Game Theory in FDE: Tech Giants Seek Insights, Enterprises Retain Control

Recently, in a conference room at Mengniu Dairy's headquarters, employees from various business departments—including sales, R&D, and supply chain—took turns demonstrating how they used WorkBuddy to transform their daily work and improve efficiency.

Beyond the competition, Mengniu is also considering further training some of these employees as Forward Deployed Engineers (FDEs) to help Agents truly integrate into corporate workflows.

Similar initiatives are not limited to Mengniu. COSCO SHIPPING Specialized Carriers began implementing an internal FDE model this year, and Cosco Ship Hold has also incorporated FDEs into its talent development strategy.

On the other side, AI vendors such as Tencent and Alibaba are sending increasing numbers of FDEs into enterprises. They hope these engineers will delve into business operations to get Agents running effectively and bring feedback on issues exposed at the frontline back to product teams for iteration.

Subtle boundaries are emerging: On one hand, AI vendors use FDEs to penetrate deeper into enterprises' core processes, aiming to expose Agents to more real-world workflows;

On the other hand, enterprises are cultivating their own FDEs, hoping to retain greater internal control over judgments regarding business, process, and data boundaries while opening up scenarios.

Where this division of labor will eventually stabilize remains unanswered.

The rise of FDEs indicates that competition in the office Agent sector has moved beyond product features to a stage where tech giants and enterprises are reorganizing their collaborative relationships.

FDEs Take the Lead

Since the beginning of this year, office Agents have become one of the most crowded battlegrounds in the AI competition among major tech firms.

Tencent launched the enterprise version of WorkBuddy in June. In July, Alibaba integrated QoderWork, Wukong, and MuleRun into "Qianwen Office" and began testing. ByteDance has also further integrated teams related to Feishu and Doubao.

Although product forms differ, the direction is converging: AI is starting to directly read files, create spreadsheets, and invoke software, further integrating into enterprises' internal knowledge, systems, and processes.

However, truly bringing AI into enterprises is not just about installing an Agent on a computer.

Apart from IT departments, business personnel clearly understand their pain points but may not grasp the capability boundaries of models and Agents. They also struggle to determine how to break down business tasks into executable AI actions. Conversely, while AI companies understand AI best, they cannot inherently know the intricacies of a specific enterprise's order flows, for example.

Therefore, a "translator" is needed between product and business—someone who can translate enterprise problems into AI-solvable tasks and apply business understanding to sustain AI iteration.

This has pushed FDEs to the frontline.

FDE stands for Forward Deployed Engineer. Compared to traditional pre-sales or implementation roles, FDEs are closer to customers' real businesses: they enter the frontline to identify scenarios worthy of AI transformation, assemble models, Agents, data, and systems into functional workflows, and bring issues exposed during the process back to the product side.

Palantir is one company that has taken this model to the extreme, with its FDEs long embedded in customer sites to solve specific problems, subsequently embedding recurring business rules, data relationships, and requirements back into the product system.

In this round of office Agent competition, this approach has more direct commercial significance: For an Agent to truly enter an enterprise's daily work, someone must go to the frontline to find scenarios, connect systems, modify processes, and drive employees to actually use the Agent.

Domestic tech giants have started offering high salaries for FDEs. Public recruitment information shows that ByteDance's "Doubao AI Large Model FDE" positions offer monthly salaries of 35,000 to 70,000 yuan with 15 months' pay, reaching an annual package of approximately 1.05 million yuan. Ant Digital Technologies' B-side FDEs earn 40,000 to 60,000 yuan monthly with 15 months' pay, while the head of FDEs at Zhipu AI commands a monthly salary of 60,000 to 80,000 yuan.

Behind the high salaries, FDEs are bearing more than just project delivery.

As office Agents begin operating browsers, reading documents, and invoking enterprise systems, details such as how tasks are decomposed, where tool invocation fails, and why users rework tasks form complete feedback loops. For AI companies, entering these workflows means their products can continuously expose issues in real tasks, allowing for ongoing optimization of the Harness, models, and products.

Thus, competition around office Agents is extending from product capabilities to real-world workflows. FDEs stand at the end closest to enterprise business, responsible for pushing Agents in and bringing back tasks and issues from the field.

But as AI companies push deeper into enterprises, another change is underway.

Tech Giants Move In, Enterprises Cultivate Internally

According to recent on-site visits by Wall Street News · All-Weather Tech, some companies have begun cultivating their own "FDEs" while introducing office Agents.

Mengniu is one such example. On August 11, All-Weather Tech observed 45 employees participating in an AI application roadshow for WorkBuddy at Mengniu's headquarters.

IT staff were not the main participants in this competition. Mengniu selected 200 "AI Pioneers" from its 28 primary business units, with only three from the Digital Innovation Department. The majority came from business departments such as sales, R&D, and supply chain.

Mengniu intentionally lowered the proportion of IT personnel primarily because business staff "bring their own scenarios." They know best which daily tasks are repetitive, which processes are most time-consuming, and which links are most error-prone. However, these individuals previously might not have known what models and Agents could achieve, nor did they necessarily possess the ability to redesign workflows using AI.

