FDE 怎么炼成：审计先行 + 30 天拆流程，读「一页 Digest」的方法论笔记

by Mycelium Protocol

---

原帖来源：一页 Digest，社交媒体帖子，讨论 FDE（Field Deployment Engineer，AI 落地工程师）的实操方法论和练成路径。本文是读后感与补充整理。

---

FDE 在一家公司要干的三件事

原帖把 FDE 的工作拆成三步，顺序不能乱：

第一步：搞懂业务现在到底是怎么运转的

不是读文档，是挖现实。写在文档里的流程和真实发生的往往是两回事。

举个例子：「收到一封邮件」这种触发点，听着特别简单。可这一封邮件背后，可能来自 40 多个不同的发件人，格式还各不相同。你以为是一个节点，实际上是 40 个分支。

这个挖掘过程有个正式名字：审计（Audit）。

第二步：判断这套智能该放在哪个环节、不该放在哪

不是把 AI 塞进所有流程，而是找到杠杆点——哪里自动化 ROI 最高、哪里人工不可替代、哪里有数据才能跑、哪里跑错了成本最大。

第三步：真正动手搭系统

前两步没做好，第三步等于在沙上建塔。反过来，前两步做扎实了，第三步才不会返工。

---

审计为什么是被低估的第一步

原帖提到一个数据：他们见过客户反馈，说这次审计带来的价值是付出成本的十倍。

这个数据不奇怪。审计做对了，等于：

• 找到了真正值得自动化的流程（而不是看起来值得的）
• 摸清了所有例外情况，避免 agent 在生产里遇到「第 41 种邮件格式」就崩
• 给后续的 agent 设计提供了真实的 ground truth

在大多数失败的 AI 落地项目里，跳过审计（或者把它做成走过场）是最常见的根因之一。「AI 不理解我们的业务」其实是「我们自己还没把业务讲清楚」的另一种表达。

---

30 天从零练起的计划

原帖给了一份三周的入门节奏：

第 1 周：挑一个真实存在的后台工作流，搭一个能真正跑通的 agent

选材范围：财务、HR、采购、物流都行。标准只有一个——是公司里真实发生的事，不是示例数据。

目标不是 demo，是「在真实数据上跑一次完整流程」。

第 2 周：开始处理意外情况

一件事做对只有一种方式，但做错的方式能有一千种。

第 1 周搭好的 agent 大概率在某些边缘情况下会挂。第 2 周的任务就是一条一条地把这些边缘情况处理掉。能接住意外情况，agent 才算真正有生产价值，而不是玩具。

第 3 周：用数字说话

衡量指标只看三类：

指标类型	说明
营收提升	因为 agent 跑通，这个流程带来多少额外收入或更快的成单速度
风险降低	减少了哪些人工失误、合规风险、数据遗漏
成本节省	人力时间折算成金额，跟 agent 运维成本比

能把这三类数字讲清楚，你对 FDE 的理解就已经超过市场上大多数人了。

---

我们的补充：为什么现在练特别值

原帖最后一句话值得单独拎出来：「这个岗位现在有多稀缺？窗口期还早，别等谁来教你，自己动手练就是了。」

稀缺的原因很具体：

复合要求高。FDE 要同时懂业务流程、会用 AI 工具、能跟业务侧沟通、还能把 agent 跑在生产里。这四个能力单独看都不难，合在一个人身上的特别少。

需求是真实的。大量企业的「AI 战略」卡在「落地」这一步——模型够用了，但没有人能把它嵌进现有业务。这个缺口就是 FDE 的市场。

窗口期有限。一旦有足够多有经验的 FDE 出现，这条路就不再是「自学就能占到先机」的阶段了。

练法和原帖一致：找一个真实业务流程，先审计，再搭，再量化。三周一个循环，做三个不同行业的案例，你就有了一份有说服力的作品集。

---

和我们之前写的 FDE 文章的关系

七月份我们写过一篇《每个人都是自己的 FDE》，讲的是 FDE 的全球背景、Palantir 起源和国内一线调查。

这篇是实操层的补充：原帖提供的三步工作法 + 审计的具体价值 + 30 天练习路径。

两篇合在一起看：上一篇解决「为什么」和「是什么」，这篇解决「怎么练」。

---

Mycelium Protocol — 追踪 AI 系统的底层演化

---

关于 Mycelium

菌丝协议。持续追踪 AI 工具、系统和实验的内容节点。

---

How FDE Skills Are Built: Audit First, 30 Days of Workflow Deconstruction

by Mycelium Protocol

---

Source: A Yiye Digest social media post on FDE (Field Deployment Engineer) methodology and the path to getting there. This piece is a reading response and synthesis.

