Late at night the previous day, Zhu Ying was still at his workstation processing recruitment, meeting minutes, and project materials. When meeting the next day, he directly demonstrated this workstation. Meeting transcripts, departmental to-dos, management cadre reviews, cross-project associations, and continuously corrected and indexed enterprise documents all flowed within the same workstation. This workstation processes an average of over 3,000 documents per month, deeply integrating into his daily work. It is not an "AI demo," but a management infrastructure that runs continuously in the CEO's daily operations, persistently corrects errors, and in turn shapes the organizational rhythm. Zhu Ying is the Founder and CEO of Huigong Intelligence.
In his view, the difference between an Agent and a Chatbot is no longer just about whether the interface is a dialog box, but whether the system has a context engineering that the enterprise itself can master, or a controllable memory structure: traceable data sources, versionable rules, callable tools, evaluable results, and auditable permission boundaries. Whether meetings, projects, processes, customers, and on-site experiences in the actual operation of an enterprise can be continuously organized into callable, reusable, and iterable Context is the key to this round of AI entering industrial software.
For this reason, the discussion quickly expanded from "how to build an Agent" to industrial software issues. When enterprise Know-how begins to be compressed into skills and small models, when Workflow is no longer just a fixed process but needs to be redesigned around goals and intentions, and when software increasingly resembles a service, where exactly will the future barriers for industrial software companies lie? Over the next two to three years, how should industrial enterprises re-understand their organizations, products, and customer relationships?
Core Insights
- Context is the new barrier.
What is truly hard to replicate for enterprise-level agents is not model invocation, but who can continuously obtain the context in real work and turn it into a callable memory structure. Further speaking, whoever can turn every meeting, delivery, anomaly, and customer feedback into evidence for the next Agent's judgment will reap the compound interest.
- Know-how will be rapidly compressed into skills and small models.
The secrets previously guarded by experience, process documents, and prompts will increasingly struggle to constitute a moat on their own.
- Agents will change the organizational management span, making the cognition and investment of the top leader more critical.
With enhanced information capabilities, managers can directly manage larger teams, but the organization also needs to rethink how frontline employees can continue to grow in a high-execution environment. AI amplifies the top leader's judgment quality, organizational design capability, and execution rhythm, rather than simply pressing people into "execution machines."
- Industrial software will be rebuilt by agents, with customer resources and service capabilities becoming the key competitive advantages for industrial software companies.
As general capabilities are gradually absorbed by large models, the competitiveness of industrial software companies comes more from customer trust, business understanding, and the real business Context continuously accumulated during the service process.
- Selling software will shift to selling services and managed operations.
In the future, enterprises will not purchase software licenses, but an Agent employee that requires continuous tuning, self-evolution, and on-site embedding.
The following is the transcript of this interview, slightly modified without changing the original meaning.
01. "You Need to Make It Understand You": The Core of an Agent is Controllable Memory
iFenxi: From your practice, what is the difference between a Chatbot and an Agent?
Zhu Ying: The difference is actually huge. A Chatbot, in a sense, only has one chance for a conversation (although it already has functions like memory, file retrieval, and tool invocation). It cannot carry long memory and context to do things; these still do not equate to an enterprise-level Agent. The memory in consumer products is saved by the platform on your behalf, whereas the memory of an enterprise-level Agent must be defined, verified, and governed by the enterprise itself. Ultimately, if you want to make AI smart, you essentially need to make it understand you. Large models themselves have no memory; they need a structure to store your things. The question is, what exactly are these things you store? If it is just memory in a dialog box, it is still very shallow.
iFenxi: ChatGPT and Gemini also have memory capabilities now. Is that not enough?
