# McKinsey Won’t Make You AI-Native. Agents Will.
**作者**: Raphaël Dabadie (YC P26)
**日期**: 2026-04-26T16:04:42.000Z
**来源**: [https://x.com/RaphaelDabadie/status/2048433094198231296](https://x.com/RaphaelDabadie/status/2048433094198231296)
---
Every company wants to become AI-native. Most are still using the same method: hire consultants, run workshops, write a roadmap, and hope the organization changes.
That model was built for a slower world.
If AI is the transformation, AI also has to be the method.
## Adoption is not transformation
Most companies confuse AI adoption with AI transformation.
AI adoption is easy. It means giving people better tools. Sales gets a writing assistant. Engineers get Claude Code. HR gets a chatbot. Everyone gets ChatGPT.
This can be useful. But the company itself remains mostly unchanged: same workflows, same meetings, same approval chains, same reporting lines, same blind spots.
The machine runs faster, but it is still the same machine.
That is not what it means to become AI-native.
Jack Dorsey and Roelof Botha describe an AI-native company as something closer to an intelligence system than a traditional hierarchy. AI is not just added to individual jobs. It changes how information moves, how work is coordinated, how decisions are made, and how the company learns.
AI adoption adds tools to the existing company. AI transformation redesigns the company around what AI now makes possible.
The same mistake happened with electricity. In the 1890s, factories replaced steam engines with electric motors, but productivity barely moved because the factory was still organized around steam. The real gains came in the 1920s, when factories were redesigned around electricity itself.
Same factory, new motor. Real gains came from redesigning the system.
As @gsivulka put it:
> “We’ve swapped the motor; we have not yet redesigned the factory.”
That is where most companies are with AI. They have swapped the motor but they have not redesigned the factory.
## McKinsey Was Never Built for AI Transformation
Neither were the other consulting firms.
Beyond being absurdly slow and costing a fortune, consulting is built on sampling by default. A small team interviews a limited part of the organization, runs workshops with selected stakeholders, synthesizes what it heard, and turns that into a roadmap.
That can work when the problem is narrow enough but AI transformation is not narrow.
It touches every function, every workflow, every team, every tool, and almost every role inside the company.
When led internally, the transformation doesn't get any simpler. A small group of key people suddenly has to understand how the entire company works, decide where AI should be integrated, convince teams to change, and keep updating the plan as models improve.
That is an impossible amount of context to hold manually.
## You Cannot Redesign What You Cannot See
This is the painful part of AI transformation.
AI-native startups get to think forward. They start from a blank page and design the company around AI from day one.
Incumbents have to think backward first. Before deciding what the AI-native version of the company should look like, they have to understand how the current company actually works, in all its messy detail.
AI-native startups start from a blank page. Incumbents have to map the old system first.
Most companies avoid this because, through a human-first lens, the task feels impossible: too many teams, too many workflows, too many tools, too much hidden context.
So they jump straight to the future: tools, pilots, roadmaps. But that skips the thing that matters most. Org charts, dashboards, and process docs are not enough. The real company lives in handoffs, workarounds, dependencies, approval loops, and context trapped in people’s heads.
If you only see a sample, you only transform the visible parts.
To redesign the whole system, you need a map of the whole system.
## AI Has to Become the Method
This is why AI transformation cannot stay human-led.
AI should not only be the thing companies are trying to adopt. It should become the method used to transform them.
Humans cannot scale their time. It would make no sense to send a small team to interview hundreds or thousands of people one by one, then expect them to connect every workflow, spot every bottleneck, and keep the plan alive as the company changes.
That limitation shaped the old world. Interviews stayed partial. Training stayed generic. Adoption was tracked through surveys, workshops, and dashboards.
AI agents do not have that constraint. They can interview 500, 1,000, or 10,000 people in parallel, structure far more context than any human team, much faster and keep updating the map as the company changes.
They can also support each person individually, based on their role, workflows, and level with AI.
So why are we still asking humans to do this part of the work that AI is obviously better suited to handle?
