On October 24th, at the invitation of the Korea Industry Intelligence Association (KOIIA), Clevi CEO Hwanho Lee delivered a keynote speech at the '2025 Core Talent Development Academy for Mid-sized Companies, Future Industry Response Strategy Seminar.'

We participated in the session 'AI Technology Adoption Cases and DX Practical Strategies' with the topic 'Physical AI Use Cases and Failure Factors and Improvement Strategies in the Introduction of RAG Agents.'

Recently, as interest in AI agents has increased, more and more companies are considering adopting AI. It is known that many institutions and companies are already piloting and operating AI.

However, unfortunately, it has been found that in many companies and institutions, the desired ROI is not achieved, or functions that worked in PoC do not provide proper answers or do not work at all in actual business or on-site operations.

Because of this information, many companies are hesitant to rush into AI adoption.

In this lecture, Clevi CEO Hwanho Lee explained in detail the limitations of RAG systems, which are a representative cause of hasty AI adoption failures, as well as open source models and small LLMs (SLMs), and why previous adoption cases failed.

Ultimately, the reason for using AI is to replace human work or to expect better performance than humans.

Currently, practitioners in companies that have adopted AI are expressing dissatisfaction that their workload has actually increased because they cannot trust the quality of AI answers and have to check each AI response one by one.

The reason is that most companies use open source SLM models that have been fine-tuned as solutions.

Intuitively, one might think that AI will become smarter by learning new things, but in fact, when AI learns new knowledge, it loses past memories, that is, arbitrary parameter weights.

The more serious problem is that no one knows what information or knowledge has been lost.

Since AI adoption is not a task that requires little resources, you must have conducted a lot of verification from the PoC stage.

Why didn't such problems appear at that time? Of course, it is because retraining was done with target information.

If you teach an AI that only knows numbers about the alphabet and then only ask questions about the alphabet, it is natural that it will answer well. But in the field, do we only need models specialized for such specific domains?

Our field faces new environments and new data every day. However, AI trained while randomly destroying existing knowledge cannot guarantee generality or accuracy.

Therefore, it is necessary to have or be able to create a from-scratch foundation large language model.

These models contain almost all the knowledge in the world, so not only is the answer quality high, but so is the accuracy.

For specialized domains, you can overcome the shortcomings of fine-tuned SLMs by using SFT (Supervised Fine-Tuning) technology, or for small amounts of data, by using RAG systems or Clevi's own developed intelligence database system, the semantic database technology.

The government also recognizes the importance of such foundation models and is providing a lot of support.

This is the National AI Project. The biggest goal of this project is to secure digital sovereignty.

We at Clevi also recognized the fatal shortcomings of using open source models, so we independently developed everything from data collection to architecture design and trained Clevi's models.

Currently, Clevi models have performance that is not at all inferior to the world's leading AIs.

In particular, since we have our own models, Clevi has no dependency issues that can arise when using external models.

Now, AI adoption is becoming a necessity rather than a choice. However, simply adopting the latest technology does not guarantee the desired results.

True innovation begins only when you accurately understand the limitations of technology and establish the optimal strategy that fits your organization's environment and goals.

Clevi, with its differentiated technologies such as its own large language model and semantic database, is minimizing the trial and error and risks that companies experience in the process of adopting AI.

We will continue to listen to voices from the field and do our best to become a trusted AI partner for more companies. We hope you will achieve true success in AI adoption together with Clevi.

Future Industry Response Strategy Seminar - Korea Industry Intelligence Association