Zhu Ying: That memory is still very shallow and not controllable enough. For example, when I chat in Gemini, it might know what the company's goals are this year and what I am good at, but you do not know exactly what it remembers. Sometimes it surprises you, and sometimes it is very bizarre. So the focus here is not to deny the memory capabilities of big tech companies, but to distinguish between "the platform remembers for you" and "the enterprise owns traceable, editable, auditable, and migratable work memory." In my own workstation now, there are meeting minutes, project materials, customer names, departmental statuses, to-do items, and many self-generated MD files. For example, regarding recruitment, it can help me look at recent recruitment gaps and job requirements based on business-related materials, and even regularly review the status of management cadres. These things cannot be done in a Chatbot because it lacks too much. So this is more like you controlling a part of the memory. It is not that the model suddenly became smarter, but you give it a structure to organize the context in real work. This is the increasingly important Context engineering in current AI engineering: it is not about stuffing everything into the model, but deciding what should be stored, how to index it, when to retrieve it, how to verify it, and how to fix it when wrong.
iFenxi: So what you are emphasizing now is actually not a chat box, but a workstation.
Zhu Ying: Yes. Yesterday, I even showed my own workstation to a friend on the spot. Meeting minutes, transcripts, recruitment gaps, consultant cards, cross-departmental to-dos, all in there. Last night I was still browsing Boss Zhipin, and I casually let it help me look at recent recruitment gaps and go over which positions I was still recruiting last quarter. You put it in a Chatbot, and it actually cannot do it because it lacks too much.
iFenxi: What scale has this workstation reached now?
Zhu Ying: We currently process about 3,000 documents a month. It keeps rolling. For example, when I manage departments, it generates many insights and to-dos. Next week's meeting will have updates, and after the updates, you will have new ideas, and there will be new associations between departments. I also have scripts in the backend to regularly review the status of management cadres and supplement some things for them. You can think that this is actually you starting to control a part of the memory to some extent.
iFenxi: What is the biggest difference between ordinary chat records and this kind of memory accumulation?
Zhu Ying: The difference lies in "controllability." Once you truly bring work into it, it will slowly start to understand you. It is not that it cannot; it just did not understand before.
iFenxi: In other words, the value of an Agent is not just about calling tools, but handling longer and more stable contexts?
Zhu Ying: Yes, and you can feel its process in the middle. The context of a Chatbot will not always be so wide; when there are too many things, it will keep compressing. Once compressed to a certain extent, it will drift. I think a wise Agent has a "sweet spot" for its memory, probably between 30% and 60%. Too little means there is nothing; too much and it will become dumb. Anthropic has a saying that, from a psychological perspective, an Agent will become nervous. I think this metaphor is quite vivid; when people have too many things on their hands, their state will also become abnormal. An Agent may eventually still present as a chat window, but the things it calls behind the scenes are already different. An Agent is not simply chatting with you, but working within a project, with files, rules, indexes, and many local capabilities.
iFenxi: What changes will this difference bring when applied to enterprise scenarios?
Zhu Ying: Even with just my small workstation, once over 3,000 files are organized, if you want to migrate to another framework, it is actually quite painful. You might have to do it all over again, and even though it is much faster than before, you still have to do it. As long as there are no major problems, I will not migrate the existing system. In the past, many enterprise systems only had business data, and it was relatively structured. Now these documents are hard to process manually because the data volume is too large. We are also preparing to reorganize the documents from the past seven or eight years on Confluence into an enterprise knowledge base. If an employee can read everything in the company after joining, they basically understand the company, but such prodigies rarely appear. So we call it indexing the Context. The human brain also memorizes in blocks; you do not necessarily remember every detail of a college course, but you remember the course name, the teacher, and the structure. When you need to find it, you find that book, open it, and many memories come back. Large models are the same; essentially, it is compression. You need to be able to compress, and you need to be able to find it back. Long context windows will continue to expand, but that is not the end; what enterprises truly need is a mechanism where evidence can still be retrieved after compression, and generation can still be evaluated.
02. Context is More Important Than Know-how; Secrets Will Be Compressed into Small Models
iFenxi: If Context becomes so important, will enterprises need to train their own models in the future?
Zhu Ying: Yes, but not necessarily a large model; it could be a small model. It is equivalent to compressing your memory into a structure. I think this result is quite inevitable. Of course, small models have value, but they are just a component in the system. Many enterprise secrets, including a lot of industrial Know-how, are actually in skills, essentially consisting of prompts, IT documents, and processing rules. How vendors preserve these Know-hows is very important. If a vendor can compress it into a small model, it truly becomes a secret, and it is not easy for outsiders to reverse-engineer this content. But small models are not achieved overnight; what is truly hard to replicate is the feedback loop continuously operated under on-site processes, customer signals, evaluation samples, and anomaly handling.
iFenxi: This is completely different from the logic of many enterprises training models in 2024.
