"Dismantling the middle platform" might be a misunderstanding. Over the past few years, the data middle platform has gone from being one of the hottest concepts in China's enterprise service market to an object of repeated questioning. Many enterprises believe the middle platform is outdated and should be dismantled. However, as large models truly enter enterprise scenarios, a more fundamental issue is brought back to the forefront: Where is the enterprise data? How is the data quality? Can this data support AI in making reliable and executable judgments?
Shulan Technology is a typical sample in this cycle of change. As a vendor that deeply cultivated the data middle platform early on, Shulan has experienced the industry peak and also undergone drastic adjustments after the market cooled down. The entire team was once reduced from a peak of 500 people to just dozens, and various voices of doubt were once heard from the outside. However, as enterprises re-emphasize data quality and rebuild the data foundation, Shulan is also climbing out of the bottom.
As the founder and CEO of Shulan Technology, Gan Yunfeng believes that AI does not replace data, but rather forces enterprises to rebuild their data capabilities. The data middle platform, which previously focused on data exchange, is evolving into an AI data foundation, undertaking high-quality datasets, ontology semantic layers, industry models, and enterprise-level agents. Shulan is no longer just a data middle platform vendor, but positions itself as a "Data + AI" vendor, focusing on the data foundation, industry models, and data agents.
In this interview, Ai Fenxi and Gan Yunfeng had an in-depth exchange on topics such as the re-evaluation of the value of the data middle platform in the AI era, data foundation construction, ontology semantic layer, enterprise-level AI business models, and changes in organizational capabilities.
Core Insights
- "Dismantling the middle platform" is a misunderstanding. With the rise of AI, the window for the data middle platform has reopened.
The data middle platform is misunderstood as backward technology. After large models enter enterprises, the first question enterprises still need to answer is: Where is the data, what is the quality, and can it support reliable output?
- The core goal of the data foundation is to consistently output high-quality datasets.
Databases, data warehouses, and lakehouses are merely data sources; the true data foundation also includes standards, processes, objects, semantics, ontology, and governance capabilities.
- The ontology and semantic layer represent the exclusive worldview of enterprise AI.
General large models lack the enterprise's own business rules, terminology systems, and object relationships. The value of the ontology and semantic layer lies in eliminating ambiguity and reducing hallucinations.
- Enterprise AI is shifting from data exchange to model exchange.
In the past, the data middle platform emphasized more on bringing data in and managing it; in the AI era, it is necessary to connect data, industry knowledge, enterprise knowledge bases, and model capabilities, allowing data to truly enter the business closed loop.
- AI Coding will depress the value of point-solution software; platform vendors must compete on methodologies and asset operations.
The value of enterprise-level AI vendors lies not only in software delivery but also in platform subscriptions, industry asset libraries, continuous operations, and the capabilities of AI architects to help customers realize AI value.
The following is the transcript of this interview, slightly modified without changing the original meaning.
01. Shulan Is Not Transforming to AI, but Advancing the Data Middle Platform into the AI Era
Ai Fenxi: In recent years, what new upgrades and changes has Shulan made at the industry and product levels?
Gan Yunfeng: Traditionally, Shulan has been doing the data middle platform. I have always liked to say that the data middle platform is purely a commercial term, not a technical system, nor any technical concept. For the past seven or eight years, we have been doing pure middle platforms and foundations. Starting from the end of 2022, we have laid out large models and agents. In 2025, our business in large models and agents has been doing very well. Currently, a large number of military industry, central and state-owned enterprises, and large well-known enterprise customers are using Shulan's data foundation and agent capabilities.
Ai Fenxi: Placed in today's AI era, how do you define Shulan?
