The Last Mile of Enterprise AI: Three Camps Go Head-to-Head Here
At the 2026 World Robot Conference, when talking about FDE (Frontline Deployment Engineer) that has suddenly risen to popularity recently, Wu Minghui, CEO of Mininglamp Technology, traced the timeline back more than a decade.
"We started conducting in-depth research on this as early as 12 years ago." When Wall Street CN · All-weather Tech asked about the differences between FDE and traditional software deployment, Wu Minghui said that both scenarios require personnel to be present at the client's site, but today's FDE needs to go deeper: connecting Agents to real business scenarios on one hand, and continuously precipitating the capabilities formed on site back to the backend on the other.
Behind this trend lies new changes emerging in the delivery model of enterprise services.
In May this year, OpenAI specially set up a Deployment Company to send FDEs to enterprises, working with clients to complete 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 connect Claude to the core systems of industries including banking, aviation, insurance and manufacturing.
On August 18, Tencent Cloud also directly renamed its original "Intelligent Agent Development Platform AI Application Engineer Certification" to "ADP Frontline Deployment Engineer (FDE) Certification".
The focus of these companies on FDE points to the "last mile" problem of enterprise AI implementation.
Large tech players including OpenAI and Tencent have begun to take on more enterprise delivery work in person, sending FDEs to the site to connect models and Agent platforms to clients' systems; enterprise service providers such as Mininglamp Technology are redoing the original customized development with AI, turning industry know-how, data governance and software implementation capabilities into individual Agents; enterprises as payers have also started to cultivate their own AI pioneers from business departments such as sales, R&D and supply chain, to build internal FDE capabilities.
AI has made coding increasingly cheaper, but has turned "who understands the enterprise, defines requirements and makes Agents actually run" into a new business.
I
No Standard Answer
FDE emerged long before the era of large language models.
Palantir has long adopted a similar model to let engineers go deep into client sites, integrating complex data, software capabilities and specific businesses. The software industry has long had implementation consultants, solution architects and on-site engineers. Therefore, after FDE regained popularity in this round of AI wave, there has always been a controversy: does it bring a brand new engineering model, or just follow the old logic of the previous generation of software implementation?
"We already have a pre-sales technical support team. I think FDE is just a trendy buzzword. But if our team believes that using this title can help sell products better, I think that's fine." A CEO of a cloud vendor in South China once told All-weather Tech frankly.
Wu Minghui believes that the traditional implementation team is only responsible for delivering a set of determined software, and may not even have a direct connection with the software R&D team; while FDE needs to continuously contribute the capabilities formed at the front line to the backend, turning problems exposed at one client site into reusable assets for subsequent use.
This is also the reason why large players such as OpenAI and Tencent have begun to send engineers directly to their clients. For example, OpenAI's job definition for FDE has clearly gone beyond the boundary of traditional "technical support". Its FDEs need to take full charge of the whole process from requirement discovery, technical scope confirmation, system design to production launch, and also precipitate the verified and working methods into tools, Playbooks and Building Blocks, then bring on-site feedback back to the product and model team.
After exposing problems through real workflows, large AI companies can further optimize their Harness, models and products accordingly. But whether this can solve the "last mile" problem of enterprise AI implementation still remains uncertain.
Wu Minghui cited an example from the advertising industry: even for the FMCG sector, the processes and operation methods of P&G and Unilever will keep evolving.
In his view, the expectation that a general foundational model can adapt to all scenarios, handle all planning tasks and decompose workflows properly is a paradox. To maintain competitiveness, enterprises must continuously create new processes and operation methods, and the working methods learned today may have already changed tomorrow.
Under such a background, FDE is facing a constantly changing Context, and it is difficult to cover all scenarios for a long time with a static "industry answer".
This also leaves room for another type of companies — software service providers.
II
The Leverage Ratio Test for FDE
Since AI evolved from chatting to "getting real work done", the market has been constantly worried about the replacement of traditional SaaS. A more popular judgment in the tech circle is that as Vibe Coding continues to lower the development threshold, enterprises will be able to write code and build Agents on their own in the future, and the standard software and a large number of customized development that originally need to be purchased will be further reduced.
This directly pushes software service providers under the pressure of AI transformation. The revenue structure formed by Licenses, implementation and customized development in the past may be impacted.
But if the logic that external FDEs can exist for a long time holds true, software service providers may thus gain new living space.
Enterprises can get models and code more and more easily, but they still need someone to understand their business changes, system structures and delivery constraints for a long time. For traditional software service providers, the industry insights, system experience and delivery capabilities accumulated through years of presence at client sites just get a chance to be translated into Agent capabilities.
Mininglamp Technology is one of such samples.
Its core businesses in the past focused on marketing intelligence, operational intelligence, data governance and enterprise software services. Entering the Agent era, it began to develop Agentic Services, aiming to further convert its original software implementation, data governance and industry service capabilities into customized Agent deliveries.
Mininglamp's recent plan to acquire control of Pulead Software also follows this logic. Pulead has long served large group enterprises in sectors such as petroleum and petrochemical, coal and power, with strong customized development and implementation attributes.
