---
id: "lib-2026-10-18-fde-ai-field-report-on-fde-work-for-smes"
type: "library"
title: "調研FDE的實戰可行性分析｜AI時代，中小企業如何存活？/ Field report on FDE work for SMEs"
lang: "ZH"
org: "淡出江湖 (空肚, Shenzhen)"
source_type: "Talk"
date: "2026-10-18"
week: "2026-W43"
url: "https://nosov.ai/library/lib-2026-10-18-fde-ai-field-report-on-fde-work-for-smes/"
original_url: "https://www.youtube.com/watch?v=6eNPYWBlEc4"
tier: "C"
tier_name: "Weak signal"
tier_reason: "A 22-minute monologue by a Shenzhen software veteran on a small channel; a plausible field account of forward-deployed AI work for small firms, but one person's experience with no client names or outcome data."
summary_basis: "transcript"
original_minutes: "22"
tags: ["fde", "sme adoption", "workflow mapping", "consulting model"]
---

# 調研FDE的實戰可行性分析｜AI時代，中小企業如何存活？/ Field report on FDE work for SMEs

Tier C (Weak signal): A 22-minute monologue by a Shenzhen software veteran on a small channel; a plausible field account of forward-deployed AI work for small firms, but one person's experience with no client names or outcome data.

Summary based on the full transcript.

## In short

A Mandarin monologue (18 October 2026) by 空肚, a Shenzhen software practitioner running a small YouTube channel, on what forward-deployed engineering (FDE) means for small and mid-sized Chinese firms. After visiting clients himself, he reports that inland firms equate AI with the Doubao chatbot and use it as a search engine, while workflows, reports and meetings stay unchanged. Three-day consultant courses produce nothing deployable. His alternative is a long, on-site engagement that maps data flows before building anything.

## Key numbers

- 3 + 2 + 3 months: on-site discovery, build, then heavy testing in the presenter's delivery model, followed by a year of follow-up
- 20% / 60% / 80-85%: share of reusable components he expects after the first, second and third-to-fifth projects
- 0.85: rough reliability he assigns to raw model output before engineering
- 2: maximum clients one forward-deployed engineer can carry in parallel, by his account

## What it reports

- Shenzhen owners understand AI and feel urgency; inland firms understand it "only as Doubao".
- A large client uses Doubao widely; asked whether company efficiency improved, the answer was "not at all". It replaced ad-heavy Baidu searches; business flow, weekly and monthly reports, meetings and training are unchanged.
- Clients say training courses do not lead to deployment and then blame their own staff.
- His model: three months living on site and talking to every role to map data flows, often manual and never digitised; two months building on his own base layer with skills, agents and harnesses to contain hallucination; three months of intensive testing; a year of follow-up. He calls it a six-month cycle to launch.
- He offers to bill two months and take the third month's fee after results.
- One client exhausts him; two in parallel is the limit. He knows "two or three" people in Shenzhen who can do this work, and they charge more than he does.
- He accepts losses now to build reusable abstractions: about 20% reusable after one project, 60% after the next, 80-85% after three to five.

## How they got it

Personal experience from client visits, told as a monologue. No client names, project figures or results are given. The digest is based on the uploader's Traditional Chinese captions, which contain recognition errors.

## Limits

- One practitioner's account on a channel with about 580 subscribers; no evidence beyond his own durations and estimates.
- Reuse percentages are expectations, not measured results.
- Large parts of the talk are general commentary on platform monopolies and the software industry.

## Why it matters for companies

The account matches what larger firms see: chatbot access improves search but leaves processes unchanged until someone maps the actual data flows. The six-month cycle and the cap of two clients per engineer explain why forward-deployed AI work is expensive and scarce, and why small firms struggle to buy it. Firms planning internal AI deployment can read this as an argument for funding process discovery before tooling.

## Key passages in translation

> 你企業的效率AI加持到了沒有他說一點沒加持到

Has AI improved your company's efficiency? He said: not at all.

> 一個項目落地的話至少前三個月要住在你那去要跟每一個人說聊聚焦你的痛點

To land one project, for at least the first three months I have to live on your site and talk to every person about your pain points.

> 再下一次我就有60%的空間當我打拼三個五個的時候基本上80%到85%的空間就有了

The time after that I have 60% in hand, and once I have done three to five, I basically have 80% to 85%.

Original: https://www.youtube.com/watch?v=6eNPYWBlEc4