From Trending Topic to Target: What Has FDE Experienced?

‘I don’t want to talk about FDE anymore,’ Huo Jia, Vice President of Alibaba Cloud Intelligence Group, told Shuzhi Qianxian at the Cloud Town Conference. ‘Now, anything you say can be nitpicked.’ At the height of FDE’s popularity, Shuzhi Qianxian reached out to several industry veterans. Some believed FDE was overly idealistic, others felt it was too early to discuss in-depth, and some claimed FDE was already outdated.

In just two months, FDE transformed from a trending topic into a target. On one side, a wave of supporters emerged: tech companies rushed to form FDE teams, and the Ministry of Industry and Information Technology issued a directive to cultivate FDE talent. Maimai data revealed a 21-fold surge in FDE job postings in the first half of the year. Proponents argue that for models to be implemented, someone must bring them into enterprise settings—and FDEs are the ‘enablers of large model capabilities.’

On the other side, skepticism arose. ‘FDE in China is just a new term for on-site grunts,’ some joked, suggesting that rebranding implementation engineers could justify raising daily rates from 1,500 to 15,000 yuan. Critics argue that on-site customization is already commonplace in China. A cloud vendor executive told Shuzhi Qianxian that FDEs ‘don’t add anything new’—they are simply client-facing service teams with a trendier name.

Huo Jia noted that FDEs in China risk becoming a ‘distorted concept,’ dismissed as old wine in new bottles or a wolf in sheep’s clothing. ‘But unless you’re on the industrial frontlines, it’s hard to answer these questions clearly,’ he said. Liu Song, Vice President of PingCAP, was blunt with Shuzhi Qianxian: discussions about FDE are still in their early stages, and large-scale validation has yet to begin. ‘All predictions and judgments carry some truth but also some subjectivity,’ he said.

So, what exactly is FDE? And why has it surged in popularity this year?

01 FDE: An ‘Old Model’ Amplified by AI

FDE, or Frontline Deployment Engineer, involves sending engineers to customer sites to collaborate on solution implementation. Instead of delivering reports, they deliver functional systems while feeding back frontline insights to refine the company’s product capabilities.

‘Some argue that FDE is just custom development or outsourcing under a new name,’ Wang Peng, a senior expert at Tencent Research Institute and author of the FDE report, told Shuzhi Qianxian. He emphasized a fundamental difference: in the past, project experience remained trapped in experts’ minds, requirements documents, and code, with extremely low transfer efficiency. Today, AI can solidify personal expertise through bidirectional distillation mechanisms and Skills, while also precipitate (chén diàn, ‘ precipitate ’ meaning ‘crystallizing’) business knowledge into ‘ontologies.’ ‘AI makes this scalable,’ Wang said.

Chart Source: Tencent Research Institute’s ‘FDE Model: Industry Observations and Practices’

‘FDE is a model pioneered by software company Palantir, not by internet giants,’ said Zuo Chun, Chairman of Sinosoft, at an AI roundtable. FDE, he explained, was designed for ‘constraint-based software.’

Software broadly falls into two categories: recording software documents enterprise activities (e.g., ERP, CRM, financial systems, banking core systems), with relatively stable and standardized requirements. Constraint-based software, however, includes military command, aerospace mission planning, factory production scheduling (APS), and financial risk control. It aims to find optimal action plans under a set of resource, rule, and risk constraints, requiring flexibility.

Palantir initially served military clients, creating a model where traditional consulting delivered reports, while Palantir delivered functional prototypes—a concept Zuo Chun called ‘the unity of knowledge and action.’ This approach earned client acclaim and gave rise to FDE.

‘FDE is clearly a team effort,’ Zuo Chun said. Palantir divided roles into two: Echo, responsible for understanding business and defining problems, and Delta, responsible for translating solutions into operational systems. In Zuo’s view, the recent hype around Agents aligns methodologically with this approach, except Delta’s role has been vastly accelerated by AI. While Palantir once relied on IBM low-code tools, today’s AI-driven interaction and experimentation cycles are incomparably faster.