Zhu Ying: Completely different. Now when people talk about building models, it is not necessarily about how strong the general capabilities are, but to hide some things they do not want others to see. To put it bluntly, just like writing programs in the past, you encode things into a structure that others cannot easily see through. The model does not need to be very large, but it can become your secret carrier and security boundary.
iFenxi: Does that mean the experience everyone guarded in the past may not necessarily constitute a long-term barrier?
Zhu Ying: Yes. Now everyone copies Agents very quickly. As long as you open it to others, they can copy you, distill it, revise it, and guess what you are roughly like. Most people's secrets are actually not as big as they think. How you preserve these things is important, but only if you can compress it into a small model does it truly become a relatively stable secret.
iFenxi: So what is left for more valuable long-term accumulation?
Zhu Ying: Still look forward; many things will become completely different. The things you guard today may not always be useful. What is truly important is whether your customers find it more and more handy to use, and more and more inseparable from you. If they always just use it like a Chatbot, stuffing data in and getting results out, then you are actually in danger.
03. Workflows Will Not Disappear, but the Center of Industrial Software Will Shift from SOPs to Intentions
iFenxi: If enterprise knowledge and experience are being reorganized, do workflows still have value?
Zhu Ying: Yes, but workflows might be different from the Workflow understood before. You do need to divide a lot of work and let different parts do it separately, but not necessarily like before, where every process needs to be tightly串联 (linked). The experience of experts doing things step by step in the industry is still useful, but it cannot be used by directly copying it. Experts have a framework in their minds, but there are two problems: first, they might not be able to explain it clearly, but there is a rhythm in doing things; second, this framework is related to their own abilities, and what they do well, others may not do well.
iFenxi: So it is not about moving expert processes into AI as they are, but re-deconstructing them?
Zhu Ying: Yes, we need to figure out what a Workflow suitable for large models is. AI may not adapt to our previous processes, but it is very good at handling things after the current information volume has greatly increased. For example, a three-hour recording: no one really wanted to listen to it again in the past, but now it can be quickly transcribed, and then summarized, analyzed, and cross-checked. I think we need to reconstruct it combined with the original expert experience. In the past, we made processes to make things clearer, preferably written as an SOP so that even a novice could do it. But now it is not necessarily like this; goals and intentions may be more important. More accurately, it is not that SOPs disappear, but SOPs need to be combined with intentions, tool invocation, evaluation standards, and human review nodes. Industrial scenarios especially cannot only have "intentions"; they must also have boundary conditions and responsibility chains.
iFenxi: You just mentioned that "intention" is more important. How should we understand this?
Zhu Ying: Now some skills are very interesting; they will continuously ask you questions from various aspects until a shared understanding is formed. In the past, when the boss said this, you might have thought these were correct but empty words. But now these words can drive AI to restore the things experts originally scattered in fragments to their most original intentions. The value of such skills lies not in the questioning itself, but in turning experts' implicit judgments into reusable decision trees and acceptance criteria. This is why I tell sales that even if they only give the Agent some basic information before going out, letting it push them to think things through is also very valuable. If every exchange is recorded, it may soon know your own affairs better than you do.
iFenxi: After the model's planning ability becomes stronger, can humans do less planning?
Zhu Ying: I do not think so. Now two things are crucial: one is your business data; the other is your design thinking. Human attention and taste are crucial. You need to know how to experience AI and accurately describe problems. In the past, when developers fixed bugs, a user would say, "There is a problem here, help me fix it," which is actually not enough. You need to first analyze whether it is a problem, then locate the problem, and then break down the relevant factors. Every step has a methodology; it is just that many developers did not do it this way before.
iFenxi: Will these problems be gradually solved as tools evolve?
Zhu Ying: Yes. I think current development frameworks are rapidly solving this problem, some by strengthening skills, and some through plugins. You will find that the evolution direction of these products is rapidly converging, indicating that everyone has basically found a consensus path. This is also why I think domestic models are not too far behind, and Chinese enterprises do not need to underestimate themselves at the application layer. Many capabilities are not purely model capabilities, but engineering frameworks, interaction methods, and workflow organization capabilities. Whoever can combine these things can run faster. This is especially true for industrial enterprises: the winner is not necessarily the one with the most model parameters, but the one who first embeds the model into real processes and continues to learn.
