(Source: China Entrepreneur Magazine)

Text | Sun Xin, reporter at China Entrepreneur

Contributing Editor | Li Yuan

Header image source | VCG

Over the past six months, a brand-new job concept from overseas — FDE — has ignited a talent war among tech giants, with million-yuan annual salaries merely the baseline.

ByteDance, Alibaba, Tencent, Kingsoft and other big companies are all going full throttle to urgently recruit FDEs. China Entrepreneur found on platforms like BOSS Zhipin and Liepin: Doubao is offering an annual package of 1.05 million yuan; Alibaba Cloud's large model FDE role pays 20,000 to 55,000 yuan per month with 16 months' salary per year; Moonshot AI's Kimi has posted an FDE position with an annual package of 1.09 million yuan. Silicon Valley insiders say overseas FDE compensation is virtually uncapped.

A source close to ByteDance told China Entrepreneur: some ByteDance FDEs stationed at client companies can earn a daily rate of 10,000 yuan.

FDE stands for Forward Deployed Engineer, meaning Forward Deployed Engineer. It was popularized by the American company Palantir, and its core is "on-site delivery" — engineers go directly into the client's site and embed standardized data analysis products into the client's existing systems and workflows.

This playbook helped Palantir's stock price climb from under $6 at the end of 2022 to $200 in 2025, and it has shown Chinese companies struggling with AI-to-B implementation a potential cure.

The AI office war that began this summer has further intensified the giants' anxiety. Mid-year, ByteDance merged Feishu, TRAE and Coze into Doubao Work in just over twenty days; Alibaba combined QoderWork, some peripheral DingTalk products and MuleRun into Qianwen Office; Tencent's WorkBuddy requires the thousands of service providers in its ecosystem to have their own FDE delivery capabilities.

Source: AI-generated

Why are the giants collectively targeting FDEs? On one hand, the AI office track has seen price wars, channel wars and product integration wars in succession. "What they are ultimately competing for is who can truly fulfill enterprise needs and standardize them into products — but there's a time window for this," said Liu Yuxuan of CDH Investments' Innovation and Growth Fund.

Amid the fierce battle, AI products also face a core contradiction in delivery that remains unsolved: "Sales can't articulate requirements clearly, and R&D can't reach the customers," a head of a leading cloud vendor summarized to China Entrepreneur.

"FDE was first proposed by Volcano Cloud, which is related to their to-C genes. In the past, sales didn't know where R&D was — they sold whatever they had. But with AI to B, especially to large B clients, you must discuss requirements, do planning and testing. If salespeople can't articulate it clearly, you put an FDE beside them — essentially a group of R&D people. So in the end, FDE is still about improving the combat capability of frontline staff," the executive said.

FDE Is Just a New Name for Sales

Although FDE is a newly trendy term, it is not a new concept. "The requirements to become a qualified FDE are extremely demanding — that's why they're expensive," said Li Feng, general manager of Kingdee's central SOE digitalization consulting department.

In his view, FDE has emerged at this point in time because AI is no longer a tool but is entering organizations in the role of a "work partner" — when productivity changes, production relations must change too.

In the industrial era, marketing, pre-sales and delivery were handled by three separate lines, whereas an FDE must be able to win deals in the market, understand the business, translate requirements into technical language, and also deploy and deliver on their own. In short, it requires a versatile talent who can take responsibility for results and be paid based on results.

This is not a problem unique to China. "Hiring an FDE in the US is just as hard," Li Feng said.

In August 2026, Zhang Jian received FDE interview invitations from both Alibaba Cloud and Volcano Engine at the same time. He recalled the interview requirements: "A role similar to a TAM (Technical Account Manager), but requiring a very broad tech stack coverage, from servers all the way to Agent platforms, with the position graded from level 1 to 10."

Zhang Jian told China Entrepreneur that FDE positions on the market actually fall into two categories: one builds vertical industry capabilities based on the company's own Agent products, requiring understanding of industry workflows and boundaries, with greater technical depth; the other, without an in-house Agent product, uses external tools to custom-develop software from 0 to 1 according to client needs.

The FDE on-site fee model in China depends on complexity and generally comes in two forms: project-based, charged per person per day; and subscription products, with services bundled into a package fee.

For example, the ByteDance position Zhang Jian applied for does not require the FDE to be stationed on-site long-term; instead, the focus is on improving the company's own Agent, with only occasional business trips. Alibaba Cloud also allows online collaborative work, whereas some startups' FDE positions do require long-term travel.