Gan Yunfeng: Now Shulan's positioning is clearer. Everyone is talking about AI. I think AI is a historical trend with no alternative as human society develops to this day. There is no need to discuss whether to do AI or not; AI will definitely come, and the future world will certainly be an AI world. So how does Shulan define itself? I think there are three elements of AI: computing power, large models, and data. I particularly like to describe data as fuel. Computing power is the easiest to understand; from electricity, chips, semiconductors to the entire industrial chain, there has been a history of hundreds of years. Large models are about general capability building, requiring massive capital investment. It is not appropriate for a company like ours to participate in the independent research and development of general large models at the current stage. Shulan has its own large model platform, but we focus more on industry models, utilizing the data assets accumulated from past data middle platforms and data foundations, as well as high-quality datasets deduced based on data assets, to provide services for AI. We have served many Fortune 500 enterprise customers, and many central and state-owned enterprises as well. Therefore, Shulan is now a Data + AI company, focusing on the data foundation, industry models, and data agents.
Ai Fenxi: Can it be understood that on one side is the data foundation, and on the other side are industry models, supporting agent applications?
Gan Yunfeng: It can be understood this way. Our positioning has two ends: one end is heavy, focusing on high reliability, high security, and deep scenarios, somewhat like Palantir; the other end is moving towards standardization, low-code, and reusability, somewhat leaning towards Databricks. Our current core positioning is to make an intersection of these two ends.
02. The Core Goal of the Data Foundation Is to Consistently Output High-Quality Datasets
Ai Fenxi: Facing agent applications, what changes have occurred in the data foundation compared to the original data middle platform?
Gan Yunfeng: In the past, everyone ignored the inherent value of the data middle platform itself. The data middle platform has gone through several stages. From the wave of start-ups in 2016 to 2018, there might have been over 300 well-known data middle platform companies nationwide. By 2022, many companies doing data middle platforms were gone or transformed to do other things. A small number of enterprises persisted, but many no longer called it a data middle platform. Shulan has always adhered to the capability building of the data middle platform, and our external slogans and directions have never changed. The data middle platform is one of the most core capabilities of the data foundation. The data foundation is not the big data platform in the traditional understanding; it requires a massive accumulation of capabilities here, such as data access standards, packaging standards, full-process controllability, data exchange, data assetization, governance, cleaning, mining, modeling, and then to upper-level business capabilities. Recently, the national level has also been talking about the data foundation. When I interpreted it internally within the company, I summarized it into two keywords: building the most valuable data foundation and high-quality datasets to provide energy for AI. This sounds simple, but it is hard to do. Therefore, I prefer to understand the data foundation as a capability system, rather than a single technical capability or technical indicator.
Ai Fenxi: In your view, what is the most important measurement goal of the data foundation?
Gan Yunfeng: If I have to name a goal, it is to output high-quality datasets. Without the support of a data foundation, I don't believe one can piece together high-quality datasets. The B-end and C-end are completely different; enterprise-level AI cannot make things up. For example, if we want AI to help us generate a report, if the data is inaccurate or no data is provided to the AI, then the AI cannot output this report. Enterprise-level reports cannot allow hallucinations to occur.
Ai Fenxi: Do databases, data warehouses, and lakehouse architectures count as part of the data foundation?
Gan Yunfeng: I prefer to treat databases, data warehouses, and lakehouse architectures as data sources of the data foundation; its core logic is to put data here. But what the data foundation bears is not just these data sources, but also a large amount of unstructured data—audio, video, voice, text, and all information that is visible, invisible, tangible, or intangible. These are all data. The capability scope of the data foundation is broader. It is not understood in the past that having a lakehouse system means it is a data foundation. There are also standards, processes, objects, semantics, and everything related to ontology that everyone is talking about now.
Ai Fenxi: Will the data foundation encounter problems similar to those of the past data middle platform? For example, when the results are not good, customers will think it is the platform's problem.
Gan Yunfeng: This problem is inevitable when enterprises procure services; in the end, they will always attribute the problem to what platform you used. But when we were doing the data middle platform in the past, the first thing we liked to do most was data摸底 (data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data摸底 -> data...... wait, I need to translate "数据摸底" properly. "data摸底" -> "data inventory" or "data assessment". Let's use "data inventory".
