This article comes from the WeChat public account: AI闹, author: Shen Yue, cover image: AI-generated
Another term has caught fire from Silicon Valley to China: FDE, literally translated as Forward Deployed Engineer.
The Silicon Valley narrative: a new role that understands technology, business, and consulting, who goes on-site at enterprises to help companies complete their AI transformation—"the last mile of AI implementation."
Job postings for FDE-related positions in the US have surged 800% year-over-year; public information shows a senior FDE's annual salary is roughly 1.5–2 million RMB; OpenAI and Anthropic have partnered with top private equity firms and investment banking giants to directly launch FDE businesses; domestic giants Volcano Engine, Tencent Cloud, and Alibaba Cloud have also started building their own FDE teams, or having agents handle FDE delivery.
FDE, like the Agent engineer a year ago, is being described as "the sexiest engineering job of the AI era."
But the information currently being conveyed lacks voices from the front lines. In China, how does an FDE who has actually entered enterprises understand this new role? What has he experienced? What are the actual implementation results?
"A Chinese FDE is more like vendor outsourced labor. After the vendor sells the product, we do the dirty, tiring work," says Shen Yue, the narrator of this episode of "AI Practice," one of the earliest FDEs in China working on the front lines, a former ByteDance employee who began receiving FDE projects in early 2025. "Back then, we weren't called FDE yet—we were called AI consulting companions."
Contrary to the industry's enthusiasm for FDE, his experience carries its own "anti-climactic" drama: he has provided on-site services to 4 companies, both private and large state-owned, with project values ranging from nearly a million to tens of millions.
Through this, he has also observed the unique uneven landscape of Chinese enterprises. "The power struggles within organizations are always more complicated than the technical solutions. Where there are people, there are politics—how would AI ever understand that?"
Below is Shen Yue's first-person account: the same FDE, China vs. Silicon Valley, two worlds.
A Week as a Firefighting FDE
Last week, I was sent to a large state-owned enterprise to put out fires.
The project was a tens-of-millions deal signed in Q4 of last year, promising to deliver 10 agents to the client within six months. But almost a year later, it still couldn't pass acceptance and had fallen into serious paralysis. The client made it clear: they were not satisfied.
My firefighting mission: figure out "to what standard our Agent needs to be delivered" before the client would accept it.
On my first day on-site, I figured out the problem. The FDE contract was signed by the big boss of this state-owned enterprise, but after signing, he handed it off to someone below to execute. As for which specific business scenarios it should be implemented in, he gave no instructions—he just felt the AI concept was good and the company should do it.
But the person below in charge of execution couldn't articulate specific requirements either. The style of state-owned enterprises is to avoid making mistakes, so he told the FDE to look around, propose things themselves, and do it themselves. So over the past year, the FDE wandered from department to department like a headless fly, "trying their luck" discussing requirements, and finally gritted their teeth and built 7 agents. I pulled up the products and took a look—they really were unusable. There was no clear requirements owner, let alone business value. No wonder the big boss was unwilling to sign off.
What's worse, the project lead originally assigned by this senior leader, for some reason, refused to manage us, and then set up his own AI transformation department within the company—doing essentially the same thing as us FDEs.
So the person I dealt with became an IT lead temporarily brought in. At our first meeting in a small conference room, he said very directly that he was purely "taking the fall," saddled with a lot of extra work for no reason. His goal was for me to wrap up the project quickly and not cause him any trouble.
He privately gave me an idea: since everyone wanted to close the case, the existing materials could be reframed into a "reporting logic"—just show the boss actual value, like making a demo, pulling some before-and-after usage data, and showing the comparison.
I had an epiphany: learning how to report upward should be recorded as one of the FDE's most core competencies.
So I proposed: could he help me connect with some business colleagues so I could get some real data? He flatly refused—no, he couldn't disturb the business teams, and told me to figure it out myself.
Then where was I supposed to find data? (laughs) I was left hanging in a vacuum, feeling like I couldn't reach the sky above or touch the ground below.
So I set my goal as: don't fantasize about solving all the problems; instead, list what can be advanced this week.
If this were a success-story article, it would probably have a dramatic reversal right about here: thanks to my efforts, the client was moved, the proposal passed smoothly, and the project entered a new phase.
But the real situation wasn't like that at all. After working for a while, the client's conclusion was still: "Your demo still isn't very convincing. You need to think about it more."
Honestly, it was quite frustrating. It reminded me of the mobile internet era, when many companies were doing digital transformation: paperless offices, buying systems, building products. The boss said to use them, but never said how to evaluate them; middle managers were unwilling to execute, and in the end the systems became "tools for filling in reports for the leadership to see" or simply fizzled out.
The logic behind this project's paralysis was exactly the same—internal authority and responsibility within the organization were never sorted out, and no one was willing to take responsibility. It's a state-owned enterprise, after all; everyone is afraid of making mistakes.
Outdone by a 60-Year-Old Principal's Three-Day Hand-Built Version
I started getting involved with FDE work in the summer of 2025. It was initially called "consulting companion," and later a new concept came out of Silicon Valley: FDE. Leave it to Silicon Valley to coin terms (laughs).
Frankly, I feel our work is still vendor labor outsourcing—nothing more than pre-sales bidding, sales signing, and post-sales delivery. It's just wrapped in an AI shell; the business model hasn't changed. Sometimes at bids, the ones competing for deals at the same table are still the same batch of companies that used to help enterprises with digital transformation.
It's just that before, when we said we did digital transformation, we couldn't even get a meeting with the boss. But now when we say we're FDEs, we can at least get a chance to chat with the boss.