---

The Three Things an FDE Does in a Company

The original post breaks FDE work into three steps, in strict order:

Step 1: Understand how the business actually operates right now

Not by reading documentation — by excavating reality. Written procedures and what actually happens are often two different things.

Example: “receiving an email” as a trigger sounds trivially simple. But that one email could come from 40+ different senders, each with a different format. You think you’re designing for one node; you’re actually designing for 40 branches.

This excavation process has an official name: Audit.

Step 2: Determine which processes get AI and which don’t

Not stuffing AI into everything — finding the leverage points. Where does automation deliver the highest ROI? Where is human judgment irreplaceable? Where does data exist to run inference? Where would a failure be most costly?

Step 3: Actually build the system

Steps 1 and 2 done wrong makes Step 3 a foundation of sand. Done right, Step 3 doesn’t need a redo.

---

Why Audit Is the Underrated First Step

The original post includes a data point: client feedback reported audit value at 10x the cost.

That figure isn’t surprising. A proper audit means:

• Finding the processes actually worth automating (not just the ones that look worth automating)
• Mapping all the edge cases before the agent hits “email format #41” in production and breaks
• Giving subsequent agent design a real ground truth to work from

In most failed enterprise AI deployments, skipping the audit — or treating it as a formality — is the most common root cause. “The AI doesn’t understand our business” is often “we haven’t actually explained our business clearly” said differently.

---

The 30-Day From-Zero Practice Plan

The post outlines a three-week starter cadence:

Week 1: Pick one real backend workflow and build an agent that actually runs on it

Candidates: finance, HR, procurement, logistics — anything that’s happening in a real company. One standard only: real data, not example data.

The goal isn’t a demo — it’s “one complete end-to-end run on real inputs.”

Week 2: Handle the edge cases

There’s only one way to do something right, but a thousand ways to do it wrong.

The agent from Week 1 will almost certainly break on some edge cases. Week 2’s job is to handle them one by one. An agent that can absorb exceptions is a production asset; one that can’t is a toy.

Week 3: Talk in numbers

Three categories only:

Metric type	Description
Revenue increase	More deals closed, faster, or additional revenue the agent enabled
Risk reduction	Fewer manual errors, compliance gaps, missed data points
Cost savings	Human hours saved, translated to dollar value, compared to agent operating cost

If you can narrate all three clearly, your understanding of FDE already exceeds most people currently in the market.

---

Our Supplement: Why Practicing Now Is Especially Valuable

The post’s closing line is worth isolating: “This position is now very scarce. The window is still early — don’t wait for someone to teach you, just start practicing.”

The scarcity is specific:

High compound requirements. FDE requires understanding business processes, knowing AI tooling, communicating with business stakeholders, AND running agents in production. Each alone isn’t hard; all four in one person is rare.

Demand is real. Many companies’ “AI strategies” are stalled at “deployment” — the models are good enough, but no one can embed them into existing operations. That gap is the FDE market.

The window is finite. Once enough experienced FDEs exist, this won’t be a “self-taught head start” phase anymore.

The practice method matches the post: find a real business process, audit it first, build second, quantify third. Three weeks per cycle, three cases across different industries, and you have a convincing portfolio.

---

How This Relates to Our Earlier FDE Piece

In July we wrote Everyone Can Be Their Own FDE, covering FDE’s global background, Palantir origins, and a survey of Chinese builders.

This piece is the operational layer: the three-step methodology, the concrete value of audit, and the 30-day practice path.

Read together: the July piece covers “why” and “what.” This piece covers “how to practice.”

---

Mycelium Protocol — tracking the deep evolution of AI systems

© 2026 Mycelium Protocol. All rights reserved.