Wu Minghui said that Mininglamp plans to first bring FDEs, Agents and related infrastructure into Pulead, and then Pulead will extend this set of capabilities to its existing clients.
But the key that determines whether this logic can be established is the leverage ratio of FDE.
Traditional customized software has always been a manpower-intensive business. A project usually requires a large number of on-site engineers. The more complex the requirements are, the more personnel will be invested. Revenue and manpower have maintained a strong linear relationship for a long time, with limited economies of scale.
For example, Pulead Software's revenue increased from 582 million yuan in 2021 to 825 million yuan in 2025, with a growth rate of over 40% during the period, but it "earns less and less as it does more work". Its net profit margin in 2025 was only 8.09%, down more than 15 percentage points from 2021.
The Agent era is trying to change the delivery efficiency.
Wu Minghui's judgment is that for standard software with clearly defined functions, the code itself will be more and more easily replicated by AI. Business scenarios where clients have not yet figured out how to operate may become more valuable parts in enterprise services.
"If you just write code with a laptop, there is no need to be on site," he said. The more important work for engineers when they are around clients is to communicate with clients continuously and figure out "how exactly the business should run".
According to his description, in the past, an implementation staff might spend two-thirds of their time communicating with clients and the remaining time writing code; now a large amount of coding work can be handed over to AI, and FDEs can spend more time on requirement judgment, process decomposition and result verification.
If one single person can complete more customized Agents and support more clients, the project-based business that used to rely on increasing manpower will get a chance to improve its leverage ratio.
III
Payer "Voluntarily Joins the Game"
AI not only lowers the development cost of external service providers, but also lowers the threshold for enterprises themselves to engage in Agent production.
On August 11, All-weather Tech saw 45 employees participating in an AI application roadshow at Mengniu's headquarters. They mainly came from business lines including sales, R&D, supply chain, milk source and marketing. Over the past three months, Mengniu has selected about 200 "AI pioneers" from 28 first-level business divisions, letting the people who are most familiar with the business learn to use Agents first, then go back to their own work to find parts that can be transformed by AI.
This arrangement has further evolved towards internal FDE.
Mengniu plans to establish an L1 to L3 certification system for AI pioneers. At the L3 level, employees not only need to be able to build Agents, but also need to understand business requirements, form solutions and continuously track the effects. Mengniu Digital Technology team has also set up a dedicated AI FDE team internally.
Small and medium-sized enterprises may take a more direct path. In Yanchi County, Ningxia, Zhang Lifei, General Manager of Xixianji, used Alibaba's AI programming tool Qoder to build a complete breeding management system in more than three months. National standards, academic papers, feed formulas, pen entry records and frontline breeding experience have been successively written into the software, and the system continues to extend along the processes of feeding, inventory, slaughtering, processing and quotation.
The cumulative cost of using Qoder for this system is at the level of tens of thousands of yuan, while the quotation of external customized solutions that Zhang Lifei came into contact with before may reach the level of millions of yuan.
In the past, small and medium-sized enterprises often faced two dilemmas: standard software can hardly fully fit their own business, while customized development is too expensive.
AI Coding begins to lower the development cost in the middle, making some highly customized software fall into the affordable range for small and medium-sized enterprises.
But it is undeniable that Zhang Lifei's advantages are also very special.
He has experience in R&D and has long been involved in business operations. He not only knows how to implement code, but also knows what a sheep eats every day, how to allocate feed costs, and what the lowest possible quotation for products is. Requirement confirmation, testing and usage can all be completed internally within the company, eliminating the repeated requirement transmission links in traditional customized development.
But this model also has its boundaries. As the system expands to more departments, testing, permission management, data security and long-term maintenance can hardly be concentrated on one single person. After AI reduces the cost of writing code, enterprises still need someone to continuously understand the business, maintain the system and take the final decision-making responsibility.
Wu Minghui also admitted that large enterprises will gradually internalize part of their FDE capabilities in the long run. However, as enterprises move from individual AI usage to organizational-level AI, they will still encounter problems such as post adjustment, incentive mechanism, system infrastructure and cost management, and external forces may still undertake the initial guidance work.
From this perspective, it is difficult for all parties to decide the winner in the short term.
Model and cloud vendors have underlying capabilities, hoping to send models and Agents into production with the help of FDEs; enterprise service providers such as Mininglamp have accumulated industry insights and delivery experience, hoping to convert their past know-how into Agents; large enterprises have begun to cultivate internal FDEs, and small and medium-sized enterprises may also directly build systems on their own with the help of AI Coding.
Therefore, FDE is more like a constantly moving boundary, and the position itself is far from being finalized.
When enterprises have dozens, hundreds or even more Agents, the truly important question is who will produce these Agents, who will be responsible for maintaining them, and who can turn one business transformation into the next reusable capability.
Around these issues, the enterprise AI market in the Agent era may have just started its new division of labor.
This article is from the WeChat official account "All-weather Tech" (ID: iawtmt), written by Zheng Minfang, published with authorization from 36Kr.