---
title: "Linghe Shuzhi: Using AI to Turn Decades of Factory Experience into Reproducible Capabilities"
author: "Unique Research"
sourcePublication: "Unique Research Substack"
originalPublishedAt: "2026-03-18T13:41:06+00:00"
canonical: "https://ffcap.cn/en/research/src-20260318-03html"
source: "https://uniqueresearch.substack.com/p/src-20260318-03html"
language: "en"
---

# Linghe Shuzhi: Using AI to Turn Decades of Factory Experience into Reproducible Capabilities

_Original · Unique Research · 2026-03-18_

_English edition note: This full translation preserves the original article's reporting, interview answers and editorial assessments as of March 18, 2026. The order-entry results, professional biographies and financing experience below are claims reported in the source, not independently verified results. Views about AI, employment and organizational change are those expressed in the original article and interview. Chinese company, person and module names are romanized where an official English name has not been established._

Unique Awards · Guest Interview

Manufacturing's Biggest Bottleneck Is Not Equipment

Linghe Shuzhi founder Chen Hongxuan: The real challenge is not putting AI into factories, but turning decades of experience into reproducible organizational capabilities

"

Manufacturing's biggest bottleneck is not equipment, but information and knowledge flows that have not been organized.

When people discuss AI in manufacturing, their first reaction still tends to be two words: automation and cost-cutting.

But anyone who has actually worked inside a factory knows that manufacturing's toughest problems often lie neither in the machinery nor in the systems—and not entirely in the data, either.

They lie in tacit experience that is difficult to explain, impossible to write down fully, and liable to go off track as soon as someone else takes over.

How should orders be prioritized? How should exceptions be classified and handled? Where should purchasing, production scheduling, quotations, and approvals face strict controls, and where should discretion be allowed? What many companies truly rely on to function is not an SOP, but the practical know-how stored in the heads of veteran workers and middle managers.

This also explains why so many manufacturing business owners today feel that AI is important, yet still cannot take the step toward implementing it in practice.

They want to adopt AI but do not know where to start. They know their processes have problems but cannot tell whether the obstacle lies in tools, organization, or their own understanding. They worry about efficiency and displacement alike. They want results but fear their investment will turn into another round of disruptive, fruitless digital transformation.

Linghe Shuzhi founder Chen Hongxuan is an entrepreneur who emerged from precisely this kind of real-world operating environment.

He was previously vice president of Chunguang Technology Group. What really pushed him from using AI to building AI was not the technology trend itself, but a central contradiction he saw in the day-to-day operations of manufacturing businesses: manufacturing's biggest bottleneck is not equipment, but information and knowledge flows that have not been organized.

If AI can take on real job functions and handle the work of organizing information, making judgments, coordinating, and triaging that people previously performed, decades of accumulated corporate experience could move from dependence on individuals to organizational capability.

That is what Linghe Shuzhi is working on.

Rather Than Selling Tools, First Help Companies Identify Which Job AI Should Take On

Chen Hongxuan noted that many manufacturing companies are not unwilling to use AI. They are stuck on three very practical problems.

First, people do not understand AI. This is not limited to employees: many managers do not truly understand what AI can and cannot do either.

Second, the strategy is unclear. Many business owners know AI matters, but do not know where to take the first step.

Third, there are misconceptions about AI. For many companies, the first reaction to mentioning AI is: "Is this about cutting headcount?" In his view, however, AI's greater value lies in improving processes and efficiency, not simply replacing labor.

So from the outset, Linghe Shuzhi did not position itself as a team selling tools. It is doing something harder, but more fundamental: first conducting an AI diagnostic for a company and identifying the scenarios genuinely suited to implementation.

This distinction matters. Manufacturing differs from many internet use cases. An internet product can begin as a broad, all-in-one tool and gradually find its use cases. Manufacturing cannot work that way. Its processes are long, its roles numerous, and its operational chains deep. Begin with vague talk of a "platform," and chances are no one will actually pay.

What manufacturers are truly willing to pay for is not a concept, but dependable efficiency gains in a specific job.

How an "Intelligent Order-Entry Clerk" Turned Skepticism into Enthusiasm

One of Linghe Shuzhi's earliest representative implementations was the order-entry process.

It sounds unremarkable, but it is actually one of the most labor-intensive job tasks in many manufacturing companies.

