This article is excerpted from *Yunqi Strategic Reference*, a publication jointly curated by Alibaba Cloud and TMTPost. Its purpose is to present the technical explorations and business practices of industry pioneers, and to discuss and exchange ideas with "digital-intelligence pioneers" who are thinking about the same questions. We hope these contents will inspire you.
In the 1990s, the byword for enterprise informatization was ERP—represented by SAP and Oracle—framing business processes with a set of standardized commercial software suites.
Over the years, systems grew heavier and data accumulated more and more, making data flow across systems and departments extremely difficult. Back then, informatization and digitalization meant buying servers, deploying systems, and doing implementation—but by the time a system went live, the original business need might already have been missed.
Thus, another force began to grow. Over the past two decades, Palantir's most distinctive capability has not been any particular technical feature of Ontology or Foundry, but its methodology: FDE (Forward Deployed Engineer), i.e., the Forward Deployed Engineer. These engineers embed themselves at customer sites and in core business operations, turning vague requirements, scattered data, and unactionable business goals into standardized tools and methodologies, delivering to customers with quantifiable business results.
From informatization to digitalization and intelligence, and up to the large model boom of the last three or four years. We see that although large model technology is beginning to penetrate enterprise scenarios, many enterprises still don't know where it should be applied, how to apply it, or how to run evals after applying it. The gap between large models as a showpiece and large models as productivity is far deeper than imagined.
Thus, the value of the FDE has been rediscovered. Many people wonder: isn't this just on-site staffing? Outsourcing? After-sales operations and maintenance? In fact, no. Today's FDE differs from the model of twenty years ago and must transform into a lighter-weight expert model: versatile talent who understands the business, the models, and data governance, carrying methodologies and reusable toolchains, rapidly validating, rapidly iterating, and rapidly delivering measurable results at the customer site.
The value output of the FDE in the Chinese enterprise market is also quietly changing.
China's informatization journey differs from that of other countries, and China-specific customer needs and market conditions have led to wide variation in the enterprise information systems that were built. This creates a problem: relying on heavily customized system integration projects, clients are used to turnkey projects, while vendors are used to dispatching implementation teams of dozens of people on-site for half a year—a heavy model with long cycles and high costs.
Today, an FDE must not only have solid technical engineering capabilities, but also the ability to reliably run systems on real data, and even more importantly, clear judgment of customer scenarios—doing the math and turning vague business demands into actionable solutions.
As Palantir CEO Alex Karp has said, large models are not unimportant, but for at least the next few years, the value lies in implementation. The FDE has become the last mile in converting the technical value of large models into business value.
Why is everyone deploying FDEs?
At present, the AI industry is seeing a wave of FDE deployment. Internationally, the OpenAI Deployment Company was founded in May 2026, led by TPG with over $4 billion raised. In the same month, Anthropic formed an AI services joint venture of roughly $1.5 billion with Blackstone Group, Hellman & Friedman, and Goldman Sachs Group. Both companies adopt an embedded engineer model and grant ownership of the final results, and the latter's deployment-led business grew 85% year over year. Within weeks, leading cloud vendors followed suit.
OpenAI, Anthropic, Microsoft, Amazon, and others have invested over $9 billion in FDE deployment projects, with thousands of engineers worldwide already dispatched into various enterprise customers. Renowned consulting firms such as Accenture, Deloitte, and McKinsey have retrained tens of thousands of internal employees and rebuilt their FDE practice systems. In addition, a group of professional, neutral FDE service companies has emerged, such as Distyl AI, Tribe AI, and Plank.
Indeed data shows that in the year up to April 2026, FDE job postings grew 729%, with salaries for these positions typically above $200,000 to $300,000 per year. In less than a year, the FDE went from a job title at Palantir to an industry worth hundreds of millions of dollars—enough to prove the potential of its commercial value.
Of course, the popularity of the FDE model is, in essence, an inevitable choice born of the need to solve the difficulty of enterprise AI implementation as AI moves from technical exploration to industrialized production.
Multiple research institutions point out that most enterprise AI projects are never actually put into use, and this phenomenon is worsening. Among them, S&P Global data shows an AI abandonment rate as high as 42%, up from just 17% a year earlier; Gartner predicts that by 2027, more than 40% of agent projects will be canceled; a report jointly released by IDC and Lenovo (2025) shows that 88% of AI proof-of-concept projects never reach production. The problem is not model capability, but the integration, evaluation, and ownership of the implicit reasoning behind enterprise decisions.
Judging from the current practices of Palantir and similar companies, the FDE is fundamentally different from traditional on-site engineers. Under the traditional software delivery model, vendors typically adopt waterfall development and standardized product output, with customers passively adapting to software features.