Wan Ruyi, Vice President of Chinasoft International and Dean of its AIGC Research Institute, believes AI has ‘amplified’ certain strengths of FDE. The most significant is ‘upfront effectiveness’: AI coding drastically shortens the path from idea to result. Previously, government and enterprise projects followed a ‘problem-driven’ then ‘test-driven’ approach. Now, with large models producing probabilistic outputs, benchmarks and test sets must be defined upfront—a requirement absent in traditional FDE.

‘FDE operates across multiple dimensions,’ said Li Gang, Vice President and CTO of Digital China Group. It can be a role, a new organizational unit (still debated internally at Digital China as a physical or virtual team), a new process (redefining boundaries between presales and product managers), or most critically, a business model. ‘Without transformation, companies like Digital China or Chinasoft have no interest,’ Li said. He believes FDE’s initial focus should be building ‘super teams.’

02 Why Has FDE Suddenly Gained Traction?

When Wang Peng began writing his FDE report in October last year, he found no domestic companies hiring for such roles, while the U.S. had already seen a surge in FDE-focused firms and talent recruitment. ‘They spotted this half a year earlier than us,’ Wang said.

Chart Source: a16z’s ‘2026 FDE Role Study’

Why did FDE explode in popularity this year? ‘Because 2024 marks the first year of AI application deployment,’ Huo Jia said. AI has evolved from a chat and entertainment tool into an early-stage productivity enhancer. Since September last year, Huo observed, companies have begun seriously considering how to integrate AI into core processes. But this is exceedingly difficult: AI talent is scarce, proven implementation cases are rare, and traditional IT project management (e.g., Harness Engineering, Loop Engineering) struggles to adapt. A highly agile development approach is needed.

Alibaba Cloud’s FDE model emerged from on-the-ground experimentation. During last year’s Cloud Town Conference, Huo Jia met with Kunlun Digital Intelligence, which aimed to integrate AI into core production processes. After visiting multiple client sites, Alibaba Cloud’s solution team realized that initial scenario assumptions often failed to materialize. Even when scenarios were viable, projects stalled without quick results.

Traditional approaches no longer worked. Instead, the team collaborated with clients to validate scenarios iteratively—a departure from the conventional presales phase of defining requirements upfront. Clients received not PPTs but functional demos, rapidly evolved into production-grade systems.

This approach worked. During a project review late last year, Huo Jia realized: ‘This feels like FDE.’ Kunlun Digital Intelligence agreed. Many clients faced three common bottlenecks: identifying viable scenarios, validating technical feasibility and ROI, and scaling pilots into production systems. Given strong client demand, Alibaba formally discussed FDE at its January 2024 strategy meeting and decided to ‘give it a try.’

Broader industry trends align with this shift. Liu Song of PingCAP noted that after March 2024, a consensus emerged: enterprise-grade productivity is AI’s main frontier. Large manufacturers, financial firms, and pharmaceutical companies are seriously advancing core operations into the ‘enterprise agent era.’

However, moving from individual to enterprise efficiency requires overcoming a hurdle. Wang Peng observed that some companies pursued ‘token maximization,’ distributing tokens liberally to boost individual productivity. While effective initially, enterprises saw no long-term gains. Liu Song’s team fell into this trap: in November 2023, their 300-person R&D team consumed massive tokens without achieving enterprise-wide efficiency gains. A June 2024 client survey revealed that Chinese financial institutions using vibe coding for over a year rarely exceeded 15% overall efficiency improvements.

Wang Peng attributed this to two barriers: organizational silos and business process fragmentation. In his view, FDE must break through these walls—it’s far more than custom development. A true Echo can articulate to client leadership how to reconstruct the enterprise using AI-native methods, a skill 99.9% of self-proclaimed FDEs lack.