What an FDE actually does is go deep into the client's site, standardize client requirements, and feed them back to product design. Zhang Lanyin, founder of Hangzhou Spatiotemporal Emergence Technology, has been an entrepreneur in the to-B track for years. She noted that an FDE must understand technology in order to spot where AI can be applied within a client's processes.

In addition to full-stack technical capability, an FDE also needs a salesperson's keen instincts.

Li Ge, an OPC entrepreneur in AI to B, describes FDEs as "patching gaps." In 2022, an office software company widely recruited a batch of CSMs (Customer Success Managers) responsible for sales, channel training and customer service. "Now they're heavily recruiting FDEs — to some extent, that's making up for the CSMs' shortcomings in AI capabilities."

Guo Qizhi, senior partner at CDH, offered his judgment — consulting firms have always separated "hunters" and "implementers," but an FDE who can discuss business processes with clients "is ideally a former top salesperson, whose technical shortcomings can now be filled by AI Coding skills. Someone who can learn programming from zero is a natural FDE candidate."

Zhang Jian, speaking from a job-seeker's perspective, has no aversion to sales targets. "If you only do technical work, the boss most likely won't understand what you're doing — you're just a cog."

This also leads to a situation where FDEs are often hired as technical roles, used as delivery workers, and ultimately evaluated as salespeople.

What Exactly Are Customers Paying For?

In the giants' imagination, an FDE can single-handedly handle business analysis, AI technical closed loops, productization and on-site communication. But after months of hype, social media platforms have become flooded with headlines like "the FDE bubble has burst."

In reality, FDE is still far from the ideal state. Guo Qizhi said that much of what many FDEs currently do inside companies boils down to two words — training: organizing AI competitions for clients and teaching AI courses.

To understand why delivery is so misaligned, one must first answer a question: what exactly are customers paying for? And why haven't customers paid yet?

"In domestic to-B delivery, the moment the client disagrees, they make you do customization — if the platform ends up as a 'heavily modified' version, that's already pretty good."

Li Ge said that to understand why FDEs are necessary at delivery sites, one must first acknowledge one thing: most Chinese enterprises do not trust AI, let alone heavily depend on any AI product on the market.

"Enterprise AI today has only achieved partial adoption." Li Feng of Kingdee divides AI implementation into four stages: individual use, partial enterprise use, full enterprise use, and society-level management.

Source: AI-generated

Although the market is full of claims like "we'll build you an ERP in three days," Li Feng's judgment is that the enterprise side has only reached the second stage. "End-to-end, full-domain — I haven't seen that yet to this day."

After all, in enterprises' view, the B side has an extremely low tolerance for error. "Customer needs are not standardized, and for AI agents in long-process business scenarios, the tolerance for error rates is very low. Even if AI has a 0.01% chance of mistakenly deleting enterprise data or getting an account entry wrong, enterprises will absolutely not accept it," said Liu Yuxuan.

Even overseas, clients are far from trusting AI. She cited the case of a semiconductor company served by OpenAI: the team produced a demo in just 4 weeks, but then spent 4 months fully annotating every output of the system with its data sources. "This isn't a practical need — it's more of a psychological need."

Alongside the lack of trust, insufficient context and fragmented data are also becoming roadblocks for general AI to enter large enterprises. Feng Luyao, founder of Keyshi Technology, is currently an entrepreneur in the FDE track and was previously a frontline to-B salesperson at a major Chinese tech company.

While working at a big tech company, Feng Luyao often encountered this same picture: hundreds of employees at a client company were all using a certain to-B office product and had built many AI agents, but the company's data wasn't stored in that software. Employees had to manually feed information into it, meaning all decisions and reviews still had to go through a human pass.

"To customers, AI seems like icing on the cake, and somewhat superfluous," Feng Luyao said bluntly.

She also encountered clients who had purchased a leading AI platform, paying several hundred thousand yuan in the first year, only to completely refuse renewal in the second year. This meant "from the big tech company's perspective, the deal became a loss-making business."

Many clients had thrown questions at Feng Luyao: Is there a way to let the boss intuitively see the results of using AI? "At least give me a platform that tells me how many tokens were used and how much efficiency improvement the company gained."

But today, the actual results of AI products are hard to display intuitively, and even the delivery process has yet to be standardized.

Feng Luyao revealed that after a salesperson at a leading AI platform closes a deal, the work is handed off to channel partners, who then try to bundle their own services into it. Although the big tech company's SOP (standard operating procedure) is standardized, the level of investment also depends on the client's brand strength.

