This article comes from the WeChat public account: 划重点KeyPoints, author: 心怡, editor: 重点君
Not long ago, a logistics company in Shenzhen approached Li Fei, a former big-tech programmer, hoping to use AI to optimize its business processes. He went to take a look: in this company of forty to fifty people, twenty or thirty of them did the same thing every day—customers would send PDFs, and staff would manually extract tracking numbers and shipping addresses and enter them into Excel. He built a demo with AI coding tools: work that took a person several minutes per document was parsed by AI in seconds, and the client signed a letter of intent on the spot. Now he spends his mornings visiting companies for research, his afternoons on-site observing, seizing gaps while salespeople take a sip of water to ask questions, and his evenings back home sorting out requirements and writing code. "I used to sit in the office waiting for others to bring me requirements; now I run all over Shenzhen."
Li Fei's new role is called FDE (Forward Deployed Engineer) in Silicon Valley—an engineer sent into the client's site to embed AI into business processes, currently one of the most sought-after types of people in the global AI industry.
In the spring of 2026, OpenAI and Anthropic almost simultaneously set up dedicated "deployment companies." Anthropic brought in Blackstone and Goldman Sachs to build an AI-native enterprise services company; OpenAI's new company has an initial investment of over $4 billion, and according to Reuters, it has promised private equity investors a 17.5% annual return over five years. A LinkedIn report from January this year showed that the number of new FDE positions surged 42-fold from 2023 to 2025, while AI engineers—themselves in the spotlight—grew only 13-fold over the same period.
The buzz quickly crossed the Pacific to China, where domestic discussion split into two camps. One camp, fixated on job postings offering million-yuan salaries, hails the FDE as "the hottest job in the AI world." The other camp asserts it won't take root in Chinese soil: clients won't pay for the exploration process, companies can't afford long-term on-site teams, and an industry where "the vendor is treated worse than a dog" can't retain versatile talent.
Many people have started debating whether the FDE can actually work in China. But this may be the wrong question.
The two camps actually share the same assumption: that the FDE must be sustained by a high-gross-margin software company, stationed at client sites for months at a time, with costs slowly recouped through platform accumulation.
That assumption doesn't necessarily hold. What has always been transplanted across the ocean is only the function—that is, the act of sending engineers to client sites to embed AI into business processes. In a different place, the function grows into a different form: in the US, it grew into a heavy model sustained by big companies; in China, it may skip that step and directly become individuals who travel light, unattached to any platform.
The most valuable knowledge hides in frontline business intuition
The FDE function itself is not controversial. 95% of enterprise AI pilots have produced no measurable financial return, even though enterprises have invested $30 to $40 billion in them. These failures rarely lose on technology.
What truly stands between AI and the business are three walls.
The first wall is language. When technical teams meet clients, their first sentence is often "What data do you have?" The business leader sitting across the table really wants to ask, "How can you help me earn five more points?" The traditional solution is to insert a product manager in between as a go-between—but every transmission distorts the message.
The second wall is more hidden. The most valuable knowledge in an industry mostly never gets written into databases; it lives only in the intuition of veteran practitioners. Under what circumstances a client's shipment can be released early, which equipment alarms can be ignored—rules like these have no documentation. They are like dark matter in physics: invisible, yet determining how the entire system operates. No matter how powerful the model, it cannot read what has never been written down.
The third wall is people. A joint survey of 2,400 employees and executives by enterprise AI company Writer and research firm Workplace Intelligence found that 29% of employees admitted they were secretly resisting their company's AI strategy—a figure as high as 44% among Gen Z; 76% of executives listed employee resistance as a serious threat to AI adoption. No matter how good the external engineer's technology, what they face is not just a technical problem but also the friction of people within the organization.
The entire job of the FDE is to tear down these three walls.
Ya Wen, who has worked as an FDE in Los Angeles for two years, says the job isn't responsible for any specific module of the product, but requires roughly 70% understanding of the entire product, and you dive into the client's workflow with that understanding. "Every day you face ambiguous requirements, like a tangled mess. The client is already anxious, but you have to calm down and untangle the threads of the business one by one."
Clay Bavor is co-founder of the American AI company Sierra, whose core business is customer-interaction AI for large enterprises. Bavor has described it similarly: the team must get close to how the client's business operates—its people and business model—achieving an understanding that, "while we wouldn't dare claim it's as deep as the client's, is approaching it," before a good product becomes possible.