Liu Song explained that AI’s rapid development caused functional overlaps and conflicts, necessitating a shift from ‘mass vibe coding’ to collaboration: unifying teams under one project and channel, and empowering Agents to autonomously drive most tasks. Under this multi-agent collaboration model, new application efficiency improved roughly 10-fold, while legacy applications saw ~4x gains due to integration with existing systems. He noted that internet firms like ByteDance broadly agree.

Liu Song believes that over time, serious, large-scale, complex tasks and organizations will adopt multi-agent systems. The FDE’s key role isn’t to complete systems for clients but to ‘teach them to fish’: start with a minimal viable scenario, then let clients independently develop broader applications.

Gartner highlights the key to ‘teaching to fish’: retaining systems to crystallize product capabilities.

03 FDE Pricing: The Industry’s Biggest Mystery

‘It’s not a question of quality but survival,’ Wan Ruyi said of FDE’s business model.

Liu Song recalled reading an industry article arguing that FDE should abandon per-diem pricing in favor of outcome-based payments. However, some industry figures remain cautious. ‘Domestic enterprise software defaults to per-diem models because of project-based structures: the client’s IT department owns the solution, defines requirements, and sets architectures, reducing vendors to ‘labor providers’ priced by man-days,’ one expert explained. ‘Only when the counterparty shifts to business departments—with vendors defining problems and owning results—can value-based pricing work.’""Huo Jia’s stance on outcome-based payments evolved. He once viewed feature-based SaaS pricing as prone to commoditization and price wars, while outcome-based models offered negotiation leverage for truly capable providers. Yet he distrusted this approach, recalling failed outcome-based AI projects from previous waves. By May 2024, he realized the issue wasn’t the model but insufficient practice: ‘You don’t know where to confidently deliver results.’ However, AI advancements have boosted faith in delivery capabilities.

Wan Ruyi outlined layers of ‘outcome-based pricing’: the baseline is accuracy/adoption rates, while the ideal is improving clients’ financial metrics. Thus, FDE’s success hinges on solving hard problems. If challenges remain unsolved, clients won’t pay premiums, and vendors won’t deploy top talent, leading to mutual degradation. ‘The downward pull is strong,’ Wan warned. Without vigilance, FDE could revert to traditional custom development or outsourcing.

This explains why OpenAI’s FDE lead insists on focusing solely on core, high-value scenarios. Yet what price corresponds to high value remains unanswered. A company that built an FDE team laments pricing dilemmas: per-diem rates are ‘financially suicidal,’ as delivered value may exceed costs by 10–20x.

Wang Peng’s report states that current FDE projects typically require million-yuan investments, commonly ranging from 1–5 million yuan in pure service fees. Few clients are willing to pay this. Wang estimates that ~10% of traditional enterprises are motivated for AI transformation, with only ~10% of those willing to pay for FDE—totaling ~1% of all enterprises, primarily industry leaders.

‘Early-stage FDE costs are higher due to additional ontology development, but success in a representative industry enterprise enables rapid replication, slashing costs,’ he said. Over time, these capabilities will crystallize into products, becoming industry standards for long-tail clients.

PingCAP currently mixes pricing models: bundling products with FDE services (products charged by deployment, FDE by project complexity) or sharing business outcomes. Overall, exploration continues. ‘FDE pricing remains the industry’s biggest mystery,’ Liu Song admitted.

Another approach is ‘letting others pay.’ The industry sees ISVs seeking better models, with potential help from tech giants. Giants deploy FDEs to drive model adoption and token consumption, incentivizing ISVs to handle projects while sharing token revenue. Li Gang compared this to mainframe sales: users once demanded free software, with vendors recouping costs through hardware margins. Today, funds might come from the model side.

Huo Jia tied the issue to competitive moats. ‘For vendors, the moat is always the product,’ he said, noting that at Alibaba Cloud, driving product adoption is a key FDE team metric. Another path is long-term service contracts. Traditionally, Chinese enterprises favored project-based contracts, but Huo observed that Beijing- and Shanghai-based startups led by post-90s teams have adopted service contracts since 2021, securing long-term client commitments and scaling to nearly 100 employees profitably after proving value.

