On the verge of the inflection point of AI industrialization, Belle Fashion deconstructs "intelligence"

In the brand retail industry, the intelligent transformation has fallen into a strange cycle.

The technology team is obsessed with using increasingly high model parameters and algorithm optimization to deconstruct standardized processes, aiming to reduce costs and increase efficiency. However, when the system is up and running, the business department inevitably complains, "The system doesn't understand human language."

Since 2023, enterprise digitalization has started to chase the wave of large models. While AI has continuously produced efficiency results on the C - end, on the B - end, there is a misalignment between business and cutting - edge technologies.

Enterprises are faced with new data governance costs, the boundaries of SaaS systems have limitations, and the ability of large models to assist decision - making remains to be verified... Especially from the perspective of medium - and large - sized enterprises, although AI can be applied to many single - point scenarios, it is difficult to find a systematic entry point.

Just when the industrial implementation of AI urgently needs benchmark cases, Belle Fashion Group (hereinafter referred to as Belle Fashion) has explored a path that returns to the essence of digitalization - letting the business be the "helmsman" of technology.

01. Systematic Reconstruction: Deconstructing Technology with Business and Laying a Solid Data "Foundation"

Belle Fashion, a fashion footwear and apparel giant with more than 8,000 stores in over 300 cities in China, has independently developed a methodology for AI implementation in the past two years.

Instead of using technology to disrupt the business, it has transformed business rules into the "native language" of AI. In the current narrative of competing large models, this reflects the enterprise's own initiative in embracing the wave.

Before deciding to make a systematic and comprehensive investment in AI, Belle Fashion first realized that the hallucination problem of large models would be infinitely magnified in industrial scenarios.

For example, when managers ask, "Why is a certain product not selling well?", AI may infer different explanations for the same data but have difficulty penetrating the essence of the business.

The core of the problem is that both general - purpose models and vertical models lack the rule anchor points of the enterprise's own scenarios.

This means that the upper limit of AI's capabilities needs to be bridged by combining business rules with data quality, engineering technology, and human experience participation.

Therefore, Belle Fashion chose to cooperate with its long - term Data + AI service provider, Deepu Technology. By integrating the model stack, based on Deepu Technology's enterprise large model Deepexi and Belle Fashion's data - based fine - tuned training inference model Deepexi - RM, a vertical model for the commercial circulation industry was developed. Through this industry model and Deepu Technology's FastAGI intelligent agent platform, the logic of Agentic AI (Agent - type Artificial Intelligence) applications was constructed to achieve application accuracy (eliminating the hallucinations of large models).

In addition, there are two key pain points in the industrial implementation of AI:

First, the data governance problem, which highly tests the organization's innovation gene, is also regarded as the "foundation" problem of AI industrialization. To achieve the asset transformation from a "data swamp" to a "data warehouse" and then to a "data gold mine", enterprises need to abandon the isolated technical perspective and incorporate data governance into the strategic framework.

In the past, Belle Fashion's data processing mainly relied on tags. However, the tagging thinking under the data warehouse cannot fully unleash the data potential in the AI era. Therefore, Belle Fashion began to design an annotation system and conduct scenario - based decomposition based on the business value density.

While deploying AI - driven governance toolkits, Belle Fashion uses the "tag + annotation" system to transform static data into dynamic business context, which is easy for AI to understand and reason. For example, the "hot - selling" tag of a pair of shoes is no longer manually defined but dynamically generated based on indicators such as real - time sales volume, try - on rate, and associated purchase rate.

Data governance is the first step for enterprises to move towards "AI industrialization". In the process of long - term application and continuous learning mechanism, the misalignment between data - driven model decisions and the business end is another point that AI is "widely criticized" for in the industrialization process.

Based on Belle Fashion's own experience, in the actual model training process, enterprises have a large amount of data and a lot of noise. They cannot feed all the data to the large model but need to balance data quality and model effects through systematic strategies.