Chen Hongxuan said that frontline salespeople at many companies spend 4 to 5 hours a day, on 3 to 4 days each week, doing the same thing: gathering information from customer emails, systems, and assorted software, then entering it into the company's own system.

The most complicated step is that customer order codes often differ from the company's internal codes, requiring extensive manual cross-referencing. The work is both time-consuming and highly error-prone.

So they built an intelligent order-entry clerk called "Xiao He."

This digital-intelligence employee automatically organizes order information and places orders. After deployment, work that previously took each person several hours a day could be completed with just 5 to 10 minutes of confirmation.

For a manufacturer, the significance goes beyond saving a few hours.

More importantly, for the first time it turns a job process that depends heavily on human experience and is prone to mistakes into an organizational capability that can be reused, verified, and continuously improved.

This is an easily underestimated aspect of implementing AI in manufacturing: what companies really want to see is not an AI that can chat, but one that can step into a job, accept an assignment, and deliver the result.

What Makes Linghe's "Digital-Intelligence Employees" Different from Traditional RPA?

Many companies still conflate Agent, Workflow, and RPA when discussing AI implementation.

But Chen Hongxuan draws a clear distinction:

RPA addresses repetitive operations;

Workflow addresses process structure;

Agent addresses complex decisions.

The difference is that an Agent does not mechanically execute predetermined steps. It continually makes judgments around a goal and iteratively adjusts its course of action.

In other words, Workflow means completing a task by following a process, while an Agent is more about choosing actions to achieve a goal.

On that basis, what Linghe Shuzhi calls a "digital-intelligence employee" is effectively a job-level Agent. It is neither an unrestricted autonomous system nor a fixed script, but a digital execution unit that performs analysis, coordination, triage, and judgment within clearly defined responsibility boundaries.

This definition is important because it answers one of manufacturers' central questions: is AI an add-on tool, or a role that can genuinely become part of a business function?

Linghe's answer is clearly the latter.

Human–Machine Symbiosis Is Not a Slogan: It Is About Capturing Experience

In manufacturing, the most valuable assets are often neither equipment nor process diagrams, but the tacit knowledge held by veteran workers and employees in critical roles.

The problem is that organizations have long struggled to retain and inherit this knowledge.

Linghe Shuzhi proposes a four-step methodology: make tacit knowledge explicit → standardize explicit knowledge → organize standardized knowledge → make organizational knowledge intelligent.

Behind this approach lies a problem larger than building an Agent: how to turn knowledge that once depended on individual experience into the company's own lasting asset.

In its product system, three modules work together to do this:

Paipai Marketplace: Like a talent market, it helps companies select, recruit, and develop AI digital-intelligence employees suited to their businesses.

Jiji Space: Stores company documents and knowledge, while continuously capturing the data, results, and experience AI generates in actual work.

Wenwen Center: Provides a unified collaboration interface between people and AI digital-intelligence employees, allowing users to ask questions, assign tasks, and obtain results as they would with colleagues.

Taken together, these three modules make it easier to see that Linghe Shuzhi wants to build more than an individual Agent tool: it is pursuing infrastructure for human–machine collaboration in enterprises.

Manufacturing Can Be Standardized—But Not in the Wrong Way

Manufacturing has always been seen as an industry where every factory is different. That is also why many people are pessimistic about turning AI into a standardized product: how could one product possibly handle so many nonstandard processes?

Linghe Shuzhi's answer is direct: do not use customization to resolve complexity; use productization to absorb it.

It divides the LinkCrux platform into a three-layer structure.

The first layer consists of standardized platform capabilities: document understanding, ERP/MES system integration, data structuring, and Agent and Workflow orchestration, for example. These are foundational capabilities almost every company needs.

The second layer consists of scenario templates. Order processing, quotation analysis, contract review, and procurement inquiries differ in company-specific details, but their underlying structures are highly repetitive. They can therefore be captured in Agent templates. By configuring data interfaces, rules, and prompts, companies can cover most scenarios.

The third layer involves a small amount of customization. Special approval chains, particular algorithmic models, or deep system integrations receive extensions, but these are still built within the platform framework rather than as projects started from scratch.

This reflects a mature product judgment: instead of forcing manufacturing processes to become standardized, standardize the underlying capabilities of AI Agents.

The Biggest Obstacle to Manufacturing AI Is Organization, Not Data

When asked about the biggest obstacle to moving manufacturing AI from POC to deployment at scale, Chen Hongxuan gave an interesting answer: it is neither data nor system integration, but organizational issues.