## The new model is software plus service
AI transformation requires a new model, built around software and service from day one.
Software layer
The software layer uses AI to rebuild the company’s operational graph: a map of how work actually happens across teams, tools, workflows, handoffs, bottlenecks, and informal habits.
This is the thinking-backward phase most companies avoid. It contains some of the most valuable context in the company, but until now it was almost impossible to collect at scale because it lived in people’s heads and daily routines.
Agents can now make that context visible, keep it updated, and use it to personalize support for every employee based on their actual work.
Service layer
The service layer keeps what was actually valuable in consulting: expert humans who bring judgment, experience, and domain expertise to a complex transformation while removing the parts that no longer make sense.
Humans decide what matters, prioritize the roadmap, validate recommendations, handle sensitive trade-offs, guide adoption, and help teams implement the changes.
They shouldn't do anything AI can do better.
Agents collect and structure the context. Humans bring judgment where it matters.
## How to make it work in practice
This is the thesis we joined Y Combinator with, and the system we are now building at Foaster.
Make the company legible
We start the engagement with 30 to 45-minute AI-led interviews that run in parallel across the organization and rebuild an operational map of how work actually happens.
Agents then reason on top of that map to identify bottlenecks, inefficiencies, hidden dependencies, and the first places AI should be integrated, aligned with the company’s goals.
Keep the transformation alive
From there, we stay embedded inside the company throughout the transformation. Agents keep the map updated and, every month, generate tailored upskilling, feedback, and recommendations for each role based on a precise understanding of how each person actually works.
This helps move the whole company in one direction while supporting everyone at individual scale.
It also changes the business model. Instead of one-off consulting projects, which are not suited to something as broad, operational, and fast-moving as AI transformation, companies pay a monthly subscription for a system that stays active over time.
That subscription gives access to the system and agents doing the work, but also to what we call Forward Deployed Advisors (FDAs): human experts sized to the company who bring the judgment, experience, and domain expertise that made consulting valuable in the first place.
