On October 24, at the invitation of the Korea Industrial Intelligence Association (KOIIA), CLEVI CEO Hwanho Lee delivered a keynote speech at the “2025 Academy for Developing Core Talent at Mid-Sized Companies, Future Industry Response Strategy Seminar.”
We participated in the “AI Technology Adoption Cases and DX Implementation Strategies” session on the topic of “Physical AI Use Cases, Failure Factors During the Introduction of RAG Agents, and Improvement Strategies.”
As interest in AI agents has recently increased, more companies are considering adopting AI. Many organizations and companies are already piloting and operating AI systems.
However, studies have found that many companies and organizations fail to achieve the desired ROI, or that functions that worked during PoCs fail to provide proper answers or operate correctly in actual projects or on-site environments.
Because of this information, many companies are hesitant to adopt AI hastily.
In this lecture, CLEVI CEO Hwanho Lee explained in detail the limitations of RAG systems, a leading cause of failed AI implementations, as well as open-source models and small LLMs (SLMs), explaining why previous adoption cases failed.
Ultimately, the reason for using AI is the expectation that it will perform people’s work or deliver better performance than people.
However, practitioners at companies that have adopted AI are increasingly voicing complaints that their workload has actually increased because they cannot trust the quality of AI-generated answers and must check each response individually.
The reason lies in the fact that most companies use solutions based on fine-tuned open-source SLM models.
Intuitively, one might think that AI will become smarter when it learns something new. However, teaching AI new knowledge causes it to lose memories from the past—in other words, arbitrary parameter weights.
An even more serious problem is that no one knows what information or knowledge has been lost.
Since adopting AI is not a small undertaking, you likely conducted extensive validation from the PoC stage onward.
Why did such problems not appear at this point? This was, of course, because it had been retrained using the target information.
You taught an AI that only knew numbers to learn the alphabet and then asked it questions only about the alphabet, so naturally it provided good answers. But in the field, do you only need models specialized for such specific domains?
Our operations face new environments and new data every day. However, AI trained while randomly destroying its existing knowledge cannot guarantee either general applicability or accuracy.
Therefore, you need to possess a From-scratch Foundation large language model or have the technology to build one.
Because these models contain almost all the knowledge in the world, they provide not only high-quality answers but also greater accuracy.
For specialized domains, you can overcome the limitations of a fine-tuned SLM by using a technology called SFT (Supervised Fine-Tuning), or, for small amounts of data, a RAG system or CLEVI's proprietary intelligence database system, Semantic Database technology.
The government also recognizes the importance of these foundation models and is providing substantial support.
This is the National Representative AI Project. The primary goal of this project is likely to secure digital sovereignty.
CLEVI was also aware of the critical limitations of using open-source models, so we trained CLEVI's models through proprietary development, from data collection to architecture design.
CLEVI's models currently deliver performance that is fully competitive with leading AI systems worldwide.
In particular, because we have our own models, CLEVI is not affected by the dependency issues that can arise when using external models.
AI adoption is now becoming a necessity rather than a choice. However, simply adopting the latest technology makes it difficult to achieve the desired results.
True innovation begins only when you accurately understand the limitations of the technology and establish the optimal strategy for your organization's environment and goals.
With differentiated technological capabilities, including its proprietary large language models and Semantic Database, CLEVI is minimizing the trial and error and risks that companies face during AI adoption.
We will continue to listen to the voices of those on the front lines and do our utmost to become an AI partner that more businesses can trust. We hope you will build the true success of AI adoption together with CLEVI.