通过同样的高管会议、状态更新、汇报线和人数计划来运营公司。发布公告 ≠ 采用。
**L1 — 个人生产力**
- **AI 能看到什么？** 每个人私有的 AI 只能看到那个人喂给它的内容。保存的提示词、草稿文件、私人知识库。没有组织级别的可见性。
- **AI 能做什么？** 帮助个人起草、总结、头脑风暴、写代码。不能对记录系统采取行动。
- **谁可以扩展系统？** 每个用户都在独立重复造轮子。高级用户是英雄，但他们的工作流会随他们一起离开。
- **组织发生了怎样的变化？** 没有变化。同样的架构图。也许新招了一个“AI 负责人”，拥有预算影响力，并为公司购买了一些 AI 产品。
**严苛测试：** 如果你最好的 AI 用户明天离职，他/她的工作流是否会留存在公司？
**常见的假阳性：** “80% 的员工每周都在使用 AI！”这可能是真的，但也毫无意义。
**L2 — 团队工作流**
- **AI 能看到什么？** 团队拥有共享的上下文，比如每个团队的 claude.md 文件、共享的提示词、特定功能的 MCP 集成。AI 只能看到团队边界内的内容。
- **AI 能做什么？** 功能性的工作流。用于销售线索挖掘、支持一级分诊、工程代码审查的 AI。在团队领域内的有界行动。
- **谁可以扩展系统？** 在团队内部，非工程师可以利用共享的工作流。但在跨团队层面，并非如此。每个职能都在私下重建同样的东西。
- **组织发生了怎样的变化？** 角色内的职能效率提升。一个使用 AI 的 CSM（客户成功经理）能处理 200 个账户，而不是 50 个。招聘放缓，但组织形态未变。角色边界依然存在。
**严苛测试：** 这一工作流是否跨越了团队边界，还是每个职能部门都在构建自己的私有 AI 栈？
**常见的假阳性：** “我们在每个部门都有 AI 工作流。”但这些工作流互不连通，所以公司只是一堆 AI 增强的孤岛，而非一个 AI 原生组织。
**L3 — 组织基础设施**
- **AI 能看到什么？** 整个组织都是可查询的。跨职能上下文可访问。核心记录系统通过 CLI / MCP / 定义良好的 API 暴露出来，并集成到一个视图上，智能体可以在此视图上采取行动，而不仅仅是观察。
- **AI 能做什么？** 智能体跨系统行动。它们更新 CRM、打开 PR、路由工单、运行分析、起草客户沟通、对账发票。跨职能但仍然是有界的。
- **谁可以扩展系统？** 非工程师不仅仅是消费共享技能——他们还在创造技能。销售代表将通话分析打包为可共享的技能。CX（客户体验）工程师将工单调查模式打包。技能在职能间横向移动。
- **组织发生了怎样的变化？** 组织架构图与 2023 年的同类公司相比有实质性不同。具体形态各异——零 PM 团队、PM 作为智能体编排者和产品策划者，或者是纯粹的角色融合为“构建者”。统一的信号是：公司已经做出了明确的结构性选择，决定 AI 如何改变谁做什么，而且这一选择是可见的。追求 Token 效率最大化，而非人数最大化，背负着令人不安的 API 账单。
**严苛测试：** 一个智能体能否在不召集跨职能会议的情况下，跨系统回答：上个冲刺发布了什么、谁要求的、发布后坏了什么、客户说了什么、公司下一步该做什么？
**常见的假阳性：** 会议记录和仪表板的垃圾堆，没有任何综合。抓取 ≠ 可读性。一个惰性的档案不是操作系统。
**L4 — 复利型操作系统**
- **AI 能看到什么？** 不仅仅是发生了什么，还有事情之间的关系。系统维护自己的上下文，以便智能体更新智能体、技能市场传播胜利并去除重复工作、系统学习应该呈现什么。抓取 + 综合 + 查询是连续不断的。
- **AI 能做什么？** 智能体在限定领域内拥有策略驱动的决策权。安全智能体检测然后验证然后修复，然后在合并步骤打开 PR 等待人工审查。针对公司最常做的工作定制的内部线束。主动移除阻碍智能体发挥作用的软件障碍。
- **谁可以扩展系统？** 非工程师发布生产级内部工具。财务人员构建自动合同审查员。客户经理（AE）在一小时内构建了一个销售工具。他们都不是工程师。他们没有提交工单。他们发现了自己的痛点，原型化了解决方案，只有在需要上线生产时才拉工程团队入伙。
- **组织发生了怎样的变化？** 等级结构向智能体工作流的“渠道管理者”坍塌。出现新的原型。薪酬/晋升明确与 AI 熟练度挂钩。从客户信号到交付的时间以小时计。