When implementation is needed, Forward Deployed Engineers (FDEs) can step in with the same operational context and go directly to the highest-value work.
Map the company, identify where AI matters, and keep the transformation alive.
In both cases, the goal is the same: agents do everything they can do better, and humans step in only where they are truly needed. That is how this model can go deeper, move faster, and serve more companies than traditional consulting ever could.
Consulting was an optimization of the capabilities we had before AI. AI now makes a completely new model possible, ironically far better suited to AI transformation itself.
## 相关链接
- [Raphaël Dabadie (YC P26)](https://x.com/RaphaelDabadie)
- [@RaphaelDabadie](https://x.com/RaphaelDabadie)
- [44K](https://x.com/RaphaelDabadie/status/2048433094198231296/analytics)
- [@gsivulka](https://x.com/@gsivulka)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [12:04 AM · Apr 27, 2026](https://x.com/RaphaelDabadie/status/2048433094198231296)
- [44.6K Views](https://x.com/RaphaelDabadie/status/2048433094198231296/analytics)
- [View quotes](https://x.com/RaphaelDabadie/status/2048433094198231296/quotes)
---
*导出时间: 2026/4/27 21:45:05*
---
## 中文翻译
# 麦肯锡不会让你成为 AI 原生企业。智能代理才会。
**作者**: Raphaël Dabadie (YC P26)
**日期**: 2026-04-26T16:04:42.000Z
**来源**: [https://x.com/RaphaelDabadie/status/2048433094198231296](https://x.com/RaphaelDabadie/status/2048433094198231296)
---
每家公司都想成为 AI 原生企业。但大多数公司仍在使用同样的方法：聘请顾问、举办工作坊、编写路线图，然后寄希望于组织会随之改变。
这套模式是为一个节奏更慢的世界而设计的。
如果 AI 本身就是这场变革，那么 AI 也必须成为变革的方法。
## 采用不等于转型
大多数公司把 AI 采用和 AI 转型混为一谈。
AI 采用很简单。它意味着给人们更好的工具。销售部门得到一个写作助手。工程师得到 Claude Code。人力资源部门得到一个聊天机器人。人人都用上 ChatGPT。
这可能有用。但公司本身基本没有改变：同样的工作流程、同样的会议、同样的审批链条、同样的汇报关系、同样的盲区。
机器跑得更快了，但它还是原来那台机器。
这不是成为 AI 原生企业的含义。
Jack Dorsey 和 Roelof Botha 将 AI 原生公司描述为一种更接近智能系统而非传统等级制的东西。AI 不只是被添加到各个岗位中。它改变了信息的流动方式、工作的协调方式、决策的制定方式，以及公司的学习方式。
AI 采用是在现有公司上添加工具。AI 转型则是围绕 AI 现在所能实现的可能性重新设计公司。
同样的错误在电气化时代也发生过。在 19 世纪 90 年代，工厂用电动机取代了蒸汽机，但生产力几乎没有提升，因为工厂仍然是围绕蒸汽组织起来的。真正的收益出现在 20 世纪 20 年代，当时工厂围绕电力本身进行了重新设计。
同样的工厂，新的发动机。真正的收益来自对系统的重新设计。
正如 @gsivulka 所说：
> “我们换掉了发动机，但还没有重新设计工厂。”
大多数公司在 AI 上的处境正是如此。它们换掉了发动机，但还没有重新设计工厂。
## 麦肯锡从来就不是为 AI 转型而生的
其他咨询公司也一样。
除了极其缓慢且费用高昂之外，咨询服务默认建立在抽样的基础上。一个小团队采访组织中有限的一部分人员，与挑选出的利益相关者举办工作坊，综合所听到的内容，然后将其转化为一份路线图。
当问题足够窄时，这或许可行，但 AI 转型并不窄。
它涉及公司内部的每一个职能部门、每一个工作流程、每一个团队、每一个工具，以及几乎每一个角色。
即使由内部主导，转型也不会变得更简单。一小群关键人员突然必须了解整个公司是如何运作的，决定 AI 应该在哪些环节集成，说服团队做出改变，并在模型不断进步的同时持续更新计划。