Let's restart A7 translation carefully:
Gan Yunfeng: This problem is inevitable when enterprises procure services; in the end, they will always attribute the problem to what platform you used. But when we were doing the data middle platform in the past, the first thing we liked to do most was data inventory. To put it bluntly, it is about what assets you have; I will figure it out first and build your native data asset catalog system. If you only have 15 tables and a few million rows of data, and you insist on asking me to produce hundreds of billions of rows of results, that won't happen. The original process of business dataization is the foundation of enterprise development. Without these, whether doing a data middle platform or AI, there is no root. Shulan also had a hard time in the past few years. Everyone had different voices about the data middle platform, but we started to comprehensively improve and become profitable in 2025. The core is that what we provide is not just a set of data platforms, but a data capability system, rather than a single data technology. After AI took off in 2023, many enterprises started to use various methods to use large models to build various AI applications, but the results were not good. The root cause is still data and data quality issues, and they began to think back: Where is my data? How is the quality? Therefore, from 2024, especially to 2025, a new round of data middle platform window period has emerged. In the past, many people called the data middle platform backward technology, which I find ridiculous. The data middle platform is inherently a commercial term, not a technical term. After AI took off, everyone realized: Where is my data? At this time, the opportunity for the data middle platform, or rather the data foundation, has come again.
03. Enterprise AI Implementation: Shifting from Data Exchange to Model Exchange
Ai Fenxi: What is the difference between this round of enterprises rebuilding the data middle platform and the construction ideas around 2020?
Gan Yunfeng: The difference is very large. In the past, when we did the data middle platform, the underlying adaptation capability was very strong. Databases, servers, chips, and operating systems, both domestic and international, could be adapted. But the problem is, we didn't know what customers wanted to do with that data. Customers spent millions or tens of millions to build things, and then you tell them they don't know how to use it in business. This was indeed the biggest confusion in the past. We have served many industry-leading customers, including manufacturing and central and state-owned enterprises, but in the past, we mostly just sold a platform to customers and did a little bit of business. For example, for some very large enterprises, the platform was only sold for over a million, because only a platform was sold without truly diving into the business. After AI took off, our distance to the business suddenly narrowed. In the past, Shulan's data middle platform construction methodology was called "data exchange as the core". In the next ten years, we want to create a data middle platform construction methodology "model exchange as the core". Data, industry knowledge, enterprise knowledge bases, coupled with general large model capabilities, can train customers' own models, and then use agents to continuously overlay Skills to do various scenarios such as risk control, marketing, operation management, supply chain, and financial and tax control. Only by taking model exchange as the core can we narrow the value gap between data and business.
Ai Fenxi: Why is it only now more possible to understand the industry? Besides Shulan doing it for a long enough time, what other variables are there?
Gan Yunfeng: Understanding an industry is multifaceted. First, our own capabilities are steadily advancing. In manufacturing, many domestic leading enterprises are our customers. Doing more will naturally accumulate some capabilities. Second, I think the most important is the enterprise customers themselves. If Shulan finds an industry expert to tell customers how their enterprise will develop in the future, will customers believe it? They have been doing it for decades or even hundreds of years. Is their understanding of the business weaker than ours? Impossible. Therefore, the enterprise's own knowledge base is the most important, it's just that these knowledge are scattered in emails, PDFs, PPTs, and various files. Third, it is the industry knowledge and capabilities that come with general large models. But if leading enterprises only use general large models, they cannot create truly valuable AI, because they are already at the top of the world. How can they win with general capabilities? Impossible.
Ai Fenxi: When implementing for customers, do you start from a single scenario or build a company-level large platform first?
Gan Yunfeng: Both exist. In the past, among large customers, half just wanted to build a data middle platform, starting from the platform; the other half were driven by business scenarios, such as doing one business scenario first. The more viable path now is to run through a single-point business chain first. Because in the past, we were very far from the business. After AI took off, it might only take one step to get there. AI not only forces enterprises to realize the importance of the data middle platform but also brings us closer to the business, making it easier to create business value.
Ai Fenxi: For different industries and different customers, how is the reusability of the data foundation?
Gan Yunfeng: The technical capabilities of the data foundation itself can certainly cover the whole industry, but it requires the continuous backfilling of industry knowledge capabilities. You can understand it as an operating system: Windows is very powerful, but why does Mac always sell for more? Because it continuously strengthens based on its own capabilities and achieves the extreme. Data technology is the same. The platform capability itself is universal, but the construction of platform tools requires the continuous accumulation of enterprise knowledge. Many capabilities in the industry asset package will eventually be precipitated into the universal foundation.
