Key Summary
- 83.5% of AI-using manufacturers operate at the department or project level
- The bottleneck isn't the model — it's the circulation of on-site data
- 'Data health checks' come before solution selection
- What to check this week: pre-implementation baselines
The government's 'Manufacturing AI 2030 Strategy,' announced in June, sets a goal of investing 20 trillion won from public and private sources by 2030 to generate more than 100 trillion won in economic added value. The plan calls for building a national manufacturing data library, a manufacturing AI foundation model, full-stack AI factories, and regional Manufacturing AI Transformation (M.AX) clusters, along with training specialized talent. The number of AI factories is targeted to grow from 102 in 2025 to more than 200 in 2026, and to 500 by 2030.
The 9th installment of Newspim's 'Reading the Economy through AI' series (by Economic Affairs Director Sung-hoon Jung) takes this figure as its starting point, drawing on seven prior case studies — steel mills, precision agriculture, drug development, semiconductor design, autonomous ships, logistics centers, and robotaxis — to identify a common formula: a five-stage cycle of measurement, learning, judgment, control, and verification/retraining. Newspim noted that the series is compiled by reporters based on analysis from an AI assistant.
"If any link in this cycle breaks — if there's no data (measurement), equipment isn't connected (control), or outcomes aren't verified (verification) — then no matter how good the model is, AI stays stuck on a dashboard screen."
Depth of use, not adoption rate
According to an analysis by the National Information Society Agency (NIA) focused on manufacturing, Korea's manufacturing sector had an AI adoption rate of 25.4% in 2023, below the all-industry average of 30.3%. Even among companies that adopted AI, the scope of use was narrow. Among manufacturers using AI, 83.5% applied it within one or two related departments or individual projects, only 13.8% linked it across multiple departments, and just 2.8% achieved enterprise-wide use. 65.9% of respondents said their company had no internal AI usage guidelines whatsoever.
Adoption did not directly translate into labor cost savings either. In the survey, 66.5% of manufacturers reported no change in labor costs after adoption, while 25.8% said costs actually increased. The article attributed this to the additional personnel needed for data preparation, system operation, and verification and training.
Data for large enterprises, cost for SMEs
The article noted that large enterprises, which possess production equipment, data centers, AI engineers, and maintenance organizations all together, can create a virtuous cycle by reinvesting AI-driven cost savings into sensors and models. Small and medium-sized enterprises (SMEs), by contrast, face simultaneous costs from connecting equipment of different ages and manufacturers, data preparation and cloud usage fees, and model updates, security, and maintenance. In the NIA survey, manufacturers cited a lack of suitable information and infrastructure (36.8%), a lack of specialized talent (34.7%), a lack of quality AI technology and services (28.3%), and a lack of funding (27.1%) as their top challenges, in that order.
OECD data, based on the most recent available year, showed that 40% of large enterprises with 250 or more employees used AI, compared to 20.4% of companies with 50–249 employees and 11.9% of companies with 10–49 employees. In the fields of autonomous robots, autonomous driving, and drones, the figures were 7.2% for large enterprises and 0.7% for small enterprises. The article noted that since these statistics are EU-based, they should not be applied directly to Korea and are cited only for comparative trend purposes.
Implications for AX at Korean Companies
The 'data health check' proposed in the article can be used directly as a self-assessment checklist within a company. The order of operations isn't choosing an AI solution first — it's diagnosing equipment connectivity, data quality, economic bottlenecks in processes, and the skill level and security posture of on-site staff. As the article puts it, "For some companies, sensors and data standardization need to come before AI; for others, the process itself needs to be simplified first" — and this diagnosis is what determines which path applies.
Mapping the five-stage cycle onto your own company's projects reveals which link is missing. Check whether sensors are continuously generating context-rich data (measurement), whether actual equipment is connected so it can execute AI's decisions under human approval (control), and whether economic performance and failures are being measured and fed back into the next model (verification/retraining). If any one of these is missing, closing that gap should take priority over swapping out the model.
It's safer not to judge cost based on initial build expenses alone. The total-cost-of-ownership items the article lists — integration, cloud, updates, security, training, and downtime — should be factored into estimates from the review stage onward, and portability rights and interoperability should be confirmed as contract terms to avoid being locked into a single vendor for data and models. Since 25.8% of manufacturers in the survey reported increased labor costs after adoption, it's also necessary to decide in advance who will handle data preparation, operations, verification, and training.
One thing to check this week is the pre-implementation baseline. For any AI project currently under review or already underway, verify whether production volume, defect rates, energy use, working hours, incidents, and maintenance costs were recorded as pre-implementation figures, and whether there's a plan to measure them again on the same basis after one and three years. Without a baseline, there's no way to prove results or learn from failures.