这是人工方式根本无法承载的庞大背景信息。
## 你无法重新设计你看不见的东西
这是 AI 转型中最痛苦的部分。
AI 原生初创公司可以向前思考。它们从一张白纸开始，从第一天起就围绕 AI 设计公司。
既有企业必须先向后思考。在决定公司的 AI 原生版本应该是什么样子之前，它们必须先了解当前公司实际上是如何运作的，包括所有琐碎的细节。
AI 原生初创公司从一张白纸开始。既有企业必须先绘制旧系统的地图。
大多数公司回避这一点，因为从以人为中心的视角来看，这项任务显得不可能完成：团队太多、工作流程太多、工具太多、隐藏的背景信息太多。
于是它们直接跳向未来：工具、试点、路线图。但这跳过了最重要的部分。组织架构图、仪表盘和流程文档是不够的。真正的公司存在于工作交接、变通办法、依赖关系、审批循环以及困在人们脑海中的背景信息里。
如果你只看到一个样本，你就只能转型可见的部分。
要重新设计整个系统，你需要一张整个系统的地图。
## AI 必须成为方法本身
这就是为什么 AI 转型不能继续由人类主导。
AI 不应该只是企业试图采用的东西。它应该成为用来推动企业转型的方法。
人类无法扩展自己的时间。派一个小团队逐个采访成百上千的人，然后指望他们把每个工作流程串联起来、发现每个瓶颈，并在公司不断变化的同时让计划保持鲜活——这是毫无意义的。
这一限制塑造了旧世界。采访始终是局部的。培训始终是通用的。采用情况只能通过调查、工作坊和仪表盘来追踪。
AI 智能代理没有这种限制。它们可以并行采访 500、1,000 或 10,000 人，以比任何人类团队快得多的速度整理远为庞大的背景信息，并在公司变化时持续更新这张地图。
它们还能根据每个人的角色、工作流程和 AI 使用水平，为每个人提供个性化支持。
那么，为什么我们仍然要让人类去做这部分显然 AI 更擅长处理的工作呢？
## 新模式是软件加服务
AI 转型需要一种新的模式，从第一天起就围绕软件和服务来构建。
软件层
软件层使用 AI 重建公司的运营图谱：一张展示工作在团队、工具、工作流程、交接、瓶颈和非正式习惯之间实际如何发生的地图。
这正是大多数公司回避的向后思考阶段。它包含公司中一些最有价值的背景信息，但迄今为止几乎不可能大规模收集，因为这些信息存在于人们的头脑和日常习惯之中。
如今，智能代理可以让这些背景信息变得可见，保持其更新，并根据员工的实际工作为其提供个性化支持。
服务层
服务层保留了咨询中真正有价值的部分：由具备判断力、经验和领域专业知识的专业人士参与复杂的转型，同时去除了那些已不再合理的环节。
由人类来决定什么重要、排定路线图优先级、验证建议、处理敏感的权衡、引导落地采用，并帮助团队实施变革。
他们不应做任何 AI 能做得更好的事。
智能体负责收集和结构化上下文。人类在最关键的地方贡献判断力。
## 如何在实践中落地
这就是我们加入 Y Combinator 时的核心命题，也是我们现在在 Foaster 构建的系统。
让公司变得清晰可读
我们在项目启动时进行 30 到 45 分钟的 AI 主导访谈，这些访谈在组织内并行开展，重建一张关于工作实际如何进行的运营地图。
随后，智能体在这张地图之上进行推理，识别瓶颈、低效环节、隐性依赖，以及 AI 应当首先切入的地方，并与公司的目标保持一致。
让转型保持活力
在此基础上，我们在整个转型过程中持续驻留在公司内部。智能体保持地图的实时更新，并且每个月基于对每个人实际工作方式的精确理解，为每个角色生成量身定制的技能提升、反馈和建议。
这有助于让整个公司朝着一个方向前进，同时在个体层面支持每一个人。
这也改变了商业模式。企业不再购买一次性的咨询项目——这种方式并不适合 AI 转型这种涉及面广、依赖运营且变化迅速的事情——而是按月订阅一个持续运转的系统。
这份订阅不仅提供系统和使用这些智能体的权限，还包括我们所说的“前线部署顾问”（Forward Deployed Advisors，FDAs）：根据公司规模配置的人类专家，他们带来判断力、经验和领域专业知识——正是这些让咨询从一开始就有了价值。
当需要具体实施时，“前线部署工程师”（Forward Deployed Engineers，FDEs）可以带着同样的运营上下文介入，直接投入到最高价值的工作中。
绘制公司地图，识别 AI 发挥价值的地方，并让转型保持活力。
在这两种情况下，目标都是一样的：智能体做它们能做得更好的所有事，人类只在真正需要的地方介入。这就是这一模式能够比传统咨询更深入、更快速、服务更多公司的原因。
咨询是对 AI 出现之前我们所拥有能力的一种优化。如今 AI 让一种全新的模式成为可能——颇具讽刺意味的是，这种模式远更适合 AI 转型本身。
## 相关链接
- [Raphaël Dabadie (YC P26)](https://x.com/RaphaelDabadie)
- [@RaphaelDabadie](https://x.com/RaphaelDabadie)
- [44K](https://x.com/RaphaelDabadie/status/2048433094198231296/analytics)
- [@gsivulka](https://x.com/@gsivulka)
- [升级至高级版](https://x.com/i/premium_sign_up)
- [2026年4月27日 凌晨12:04](https://x.com/RaphaelDabadie/status/2048433094198231296)
- [44.6K 次浏览](https://x.com/RaphaelDabadie/status/2048433094198231296/analytics)
- [查看引用](https://x.com/RaphaelDabadie/status/2048433094198231296/quotes)
---
*导出时间: 2026/4/27 21:45:05*
