[Edaily Reporter Shin Yeong-bin ] As the artificial intelligence transformation (AX) in the industrial sector gains momentum, there is an increasing trend of linking these efforts to performance outcomes. Against this backdrop, calls are growing for companies to redesign their workflows and systematically measure performance both before and after AI implementation.

Kim Ju-ho, a professor in the Department of Computer Science at KAIST, said in a recent interview with E-Daily, “Companies that have entered their second or third year of AI adoption are now at the point where they must demonstrate a return on investment,” adding, “This is a sign that AI adoption has moved from the tool-deployment stage to the performance-management stage.”

Professor Kim is a member of the UN’s Independent International Panel on AI Science and an outside director at Multicampus Corporation(067280). Also present for this interview was Ko Min-jeong, CTO of Multicampus Corporation and Head of the AX Learning Innovation Center. CTO Ko previously served as Head of the Information Strategy Group at Samsung SDS and Head of the Platform Team at Multicampus Corporation.

AI Code Survival Rate at Just 19%… Validation Bottlenecks Determine Outcomes

Professor Kim noted that it is difficult to judge AX performance based solely on easily quantifiable metrics such as token usage, the number of AI licenses, or the number of prompts. He explained that one must examine whether AI has eliminated actual bottlenecks in work processes and whether throughput and the quality of outputs have improved.

Just because AI processes tasks quickly doesn’t necessarily mean an organization’s productivity increases. This issue became clearly evident in an analysis of a development team at a global financial firm in which Professor Kim participated.

“While only 19% of the code written by AI survived, the survival rate for code written directly by humans was 42%,” Professor Kim explained. “Even if code writing speeds up, if bottlenecks occur during the verification stage, the total time required for the task ultimately remains unchanged.”

Even within the same organization, performance varied significantly depending on how AI was utilized. In that analysis, the productivity gap between the top 25% and bottom 25% of employees reached 3.5 times. Professor Kim attributed this to differences in “how AI was used,” rather than the volume of AI usage.

“Break Down Work and Reorganize the Roles of People and AI”

He proposed “work redesign” as a solution. This involves breaking down the work of individuals and teams into task units and then reassigning tasks that require human judgment to those that can be handled by AI. This must be followed by process innovation that transforms interdepartmental workflows, approval processes, and decision-making structures.

Professor Kim stated, “According to a 2025 McKinsey survey, among approximately 25 organizational attributes, the factor most strongly correlated with AI’s contribution to revenue was whether the organization had fundamentally redesigned its workflows.” He added, “However, only 21% of organizations reported having actually accomplished this.”

He identified “diagnosis” as the first step in workflow redesign. This involves first identifying exactly how employees spend their time and pinpointing where repetitive tasks and bottlenecks occur. Other factors to examine include the ability to delegate tasks to AI and verify results, the suitability of AI for specific tasks, data accessibility, and security regulations.

CTO Ko explained that demand for AX competency assessments is also growing in the corporate sector. Multicampus Corporation assesses the AX levels of individuals and organizations based on the eight core competencies required for AI utilization. In particular, there is a high demand—especially among large corporations and public enterprises—to identify disparities in AI utilization across departments and job roles.

Recently, there has been growing interest not only in pre-training assessments to gauge initial proficiency levels but also in post-training assessments to verify actual changes in competency following training.

Extending Beyond the Classroom to Real-World Work Tasks

Training methods are also evolving. Multicampus Corporation is expanding its training programs where corporate employees first practice “agent orchestration”—creating Agent AI agents themselves and coordinating multiple agents—and then apply these skills to their own real-world work

tasks

during hackathons.

CTO Ko stated, “Even if people learn how to use AI, if they don’t have the opportunity to apply it to their own work, the training ends with just improving individual skills.” She added, “The process must connect identifying real-world problems, applying AI, experimentation and validation, redesigning work processes, building Agent AI, and scaling across the organization.”

In actual training settings, there have been cases where work hours were significantly reduced. According to Multicampus Corporation, a manufacturing site implemented a system that connects internal data to automate processes ranging from checking parts inventory to generating purchase and order proposals. The time required for these tasks was reduced to one-fifth of what it had been.

In tasks involving the analysis of customer reviews and Voice of the Customer (VOC), a process that previously took about 22 hours was reduced to less than one hour after the introduction of an AI dashboard. A “Team Agent AI” for reviewing new business opportunities—in which multiple Agent AI agents divide and analyze market opportunities, profitability, and risks—was also developed during the training program.

Multicampus Corporation also plans to launch an enterprise-level “AI Agent Hub” service that will allow employees to share and reuse the AI agents they create within the organization.

CTO Ko stated, “To transition into an AI-native company, changes across the board—including education, organization, processes, systems, and data—are necessary,” adding, “We must also establish a system that allows employees to create, share, and reuse Agent AI themselves.”

Beware of ‘Skill Debt’ Created by AI Dependency

However, we must also consider the side effect that human capabilities may weaken as AI usage increases. Professor Kim referred to this as “skill debt.”

Citing a 2025 study analyzing 1,443 examinations at four endoscopy centers in Poland, he explained, “When skilled specialists who had been using AI stopped using it, the adenoma detection rate dropped from 28.4% to 22.4%.”

He added that in another experiment where developers trained a new library, the comprehension scores of the group using AI were 17% lower than those of the group that did not use AI.

Professor Kim stated, “What matters is how AI is used,” adding, “We need training that encourages cognitive engagement—where users verify the AI’s answers themselves rather than simply accepting them at face value.”

AX Education Enters the Era of Performance Management… Work Transformation Is Key

Both experts agreed that the criteria for measuring the performance of AX education must change going forward. CTO Ko emphasized that instead of focusing on completion rates or satisfaction levels, we should look at actual changes that occur after training—such as work hours, productivity, and the quality and accuracy of deliverables.

Professor Kim explained that performance metrics can be divided into three stages: time, output, and quality. He considered the final stage to be particularly important.

Professor Kim said, “A more important question than ‘How many hours were saved?’ or ‘How many people’s work did AI take over?’ is ‘What new tasks have emerged thanks to AI?’” He added, “The unit of measurement for training must shift from people to work.”

He continued, “While we’ve focused on how many people have completed the training and what their satisfaction levels are, we must now look at what has changed in specific tasks.” He emphasized, “Organizations that measure AI’s performance by the reduction in headcount versus those that measure it by increased capabilities will find themselves in completely different positions a few years from now.”