For customers such as governments and state-owned enterprises, the high degree of customization also leads to extremely high costs. FDEs possess strong technical penetration and business deconstruction capabilities, able to go deep into customer sites, share the same data environment and the same business KPI with customers, and rapidly test and iterate adaptations in real business flows.
The deeper significance of the FDE lies in effectively breaking the deadlock of "those who understand technology don't understand the business, and those who understand the business don't understand technology." By moving engineering capability forward to the decision-making front line, technology enablement transforms from the vendor's standardized delivery into an organizational capability co-created with customers and growing on demand.
Therefore, the FDE has become indispensable strategic human capital for enterprises implementing AI.
FDE implementation in the Chinese market still faces challenges
Currently, domestic enterprises are also rushing to learn Palantir's FDE model. Some say this cannot be learned in China, because Palantir's customers are all the US military and government. Yet in the most recent quarter, Palantir's civilian commercial revenue grew 149% year over year, which at least shows one thing: while different customer environments may call for different product features and interfaces, the core methodology for users should not differ.
But what cannot be ignored is that FDE implementation in China still faces many underestimated challenges.
One lies in the mismatch of business models.
For a long time, China's enterprise services market has been accustomed to outsourcing models priced by person-days. Clients expect on-site engineers to be on call to complete assigned tasks, rather than to deeply participate in business decisions and provide strategic advice. As a result, when FDEs enter customer sites, they are easily treated as premium outsourced labor, and their business insight and solution design capabilities—which they should be exercising—are drowned out by trivial daily demands.
Some independent FDEs point out that the biggest problem at present is that labor costs are still too high: doing a project requires going deep inside an enterprise, figuring out an industry on one's own, and teaching employees how to use AI. Meanwhile, within enterprises there is often a conflict between frontline business needs and IT department responsibilities—either the requirements are too specific and change too fast, and FDEs often get stuck in back-and-forth communication and repeated confirmations, greatly compressing the time actually available for technical work.
On the other hand, although the FDE is a capability Palantir has honed for twenty years, it remains an emerging role in today's AI era, and a mature talent development system and industry consensus have not yet formed in China. Every company's FDE team is groping for its own working methodology, and repeated missteps are common.
A Gartner analyst predicts that 70% of enterprises will eventually abandon building AI agents through FDE-led projects, due to high vendor costs and a lack of internal skills—especially with engineers who only optimize platforms for the enterprise and leave nothing behind. The recommendation is to form neutral vendor teams that place code in the customer's codebase and document all patterns, so that internal teams can take full control.
Over the past two years, several cloud computing companies and large model firms in China, including Alibaba Cloud, have begun systematically building FDE teams. Some leading vendors have started to consolidate the deployment playbooks, tuning guides, and troubleshooting solutions accumulated by their FDE teams across various industry scenarios into open documents for industry-wide sharing and reference. An increasing number of experienced consultants and R&D engineers are also transitioning to become independent FDEs. An FDE methodology suited to China's industrial landscape is gradually taking shape through practice.
In summary, an FDE needs the following characteristics: first, a deep understanding of the customer's industry business logic and the ability to translate it into business code; second, hands-on engineering capability to quickly experiment and adjust in real data environments; third, the narrative ability to convert technical value into ROI stories that business leaders can understand.
Alibaba Cloud's FDE Capabilities and Practices
Currently, the state of FDE adoption in China can be described as: the direction is right, but it is far from the stage where success stories come easily.
From the customer's perspective, having a model does not mean producing results. Beneath the model, an execution layer is needed to transform technical capability into business value. This execution layer can rely on the customer itself or on a third-party FDE team. The model, the FDE, the data, and the context together determine the final outcome of an enterprise's AI adoption.
Alibaba Cloud's FDE capabilities were honed through its customers' repeated scenario-specific needs, built on its own mature full-stack AI cloud technology capabilities and solutions experience.
Take the CCTV Sports app's World Cup "AI Viewing Assistant" as an example. The World Cup has a schedule that waits for no project process: matches happen continuously, schedules, players, and scores change at any moment, and the timing of user questions may be out of sync with data updates. The system must answer promptly while relying on official sources to prevent the large model from fabricating content. The Alibaba Cloud FDE team pushed the application into production in about a week, letting real users and real-time data enter validation as early as possible.
Going live in a week is certainly fast, but speed is not the whole point of this case.
Traditional projects prefer to eliminate uncertainty as much as possible before delivery, whereas this project turned real usage itself into part of the validation. The system went live with the matches first, then continuously optimized and adjusted based on issues encountered on the ground, as long as outcomes remained controllable.