Ai Fenxi: To summarize Shulan's evolution path, how would you概括 (summarize)?
Gan Yunfeng: We are walking a complete chain: from the earliest construction of data service warehouses, to the formation of the entire data middle platform, to the formation of the AI data foundation platform, to the construction of the ontology semantic layer, and then to the construction of the agent system. This is a full-chain, full-stack evolution, while positioning ourselves as an infrastructure vendor. I think vendors like us have more vitality. If you cut in from a single point and want to learn Palantir, it is hard to get through. In the early years, who really succeeded in learning SaaS services in China? None. Without platform support, a single business SaaS is hard to grow bigger and stronger. You must go through the whole chain to form true basic service capabilities to have a future.
04. The Ontology and Semantic Layer: The Exclusive Worldview of Enterprise AI
Ai Fenxi: Neither the semantic layer nor ontology are new concepts. Why are they now receiving large-scale attention?
Gan Yunfeng: On the one hand, it is indeed because the industry has seen companies like Palantir. Five years ago, Palantir was actually not optimistic in China. Everyone thought it was a project-based company and impossible to learn; now its market value has risen to hundreds of billions of dollars, and everyone goes to learn it. But returning to the domestic business facts, the root cause is still the lack of attention to basic capability building in the past. Many customers bought the data middle platform, wanting to blow a big bubble at once, wanting business growth of 30% or 50% this year. How is that possible? Things like the semantic layer and ontology were already completely defined in Shulan's platform very early on, such as business understanding, semantic definition, and unified definition of the same object in different systems. But no one paid attention in the past, because this was not a KPI, nor what customers wanted to understand. Customers would rather buy something that generates value the next day. Now that AI has taken off, everyone finds that without these basic capabilities, AI will have a large number of hallucinations, so they start to pay attention again to things like ontology and the semantic layer, which were considered thankless tasks in the past.
Ai Fenxi: In Shulan's understanding, what problems do the ontology and semantic layer solve respectively?
Gan Yunfeng: What we are doing in AI now, a very core part is the ontology and semantic layer. Because if you want to truly eliminate AI hallucinations, it is impossible to achieve it without doing this layer of capability well. The biggest shortcoming of general large models is that they lack an enterprise-exclusive worldview. It has a general worldview, but every enterprise is different, and there must be the enterprise's own exclusive worldview. Facing complex heterogeneous systems, enterprise terminology, and business rules, if there is no such layer, reasoning will break. We call the semantic layer a unified translator plus a business map, unifying scattered business concepts, data rules, and processes into a structured model that machines can understand without ambiguity. The semantic layer lets AI understand what the business is, and the ontology lets AI clarify the logical action relationships between businesses.
Ai Fenxi: Will ontology be just a phased concept like the middle platform back in the day?
Gan Yunfeng: I don't think so. As long as it is not deliberately misinterpreted domestically, it is like the master data concept in the early years. As long as data exists, ontology will definitely exist. Just like if you have ERP, master data will never disappear. Even if one day it is no longer called master data, but called something else, it doesn't matter. Ontology describes the essence of things, describing all causal relationships on the things themselves. This never changes. It is not a simple technical term, nor a commercial term; it is business logic.
Ai Fenxi: In the AI era, is ontology the only path? Or can it be achieved in other ways?
Gan Yunfeng: Ontology is not the only path. Palantir calls it ontology, and Shulan can also not call it ontology; the name is not important. But I think in the AI era, things built based on this concept are a core support point. If you want to improve AI quality and achieve precise support and services, this thing is always needed. No matter what it is called, ontology itself is a conceptual entity, a conceptual entity containing business logic composed of a set of technical systems. Its core is to connect all things, and many connections are not necessarily physical connections, but also logical connections. Whether you call it ontology or object, it doesn't matter. It's just that the word ontology feels particularly right; at a glance, it is the essence of things.
05. AI Coding Will Not Replace Platform Methodologies and Asset Operation Capabilities
Ai Fenxi: AI Coding capabilities are advancing rapidly. Which of the data foundation capabilities accumulated by Shulan in the past will be retained?
