[Edaily Reporter Yun Junghoon ] “Whether an AI model ranks first or tenth, in the real-world workplace, it’s merely the difference between a 99 and a 98.”

Yang Seung-hyun, Chief Operating Officer (COO) of KONAN TECHNOLOGY INCORPORATION(402030), met with E-Daily recently at the company’s headquarters in Seocho-gu, Seoul, and emphasized that the competitive landscape of the generative AI market is fundamentally shifting. This means that the era of performance competition has given way to a new era where the decisive factor is how quickly and reliably AI can be applied to real-world work environments. COO Yang, a founding member of KONAN TECHNOLOGY INCORPORATION, is a first-generation AI expert with a Ph.D. in Computer Engineering from Seoul National University; he also served as Chief Technology Officer (CTO) of SKT from 2023 to 2025.

COO Yang explained, “In the early days of generative AI, benchmark competition over model performance was fierce, but now the performance gap between open-source and commercial APIs is rapidly narrowing,” adding, “Even the latest open-source models are catching up to the leading models within a 3- to 4-month timeframe.”

Big Tech Enters the Public AX Race… The Decisive Factor Is ‘On-Site Responsiveness’

Recently, global Big Tech companies such as OpenAI and Google have been signing memorandums of understanding (MOUs) with South Korean local governments and public institutions one after another, exploring entry into the public sector market. However, COO Yang predicted that these companies would face significant challenges in building large-scale AX systems within actual work environments.

The reason is that data, security environments, and existing business systems vary widely from one public institution to another. While Big Tech companies have strengths in providing AI models, cloud services, and APIs, they are at a relative disadvantage when it comes to customized implementation work—such as organizing institution-specific data and integrating it with existing systems.

In reality, the data accumulated by public institutions comes in a variety of file formats, and it is not uncommon for PDF documents to be entirely absent. He noted that AI models that functioned properly in a proof-of-concept (PoC) environment often encounter unexpected errors when deployed in actual operational environments, where thousands of users access the system simultaneously and strict security measures are required.

Secured 3 out of 5 Power Generating Companies… The Secret Lies in ‘Data’ and ‘On-Site Experience’

Conan capitalized on this gap with data technology and field experience accumulated over more than 20 years. The key lies in its “data pipeline” technology, which organizes documents and data scattered across public institutions so that AI can read, search, and summarize them. The company explains that, having long operated a search engine business, it has accumulated the know-how to organize and index unstructured data to enable searching and summarization.

Another competitive advantage is the company’s ability to develop its own large language model (LLM) and directly address on-site issues. Since 2023, Conan has been a pioneer among Korean companies, purchasing over 10 billion won worth of NVIDIA H100 GPUs to pre-train its own foundation model.

COO Yang stated, “In the early stages, when we simply utilized open-source models as-is, there were limitations in modifying or retraining the model itself when problems arose in the field,” adding, “We responded by retraining the model and augmenting the data whenever issues occurred on-site.”

This responsiveness built trust with early customers and served as a reference for securing subsequent contracts. Following Korea Southern Power, Conan built a hybrid generative AI platform for Korea Western Power that utilizes H200-based infrastructure and a Korean-specialized LLM. For Korea Midland Power, the company built a system that integrates and archives equipment failure histories and inspection reports, enabling interactive searches and the automatic generation of reports.

Another key competitive advantage is the company’s adoption of an on-premises approach—including the direct on-site deployment of GPU servers—tailored to the highly security-sensitive nature of public power generation utilities. The company explains that this approach is highly preferred by public institutions because it eliminates the need to transfer data externally, thereby reducing the burden of external cloud service fees and token usage costs.

“Version 2.0 is where people do the work and AI assists… Version 3.0 is where AI does the work and people only approve it”

COO Yang views “AX 3.0” as the next key battleground.

AX 1.0 is limited to asking questions and receiving answers from a chatbot, while AX 2.0 involves using AI as a tool to automate parts of tasks previously performed manually by Saramin, thereby saving time. AX 3.0 is the stage where AI agents proactively perform tasks, and Saramin only reviews and approves the final results. He explains that in the AX 3.0 era, it will be possible to reduce tasks that used to take a year to just one week, and tasks that took a day to just one hour.

COO Yang emphasized, “In the future, only companies that bring the incredible productivity of the ‘AX 3.0’ era—where humans give instructions and AI does the work—to the front lines will survive,” adding, “We will standardize customized integration projects as if they were off-the-shelf products, strengthen our fundamentals, and establish ourselves as the No. 1 AX company in both the public and private sectors, boasting high profitability and technological capabilities.”