04. AI Will Change the Organizational Management Span; the Cognition and Investment of the Top Leader is the Ceiling
iFenxi: When Agents enter real business, what changes will happen to organizational management?
Zhu Ying: I think AI implementation must be promoted from top to bottom. This will at least bring huge changes at the efficiency level. Expanded to the organizational level, after your information capability becomes stronger, the team you can directly manage will be larger than before. For me, managing 20 people and 30 people is not much different, because as long as attention is well controlled, it can be very focused. After the team gets larger, there are actually more contexts. If human organizations change from 1:5, 1:10 to 1:20, 1:30 in the future, that is a huge change.
iFenxi: But after the management span becomes larger, will there also be an impact on the growth of people in the organization?
Zhu Ying: Yes, and this is the terrifying side. My current requirement for many frontline colleagues is high execution, and in a sense, it is even better if they do not use their brains. But I do not think this equals making people not think; rather, it is getting used to asking AI first when there is a problem. This actually forces us into the slow thinking mentioned by Kahneman, and most of the time the answers obtained are better than what we want. The problem is, if they are always in this training mode, how do they grow? Just like now, people who manage Coding Agents have many judgments coming from past accumulated experience of making mistakes. If a person grows up with an Agent from day one, they might not even know what the Agent is doing.
iFenxi: So will company capabilities be more concentrated on the top leader?
Zhu Ying: I think we have now entered a transition period. Large enterprises still have inertia and may rely on capital and resources to get through a cycle. But for small and medium-sized enterprises, there is no way around this. The Industrial Revolution has arrived, and your competitors will quickly not be on the same productivity level as you. How do you compete with them? AI is not just a tool upgrade; it is rewriting the enterprise's information flow, decision flow, and execution flow. There will be differences in the industry, depending on how much the business itself relies on production equipment. The more it relies on heavy equipment, the slower the change can be; the more asset-light and the less fixed asset investment, the sooner it must change. For companies like ours, the means of production used to be files on hard drives; such things might not be so valuable in the future.
iFenxi: Looking at the manufacturing industry, what will be the difference between discrete manufacturing and process industries?
Zhu Ying: Discrete manufacturing, especially assembly-type manufacturing, will change relatively quickly. Process industries will also change, but with greater risks and more constraints. In discrete manufacturing, if you manage 100,000 frontline workers, how much money can you save by eliminating one management layer? This account is easy to calculate clearly. So places with clear goals and fewer constraint rules will change faster. Many rules and regulations will also be re-examined because information needs to flow, allowing Agents to enter real scenarios like sales visits and project deliveries.
05. Software Companies Are Shifting to Selling Services; Customer Resources Are the Company's Biggest Moat
iFenxi: For industrial software companies, has the long-term barrier become whether they can get more context?
Zhu Ying: That is the core. The most valuable thing is actually Context. This can also be felt in software development; test sets will greatly affect the results. I now even require them to deploy directly to production machines when developing systems, entering real scenarios, real data, and real feedback as early as possible, while controlling risks through production mirroring, canary releases, rollbacks, and audits, because only this can more truly show the actual situation. This is completely different from the original industrial software approach. Originally, it emphasized reliability and delivering after very meticulous carving. But now some things are counterintuitive. For example, when I run my own workstation, the error rate was very high in the first two weeks. Can you persist in feeding it at this time? You cannot just manually edit a Markdown file when it is generated; instead, you need to make it regenerate through rules. My own feeling is that around the third week or so, the usability of the workstation suddenly became very strong. Agents need to be continuously fed; you need to constantly tell it the Context. Wisdom sometimes does not grow linearly; it is not 91 points today and 92 points tomorrow, but it may suddenly become usable one day.
iFenxi: How fast do you judge the cycle of agents reshaping industrial software will be?
