Forward Deployed Engineers Are the 'People Who Hear the Gunfire' in AI Rollouts
Using "On Practice" to Analyze the Forward Deployed Engineer: The Person Who Hears the Gunfire in the AI Era
With the rise of AI technology, a role has been repeatedly mentioned in recent years: the Forward Deployed Engineer (FDE). This role is neither a development engineer nor a deployment engineer. Instead, FDEs are embedded directly at customer companies, integrating AI into the customer's real production scenarios.
This role solves a problem: AI models are getting stronger and stronger, but they have not realized value in customers' real business scenarios. The reason is the lack of a role responsible for feeding the user's real, production-environment, broad-sense knowledge into AI.
Some time ago I had not fully thought through some of the responsibilities of this role. Then I re-read "On Practice" from the Selected Works of Mao Zedong, reorganized my doubts, and gradually arrived at some answers.
1 Starting with "On Practice"
1.1 Three Pairs of Concepts
1.1.1 Practice and Cognition
"On Practice" states that "the production activities of human beings are the most fundamental practical activities and determine all other activities." "Human cognition mainly depends on material production activities."
The relationship between practice and cognition can be summarized in four layers. First, practice is the source of cognition. Second, practice is the motivation of cognition. Third, practice is the purpose of cognition. Fourth, practice is the standard for testing cognition.
The first layer answers where cognition comes from. The second layer means that when practice cannot move forward, practice forces us to understand the nature of things more deeply. The third layer means the ultimate purpose of improving cognition is to apply it in practice. The fourth layer says "the only standard of truth is social practice."
1.1.2 Perceptual Knowledge and Rational Knowledge
"On Practice" points out that cognition has two levels. The first level is perceptual knowledge, and the second level is rational knowledge.
Perceptual knowledge is the feelings and impressions we obtain through what we hear, see, and touch. Feelings and impressions are usually superficial, one-sided, and temporary.
Rational knowledge is what we obtain by refining and processing the feelings and impressions we have repeatedly gained, using judgment and inference, to arrive at theory that conforms to reason. This is rational knowledge.
At this point cognition rises to a new height. It is not an increase in quantity, but a change in nature. We begin to grasp the essence of things.
1.1.3 Direct Experience and Indirect Experience
Experience gained by personally participating is called direct experience. Experience others have gained, which we acquire later through learning, is called indirect experience.
"On Practice" states that "all genuine knowledge originates from direct experience." Of course, it is impossible for us to personally participate in everything. In fact, "most knowledge is indirect experience." We need to know that indirect experience also originates from the direct experience of those who personally participated.
These two kinds of experience are not opposed. If we only have direct experience, then we must handle everything personally. If we only have indirect experience, then we become book-worshippers. The two kinds of experience must be combined.
1.2 Two Intellectual Leaps
1.2.1 From Seeing to Understanding
The movement from perceptual knowledge to rational knowledge, from superficial understanding to essential understanding, is called by "On Practice" the "dynamic leap."
This step does not arise from thin air. The perceptual materials need to be concrete, genuine, and comprehensive. If the materials themselves are vague, false, or one-sided, then correct rational knowledge cannot be refined from them.
When handling complex tasks, due to various objective constraints, it is sometimes difficult to obtain comprehensive perceptual knowledge in one go, so we have to resort to the method of blind men touching the elephant. But we must also repeatedly, from multiple dimensions, and honestly touch the elephant, gradually understanding the whole picture.
1.2.2 From Understanding to Doing
When cognition reaches rationality, we have only completed half the journey. "On Practice" says: "The dynamic role of cognition is not only manifested in the dynamic leap from perceptual to rational knowledge, but more importantly, it must also be manifested in the leap from rational knowledge to revolutionary practice."
This tells us that a plan can be self-consistent in a meeting room, the logic can be closed-loop, but whether rational knowledge holds in the real world can only be known by putting it into the real world.
1.3 Cycling Repeatedly
From practice to perceptual knowledge, then to rational knowledge, and finally back to practice, that is one cognition cycle. We may go through this cycle multiple times when solving complex tasks.
The results of the current round of practice may overturn the rational knowledge of the current round. The results of the next round may overturn the rational knowledge of the previous round or several rounds before.
Is this wasted effort? No. At least we know which paths do not work, which lets us start the next round from a higher starting point.
The rational knowledge gained in each round is relative truth. "On Practice" says "the sum of countless relative truths is absolute truth."
2 The FDE Role
2.1 The Relationship Between FDE and the Development Role
The FDE is stationed at the customer's company, facing and collecting the customer's specific problems, going deep into the customer's real business processes. While talking with business staff, they write code (possibly in the customer's systems), build AI workflows, write scripts, embed agents into real business scenarios, and obtain quantifiable results.
