Yoon Sung-ho, CEO of MakinaRax, unveiled a roadmap for achieving fully autonomous factories within three years, targeting the 'long tail' problem in industrial settings that frontier AI models struggle to address. He emphasized that the key to competitiveness in the era of physical AI lies not in the number of AI systems adopted, but in a company's ability to own and control its data, models, and execution processes, which he termed 'corporate AI sovereignty.'
Speaking at the 'Attention 2026' industrial AI conference held on September 3 in Seoul, Yoon stated, "An accuracy of 80% is not sufficient for application in the field. We should ask whether AI is genuinely creating competitive advantages for companies, rather than how many AI systems have been implemented."
Yoon opened his keynote by questioning why productivity has not improved despite the explosive advancement of AI technology. He noted that while the duration of tasks completed autonomously by AI has decreased from years to mere seconds or hours, a survey by the National Bureau of Economic Research (NBER) revealed that about 90% of corporate executives believe AI has not had a measurable impact on productivity over the past three years. He cited McDonald's withdrawal from AI-driven drive-thrus as an example, stating that the complexities of noise and intricate orders in fast-food settings were too challenging for AI.
He identified the long tail problem as a significant issue in business operations. The long tail refers to the challenge where AI can effectively handle a small number of frequently occurring situations but struggles with the countless rare exceptions. Yoon explained that while frontier models can process general requests like drawing trend lines based on sensor values, they fail to address specific on-site demands, such as what to do if a particular equipment's thermistor value exceeds 40 or if operations should be halted to call a technician.
This limitation arises because the expertise accumulated by on-site professionals over decades has never been made available online. Yoon pointed out that frontier models remain stuck at an accuracy level of 80%, unable to solve the 20% of problems that constitute the long tail, asserting, "An accuracy of 80% is not sufficient for application in the field."
Yoon proposed a solution based on the three pillars of corporate AI sovereignty: ownership, independence, and control. Ownership ensures that proprietary data and the judgment criteria of on-site experts are retained as internal assets without external leakage. Independence means not being reliant on specific foundation models or GPU/cloud services. Control involves documenting the entire process of how AI makes decisions and executes actions, including the associated costs.
He explained, "Through these three pillars, we aim to create a structure where all knowledge and experience within a company are accumulated in controllable AI." As a foundation for this, he introduced 'Runway,' an AI operating system that integrates data centers, factory servers, and edge devices to develop, deploy, and operate AI models and agents within a closed network.
Yoon also shared some on-site achievements: annual savings of over 1,000 hours in design document reviews, AI applied to over 1,400 robots across six automotive production plants, and participation in the first joint training exercise (UFS) with the U.S. military, where AI was piloted in battlefield networks. MakinaRax reported that its field deployment engineers (FDE) have invested 70,000 hours supporting over 80 customer sites across 32 cities in five countries, with more than 6,000 AI models currently operational in industrial settings.
During the keynote, Professor Jang Young-jae of KAIST and CEO of Dime Research proposed that a 'dark factory' based on embodied intelligence that understands space and time could reduce equipment investment and labor costs by 50%. Kim Young-ok, Chief AI Officer of HD Hyundai, discussed plans for an intelligent autonomous shipyard, while Kim Sung-jin, Chief Digital Officer of Korea & Company, presented a case study on tire pattern generation AI developed in collaboration with MakinaRax.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.