You can sell the product without a forward deployed team; but to release the technology's value quickly and at sufficient scale, the FDE is "an important catalyst."
Palantir spent 20 years proving this path works—and how expensive it is
The FDE was born at Palantir, an American big-data analytics company co-founded by PayPal founder Peter Thiel, which has long integrated and analyzed confidential data for governments and large enterprises. When it was founded in 2003, its first clients were the CIA, the FBI, and the military. These clients couldn't articulate what they wanted; their data was confidential, their processes were confidential, even the problems they faced were confidential—the standard playbook of sending requirement documents and holding remote meetings completely failed. Palantir's solution was to send engineers directly to the site for long-term deployment, finding bottlenecks and building prototypes on the spot—hence the evolution of the FDE title.
This model was looked down upon in Silicon Valley at the time: it required large numbers of engineers, expanded extremely slowly, and looked just like outsourcing. Palantir's ledger was a cycle: problems solved by FDEs on-site were brought back to headquarters and consolidated into the platform, so the next client needed less customization; front-load the cost to win lighthouse clients, then scale through the product later. The price was time. Palantir was founded in 2003 and didn't achieve its first full-year GAAP profit until 2023—a full twenty years.
Sierra, in the AI field, "rediscovered and borrowed from Palantir" this forward deployed approach. To build its first product, the team embedded engineers inside clients; founding engineer Mahi was at one point effectively an employee of client Weight Watchers, even receiving emails during their performance review season.
Once you get through it, the returns start to materialize. In the US, the median annual salary for an FDE is about $210,000, and senior FDEs at frontier AI labs can earn total packages exceeding $780,000. Former Palantir employees have cumulatively founded more than 100 companies, raising $11.6 billion; defense tech company Anduril alone is now valued at $61 billion, earning Palantir the title of "the ultimate founder factory."
But this ledger depends heavily on the commercial soil. The decades-long habit of the American consulting industry is billing by person-day; clients assume that "sending experts to understand my business" must be paid for—Accenture's annual consulting revenue exceeds $30 billion. A mature subscription model lets vendors invest heavily in year one, recouping the money through subsequent renewals; and the capital markets can afford twenty years of patience. Palantir's former CFO Colin Anderson put it bluntly: FDE economics only work on large contracts; deals of $20,000 to $100,000 don't pencil out.
No one disputes the FDE role itself; the real controversy is who pays to sustain the people doing this work: whether anyone is willing to pay for every day of their on-site presence matters more than how strong the engineer is.
In 2026, from OpenAI and Anthropic to ByteDance, Ant Digital Technologies, and Zhipu, frontier AI companies are collectively reviving this heavy model, because AI's deployment predicament is exactly the same as what Palantir faced with the CIA. The models are ready-made, but every company's data, permissions, workflows, and unwritten rules are black boxes—the only option is to send people in. The difference this time: no one has twenty years to endure. According to financial documents obtained by Fortune, OpenAI's 2025 revenue was about $13.1 billion with operating losses exceeding $20 billion, yet its deployment company has already promised a 17.5% annual return over five years. On one side is the financial rhythm of rushing toward IPO; on the other is the physical rhythm of living at client sites for months at a time. The two rhythms are in conflict.
China's soil can't sustain the heavy model; what grows first is the light model
By Palantir's ledger, the FDE indeed doesn't work in China.
On Liepin and BOSS Zhipin, ByteDance's "Doubao AI Large Model FDE" positions are listed at a monthly salary of 35,000 to 70,000 yuan, with 15 months' pay per year, topping out at 1.05 million. Yet in the same market, fresh graduates in domestic delivery and implementation roles earn roughly 100,000 to 150,000 yuan a year, and at the veteran ERP vendor Yonyou, implementation and operations roles pay 5,000 to 8,000 yuan a month. For the same kind of work—"getting software into the customer's site"—the pay gap spans an order of magnitude. What drives the gap is less about technology than about where the job sits on the pecking order. R&D, algorithms, product, pre-sales, implementation, on-site support—on-site support ranks dead last, known in industry jargon as "the vendor is lower than a dog."