Many manufacturing business owners instinctively ask: "Is AI adoption slow because we lack data?" In practice, the hardest part is often breaking through their entrenched assumptions about tools, organizations, and processes.

Another critical issue is the "last meter": even when the ROI is clear and a solution can reduce costs, implementation will still fail if frontline employees find it troublesome or difficult to use.

Many supposed data problems are merely symptoms. At a deeper level, information and knowledge flows often have not been organized, while middle-management coordination and the organizational structure itself may create invisible resistance.

This observation highlights something a technology-centered perspective often overlooks: whether AI ultimately takes hold in manufacturing is frequently an organizational question, not a technical one.

AI Cannot Replace People, but It Will Redefine What People Should Do

One of manufacturing's most common anxieties is that AI will replace workers.

But Linghe Shuzhi has consistently emphasized human–machine symbiosis, not replacement.

In its view, AI's value is not to push people out of their jobs, but to free them from repetitive operations so they can spend time on higher-value work: customer communication, order analysis, exception handling, relationship management, and complex decisions.

Chen Hongxuan specifically noted that the closer the work gets to relationship management, on-site trade-offs, and complex decisions, the more important people become.

These capabilities come from long-term practical experience and are difficult to genuinely train into large models.

This also defines a realistic boundary for manufacturing Agents. Not every job is suited to being turned into an Agent. Jobs better suited to job-level Agents tend to have several characteristics: complex decision logic, clear responsibility boundaries, a need to obtain information across systems, reliance on experience-based judgment, some room for error, and results that can be reviewed.

Open Ecosystems Will Not Undermine Them: The Real Barrier Is Connecting to Actual Processes

When the discussion turned to open Agent ecosystems such as OpenClaw, Chen Hongxuan's position was equally clear: welcome openness and embrace the ecosystem.

But he also pointed out that the barrier in manufacturing AI has never been whether someone knows how to use models or Agents. It is whether they can connect those models and agents to actual processes.

Many critical insights cannot be trained into a large model on their own: the actual weighting of decisions inside a factory, relationships in purchasing and supply chains, or the logic behind on-site trade-offs. These come from years of frontline practice and peeling processes apart layer by layer.

Linghe Shuzhi's real barrier to entry is therefore not just technical implementation, but the combination of industrial operating knowledge × engineering execution capability.

This is also reflected in the team's composition.

Founder Chen Hongxuan previously served as vice president of Chunguang Technology Group. Having worked on the front lines of manufacturing for years, he brings experience across the full industrial chain, from factory management to global market expansion. He has led multiple projects involving enterprise cloud adoption, digital transformation, and cross-border business expansion. For him, manufacturing AI is not a technical upgrade on paper, but a solution that grows out of real operating problems.

CTO Zou Wuhe provides the capability at the other end. Described as one of the first senior artificial intelligence engineers and a PhD in physical electronics from the Chinese Academy of Sciences, he previously led the large-model team at NetEase Interactive Entertainment's AI Lab and worked as an AI algorithm researcher at Alibaba's Quark Search. His long-standing focus is NLP, knowledge graphs, and multi-agent systems, with expertise in turning frontier AI technology into product capabilities deployable at scale.

Chief Business Officer Hua Luke adds ToB commercialization and ecosystem development. A serial entrepreneur, he previously served as general manager for ecosystem and channels in East China at ByteDance's Feishu, a B2B sales specialist at Alibaba, and a senior business development specialist at Alibaba Cloud. He has extensive experience in ToB commercialization strategy and ecosystem building, and has also led financing projects worth hundreds of millions of yuan.

In a sense, this combination explains why Linghe Shuzhi is not simply a technical team that knows how to build Agents, but a team that understands manufacturing, products, and how enterprises actually buy and implement solutions.

This is also why it does not see open ecosystems as a threat. Instead, it answers the question at a more concrete level: open tools will certainly proliferate, but those who can bring them into actual production environments will be the ones able to deliver manufacturing AI.

If manufacturing digitization over the past few years has largely meant moving processes into systems, the current wave of manufacturing AI is attempting to bring experience, judgment, and collaboration into those systems as well. That may be the truly difficult part—and the part truly worth doing.

Selected Interview Q&A

Q1: What made you move from using AI to building AI, going from vice president of Chunguang Technology Group to an AI entrepreneur?

