In China an FDE isn't the enabler of the Silicon Valley story — it's the vendor putting out fires for state-owned and private firms
The real work of a Chinese FDE isn't arriving with an AI plan to empower the enterprise; it's firefighting inside state-owned and private companies — internal politics, data governance, people who won't cooperate. The technology turns out to be the easy part.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
FDE is not a new job title, it's an old intelligence-agency embed practice
Palantir put forward the FDE role roughly 20 years ago, initially building custom data-analysis software for US intelligence agencies. Because the data was classified, the engineers went straight into the client's own offices — working the same hours, sitting in the same meetings. In the AI era, model capability has grown sharply while enterprises don't know how to use it, so OpenAI, Anthropic and others have been standing up FDE teams. Postings worldwide have grown several dozen times over in the past two years, salaries routinely run to a million a year, and the job has been called the sexiest in Silicon Valley.
— YaxianIn China the FDE is not an enabler, the FDE is simply the vendor
Shen Le says the domestic FDE is nothing like the media picture of someone arriving with an AI solution from a position of authority to empower the client. The real identity is that of the vendor. Getting the position right means keeping your own boss, your own management and every key stakeholder on the client side well looked after — otherwise the project doesn't move. On top of that you spend your time digging out of the holes left for you by sales, the commercial team, and the pre-sales solution engineers.
— Shen LeAn IT lead who only wants to sign off can't move a single business person
On one state-owned enterprise project Shen Le was embedded in, leadership rejected every proposal and the client wanted to cancel the order. He found the core blockage: the interim IT lead's only goal was to get the thing signed off, and wanted a demo to prove value — but had no authority to mobilize the business teams. That left Shen Le unable to get near the real business requirements, unable to build a workable scenario, and therefore unable to demonstrate to senior leadership that the delivery standard had been met. Each link jammed the next, and the project walked into a closed loop.
— Shen LeThe biggest holes are people holes, and a leader who can summon people clears them
A week later the project owner was replaced. The new lead didn't understand AI, but could mobilize the business teams, and found Shen Le eight or nine counterparts across business departments — adding each one on WeChat, calling them, going to their desks to collect materials. Shen Le took sole ownership of one business line and picked a department whose job was writing materials for leadership as the MVP showcase. The project started to move. As he puts it, the biggest holes are still people holes: once the people are handled, the rest is manageable.
— Shen LeThe boss doesn't want AI efficiency, he wants 60,000 people cut to 3,000
At a home-renovation and building-materials company, the boss — born after 1985, in a declining industry and deeply anxious — wanted AI to deliver revenue growth and cost cuts. His example: ideally the company currently has more than 60,000 people, so with AI, could that be cut to 3,000? Shen Le was taken aback. The industry solution consultant sitting next to him didn't reject it outright, but answered with a piece of verbal artistry: rather than focusing on cutting headcount, think about whether the existing 60,000 people could create the value of 12,000.
— Shen LeA probability model can't get into a production line that demands accuracy
Shen Le points out that one reason so many enterprises won't cooperate is that today's model and Agent technology falls far short of the conditions for large-scale production use. A model is fundamentally a probability model, while large-scale production applications inside an enterprise don't accept probability — what they want is accuracy. So many of the projects an FDE runs are really just laying the foundations and trying things out; there is no version where the enterprise is transformed overnight the moment something goes live.
— Shen LeIn their own words · checked verbatim
The biggest holes are still people holes, still people holes — as long as you've handled the people, the rest is actually fine.
最大的坑还是人的坑吧，还是人的坑，只要是把人搞定了，后面的倒还好
Shen Le18:27
My ideal is: I've got more than 60,000 people in the company right now — once I've got AR, can you cut that down, cut that down for me to 3,000 people?
我的理想是我现在公司里有60000多人，你是不是有了AR之后能给我能给我砍到3000人
Shen Le21:49
In terms of how the model itself works, at the end of the day it's a probability model. And if you want a large-scale, production-grade application in an enterprise, they absolutely will not accept probability as the outcome. What they want is accuracy.
模型本身的原理上面，它毕竟是概率模型。而企业你想大规模生产级应用的话，是绝对不接受概率这个结果的。他要的是准确度。
Shen Le28:02
FDE is already a big trend now — we just have to catch this wave of upside. It's definitely going to be huge.
现在FDE已经是一个大趋势了，这波红利我们只要抓住。一定会很好大的。
Shen Le30:14
Figures
| When the FDE concept was proposed | About 20 years ago, originated by Palantir | 4:48 |
| Growth multiple in FDE postings worldwide over the past two years | Several dozen times over | 5:50 |
| FDE salary level | Routinely a million RMB a year | 5:50 |
| Headcount the private-company boss imagined after AI | From more than 60,000 people cut to 3,000 | 21:49 |
| Documents adjusted for the knowledge base | More than 100 documents, taking a week | 24:00 |
| Accuracy required for legal review | 100% required; 85% won't do | 27:01 |
Glossary
- FDE / Forward Deployed Engineer
- An engineer embedded on the client's site to make AI work there, writing code and untangling the business process alike.
- MVP / Minimum Viable Product
- A product prototype built at minimum cost that is still enough to validate the value.
How to listen
Delivery engineers working on AI deployment projects, pre-sales and solution consultants, enterprise digitalization leads, and technical people thinking about moving into an FDE role.
The smart-glasses ad at the top of the episode is skippable.