04 Why is FDE more arduous than ERP implementation?

The first step for FDE is to select the scenarios. Li Gang believes that the biggest gap currently lies in knowledge processing. "Over the past few decades, there has only been one truly remarkable product in the knowledge domain: Office. However, it only addresses knowledge recording and sharing, and falls short in processing knowledge." Large models have opened up this market, with explosive growth expected in knowledge-intensive, high-wage scenarios; coding represents another major area.

Beyond scenarios, a crucial task for FDE is to construct enterprise knowledge that has not been fully digitized in the past, similar to Palantir's ontology. Ontology may sound a bit 'mystical,' but it is essentially an 'enterprise business manual' prepared for AI: what exists within the enterprise, what each element means, how they relate to each other, and what rules must be followed.

Does ontology have to be heavy? Zuo Chun advocates for lightweight ontology. During the era of knowledge graphs, major companies built graphs at the enterprise level, "requiring hundreds of people just to build the graph library—this path is simply untenable." He believes that FDE should determine the complexity of the ontology based on specific goals, building them one goal at a time. AI has also made ontology more affordable. Wang Peng said that in the insurance industry projects they have been involved in, experts "do not manually create the ontology but only review and fine-tune it," reducing the workload by one or two orders of magnitude.

In terms of pace, Huo Jia said that the biggest difference between current FDE and traditional approaches is that FDE starts without waiting for a Statement of Work (SOW) and seeks to make collaboration as flexible as possible. "In the past, to sell man-days, everyone liked to extend project timelines; FDE aims to shorten them." In a project with a central enterprise, the initial plan was for the first launch to occur after three months. "This would have been a 100% failure." Huo Jia directly met with the chairman of the central enterprise and ultimately changed the delivery to once a month, then later to once a week.

Chart source: Tencent Research Institute's 'FDE Model Industry Observation and Practice'

Through the Pingkai Loop multi-agent collaboration platform, Liu Song's team demonstrated FDE's specific actions on-site in a project for a financial institution. The client wanted to develop a large-scale securities business software and initially tested the waters by assigning a component to be jointly developed by their in-house team and the Pingkai Xingchen team. This client had been using vibe coding for a year but encountered a series of issues such as distorted requirements, uncontrolled code quality, and lack of standards.

The team first clarified the working mechanisms and capability boundaries for the client. After the FDE team entered, they first established standards for development governance and "usability," using agents for automated quality inspection. During the requirements phase, FDE and AI jointly conducted "Socratic" questioning of the business side, probing boundary conditions and exceptional scenarios; during the development phase, each AI output did not exceed three days of work, leaving space for human oversight and using test-driven development. After a month, the component successfully completed a closed loop, with the entire system, including the knowledge base middleware MeshX, deployed at the client's site. The experience was distilled into a playbook, and the client team took over the development of the next module.

The final hurdle lies in organization. Liu Song's team once tried to promote human-agent collaboration to financial and medical clients, who requested "respect for existing management norms, business processes, and division of responsibilities." Therefore, they transformed multi-agent capabilities into a PaaS, breaking it down into atomic capabilities and embedding them into the client's existing system through interfaces. The final interface remained unchanged, while the underlying layer was replaced with a platform dominated by large models and agents.

Wang Peng analyzed that AI is essentially eliminating the friction of layered communication within enterprises, and middle management is likely to resist such changes. FDE must first persuade the top leadership. Huo Jia's experience is that both parties must function as one team: "Using the traditional client-vendor approach will likely lead to project failure." The industry has observed that, fortunately, in this round of AI implementation, many directors born around the 1970s, 1980s, and even 1990s are taking charge—"they really want to get things done."

Liu Song recently attended a meeting where many CIOs with ERP implementation experience agreed that AI implementation, like ERP implementation in the past, is a very arduous process. In his view, it is even more arduous than ERP because FDE ultimately needs to embed a digital workforce system into the client's organization. While ERP reshaped processes and data, this time, agents will ultimately reshape enterprise work methods, decision-making mechanisms, and organizational structures.