In most industries covering product R & D, supply chain, retail, and marketing, including the fashion footwear and apparel industry, the complexity of business reasoning far exceeds imagination. From market trend prediction, flexible supply chain scheduling, dynamic pricing, to inventory sales analysis, each decision - making link requires the integration of experience judgment and rational data analysis. This multi - attribute nature also makes it impossible for ordinary data to form a direct decision - making - relevant connection. Therefore, building a business thinking chain has become the key breakthrough.

From Belle Fashion's practice, the core reasoning challenges for AI to implement decision - making currently lie in aspects such as multi - modal decision - making integration and the construction of a rapid response mechanism. Therefore, while independently building a vertical large model, optimizing the data warehouse management mechanism for key scenarios, precipitating the enterprise knowledge platform, and developing human - machine collaborative AI Agents can initially transform the experience - driven thinking process into a traceable, calculable, and optimizable intelligent decision - making system.

02. Intelligent Data Warehouse: From Executing Rules to "Creating" Rules

In Belle Fashion's intelligent data warehouse practice, AI is reconstructing the "activity" standard of data.

The traditional data warehouse model is essentially a container of static rules. The business end needs to exhaustively list analysis dimensions, so that the so - called intelligence still cannot break through the upper limit of human understanding. Therefore, the indicator system established by Belle Fashion in the past digital transformation is far from sufficient to support large models.

At the same time, enterprise - level artificial intelligence application providers in the market, such as Deepu Technology, which cooperates with Belle Fashion, are also iterating their enterprise data intelligent management products. Under the concept of the Data Fabric architecture, Deepu Technology has developed an intelligent data platform that integrates data lakes and warehouses, enabling the new data management mode to be more agilely and efficiently integrated into large - model - driven data analysis and intelligent execution.

From Deepu Technology's technical path, we can also observe the necessity of enterprise data reconstruction in the intelligent era.

In the past, the traditional data warehouse that emphasized digitalization focused on pre - determining business query rules. That is, under manually defined static query rules, business logic was solidified in the pipeline, and then subsequent business insights were carried out. This method is similar to the pre - training process in the field of large models.

Today's intelligent data warehouse focuses on real - time business analysis and application decision - making, allowing the model to dynamically generate rules and store business knowledge in the model parameters and vector space. It emphasizes the post - training process of quickly generating and executing business insight rules.

Therefore, Belle Fashion's reconstructive measure this year is to use various indicators and their related meanings on the indicator platform as cold - start data to generate business context, enabling the model to learn the business rules behind the indicators and the relationships between indicators, thus demonstrating and realizing the intelligent data warehouse.

When AI's ability goes beyond executing manually defined rules and rises to the level of understanding business and "creating rules", the collaboration between the business end and the management end will become more efficient.

For example, in a common standardized collaboration scenario, before making a decision, the management wants to see the data, and the business end needs to retrieve data indicators. In the traditional middle - platform system, there is often a gap between business actions and data feedback: it is necessary to submit requirements and manually connect systems, and the business end has difficulty accurately understanding how much data can support the management's analysis needs. The process of large models creating rules breaks this "deadlock" in the business process.

The workflow of "intelligent data warehouse + large - model understanding" transforms business actions into rule generators. Through data engines such as real - time graphs, dynamic indicators, and business instruction encapsulation, each operation at the business end can become a training data source, realizing the transformation from demand - responsive analysis to operation - guided decision - making. The data - driven decision - making required by the management is no longer limited to just "looking at the data".

If we use the traditional Chinese medicine concepts of "observation, auscultation, inquiry, and palpation" as an analogy, in the past, it was just "observing" the data, and now it has gradually transitioned to "inquiry and palpation".

Belle Fashion's re - thinking of the middle - platform scenario also helps the brand retail industry observe the essential logic of current digital efficiency improvement: technology should not be used to disassemble the business; instead, people should start from the business scenario to disassemble technology.

Based on this purpose, in the co - creation case with Deepu Technology, Belle Fashion trained decision - making AI and execution - oriented AI respectively, disassembled the traditional solidified digital process, moved the middle - platform capabilities to the back, and flexibly called API interfaces through AI's learning ability in the front, enabling the flexible reuse of middle - and back - end capabilities.