Gan Yunfeng: Platform capabilities will be completely retained. I don't believe AI can directly help enterprises write a truly similar data platform product. But indeed, a large amount of work no longer needs to be done by humans. Even if a bug occurs on-site now, it can be based on the agent system, inputting the requirements to complete self-iteration, without needing five people to work for three days for one bug like in the past. If traditional software enterprises do not fully embrace AI, they will be in a very embarrassing position, and a large number of pure software capabilities will definitely be replaced. But the methodologies behind the platform, product architecture, accumulation of customer scenarios, and judgments on the future enterprise data system—these cannot be replaced by AI casually writing code.
Ai Fenxi: Then what is the part of Shulan that is not easily replaced?
Gan Yunfeng: Things at the methodology and cognitive level will always exist. If you use AI to design something, you always have to tell it what you want. Enterprises need to develop, survive, grow bigger and stronger, and want to be the world's first for 50 years. How to do it? This is the core capability that enterprises like us need to accumulate most in the future. When enterprises choose Shulan, the most core thing is not a single tool, but that we have a whole set of mature methodology systems, and this system is continuously used by a large number of Fortune 500 customers. When we talk to customers about successful cases now, we don't necessarily have to talk about how complex it is; you just tell them which customers are using it.
Ai Fenxi: If the delivery cost for each customer is greatly reduced by AI, will the value of a unified platform weaken?
Gan Yunfeng: I have a counter-intuitive judgment: in the next 5 to 10 years, what enterprises need to build most may be exactly the data middle platform that everyone disliked the most in the past (perhaps everyone doesn't call it by this name). Single capabilities are not important; you need to build a real-time processing capability, and AI might be able to do it quickly. But speaking from the national level about the data foundation, from the unified technical architecture, engine to technical components, it will not change with the enterprise, but will continuously iterate with AI enhancement to undertake the explosive growth of data and business demands for data. Why are many AIs getting harder and harder to do now? Because everyone is doing single points. After tasting the sweetness of single points, they find the ROI is very low. What enterprises want is integration, not three isolated systems. You let three people use AI to write the financial system, supply chain system, and operation system respectively. All three can run, but what the enterprise wants is not three runnable points, but integration. Data is different from other things; the more it aggregates and the stronger its linking capability, the more valuable it is. From design, production, verification, mass production, delivery to the supply chain, the whole chain of manufacturing must be thoroughly connected at the data level.
Ai Fenxi: If the future is a large platform, where is Shulan's opportunity?
Gan Yunfeng: This question can be viewed more broadly. In the past, our main competitors were cloud vendors like Alibaba and Huawei. But for such cloud vendors, the computing power foundation is the root, and the data middle platform is an accessory. General large model vendors also have their own problems. They have strong model technical capabilities, but lack a complete data governance, service support, and management foundation for individual enterprises. They don't understand enterprise data asset construction, so it is hard to solve the problem of AI trusted closed-loop execution. They can only do general applications, and every enterprise is different. Traditional data platform vendors have a solid foundation, but many are unwilling to do time-consuming and labor-intensive ontology, semantics, agents, and knowledge orchestration. Knowledge graph vendors have strong graph reasoning capabilities, but lack full-domain data integration capabilities and the data's own system capabilities.
Ai Fenxi: What is Shulan's position for itself?
Gan Yunfeng: We define ourselves as: natively evolved from the data middle platform, with the ontology and semantic layer as the core, facing privatized and highly compliant industries, providing a one-stop full-stack product vendor for the underlying lakehouse data foundation, AI semantic hub, and enterprise-level agents. Many of our capabilities are open, not just Shulan's own capabilities. For example, for large models, we will connect all mainstream large models. Customers can use whoever they want, or build their own.
Ai Fenxi: From a business model perspective, will there be new changes in the future? Software License prices are getting lower and lower, and pure services are also hard to have high unit prices like Palantir's FDE. How does Shulan consider this?