05 Who are the players in the FDE arena?

The FDE arena consists of participants from various backgrounds.

One group comprises large model companies and cloud giants who engage in FDE to drive model usage, token consumption, and cloud resources, while also valuing the feedback of frontline needs to their AI products. OpenAI and Anthropic have respectively partnered with investment firms and consulting companies to establish FDE companies, venturing into high-value scenarios like finance for validation and product development. Alibaba Cloud engages in FDE, with real-time feedback of frontline needs to product teams such as Qianwen Office and Qoder.

Another group includes product platform companies like Pingkai Xingchen, which started with the open-source database TiDB and has extended into AI in recent years, launching the multi-agent collaboration platform Loop and the knowledge base middleware MeshX. They use FDE to validate and implement their multi-agent product systems in key industries such as finance, healthcare, and manufacturing. "In theory, early on, we help clients establish benchmark scenarios and use Pingkai Loop effectively. In the future, clients will no longer buy our FDE services but only the product system. Their own FDE teams can leverage the Loop multi-agent platform to build their digital workforce systems more quickly."

There are also application-focused ISVs and large outsourcing firms developing agent platforms. For them, FDE itself is a business opportunity, with the key being whether it can enhance profit margins.

Where do FDE talents come from? Palantir-style FDEs have two roles: Echo and Delta. Wang Peng said that Palantir has a saying: those who can be Echoes are "like mini-CEOs." They must possess cross-industry and cross-domain knowledge, understanding both business and AI.

Zuo Chun also values Echoes more. In his view, this role is highly practical, and practitioners may be slightly older. More critical are initiative, learning, understanding, and communication abilities. If their knowledge is too narrow, "after a long discussion, the client still won't recognize you." Deltas can be younger, as new technologies can be acquired through training. Liu Song observed that currently, those who can competent (shoulder the responsibility) are generally mid-level or above with strong AI sense; junior personnel find it difficult to take on the role.

Teams are also formed differently. Alibaba Cloud does not distinguish between Echo and Delta, instead "pushing forward" solution architects with technical backgrounds who have long worked with clients. Huo Jia found that the ability to be a good FDE depends not only on technical understanding but also on communication skills—"whether you can listen and articulate clearly"—with the most basic aspect being telling the client whether something should be done and its cost.

Nearly everyone emphasizes initiative. FDEs who can truly solve problems are not necessarily the most experienced but may be the most proactive.

FDEs do not necessarily all come from external sources. Wang Peng observed that in AI-native companies like large model firms and autonomous driving enterprises, "everyone is an FDE," eliminating the need for third parties; traditional industries will also gradually develop such roles internally. He cited the example of "Director Xie," who went viral recently: during this year's flood season, this deputy director of the Miyun Branch of the Beijing Municipal Commission of Planning and Natural Resources purchased 1 billion tokens, spent nearly a month, and used AI to develop a flood prevention app. Warning dissemination and crowd evacuation statistics were completed in real-time through the app. "He is essentially playing the role of an FDE, which is the ideal state." In Wang Peng's view, what vendors should truly do is lower the AI and programming thresholds enough so that more and more industry experts can personally serve as FDEs.

This also raises new questions for education. FDE is difficult to categorize into existing disciplines: developing financial software requires understanding both economics and technology, but disciplinary setups can only change gradually. Therefore, companies currently rely mainly on on-the-job training. Wang Peng judged that individuals like Echoes, who require cross-disciplinary, cross-industry, and innovative thinking, are "almost impossible to cultivate" within the existing education system. He noted that an increasing number of FDE talents overseas are being cultivated from high school students: they have a solid foundational knowledge and have not yet been constrained by industry processes and habits—"China will also follow this trend."

Liu Song believes that the number of successful AI projects that can be implemented in the coming years depends on the availability of qualified FDEs.