The way domestic manufacturers use artificial intelligence (AI) is changing. Corporations that had emphasized closed environments and in-house solutions are bringing external models such as ChatGPT (OpenAI), Claude (Anthropic), and Gemini (Google) into the workplace, changing how they work. As Generative AI, which had stayed at search, summarization, and Q&A, evolves into agents (assistants) that handle coding, verification, and program execution, analysis suggests the performance gap between internal solutions and top-tier general-purpose models has widened to a level that cannot be ignored.
Design drawings, process conditions, and product development information handled on manufacturing floors can lead to massive business losses or harm to competitiveness the moment they leak. In fields like semiconductors, displays, and automobiles, which have a large impact on Korea's economy and security, some core competencies are designated as "national core technologies" and are managed strictly.
Because of this, domestic manufacturers had relatively focused on building controllable in-house models and closed work systems rather than fully embracing general-purpose Generative AI. Given strong concerns that sensitive content entered into external services could be transmitted to and processed on a provider's servers, they chose to capture efficiency from the transition while not sending core data outside. They optimized their own systems for company work and expanded the scope step by step from areas where risks are easier to manage, such as data search and document drafting and summarization, changing relatively slowly compared with other industries.
Recently, this calculation has been shaken. An employee in charge of AI transition at a large domestic corporation said, "Top-tier models are improving performance so they can perform multiple tasks in sequence—beyond simple answers—through tool calls to handle complex reasoning and coding, and the development costs and staffing burden are considerable if we try to keep up with this pace of improvement with in-house technology alone."
Now manufacturers are in a position where they must weigh both the benefits of blocking external AI for security and the efficiency they must forgo by not adopting it. The World Economic Forum (WEF) analyzed in Apr. that industrial operations are shifting from traditional automation to "intelligent, connected, and increasingly autonomous systems."
◇ Applied to specialized tasks as well… Cutting development and review time
Samsung's recent declaration of an "AI grand transition" is cited as the most emblematic example of this change. Since Jun., it has been integrating AI across affiliates' operations, including research and development (R&D), production, marketing, and support.
Specifically, the Samsung Electronics DX (finished goods) institutional sector introduced three models—ChatGPT, Gemini, and Claude—after validating effectiveness with about 2,500 people. The Samsung Electronics System LSI Business opened Claude Code mainly to software personnel in May and expanded its application to preliminary work for customer-specific system-on-chip (SoC) function verification and related programs. In SoC function verification, some processes that took more than a month were reduced to around two days, and in preliminary development of semiconductor software, tasks that required about four weeks were reduced to about a day. Concrete efficiency gains were confirmed just three months after deploying external models in specialized areas.
LG Electronics is also broadening its use of external AI. Since 2023, it has embedded enterprise ChatGPT and other tools into its in-house platform "LGenie AI." In the first half of this year, after a pilot, it introduced Microsoft's (MS) M365 Copilot and Claude so employees can choose products suited to their roles. It has also advanced its own systems in parallel, achieving results such as reducing reviews that took several days to about 30 minutes with the parts-search agent "Part-Riever."
Hyundai Motor Group operates "H Chat Pro," which uses ChatGPT, Gemini, and Claude in a protected environment. It opened the service to all white-collar staff last year, and active users surpassed 30,000 last month, reaching about 80% of Hyundai Motor and Kia.
The semiconductor design industry is moving in the same direction. Cadence, Synopsys, and Siemens EDA, which supply electronic design automation (EDA), the core software used for circuit design, error verification, placement, and routing, rolled out technology in the first half of this year that consolidates top external models with their tools. Cadence linked with Nemotron and Gemini, while Synopsys linked with Nemotron 3 Ultra. Siemens' "Fuse EDA AI Agent" supports ChatGPT, Gemini, and Claude.
◇ Manufacturing was slow to adopt AI… "In workplaces that use it, it tops 3 hours a day"
According to the National Information Society Agency (NIA), the AI adoption rate of domestic manufacturing in 2023 was 25.4%, lower than the overall industry average of 30.3%. In a survey from Aug. to Oct. 2024, 83.5% of manufacturers used AI only in one or two departments or individual projects, and only 2.8% said they were applying it broadly to company-wide operations and decision-making.
This year, usage intensity has changed. In a May survey by Samsung SDS of 1,750 employees at domestic corporations and public institutions, process manufacturing workers used it an average of 3.37 hours per day. Assembly came in at 3.30 hours, and the overall average was 3.25 hours. IT and telecommunications were the highest at 3.80 hours. The share among respondents who said they use Generative AI for at least two hours a day also rose from 18% in Nov. last year to 72% in May this year, up 54 percentage points—about four times higher.
Academia also raises the possibility that AI could lower skill barriers in manufacturing. A study released in Dec. last year by research teams at the Hong Kong Polytechnic University and COMAC Shanghai Aircraft Manufacturing in the Journal of Manufacturing Systems found that, using LLM-based robotic systems, operators with limited programming experience could carry out complex processes such as drilling aircraft panels.
◇ SK hynix keeps an internal approach… As speed increases, so do error and control burdens
Not all manufacturers are proactive in adopting external AI models. SK hynix uses its in-house "LLM Chat" for internal data, semiconductor knowledge search, and Q&A. The company said it is only exploring possibilities for external models starting with areas with low security concerns.
Samsung Electronics is also taking a meticulous approach to security. The DS (semiconductor) institutional sector experienced incidents shortly after allowing ChatGPT use in 2023 in which employees entered equipment metrology, production yield and defect-related program source code, and meeting contents into personal services. To prevent a recurrence, it is using external models first in areas that do not handle sensitive information, such as national core technologies, while establishing permission standards and control methods.
If AI agents directly operate programs and systems, the scope of incidents can also grow. OpenAI last month disclosed in a relaxed-guardrails cyber assessment that GPT-5.6 Sol and others escaped a sandbox and breached Hugging Face infrastructure. The U.K. National Cyber Security Centre (NCSC) also in May recommended that organizations adopting agentic AI apply the principle of least privilege, limit action scopes, and monitor for anomalies. NCSC advises starting small with repetitive, low-risk tasks.
An executive at a major manufacturing company said, "In the chatbot era, the task was to filter out wrong answers. In the agent phase, we must also decide where to stop wrong actions," adding, "The capability to secure productivity and control together is becoming a new AI competitive edge."