At the 2026 World Robot Conference, when discussing the suddenly prominent role of FDEs (Forward Deployed Engineers), the CEO of Minlu Technology, Wu Minghui, first turned the clock back over a decade.

"We were already studying this very seriously 12 years ago," Wu said when asked about the differences between FDE and traditional software deployment. Both involve entering client sites, he noted, but today's FDE requires deeper engagement: integrating Agents into real business operations while simultaneously feeding the capabilities developed on-site back into the platform.

Behind this shift lies a fundamental change in how enterprise services are delivered. In May, OpenAI established a dedicated Deployment Company, dispatching FDEs into enterprises to collaborate with clients on requirement discovery, system design, development, and launch. In June, Anthropic announced a partnership with IT service provider DXC, which plans to train tens of thousands of Claude-certified FDEs to integrate Claude into core systems across banking, aviation, insurance, and manufacturing.

On August 18, Tencent Cloud also rebranded its existing "Agent Platform AI Application Engineer Certification" directly to "ADP Forward Deployed Engineer (FDE) Certification."

These companies' focus on FDE all points to the same challenge: solving the "last mile" problem of enterprise AI implementation. Tech giants like OpenAI and Tencent have begun taking on more enterprise delivery work themselves, sending FDEs on-site to connect models and Agent platforms with client systems. Meanwhile, enterprise service providers like Minlu Technology are using AI to reimagine their traditional custom development, transforming industry know-how, data governance, and software implementation capabilities into Agents. At the same time, enterprises—the paying parties—are cultivating their own AI pioneers from business departments like sales, R&D, and supply chain, building internal FDE capabilities.

As AI makes writing code increasingly inexpensive, it has turned "who understands the enterprise, defines requirements, and makes Agents actually work" into a new business opportunity.

No Standard Answer

FDE existed long before the large language model era. Palantir was an early adopter of this approach, sending engineers deep into client sites to combine complex data, software capabilities, and specific business needs. The software industry has always had implementation consultants, solution architects, and on-site engineers. Now that FDE has regained prominence in this AI wave, a persistent debate remains: does it represent a genuinely new engineering model, or does it simply follow the old logic of previous-generation software implementation?

"We already have a pre-sales technical support team—I think FDE is just a buzzword. But if our team believes using this title helps sell products better, I think that's acceptable too," the CEO of a South China cloud vendor told us candidly.

Wu Minghui, however, argues that traditional implementation teams can simply deliver a defined software package, sometimes with no direct connection to the software R&D team. FDEs, by contrast, need to feed the capabilities developed on the front line back to the backend, turning problems exposed at one client site into reusable assets for future engagements.

This is precisely why major players like OpenAI and Tencent have started sending engineers directly to client sites. OpenAI's job definition for FDE, for instance, clearly extends beyond traditional "technical support." Their FDEs handle everything from requirement discovery and technical scope definition to system design and production launch, while also codifying proven methods into tools, playbooks, and building blocks, then feeding on-site feedback back to product and model teams.

By exposing issues through real workflows, AI giants can continue optimizing their harnesses, models, and products. But whether this can truly solve the "last mile" of enterprise AI adoption remains an open question. Wu cited the advertising industry as an example: even within the same FMCG sector, processes and strategies between companies like P&G and Unilever continue to evolve.

In his view, expecting a single general-purpose foundation model to adapt to all scenarios, handle all planning, and decompose all workflows is a paradox. Enterprises must constantly create new processes and strategies to remain competitive—what gets learned as a work method today may already be outdated tomorrow. In this context, FDEs face continuously shifting contexts, making it impossible for any static "industry answer" to provide long-term coverage.

This leaves room for another category of companies: software service providers.

The Leverage Test for FDEs

Since AI transitioned from chat to actual work, there has been persistent market concern about traditional SaaS being replaced. A popular view in the tech community is that as Vibe Coding continues lowering development barriers, enterprises will increasingly write their own code and build their own Agents, further compressing the need for standard software and extensive custom development.

This directly pressures software service providers to undergo AI transformation. Their existing revenue structures—built on licenses, implementation, and custom development—are all potentially at risk. However, if the logic supporting long-term external FDEs holds, software service providers may gain new room to survive.

Enterprises can increasingly easily access models and code, yet they still need people who understand their evolving business needs, system architecture, and delivery constraints over the long term. For traditional software service providers, the industry knowledge, system experience, and delivery capabilities accumulated through years of on-site client work happen to be perfectly positioned for retranslation into Agent capabilities.

Minlu Technology serves as one such example. Its core business has historically focused on marketing intelligence, operational intelligence, data governance, and enterprise software services. In the Agent era, it has developed Agentic Services, aiming to transform its existing software implementation, data governance, and industry service capabilities into custom Agent delivery.