It is worth mentioning that in the decision - making AI, Belle Fashion introduced a multi - modal model stack to build a physical constraint system for B - end decision - making, thus curbing the hallucination drift of AI in business scenarios and further contributing to the reconstruction of intelligent approval processes, such as contract compliance review. The efficiency value it brings is also referential for the industrial implementation of AI in other industries at present.

03. Agentic AI: The Last Mile of Large - Model Industrialization

With the popularity of Manus, the capabilities and application potential of Agentic AI have also come into the public view.

In fact, in the eyes of the earliest batch of enterprises in China that embarked on the path of AI industrialization, Agentic AI is not a new term but the inevitable path for the last mile of AI industrial implementation.

Take Belle Fashion's optimization plan for the replenishment process as an example. In the past, the headquarters would send a replenishment list to the regions via email every week. After multiple levels of digital approval and transmission across multiple systems, there was a certain degree of chaos and lag in the material versions.

Now, Belle Fashion has built a lightweight application based on enterprise business collaboration platforms such as DingTalk, conveyed the entire chain in an embedded way, and then automatically triggered system instructions and approval processes through the intelligent means of Agentic AI, improving real - time efficiency.

The advantages of Agentic AI are not only to release productivity through automated instructions but also to converge the file formats and actions originally scattered in various systems into a coherent data stream, making the entire execution process traceable, intervenable, and capable of intelligent judgment by large models.

Through Agentic AI, Belle Fashion directly transforms the "operation traces" most valued in the enterprise management process into digital trajectories, and then provides "cell - level" data nutrients based on real - world business scenarios for subsequent model training. This is the charm of AI assistants in industrialization. Especially in the brand retail field, the collaborative efficiency between the traditional "people - goods - place" and the new digital organizational management can often be the key factor determining the success or failure of an enterprise.

It can be seen that intelligent reconstruction requires both clear business logic and attention to the value of people and the technology foundation.

Whether it is the openness of DingTalk's intelligent application ecosystem or the underlying architecture upgrade of Deepu Technology from a middle - platform to an intelligent - driven platform, the results of systematic practice are based on Belle Fashion's screening of suppliers' business understanding ability, rather than simply considering the advancement of their technology.

In the future, Agentic AI is expected to unlock the "lock" of AI industrialization and dominate the form of large - model ToB applications for some time. In this process, in addition to systematic reconstruction, enterprises should also consider three differential barriers:

1. In terms of IT infrastructure construction, the coupling depth between the technical architecture and business logic.

2. The engineering ability of industry know - how.

3. When choosing technology partners, their product forms should be highly matched with the enterprise's own commercial value points and be able to co - create in the long term.

After meeting these three elements, the core value of AI industrialization will not be the accumulation of data assets but will be like a "senior store manager" who can comprehensively perceive the subtle changes in the business, adapt to and improve business connection points in real - time, respond to needs more accurately, and enhance enterprise competitiveness.

Technical solutions are designed to stitch up business faults and connect people, rather than rigidly pursuing a standard digital experience. Deepu Technology can quickly encapsulate its understanding of the fashion industry from long - term cooperation experience into standardized products in a short period when Agentic AI has not yet become the mainstream form of enterprise intelligence, which also reflects the co - evolution of business and technology in ecological collaboration.

In the era of industrial intelligence, the role of technology suppliers will no longer be limited to the traditional customized SaaS sales model. Letting cutting - edge technologies "bend down to listen to demands" is a necessary pre - foundation step.

From Belle Fashion's case, there is no shortcut for this transformation. The process of global reconstruction and data governance may be arduous and unnoticed, but the innovation vitality and business competitiveness of an industry often come from this.

Today, Agentic AI has shown us that the best intelligence is not to start generating and creating right away but to first reconstruct the way enterprises understand the world and people, and then create. The reason why the AI application forms with the highest industrial value at present are still AI Agents and AI Copilots is that "assistance" is the theme of intelligent transformation.

The most profound industrial upgrading often starts with respect for traditional business, rather than the solo dance of emerging technologies.