Mengniu chose to train business personnel first, then seek parts of real work that could be transformed by AI.

According to Mengniu, the company plans to cultivate a group of its own "FDEs" from these business staff and establish an L1 to L3 AI Pioneer certification system. L3 requires not only proficiency in using AI but also the ability to understand business needs, build solutions, and continuously monitor results. Meanwhile, Mengniu's digital technology team has already established a dedicated AI FDE team internally.

Behind this lies a practical cost calculation.

AI demands within large enterprises are often highly dispersed, with new scenarios constantly emerging in sales, R&D, supply chain, finance, and other departments. If external engineers must re-understand the business, build solutions, and connect systems for every new scenario, delivery costs become difficult to amortize as scenarios multiply.

Xu Feixiong, head of Mengniu's Low-Temperature New Retail business, admitted that business teams bear performance pressure and cannot simply add a batch of IT staff for AI efficiency gains. More often, existing business personnel must learn to use and modify AI themselves.

But the deeper significance may lie in the desire to embed this capability within the enterprise's own system. As Agents begin to enter core processes like orders, supply chain, R&D, and finance, FDEs are getting closer to the enterprise's own data and business assets. It is difficult to leave control of this capability entirely to external teams in the long run.

Of course, security and data boundaries are also within the considerations of tech giants.

Min Liming, an expert in Tencent's Smart Retail industry solutions, told All-Weather Tech that the previous phase focused mainly on getting employees to use WorkBuddy to solve personal repetitive tasks. The next step enters the "deep water zone."

"The next phase enters the deep water zone, which is frankly very challenging," Min stated. Tencent will jointly evaluate with Mengniu's digital technology team which systems and data can be opened to AI, prioritizing those with the highest data value, and gradually connect them via methods like MCP. Meanwhile, model calls must pass through Mengniu's internal large model security audit gateway, ensuring data connections remain within secure boundaries.

Mengniu is not an isolated case.

COSCO SHIPPING Specialized Carriers began implementing an internal FDE model in March this year, directly assigning 18 digital professionals to reside at the business frontline. They participate in business meetings, visit posts, and streamline specific processes such as bill of lading handling, voyage scheduling, container management, and semi-submersible ship stowage, breaking down barriers between "technology" and "business." Cosco Ship Hold has also written FDEs into its corporate talent development direction for 2026.

These changes outline a more subtle relationship in the implementation of enterprise AI: Tech giants use FDEs to go deeper into enterprises, while enterprises hope to keep the most core business understanding in their own hands.

Co-construction: Perhaps a More Realistic Solution

In this light, FDEs do not necessarily require a choice between "vendor dispatch" and "enterprise self-build."

As Agents enter deeper business processes, a more realistic cooperation method may be for both sides to jointly build FDE capabilities.

AI vendors possess engineering capabilities such as Agent Harnesses but need real workflows to continuously expose issues. Enterprises understand their own business processes, data permissions, and organizational rules best but need external technology to make these scenarios truly operational.

Neither side can easily complete the full chain from scenario discovery and system integration to continuous iteration alone.

For AI vendors, this co-construction has another layer of practical significance.

Office Agents must ultimately remain a product business, not a labor-intensive business reliant on constantly increasing onsite engineers for expansion.

The more important value of FDEs entering enterprises lies in first identifying problems that standard products cannot temporarily cover, and then gradually embedding these issues into the Harness, Skills, connectors, and product capabilities.

Ideally, while the same type of problem requires repeated debugging by FDEs at the first customer, more parts should be directly reusable by the second and third customers.

Only in this way can frontline delivery continuously feed back into the product, making the marginal returns of office Agents "thicker" as the number of customers increases.

From this perspective, co-building FDEs may first suit large enterprises like Mengniu. Large enterprises have more complex systems, longer processes, and more real scenarios, and are better positioned to allocate their own business, data, and technical personnel for long-term collaboration with AI vendors.

Large enterprises can also serve as a "test bed" for the continued productization of office Agents. Under the premise of legal compliance and clear data boundaries, specific problems solved jointly by FDEs and enterprises can continue to be embedded into Harnesses, Skills, and product capabilities.

What Agents learn thereby is not the "family secrets" of any single company, but methods for handling a class of enterprise problems.

Processes that today require FDEs to repeatedly run through with a large enterprise, if they can be directly handled by the product in the future, will not require replicating the same heavy delivery model for small and medium-sized enterprises.

To some extent, co-building FDEs may be the more realistic middle ground in this subtle game: AI vendors continue to move forward, while enterprises do not have to completely surrender themselves.

Risk Warning and Disclaimer

The market involves risks; investment requires caution. This article does not constitute personal investment advice, nor does it consider individual users' specific investment goals, financial status, or needs. Users should consider whether any opinions, views, or conclusions in this article align with their specific circumstances. Investment based on this content is at the user's own risk.