[Edaily Reporter Shin Yeong-bin ] Artificial intelligence (AI) automatically generates drafts of assessment plans that teachers repeatedly prepare each semester. The system works by allowing teachers to select a subject and assessment method, after which the AI recommends assessment criteria and even completes the template.

A proposal has been put forward that the transition to artificial intelligence (AI) in the public sector—known as “AX”—must move beyond simple chatbots or document summarization to the stage of “Agent AI” that performs actual tasks. The explanation is that public institutions must not stop at merely adopting AI but must also develop the capability to deploy agents in the workplace and operate and control them reliably.

Lim Eun-young, Head of the Agent AI Business at LG CNS (#LGCNES), made these remarks during a presentation titled “Beyond Generative AI to Agent AI: A New Direction for Public Sector Innovation” at the “Public AX Innovation Technology Forum,” a side event of the Public AI Expo held on the 24th.

Lim introduced the Gyeonggi Provincial Office of Education’s digital platform as a representative example. The Gyeonggi Provincial Office of Education has built a service that reduces teachers’ repetitive administrative tasks through an agent that generates assessment plans. The aim is to alleviate the burden of repetitive tasks—such as filling out student records, handling academic affairs, and preparing assessment materials—so that teachers can focus more on lesson preparation.

He also presented a case study of an agent that collects trends from the internet. Public institutions often restrict internet access on their internal networks. Lim introduced an agent that crawls the internet for trends of interest, generates relevant reports, and delivers them via internal network email. This approach supports staff members who need to quickly identify incidents, accidents, or policy trends by enabling them to efficiently utilize external information.

Mr. Lim pointed out that it is difficult to achieve results by simply replacing existing tasks with AI. He explained that while public institutions continue to cite budget and personnel shortages, their concerns are shifting depending on their level of AI adoption. Institutions adopting AI for the first time struggle to find use cases, while those that have already adopted it are grappling with security and operational control issues.

“No matter how advanced AI technology becomes, it remains a probabilistic value between 0 and 1,” he said. “The accuracy of its responses could be 97% one day, 88% the next, and 60% on another.” He emphasized, “A strategy for how to control and operate these systems must be established first.”

Mr. Lim believed that to expand the use of Agent AI in public institutions, a platform-based approach is necessary rather than having individual departments develop agents independently. He explained that if the procurement, finance, and general affairs teams each develop agents using different technologies, future operation and management could become difficult.

To address this, he proposed building a common vector database (DB) or graph DB so that AI can utilize documents held by public institutions, and creating agents based on standard development methodologies. These agents should then be connected to existing business systems via application programming interfaces (APIs) or the Model Context Protocol (MCP). He added that integration capabilities, such as robotic process automation (RPA), are also necessary to link with public institutions’ legacy systems.

LG CNS also introduced its Agent AI platform, “AgentWorks.” This platform preprocesses public institution documents to make them usable by AI and supports performance comparisons following model training and fine-tuning. It provides development tools that allow business users to easily create agents using only prompts, as well as tools for professional developers to build agents using LangFlow, LangGraph, Python, and other technologies.

It also includes a “router” feature that selects the appropriate AI model based on the difficulty of the question. Simple questions are handled by low-cost models, while complex questions are routed to inferential models. Manager Lim emphasized the importance of the “hub” feature, which tracks which model the agent used, how much GPU resources were consumed, at which stage it failed, and the quality of the questions and answers.

Security governance was also identified as a key challenge. Mr. Lim stated that when public institutions adopt AI services, they must also consider prompt monitoring, personal information masking, AI red teams, model safety verification, and anomaly detection. He explained that since public institutions handle critical national information, it is just as important to establish standards for safe usage as it is to quickly provide convenient technology.

“This year is the year we transition agents into live operations,” said Im. “While technologies such as models, skills, and tools are important, the ability to deploy and operate them in actual work environments, along with an AI-native strategy, will determine our success.”