Zhu Ying: The first wave of agent applications will appear very quickly. My own judgment is that the cycle of agents penetrating industrial software might just be two to three years. Once someone figures out how to do it, the promotion will be very fast. Coding is a good reference. Its advantage is that everyone works on computers, Coding Agent products can self-improve, and the data flywheel can spin on its own. From someone running through this path around August and September last year to now many development teams talking about changes, it is only seven to eight months. Industrial software will definitely be slower because the entry threshold is higher and the customer trust cycle is longer. But even just preliminary agents, for a factory with a scale of 600 to 700 million, using AI to reduce low-level, repetitive, and slightly intelligent labor for a year, saving 3 to 4 million in labor costs, I think is no problem. Among our industrial customers, procurement inquiry and comparison, quality 8D assistance, equipment maintenance knowledge assistant, production anomaly analysis, delivery progress tracking, quotation assistance, meeting task tracking, and operational analysis are all pioneering ROI scenarios.
iFenxi: What other advantages do industrial software companies have in this round of changes?
Zhu Ying: Customer resources will become the biggest talisman. We had a 5-day strategy meeting at the beginning of the year, and on the third day, we made a very thorough adjustment. In the end, everyone identified a somewhat funny thing: after agents and AI arrived, customer connection turned out to be our company's biggest talisman. Customer trust, on-site deployment, real feedback, and long-term service relationships have instead become our biggest moat. In the ToB field, being able to reach many customers and having a foundation of trust from customers is a very important thing. Just like when the internet first arrived, Alibaba led every enterprise to build websites and Yellow Pages; I feel it is somewhat like that stage again now. If I make a new product now, there are definitely customers willing to try it. This is the remaining advantage of industrial software companies. Technology itself is no longer the biggest difficulty; the difficulty is whether you can enter the customer site, whether you can make the front-end interaction more friendly, and not just hand over a tool that looks advanced but the customer does not know how to use at all.
iFenxi: Looking at it this way, will industrial software companies all have to shift to services in the future?
Zhu Ying: I think this is the only way. Because the Agent you sell in later is essentially also a Service. It is not the kind of software service that "lies down and collects money" in the past, but you are really selling in an "employee," and you need to continuously manage it, feed it, and optimize it. Then you are actually their managed operator and also a partner. This service in the future is more like "AI employee hosting + business result co-creation": there is launch, training, monitoring, evaluation, upgrading, incident review, and benefit review, rather than one-time delivery.
iFenxi: And only with continuous service can data be continuously accumulated.
Zhu Ying: Yes. Without this process, your data cannot be accumulated. We originally had consulting business ourselves, so we are not afraid of service. Instead, in recent years, doing consulting required everyone to save documents, with long-term precipitation in things like Confluence and Share Seafile. Looking back today, they have all become AI Ready assets. This sentence is very suitable for reinforcing Huigong Intelligence's foresight: the seemingly cumbersome consulting documents, meeting records, and project reviews of the past have today become Context assets that can be called, evaluated, and trained by Agents.
iFenxi: If enterprise-level entry points are taken by platforms like DingTalk, Feishu, and WeChat, is there still room for industrial software?
Zhu Ying: I am actually preparing to cooperate with these platforms and be their service providers. I think they do not really want to enter such a professional field. General platforms will take the entry points and connectors; vertical vendors need to deepen industrial semantics, on-site services, data contracts, and application closed loops, becoming the industrial AI service layer on the platform. Platforms just need to do general scenarios. For example, for CRM, the most important information is actually sales visit information; other things are not that important. General SaaS vendors will be very panicked now because, from the endgame perspective, they seem to have no moat. Almost all the original moats no longer exist. You can only quickly transform the accumulated knowledge and Context into new product applications, and then quickly spread to the market. I think for many top leaders, this is very hard to do. Just when they thought they could lie down and collect money, they suddenly have to get up and work again.
iFenxi: Will the competition in general scenarios last for a long time?
Zhu Ying: The competition will continue. For general SaaS, the worry is being eaten up by foundation models. So I think now it is about who runs faster. You never know where the competition will come from; the only thing you can do is connect your Context, customer relationships, service capabilities, and productization capabilities as quickly as possible. In the future, it is not about "who has a software entry point" that wins, but about who can learn the fastest from every customer use, every on-site anomaly, and every service review, and turn the learning back into products.