A: It was through the actual operations of a manufacturing business that I truly recognized the opportunity for AI. At Chunguang Technology, I experienced the 2019 trade war and the surge in orders during the 2020 pandemic, and participated in production scheduling, capacity management, and organizational system building. Those experiences made me deeply aware that manufacturing's biggest bottleneck is not equipment, but information and knowledge flows that have not been organized. If AI can enter real jobs and handle this information-processing and decision-making work, we have an opportunity to turn decades of company experience into reproducible organizational capabilities. That is why I decided to start this business.

Q2: What is the fundamental difference between a digital-intelligence employee and traditional RPA / automation tools?

A: RPA addresses repetitive operations, Workflow addresses process structure, and Agent addresses complex decisions. Automation tools such as Workflow complete tasks by following a process, while an Agent chooses actions to achieve a goal. On this basis, our digital-intelligence employees are job-level Agents. Their objectives are bounded by job responsibilities, and they have controlled decision-making and tool-calling capabilities. They are neither unrestricted autonomous systems nor fixed scripts: they undertake analysis, coordination, triage, and judgment within explicitly defined responsibilities.

Q3: Can you give an example of a customer moving from skepticism to enthusiasm?

A: One of our earliest representative implementations was order entry. At many companies, frontline salespeople spend 4–5 hours a day on 3–4 days each week gathering information from customer emails, systems, and various software tools, then entering it into the company's system. The most complicated part is that customer order codes differ from internal company codes, so extensive manual cross-referencing is required and errors are very easy to make. We built an intelligent order-entry clerk called "Xiao He," which automatically organizes order information and places orders. After deployment, what had taken each person several hours a day needed only 5–10 minutes of confirmation.

Q4: Veteran workers' experience is manufacturing's most valuable tacit asset. How does AI learn it?

A: Our core methodology is: make tacit knowledge explicit → standardize explicit knowledge → organize standardized knowledge → make organizational knowledge intelligent. By participating in real workflows, AI can also continually accumulate rules and experience as it executes tasks, turning knowledge that previously depended on individual experience into a company asset.

Q5: Manufacturing scenarios are complex and highly nonstandard. How do you balance standardized products with customized requirements?

A: Our approach is simple: do not use customization to resolve complexity; use productization to absorb it. LinkCrux uses a three-layer structure. The first layer is standardized platform capabilities, the second is scenario templates, and only the third is a small amount of customization. For order processing, quotation analysis, contract review, and procurement inquiries, we can typically cover most requirements through templates. We then add extensions for the remaining small number of complex system integrations or specialized algorithms. Put simply, we do not standardize manufacturing processes; we standardize AI Agent capabilities.

Q6: What is the biggest obstacle to moving manufacturing AI from POC to deployment at scale?

A: The biggest obstacle is actually organizational. Many people assume there is not enough data, but often the company does have data: its information and knowledge flows simply have not been organized. Another critical issue is the "last meter." Even when the ROI is clear, implementation will still fail if frontline employees find the solution troublesome or difficult to use. Much of the resistance is fundamentally not technical. It concerns decision-makers' understanding, middle-management coordination, and the organization's capacity to adopt and sustain the solution.

Q7: Manufacturers commonly worry about AI replacing workers, while you advocate human–machine symbiosis. How do you resolve that tension?

A: We have always emphasized that AI's value is not replacing people but freeing up their time. In the order-entry scenario, for example, AI handles large amounts of repetitive work, while people can spend their time on higher-value tasks such as customer communication, order analysis, and exception handling. The closer the work is to relationship management, on-site trade-offs, and complex decisions, the more important people become. Those capabilities come from long-term practical experience and are difficult to genuinely train into large models.

Q8: If manufacturers build their own AI workflows using open-source Agent frameworks, how much would that affect Linghe Shuzhi?

A: The barrier in manufacturing AI is not knowing how to use models or Agents, but knowing how to connect models and agents to actual processes. Many critical insights cannot be trained into a large model: the actual weighting of decisions in a factory, purchasing and supply-chain relationships, and the logic behind on-site trade-offs, for example. These come from years of frontline practice and peeling processes apart layer by layer. Linghe's barrier to entry is not technology alone, but the combination of industrial operating knowledge and engineering execution capability.

This document is original content by Unique Research.

---

Original publication: https://uniqueresearch.substack.com/p/src-20260318-03html
On-site reading page: https://ffcap.cn/en/research/src-20260318-03html