Gan Yunfeng: I think the window period for enterprise-level large models is arriving, but the current usage of many enterprise-level large models is still intelligent data querying and document Q&A. After customers use it, they feel that large models are not that easy to use. The core problem is very clear: data silos, semantic ambiguity, and hallucination problems. We will gradually switch these capabilities from simple intelligent Q&A and document Q&A to the platform level through the data foundation, making up for the underlying shortcomings. This approach is effective for large enterprises, but not necessarily for small and medium-sized enterprises, because large enterprises are willing to invest money in this, while small and medium-sized enterprises want immediate results.
Ai Fenxi: Will the AI data foundation become the new standard for enterprise digitalization?
Gan Yunfeng: I firmly believe in this path. In the future, if enterprises want to do data businessization and turn a large amount of data into business to bring value-added and grow bigger and stronger, they need to build an AI data foundation. No matter how many applications AI has made, hundreds or thousands, enterprises will definitely find that they lack a data foundation when they look back. Just like before the early ERP came out, there were many independent vendors for human resources, finance, and supply chain. Later, the platform unified the world. It was the same in the e-commerce era; there were many vertical categories, and later platforms like Alibaba, Taobao, and JD.com rose. AI is the same. No matter how far the road goes, it will eventually return to the construction of the AI data foundation. AI without fuel cannot run.
Ai Fenxi: How will this change Shulan's business model?
Gan Yunfeng: As enterprises pay more and more attention to AI, for us, function-point-based project delivery will gradually become less attractive as AI capabilities enhance, because anyone can do it. Then we will shift our focus to product capability subscription plus continuous asset operation. Product capability subscription is the foundation, and continuous operation of data assets is more critical, because data is the fuel of AI and the means of production. AI is a productivity tool, and the innovation of means of production is an important carrier and support for productivity iteration. Therefore, Shulan's business model will revolve around data as the means of production and AI as the advanced productivity tool, still taking the data foundation as the root, and expanding on how high-quality data drives AI application scenarios.
Ai Fenxi: With AI capabilities so strong, will Shulan still need to expand to a team of five or six hundred people or even larger in the future? What changes will happen to the team composition?
Gan Yunfeng: Regarding this question, I think with the growth of the company's business scope and revenue, there will definitely be many people, rather than the idea that a few or even dozens of people can run a company with a market value of hundreds of billions. But the powerful capabilities of AI, especially capabilities like programming, will greatly reduce the demand for personnel in this area, but more people may be needed in other areas. Look at the most advanced AI companies internationally; has the number of employees decreased? Or most of them are recruiting more and more. Because software code and delivery efficiency have improved, but more people are needed who can truly turn AI into customer business value. First, Shulan needs to steadily advance its methodology capabilities in the future. Enterprises need data-related methodologies, which can guide the direction of digital transformation without deviating from the course. Methodologies have their own understandings, but an input is needed. Second, Shulan needs to improve its ability to implement AI on the enterprise business side. In the past, we had data architects. Now I prefer to call them AI data architects or AI business architects. This will be the main part in the future. It is a bit like the logic of IBM doing consulting and services for Huawei in the early years. I think the vast majority of enterprises do not have the ability to complete this AI transformation themselves; they need external forces.
Ai Fenxi: Is this role a bit like Palantir's FDE?
Gan Yunfeng: The logic is similar. The core of an AI architect is to help customers implement AI capabilities and improve business value. But this requires a strong professional knowledge background; a general small team is not enough. A person doing real estate going to serve a precision manufacturing enterprise completely doesn't understand the business. So in the future, it might be around different industries, forming small teams of three to five people to improve professional service capabilities and create value for customers.
Ai Fenxi: What are the capability requirements for such people?
Gan Yunfeng: I think there are three cores. First, to understand the relationship between data and business, this is the foundation. AI is a productivity tool; it uses computing power and large models to turn data into business and improve business value. Many people always separate AI and data, but what is computing power used for? Large models running around are essentially still processing data. Data is fuel, and everything is to turn data into business. Second, to understand AI as an advanced productivity tool and master its main characteristics. If you can't even master the tool yourself, how can you help others? Third, to have extremely strong industry knowledge. The first capability is the foundation; without the first capability, the latter two are empty. The B-end cannot imagine business detached from data; everything must be deduced based on the enterprise's own data capabilities.