AI agent connected to corporate system… Rapidly expanding scope of work

(Seoul = Yonhap News) Reporter Hyunmin Yoo = Recently, competition in artificial intelligence (AI) is moving from creating smarter models to directly acting and achieving results in actual work and industrial settings.

As AI agents are being deployed one after another in corporate work, and the semiconductor and security technologies that support them are becoming more sophisticated, the competition surrounding so-called ‘executing AI’ is in full swing.

◇ Beyond answers, even work… AI agent digging into corporate sites

There are increasing cases where AI agents go beyond answering questions and directly perform multiple tasks.

KT unveiled 'Agentic On', a platform that supports work automation by connecting AI agents to a company's existing data and work system.

It is characterized by having multiple agents divide their roles into tasks that require multiple steps, from document review to approval request to entry into the business system. Repetitive tasks allow the entire process to be executed with just one instruction based on a process performed once.

The role of AI is also expanding in the medical field.

KT decided to establish a 'medical standard model' that applies AI to all hospital operations and apply AI to treatment administration, patient guidance, and medical data utilization. We are also pursuing a plan to have treatment administration agents analyze treatment conversations, write treatment notes, and link them to electronic medical records and examination and prescription systems.

As AI agents are connected to existing business systems and process actual work, the scope of use is expanding from information search or document creation to task automation.

◇ AI works directly… Identity and authority management become more important

In order for an AI agent to perform actual work, it needs permission to access data and work systems. Accordingly, the importance of security technology that manages what AI can do and to what extent it can act is also increasing.

RaonSecure presented a blueprint for 'agentic security' that grants AI agents a digital identity, like a human employee ID, and verifies and controls the agent's identity, authority, and behavior.

Kakao also decided to expand the safety evaluation of AI models before and after distribution from language models to multimodal models and AI agents.

The government also plans to operate a public-private consultative body to prepare a national AI safety comprehensive plan within the year.

As the scope of AI's ability to judge and act on behalf of humans expands, not only AI's performance but also what authority it grants and how to verify its actions are emerging as key challenges.

◇ Domestic NPU/lightweight speed battle… AI extends to scientific research

As the use of AI increases, there is a continuing movement to improve computational efficiency to support it.

LG U+ has begun developing token optimization technology to reduce graphics processing unit (GPU) usage and power consumption with Opt AI, a company specializing in AI model optimization. Previously, the two companies applied lightweight technology to the ExaOne-based small language model to reduce power consumption by 78% and model size by 82%, and in server GPU, they secured technology to increase token processing of the same GPU by up to 4 times.

There is also ongoing movement to verify the performance of domestically produced AI semiconductors in actual industrial sites. KEPCO decided to verify the performance and reliability of domestically produced NPUs in the actual power field using 'K-Perf', a performance indicator of domestically produced neural network processing units (NPUs).

Upstage also unveiled the small language model 'Solar Mini 4' and began to improve model efficiency by selectively performing only necessary operations. Among the total 35 billion parameters, a Mix of Experts (MoE) structure was applied that selectively activates only 3 billion parameters in the actual inference process.

As the infrastructure for utilizing AI in actual fields is becoming more advanced, the field of use itself is expanding beyond industrial sites to include scientific research.

At 'AI4Sci Korea 2026' held last month, domestic and foreign researchers and industry experts participated to discuss scientific discoveries using AI and autonomous research agents. Methods of using AI in the research process, such as hypothesis setting and experimentation, beyond analyzing research data were also discussed.

The government also prepared a 10-year R&D and verification roadmap for each of the 12 missions of ‘K-Moonshot’, which uses AI to solve scientific challenges. We set a goal of investing a total of KRW 1.3465 trillion, including KRW 560.4 billion, in core projects related to the K-Moonshot mission next year, and will promote the use of supercomputer unit 6 and the development of specialized AI agents for each mission.

As AI goes beyond being a tool to provide information and is used to perform tasks and engage in scientific research, the standards for AI competition are also changing.

In addition to creating a larger model, how reliably AI can be used in the actual field and how efficiently authority, cost, and power can be managed are emerging as important factors.