I believe that in one cognition cycle, the FDE reaches at least the perceptual knowledge level. Being on site, they turn what they see, hear, and ask into feelings and impressions of the business. Of course, this does not mean the FDE doesn't need to refine and summarize. They also sort out problems clearly and write them into plans. But the further abstraction into theory is mainly the development role's work.
When the FDE submits and feeds back problems discovered at multiple customer sites, the development role sees the commonality, abstracts, and solves the problem at its essence. I believe that in a cognition cycle, the development role focuses on rational knowledge.
When development completes the feature iteration and informs the FDE, the FDE returns to the customer site to verify the new features and see whether the previous problems are truly solved.
This is one cognition cycle. The FDE derives perceptual knowledge from practice, development refines perceptual knowledge into rational knowledge, and the FDE puts rational knowledge back into practice to verify it. This is how FDE and development cooperate, completing many rounds of cognition cycles, moving up round by round. Only then can the flywheel turn.
2.2 FDE Responsibilities
The FDE needs to plant themselves in the business department long-term, not sit in the R&D department. They need to be familiar with business terminology, pain points, and processes, and be in sync with the business department. This makes the business department feel like they are charging into battle together, rather than "I am here to instruct you." If genuine materials cannot be obtained, perceptual knowledge cannot be established, let alone correct rational knowledge.
The FDE needs certain development capabilities, rather than thoughtlessly throwing problems to development. First, when discovering process bottlenecks or data silos, small needs like these can be rapidly verified through development tools or writing scripts. Second, when discovering large needs, sort them out clearly and output a plan to hand to developers.
The FDE needs continuous iteration, accompanying the business, and following up on metrics, rather than leaving after the work is done. They must be responsible for business outcomes. This is practice testing rational knowledge. Whether it is empty talk or genuine insight is shown in practice.
This role is not only for external customers. Internally, the same applies: relative to the R&D department, departments like HR and finance are also "external." They equally need people to sink in and do AI upgrades.
2.3 Context That Is Precious and Expensive
Context is precious because it is the business's real data and real solutions. It is the foundation of perceptual knowledge and directly determines whether rational knowledge can take shape. This part of the data cannot be searched from public channels or bought. It is a true valuable asset of the company.
Context is expensive, reflected in at least four aspects.
First, before AI, the data we referred to by default meant structured data, such as a clearly structured document, or data with clear rows and columns in a database. With AI, process data has also been included in the scope of data and begun to generate value, such as 2 hours of meeting video, 150 minutes of meeting recordings, 2000 rounds of work chat records.
Second, the Tao Te Ching says "the Tao that can be told is not the eternal Tao." True laws are hard to express in words or language. Some tricks or best practices in work are very hard to be deposited as data; they exist in the heads of experienced business personnel.
Third is the data maintenance problem. One is that all things evolve in the way of entropy increase, from order to disorder. In technology systems this lands as feature creep and documentation decay. Features or business change a little every day, data is not maintained accordingly, and over time it lags far behind the real business. Two is that data exists in multiple versions. The logic of version one and version two may be completely different, yet both are fed into the knowledge base at the same time.
Fourth is the FDE's authorization problem. Internal company data will have confidential parts. Which data the business department is willing to cede and share is not something the FDE can decide. It requires authorization from a higher level of the company.
2.4 How to Obtain Context
Since context is so precious and expensive, obtaining it is necessary but also difficult. Is obtaining context difficult in terms of technology? Not entirely. Obtaining context is mainly difficult across three dimensions: technology, communication, and management.
The technology dimension requires the FDE to have certain development capabilities, to develop convenient tools, such as development platforms or scripts, which are helpful for data acquisition.
The FDE, of course, writes code, but this is not the main work. The main work is to get along well with business personnel, identify bottlenecks and silos, and obtain best practices. Without good communication skills, this cannot be done.
The management dimension: first, push to establish a data version maintenance system, such as the update timing and responsible-person mechanism for the knowledge base. It is the same principle as when we write technical plans, where we designate a responsible person and establish a maintenance record table at the beginning of the document. Second, the FDE's authorization must be clear. A title without authority carries no weight. The FDE needs certain authorization to give opinions on the original business processes, and the customer company's senior management must also take these opinions seriously. This needs to be decided by higher levels of both companies.
3 Article Summary
Why is it still so difficult to deploy when the model is so strong? Because what the model lacks is not capability, but context. And context can only come from practice. This step cannot be replaced by the model for humans. So this role will always exist, because someone must always be at the scene to hear the gunfire.
Top 1 from juejin.cn, machine-translated. The original thread is authoritative.
The step of 'bringing the customer's real knowledge into the AI' — the biggest fear is that the knowledge stays only in this FDE's conversation with the AI, and when someone else goes to the same customer, they have to feel around again from scratch. Later we required that business agreements and pitfalls found on site be written into the project that same day as rules for the AI to read, so when the next person is assigned, the AI reads them first. People rotate, but this set of rules keeps growing.