The success or failure of the AI transition hinges not on technology, but on understanding reality.

At the 'NetApp INSIGHT Xtra' event held at the Westin Seoul Parnas on the 3rd, a panel discussion was held to discuss the concerns and solutions of the domestic AI industry under the theme of 'Finding Answers in the Field: The Reality and Risk Management of AI Transformation.'

The panel discussion was moderated by NetApp Managing Director Eunseop Kim, and panelists Youngmin Seo, DS&G Managing Director, and Jahyun Kim, Upstage AI Director, pointed out the misunderstandings and limitations that companies face in the process of adopting AI.

The panelists unanimously emphasized that “the AI transition cannot be solved simply by introducing the latest models or GPUs.”

Director Kim Ja-hyun said, "Many companies mistakenly believe that a single, well-crafted LLM can solve all their internal problems." She added, "The model is merely the human brain. It must also be equipped with the systems, data, and processes that actually perform the work."He said.

He emphasized the importance of agent AI that connects actual work flows and data, saying that it is difficult to achieve the 99% or higher accuracy that companies expect with just a RAG-based platform.

Executive Director Seo Young-min cited "data reality" as the starting point for the AI transformation.

He pointed out that, “Customers worry about purchasing GPUs first, but in most cases, the data is scattered and in various formats,” and that, “If data quality and governance are not prepared, the project will inevitably be delayed no matter how quickly GPUs are introduced.”

The key to AI success lies in establishing a data strategy before infrastructure.

A realistic assessment of the risks of document automation and OCR projects followed.


Director Kim Ja-hyun explained, “The existing OCR method had a significant cost burden due to document-specific model learning and limited scalability,” and “The next-generation document AI that combines vision and language models can process various documents without separate learning, and enables inference beyond simple extraction.”

This was presented as an alternative that could simultaneously solve the problems of cost, accuracy, and scalability.

Regarding the case where the pilot (PoC) was successful but the operational stage was shaky, Executive Director Seo Young-min said, “ This is because the topology of the pilot and operating environment are different,” he diagnosed.

Failure to account for the differences between single-GPU-based testing and multi-node operating environments can result in performance degradation and failures.

He emphasized, “From the outset, we need infrastructure design that is operationally sound and clear operating procedures.”

'Focus and selection' were suggested as strategies to accelerate the AI transition while reducing risk.

Director Kim Ja-hyun said, “If you try to solve all problems at once, the probability of failure is high,” and added, “The most realistic way is to select one core problem that can be solved, achieve results in a short period of time, and expand based on that.”

The panelists identified GPU-centric thinking, a separation of proof-of-concept and operations, and lack of ownership as common patterns in failed AI transformations.

Conversely, success stories differ in that they start small and have a design that considers operations and a clear accountability structure.