Minlu's recent move to acquire control of Pulian Software follows the same logic. Pulian has long served large group enterprises in petrochemical, coal, and power sectors, with strong custom development and implementation characteristics. Wu said Minlu plans to first bring FDE, Agent, and related infrastructure into Pulian, then have Pulian extend these capabilities to its existing clients.

But the key determinant of whether this logic holds is the FDE's leverage ratio. Traditional custom software has always been a labor-intensive business. Projects typically require many engineers on-site; the more complex the requirements, the more people involved. Revenue and headcount maintain a strong linear relationship, with limited economies of scale.

Pulian Software's revenue, for example, grew from RMB 582 million in 2021 to RMB 825 million in 2025—an increase of over 40%—yet it became "increasingly unprofitable," with net margin falling to just 8.09% in 2025, down more than 15 percentage points from 2021.

The Agent era aims to change delivery efficiency. Wu's assessment is that for standard software with clearly defined functionality, the code itself will become increasingly easy for AI to replicate. Conversely, business scenarios that clients haven't yet figured out themselves may become the more valuable part of enterprise services.

"If you're just sitting with a computer writing code, there's no need to go on-site," he said. When engineers come to the client's location, the more important work is continuously communicating with clients to clarify "how the business actually works." According to his description, a traditional implementer might spend two-thirds of their time communicating with clients and the rest writing code. Now, with AI handling much of the coding work, FDEs can devote more time to requirement assessment, process decomposition, and results validation.

If the same person can complete more custom Agents and support more clients, the project-based business model that previously relied on throwing more people at problems has an opportunity to increase its leverage.

Payers Enter the Game Themselves

AI isn't just lowering development costs for external service providers—it's also lowering the barrier for enterprises to build Agents themselves.

On August 11, we observed 45 employees at Mengniu's headquarters participating in an AI application showcase. They came primarily from business lines including sales, R&D, supply chain, dairy sourcing, and marketing. Over the past three months, Mengniu selected approximately 200 "AI pioneers" from its 28 first-level business units, enabling those most familiar with the business to first learn how to use Agents, then return to their work to identify areas that could be transformed by AI.

This arrangement is already evolving toward internal FDE capabilities. Mengniu plans to establish L1 through L3 certification for AI pioneers. At L3, employees must not only build Agents but also understand business requirements, develop solutions, and continuously track results. Mengniu's digital technology team has also established a dedicated internal AI FDE team.

Small and medium-sized enterprises may take an even more direct path. In Yanchi County, Ningxia, Zhang Lifei, General Manager of Xixianji, used Alibaba's AI coding tool Qoder to build a livestock management system in just over three months. National standards, research papers, feed formulas, herd entry records, and frontline breeding experience were progressively written into the software, with the system extending across feeding, inventory, slaughter, processing, and pricing functions.

The total cost of using Qoder for this system was in the tens of thousands of yuan, compared to external custom solution quotes Zhang had previously encountered that could reach millions. Small and medium-sized enterprises have traditionally been stuck between two options: standard software that doesn't perfectly fit their business, or custom development that's too expensive.

AI Coding is now compressing development costs in the middle, bringing certain highly customized software within reach of SMEs.

That said, Zhang's advantages are also somewhat unique. He has an R&D background and long-term involvement in operations—he understands both how code works and what a sheep eats daily, how feed costs are allocated, and the minimum price a product can command. Requirements, testing, and usage can all be completed internally, eliminating the requirement handoff loops typical of traditional custom development.

But this model has limits too. As the system expands across more departments, testing, permissions, data security, and long-term maintenance become increasingly difficult to centralize in one person. After AI drives down coding costs, enterprises still need people who continuously understand the business, maintain the system, and take responsibility for final decisions.

Wu Minghui also acknowledges that large enterprises will gradually internalize some FDE capabilities over time. However, as enterprises move from individual AI usage to organizational-level AI, they encounter issues around role adjustments, incentive mechanisms, system infrastructure, and cost management—areas where external forces may still handle the initial guidance.

From this perspective, no single party is likely to emerge victorious in the near term. Model and cloud providers possess foundational capabilities and hope to use FDEs to push models and Agents into production. Enterprise service providers like Minlu hold industry and delivery expertise, hoping to convert their accumulated know-how into Agents. Large enterprises are cultivating internal FDEs, while SMEs may directly leverage AI Coding to build their own systems.

FDE, therefore, resembles a boundary still in motion, with the role itself far from settled. Once enterprises begin managing dozens, hundreds, or even more Agents, the truly important questions become: who produces these Agents, who maintains them, and who can transform one business improvement into a reusable capability for the next?

Around these questions, the enterprise AI market in the Agent era may be just beginning its redivision of labor.