In the PetroChina Kunlun Digital Intelligence project, the FDE team, together with Kunlun Digital Intelligence, went to the customer's front line for research. Based on ten narrowed-down scenarios, they ultimately chose anomaly alerts for the atmospheric and vacuum distillation unit—the first process in energy refining—as the breakthrough point for AI adoption. Alibaba Cloud adopted a general-purpose large model combined with a traditional time-series database for analysis and prediction, providing suggestions to operators, and then used Skill distillation to incorporate decades of tacit experience from veteran masters into the model. This approach not only greatly reduces the workload of manual review, but also enables frontline operators to genuinely and continuously use AI effectively.
The atmospheric and vacuum distillation unit is the first process in refining production, and all subsequent stages are built on its stable operation. When AI applications enter such systems, they face not only whether a single answer is accurate, but also continuous operation, anomaly handling, and coordination with existing systems. According to customer feedback relayed by Alibaba Cloud, this was the first AI application running 24 hours a day in a production system within its infrastructure. At this point, what the FDE needs to prove is no longer whether the model can answer, but whether AI can stay in the core business.
In the collaboration with Peking University International Hospital, the FDE team completed full-chain delivery of AI agents for 7 major patient service scenarios in just 3 weeks—from architecture design and joint debugging of 24 business system interfaces to the gray-scale launch of the WeChat mini-program "Xiao'an" AI health assistant, with intent understanding and recall accuracy both reaching 90%, enabling 1.5 million patients to enjoy 7×24-hour one-stop intelligent services covering "registration - guidance - report interpretation - follow-up." It validated the MaaS delivery model of "small-scope rapid deployment, standardized solutions replicated across hospitals," providing a referenceable implementation sample for medical AI patient services to move from single-point pilots to large-scale, inclusive adoption.
Through multiple real-world engagements, Alibaba Cloud has built a relatively complete FDE capability model: technical engineering capability, communication capability, scenario understanding capability, and learning and adaptation capability.
For example, on the technical engineering side, one must understand large models (when to fine-tune, when to post-train, when to use LoRA, etc.), even be able to build models, write Python code, be familiar with mainstream open-source frameworks, and possess a certain ability to evaluate models and AI agents. On the soft-skill side such as communication, one needs requirements understanding, solution design, and execution capabilities—able to handle vague directions given by customers as well as identify and correct customers' mistaken technical routes.
On scenario understanding, the three projects above certainly do not present a standard route that all enterprises can copy, but they do reveal some common signals of high-value scenarios: the AI application can be Always On, usage continuously generates new high-quality data, and after the foundation model is upgraded, business results continue to benefit.
In addition, the Alibaba Cloud FDE team has a criterion for judging commercial value: the business value must exceed "total token cost × 10" to be worth doing (total token cost includes all expenses such as labor and computing resources); otherwise, it is regarded as a "flash in the pan" with no long-term significance. Moreover, even at the POC stage, a key number must be answered: what is the daily throughput? If this value cannot be estimated, the scenario is not feasible.
FDEs' learning capability also needs to keep up quickly. On one hand, customer needs and technologies are changing rapidly, and solutions are constantly revised along the way; on the other hand, as foundation model capabilities improve rapidly, FDEs also have their own career development path.
For a broader range of industry customers, the FDE methodology is being validated as a replicable implementation paradigm.
For example, the data flywheel logic from the media industry can be replicated to all content platforms; the Skill distillation path from the energy industry can be transferred to manufacturing, electric power, and other industrial scenarios; the delivery model of "small-scope rapid deployment, standardized solutions replicated across hospitals" from healthcare can be extended to education, government affairs, and other public service scenarios... But the core of replication is not copying technical solutions, but rather that methodology which remains constant amid all changes.
Conclusion
China's rich industrial categories, diverse enterprise organizational forms, and complex data environments mean that FDEs must cross the river by feeling the stones in local practice. Perhaps it is precisely this exploration full of uncertainty that gives FDEs in China the opportunity to truly evolve from a job title into a new hybrid profession that integrates technology, consulting, training, and organizational change. Only when more and more enterprises gain the ability to independently harness AI thanks to FDEs will this role have fulfilled its mission.
As the era of large models competing on implementation quietly arrives, the Alibaba Cloud FDE system reveals a simple truth: the core of AI implementation lies not in the model layer, but in the execution layer beneath the model. This execution layer requires not only engineering hard power, but also reverence for industry know-how, insight into organizational pain points, and the courage to say "no" to unreasonable customer demands.
As Palantir's market-cap myth is proven again and again, the FDE practices of "Alibaba Cloud and its peers" may well be the key step for China's large model industry to mature and move from showing off technology to producing real productivity.
This article is excerpted from the 22nd issue of "Yunqi Strategic Review"
Scan the QR code to view the latest magazine
↓↓