My first FDE client was a private education group with annual revenue of about 1 billion. They bought Volcano Engine's Agent development platform, and we started by teaching them to build on the Agent platform. The company even held a training camp, selecting over 100 teachers and administrative staff to participate—all tech novices with zero background.
But this training camp truly shocked me.
There was a woman in her 20s, in a finance role, who on her own built a financial reconciliation Agent—a very complex workflow that took me two hours to even partially understand. But it ran without any problems. I later heard the company adopted her product and even promoted her.
There was also an art teacher who built a live-streaming customer service Q&A Agent. I read his product documentation—it was better than most product managers'. I later heard Geely poached him to be their AI lead. I specifically asked him whether he had any relevant background before. He said no—he studied art in college, an art student.
AI will definitely unleash enormous human potential; 80% of product managers will be eliminated.
After the training camp, we moved into FDE work. This company had a clear pain point: they frequently take over schools, and each takeover requires a research report, with 4–5 principals participating on-site, taking at least two weeks. They hoped AI could help improve efficiency.
With three to four of us working for three months, we ultimately delivered four Agents covering the full pipeline of intelligence gathering, data organization, report generation, and principal recommendations, cutting the overall time from two weeks to three days.
My first FDE deal gave me quite a sense of accomplishment when completed, but after actual rollout, we found that nobody used it.
The company required principals to digitize everything—recording audio, filling out forms, uploading—but most people simply didn't have that habit and found it troublesome, which led to incomplete information input and substandard output quality.
Among them was a principal in his 60s who had attended my training camp. In three days, he built a simplified Agent that only did template generation—partial AI-ification, with the rest still filled in manually. He shared this Agent with other principals, and everyone loved using it, feeling it was better.
A professional team like ours, spending hundreds of thousands and three months building something, outdone by an old principal's three-day hand-built version—this really frustrated me (laughs).
My conclusion now: the biggest problem was "over-engineering."
Why is the old principal's product so good to use? Because he doesn't require other principals to change how they work—there's no barrier to adoption. Meanwhile, we were too eager to pursue "full-pipeline AI-ification" and ignored users' actual scenarios and habits.
This made me reflect on a detail: the people we originally surveyed for requirements weren't frontline principals, but the company's digitalization team—there was a layer in between. This is the biggest difficulty of being an FDE: you may never even meet the actual stakeholders.
The client did end up paying, but according to their internal assessment: the project was a failure.
I also think it was quite a failure.
Help me lay off 3,000 people directly
Most owners of private companies care about only one question: I'm spending several million—can you help me increase revenue, or cut headcount? As for whether it's AI or not, honestly, they don't really care.
I remember a private home renovation company in Guangdong that has been hit hard by the real estate downturn in recent years, with revenue steadily declining. The old man had no energy left to reform, so the second-generation heir took over—an 85-born woman who made sweeping changes and pushed for AI. At our first meeting, she was very forceful and stated her demand directly: the company has over 6,000 people now, and she hoped that after the FDE came in, we'd help her cut it down to 3,000.
Later I heard her husband also works at this company—he's the head of HR, and the couple is pushing this together.
This project ran for three months, and it was truly grueling. I made four round trips from Beijing to Guangdong, and the travel costs couldn't even be covered by the project revenue. Everyone assumes FDEs earn a lot, but our real income isn't high—many projects are completely in a loss-making, thankless state (laughs).
Every time we gave a report, we got scolded. She was dissatisfied with the progress and with the value we demonstrated. She kept saying: the value I want to see is how many people this AI agent can replace once it goes live. We kept trying to persuade her to lower her expectations.
How could AI possibly replace 3,000 people overnight? In the end, we deployed 7 agents, most of them for improving quality, and only one sales agent might actually help increase revenue.
The bosses I've seen are all anxious, rushing to have AI replace people. But my understanding is that AI can currently only replace tasks, not jobs.
Take a legal position, for example: AI can take over "reviewing contracts," but the job also includes a whole series of tasks like attending negotiations, organizing and filing documents, and handling emergencies.
Only when AI can complete more than 80% of a job's core work can it fully replace a person—we're nowhere near that now.
Also, many of the companies I've encountered haven't even laid a solid digitalization foundation. No matter how powerful AI is, it can't be deployed—actually, AI places extremely high demands on an organization's maturity.
Doing FDE work in China, I've summed up a few points.
1. With state-owned enterprises, you win over people; with private companies, you get things done;
2. Ideally, go straight to the person who signs the contract or the one who makes the final call;
3. The company should have some digitalization foundation;
4. The boss should maintain a certain degree of patience, rather than expecting results in three months.
In the short term, FDE work definitely has value, but in the long run it's a transitional role.
On one hand, large enterprises will internally develop their own AI deployment capabilities; on the other hand, in China, many FDE projects are taken on by cloud vendors' resellers, with too many layers in between, so some projects are doomed to fail from the start.
Frontline sales, in order to sell products, make too many unrealistic promises, and we FDEs have to fill the holes, clean up the mess, and do the dirty, exhausting work. Unlike Silicon Valley, where an FDE can tell a client directly, "This won't work."
Oh, right—I just got notified that next week I have to go rescue that large state-owned enterprise again. Apparently the project has taken a turn for the better.
The person originally in charge of our project—remember how he went off to start his own thing? He got chewed out by the big boss and told to clean up the mess. So after going full circle, we're back at square one. Let's see if we can push the acceptance through this time (laughs).
But no matter how AI evolves, one thing won't change: whether it's state-owned or private enterprises, the internal power struggles over responsibilities within an organization are always more complicated than the technology. Wherever there are people, there are politics—how would AI ever understand that? (laughs)