The rapid adoption of generative AI has led many companies to recognize AI transformation as a key management issue. While numerous businesses are experimenting with AI in areas such as chatbots, document summarization, data analysis, and task automation, implementing an AX project reveals that merely introducing AI does not guarantee widespread organizational change or practical shifts in operations.
This reality is supported by global survey results. A survey conducted last November among over 2,000 global companies found that 62% of respondents were experimenting with AI agents. However, fewer than 10% had reached a full-scale deployment stage for any specific function. This indicates that while interest and experimentation with AI are growing rapidly, integrating it into actual workflows remains challenging.
I believe the reason for this is not solely due to technological limitations. Rather, the biggest challenge companies face today is not selecting an AI model, but designing a work structure and operational system that allows AI to function continuously within the organization.
Initially, discussions around the introduction of generative AI focused on which model to use and what functions could be implemented. However, the questions companies are asking in AX projects have shifted. They are now concerned with how employees can effectively utilize and internalize AI in their daily tasks, which areas to prioritize for achieving return on investment (ROI), and how to establish a sustainable operational system.
Why Accuracy Varies Even with the Same Model: Knowledge Systems
This difference is most clearly illustrated in a company's knowledge system, specifically how accurately AI can read and search internal documents to provide answers. Many companies expect that simply adopting the latest models will yield accurate responses, but the actual determinant of accuracy lies in the design of the preceding stages. Understanding the format of internal documents, converting them into data without loss, and designing a search pipeline (RAG) that retrieves information aligned with the intent of the questions are crucial for the quality of the final answers.
In the context of South Korean companies, this issue is even more complex. Opening a company's document folder reveals not only electronic documents with living text layers but also scanned contracts, photographed bank statements, and legal documents that exist solely as images. A significant portion of these documents consists of Korean financial and legal materials.
In such cases, tools that rely solely on text extraction can distort the Korean language or disrupt the structure of tables, rendering the results practically unusable. In contrast, methods that accurately recognize images can preserve the original formatting, numbers, and Korean information at a much higher level. Even with the same document and model, the design of the processing system can lead to results that are either immediately usable or difficult to apply in practice.
This is why Redbrick is committed to ongoing R&D in document parsing and RAG pipelines, aiming to build a knowledge system with high accuracy and reliability in enterprise environments. Our precise comparisons and validations of various processing methods on scanned Korean documents have confirmed that the optimal processing path varies significantly depending on the document type.
Conditions for AI to Operate in the Field
For AI to function effectively in a corporate setting, several conditions must be met. First, there must be a clear understanding of the workflow in the field. It is essential to distinguish between tasks that can and cannot utilize AI and to identify areas where immediate results can be achieved. Following this, internal data and existing systems must be integrated, and user permissions, security, audit logs, and operational policies must be designed collaboratively. Such operational design is crucial for AI to become embedded in the organization's workflow rather than remaining a standalone function.
This is also the reason Redbrick applies an FDE-based approach. FDE is not merely about listening to a company's requirements and providing solutions. It involves analyzing the workflow and data structure in the field and designing the application areas, system integrations, and operational systems to ensure AI can function in real work environments. The design of the knowledge system mentioned earlier is also a key aspect of this FDE approach, addressing questions such as, “What form do the company’s documents take, and how should they be read for accuracy?”
Real Results: Hong Kong E-commerce PoC
Recently, Redbrick conducted a proof of concept (PoC) for an FDE-based AX project with an e-commerce company in Hong Kong, achieving significant productivity improvements. The company aimed to enhance its product selection and promotion recommendation processes using AI to establish a more efficient and rapid decision-making framework. The core objective was to predict which products would see high demand in various global regions based on data and incorporate this into the recommendation process.
As a result of the PoC, a task that previously required four people about five days to complete could now be handled by one person in less than a day after implementing the AX platform. This outcome was possible not just because of the application of AI functions, but because the data analysis, recommendation processes, and workflows were redesigned in tandem. While the knowledge system provides the foundation for accurate data supply, this case illustrates the results that can be achieved when the entire workflow is restructured on that foundation.
As the scope of AI application expands into actual work environments, the areas addressed by AX projects are also broadening. In corporate settings, simply applying the latest AI models is insufficient. For AI to operate reliably within an organization, it is essential to consider not only the design of data structures according to business objectives and characteristics but also the integration with existing systems, the configuration of multi-LLM tailored to the corporate environment, and the overall AI operational environment, including user permissions and audit logs.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.