**严苛测试：** 给我看一个因为系统从之前的运行中学习而变好的工作流，而不是因为某个英雄人物手动改进了它。另外：展示一下上个季度由非工程师发布的三个生产工具。
**常见的假阳性：** 智能体泛滥。一百个脆弱的自动化不等同于一个复利型操作系统。L4 需要受管理的复利（生命周期、可观测性、评估），而不是混乱的扩散。没有压缩纪律，工厂就会堵塞。
**L5 — 几乎自动驾驶的组织**
一个干净的操作定义（需注意的是，我意识到 L5 尚不存在，所以我在描述我认为它可能的样子）：L5 组织的核心操作闭环可以感知现实、诊断问题、启动工作、在授权范围内执行、更新共享记忆并改进未来的行为——而人类负责治理战略、品味、风险、价值观和异常情况，而不是亲自运行这些闭环。
**六个 L5 标记：**
1. 系统在未被询问的情况下注意到重要的事情。
2. 系统综合多个上下文来源。
3. 系统决定是否需要采取行动。
4. 系统在授权范围内采取行动。
5. 当不确定性或后果超出其权限时，系统上报。
6. 系统更新共享记忆，以便未来的行为得到改进。
通过这四个问题的视角来看：
- **AI 能看到什么？** 生成式——系统提出自己的问题，识别自身知识的缺口，提出调查并运行它们。
- **AI 能做什么？** 拥有针对新决策的授权，而不仅仅是执行既定策略。L4→L5 的飞跃：在 L4，系统的改进是因为人类指示它改进；在 L5，是因为它意识到它应该改进。
- **谁可以扩展系统？** 非工程师直接贡献于面向客户的产品本身，或者产品被重塑以至于任何人都可以在不编写代码的情况下扩展它。“内部工具”和“产品功能”之间的界限消融了。
- **组织发生了怎样的变化？** 真正的流动。智能体是拥有有意义授权的组织成员。组织自我修改或提议角色变更、团队边界转移。入职变得由系统驱动。机构知识在过渡中完美存续，因为它存在于系统中，而不是任何个体中。
**严苛测试：** 最近有什么重要的事情是公司在没有人类发起流程的情况下注意到、决定、行动并从中学习的？不是阈值警报；不是配置好的自动化；不是智能体总结人们提出的内容。而是系统综合出来的、人类尚未将其构建为问题的事情。
**常见的假阳性：** “虚假自主”模式。公司声称是自动驾驶行为，但系统只是在执行预配置的规则或显示基于阈值的警报。人类仍然在做所有的“感知”工作。区分真正的生成性行为与美化的可观测性，是这一层面上有待解决的挑战。
有趣的是，一家公司很少在所有四个问题上都处于同一水平，但这种不对称告诉你下一次干预应该集中在哪里。有时 AI 看到的很多但做不了多少。AI 可能做得很多，但只有工程师能真正扩展它。组织架构图可能改变了，但底层基础很薄弱。
Steve Blank 曾说过，初创公司不是大公司的缩小版。同样，一家被“AI点化”的公司也不是旧公司的 AI 辅助版本。它们是围绕新操作系统重建的组织。我们仍在学习这究竟是什么样子的，但那些充满好奇并积极参与的人将获得复利优势，因为他们能迅速在堆栈中向上攀升，并在业务运营（以及希望还有利润率）中看到现实世界的影响。
如果你正在学习以这种方式建立公司，我很乐意与你交流！
## 相关链接
- [Ann Miura-Ko](https://x.com/annimaniac)
- [@annimaniac](https://x.com/annimaniac)
- [166K](https://x.com/annimaniac/status/2050225284277026990/analytics)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [10:46 PM · May 1, 2026](https://x.com/annimaniac/status/2050225284277026990)
- [166.5K Views](https://x.com/annimaniac/status/2050225284277026990/analytics)
- [View quotes](https://x.com/annimaniac/status/2050225284277026990/quotes)
---
*导出时间: 2026/5/2 12:01:15*