"Reporting is chaotic, KPIs are chaotic, and even work goals are chaotic," Li Ge said. "The data integration module can't be sold by the collaborative office tools sales team, because the revenue from selling data integration is credited to a different group of people."

In her view, big tech companies still take the product as the subject and work backward to B-end business processes, rather than working backward from B-end business processes to the product.

The size of an FDE team is the inverse of model capability

After more than half a year as an entrepreneurial FDE, Feng Luyao has truly felt the harshness of reality. Yaoshi Technology's clients fall into two categories: one type explicitly requires getting everyone using AI first, in pursuit of cost reduction and efficiency gains; the other type wants a complete platform, more like traditional software development work.

There is also a real temperature gap between the hot market and the cold reality of enterprises. A manufacturing business owner in Suzhou told Feng Luyao that smart manufacturing is still hard to achieve through AI intervention. "Everyone says AI can bring earth-shaking transformation, but at least among the companies we've encountered, we haven't seen it. It's always the vendors trying to educate the market—customers simply don't have the corresponding demand."

More thorny than market education is the internal redistribution of interests that B-end deployment inevitably touches. After all, every deployment project cannot avoid the internal balance of interests within the client's organization.

Which departments' data will be opened up, which roles' existing verification processes will be simplified, and who signs off on the final launch results. "You need to write these things clearly; then it's up to the boss to decide. If you can't write them clearly, even the strongest technology will be worn down to death in meeting rooms," Li Feng said.

Even when doing AI to-B delivery, there are inherent differences between domestic and overseas models. Liu Yuxuan has a vivid sense of this: "Overseas, it's mostly standardized SaaS products; in China, project-based delivery is always needed."

Liu Yuxuan once spoke with the head of FDE at OpenAI, whose team had only two people in its early days—"the company's attention was all on to-C products like ChatGPT"—and FDE was only formally elevated once to-B became a strategic must-answer question. She revealed that even though OpenAI's FDE has achieved very high per-capita efficiency, as of the end of 2025, the entire team was only around 50 people.

In contrast, big tech companies in China are building delivery and ecosystem service teams of several hundred people. Behind the huge gap in manpower investment—beyond enterprise trust logic and private data management habits—the root cause also lies in the products and models themselves.

"Model capability still isn't good enough, and delivery is still being done with a SaaS to-B mindset. Model capability directly determines how many people are needed as a safety net after FDE products are standardized. As models' long-horizon task accuracy continues to improve, FDE per-capita efficiency will keep rising."

According to her observation, Zhipu currently basically only outputs standardized APIs, backed by a strong model foundation. "They're being chased by money every day. The more people a company piles up on-site, the more it shows that its products and models are inadequate."

Liu Yuxuan emphasized: the true purpose of FDE is to "bring requirements back into the product in a standardized way." If it's merely on-site one-off project delivery, it's no different from outsourcing.

The next form of FDE: human + AI agent, one-person company

"From being deified, to coming down off the pedestal, and even starting to be stigmatized—this reflects the extreme irrationality of the market itself," said an industry insider, who believes the core of FDE hasn't changed. Going down to the business front line and then extracting common needs—this still holds true. "The combination of humans and machines is an inevitable trend."

"FDE itself is not a pseudo-concept, but if you treat it merely as a door-knocking tool for sales and customer acquisition, it becomes a pseudo-concept," said Guo Qizhi. Every company's business processes, management style, and data foundation vary enormously. "Big tech companies naturally pursue generalization—one product they'd love to sell to the whole world. To deeply solve actual business problems, you often still have to turn to third parties."

This has allowed a group of small players to grow in the gaps that big tech companies cannot deliver to.

Feng Luyao's team initially had only three people: she was responsible for diagnosis, producing solutions, and "cooking up" demos; an engineer handled deployment and implementation; and another person handled customer acquisition. The client profile is very clear—companies of 50 to 300 people with no strong technical team, or poor alignment between business and technology. Pricing starts at 58,000 yuan to transform one business process; for two processes plus consulting and training, they charge 129,000 to 159,000 yuan per year.

"The best people to do FDE are the OPC (one-person company) crowd, because from sales to implementation and delivery, it's all one person," shared a netizen on a social media platform.

Zhang Jian has also noticed a trend: a group of programmers with more than ten years of development experience are transitioning into "one-person companies." They personally complete only 30% of the architecture design and technology selection, leaving the remaining 70% of execution work to AI.

Source: AI-generated