The roots of this pecking order lie in the payment habits of China's enterprise services market over the past few decades. Customers are willing to pay for tangible hardware like servers and server rooms, willing to pay for ROI figures written into contracts, but unwilling to pay for software, and even more unwilling to pay for the process of "understanding the business" itself. Software is often discounted and bundled into hardware integration deals, while consulting is treated as a free giveaway before signing a contract. The institutional system has nailed this habit in place: standard technical service projects are legally required to use the lowest-bid-wins evaluation method, where the bidder with compliant functionality and the lowest price wins. The price war has reached absurd levels—in 2017, Tencent Cloud won the Xiamen municipal government extranet cloud services project with a bid of 0.01 yuan, for a project with a budget of 4.95 million yuan. Under this system, the three months an FDE spends understanding a customer's business never appear on any feature checklist. The "fish out of water" argument gets all of this right.
But Li Fei, the man running all over Shenzhen mentioned at the start, does the math differently. "It used to take a year or two and several million yuan to build an enterprise management system. Now, with AI programming, we can deliver a custom system in two months."
Palantir needed twenty years of platform accumulation to amortize the cost of customization; AI has smashed straight through that cost. Since customers won't pay for the process, the lightweight model uses AI to compress the cost of that process to nearly zero—making exploration so cheap that no one needs to foot the bill. The role of sustaining FDEs has thus shifted from the company to AI itself.
Behind this is not just AI writing code faster; the deeper reason is the explosion of AI agent infrastructure. Coding agents have taken over the entire pipeline from writing code to testing and deployment; standard protocols like MCP let models connect directly to enterprise internal systems; and ready-made orchestration frameworks chain pre-sales demos, requirements gathering, and prototype delivery into a pipeline that one person can orchestrate. In the past, a company needed three separate teams—product, pre-sales, and implementation. Now these stages are linked together by a group of agents. One person plus a group of agents is a complete delivery line.
On delivery speed, a few numbers serve as reference points. With the help of its forward deployed engineering team, Sierra got retailer NEXT live in about six weeks and healthcare company Cigna live in about 58 days.
The reason deep customization can move this fast is tied to coding agents and the overall pace of progress. Chen Mo at an AI company in Shanghai represents a different model. He defines himself as "a general steward who handles everything except writing code and admin," combining pre-sales, product, and project management in one person. When hiring FDEs, he looks at only two things: learning ability, and the ability to grasp the essence of things. "Employees on the team who studied design or foreign languages have all performed outstandingly," said one industry practitioner. "Sitting in an office in Beijing, you cannot design a solution for a factory in Dongguan."
China didn't wait for credit cards to become widespread and jumped straight into mobile payments; likewise, China may not wait for a Palantir that can afford an FDE team to emerge, and instead jump directly to the next form: individuals not affiliated with any platform, armed with AI.
A role whose goal is to make itself obsolete
Every time an FDE solves a problem, the solution gets accumulated, and the next customer needs less customization. The better the job is done, the less the industry needs the role. Its endpoint is eliminating the need for itself.
Yawen values this experience more as a foundation for entrepreneurship. FDEs have seen too many successes and failures, and deeply understand user pain points and implementation details. "In the AI era, the barrier to entrepreneurship keeps getting lower. If you've been an FDE, you know exactly what the market lacks." This is the same logic behind Palantir alumni founding a hundred companies—except in China, this path doesn't require passing through a Palantir first.
The rise and fall of the FDE is a footnote to the redivision of labor between humans and large models. Models take everything that can be written down—code, documents, proposals—generating them cheaper by the day; humans hold onto everything that hasn't been written down—the unwritten rules of the shop floor, the antibodies of the organization, the intuition that "this order must be handled today."
The dividing line of the future may not be drawn between "technical roles" and "business roles," but more likely between those who can collaborate with AI and those who cannot. The scarce resource is shifting from the people who build models to the people who bring models into the workshop. Debating how long the FDE title can survive in China is not very meaningful; what's worth watching is the form in which the people fulfilling this function are emerging.
Li Fei has already received his Harvard acceptance letter, but he has decided not to go yet, choosing to seize the one-to-two-year window of opportunity in traditional enterprises' AI transformation. Yawen, meanwhile, says no one can accurately predict the future of any profession. "All we can do is see clearly the direction it's heading, and jump onto the next wave at the right moment."
The FDE is not the last role to be born from the redivision of labor in the AI era—it's just that this time, it stands first on the dividing line.
*Li Fei, Yawen, and Chen Mo in this article are all pseudonyms. Their experiences draw on reports from Jiemian News (China News Economy), Reuters, and Fortune, public reports released by LinkedIn, MIT, and Writer, as well as public job listings from Liepin and BOSS Zhipin.