In January 2025, news broke that the Ministry of SMEs and Startups had formed the 'SME Manufacturing Artificial Intelligence Innovation Task Force' to support the introduction of AI in the manufacturing industry. The task force is working with the Small and Medium Business Technology Information Promotion Agency to establish policies for the introduction of AI and to develop a systematic support policy for fostering smart manufacturing companies.
The task force is working with the Small and Medium Business Technology Information Promotion Agency to establish policies for the introduction of AI and to develop a systematic support policy for fostering smart manufacturing companies.
The TF also announced that it plans to establish a new designation system for companies specializing in smart manufacturing technology and conduct various activities, including support for overseas expansion.
The effects of introducing AI in the manufacturing industry are not limited to reducing the defect rate, but are also leading to an overall improvement in productivity, cost reduction, and quality improvement for manufacturers.
For example, a system is being established that can predict and solve various problems that occur in manufacturing sites through an AI-based analysis and decision-making support system.
In addition, the government is promoting digital transformation (DX) and artificial intelligence transformation (AX) to successfully introduce AI in the manufacturing industry, which is expected to strengthen the competitiveness of the entire industry. The government also plans to build a smart manufacturing AX platform by 2026, which will enable manufacturers to operate more efficiently using AI.
In this article, we have summarized the essential guide for successful implementation of AI in the manufacturing industry.
Amidst the rapidly changing global manufacturing environment, manufacturing companies are facing more serious management challenges than ever before.
In particular, the recent increase in global supply chain uncertainty has clearly revealed the limitations of the existing experience-based decision-making method.
As of 2023, the average inventory turnover rate of domestic manufacturing companies has decreased by 15% year-on-year, which is analyzed to be due to inaccuracies in demand forecasting. Moreover, the increased volatility of raw material prices is increasing the difficulty of managing manufacturing costs.
In this situation, the introduction of AI technology has become a necessity rather than an option, but most of the AI solutions currently on the market have been developed by large companies, which is quite far from the reality of the manufacturing industry.
Market research results show that the average deployment cost of AI solutions in the manufacturing industry is three times the annual IT budget of a typical company.
The biggest challenges that manufacturing companies face when trying to introduce AI can be divided into four main categories.
First, there is the burden of initial investment costs. Building a typical AI system requires a significant initial investment, including the cost of building a hardware infrastructure, software licenses, and system integration.
This is a significant burden when considering that the average annual IT investment budget for a typical company is less than 1% of its sales.
Second, it is difficult to secure specialized personnel. In order to effectively operate an AI system, specialized personnel such as data scientists and machine learning engineers are required, but most of these talents tend to be concentrated in large companies or IT companies.
Third, there are issues with the quality and integration of data. Many companies still use manual data management or disconnected systems, making it difficult to secure high-quality data needed for AI model training.
Fourth, there is the issue of connectivity with existing systems. Most mid-sized companies already operate existing ERP or MES systems, and the technical complexity and costs of integrating with a new AI system are a major burden.
A systematic and step-by-step approach is essential for the successful introduction of AI in the manufacturing industry.
In the first step, which is the establishment of a basic data infrastructure, the core business processes of the company must be analyzed to identify the essential data points and establish a system for effectively collecting them.
The most important thing in this process is to focus on core data that can create substantial value rather than pursuing a perfect system.
For example, it is more effective to collect only the key process data that directly affects quality rather than collecting all the sensor data from the production line.
In the second stage, the pilot project implementation stage, the system is verified on a small scale by selecting a specific product or process before the company-wide introduction. At this stage, it is important to select an area where the ROI can be clearly measured.
Finally, in the expansion and advancement stage, the scope of application will be gradually expanded based on the results of the pilot project, and more sophisticated AI models will be introduced.
Considering the limited resources of a typical company, ensuring investment efficiency in the process of introducing AI is a very important task.
The first way to do this is to strategically use cloud-based solutions. Using cloud services minimizes initial hardware investment and allows for a flexible cost structure based on usage.
In fact, companies that have adopted cloud-based AI systems report an average initial investment cost reduction of 40% compared to on-premises methods.
Secondly, it is to actively utilize the open source ecosystem. In recent years, the maturity of open source AI frameworks has improved significantly, and by using them, you can build a competitive system while reducing licensing costs.
Finally, establish a phased investment strategy. You can minimize risk by implementing only essential functions in the early stages and gradually expanding investment after the performance has been verified.
To successfully build a demand forecasting AI system, securing high-quality data is the most important prerequisite. Many companies are struggling at this stage due to the lack of a systematic approach.
First, companies need to secure at least three years of historical sales data. At this time, they need to collect not only simple sales volume but also various variables that affect demand, such as sales price, promotion information, seasonality, and competitor trends.
It is very important to adhere to industry-standard formats when collecting data. This is a key factor that will facilitate future system expansion and linkage with external data.
In fact, companies with a standardized data structure have been shown to have AI model accuracy that is 25% higher on average than those without.
The collected data must undergo a thorough preprocessing process. In addition to basic preprocessing such as removing outliers, handling missing values, and scaling, it is necessary to apply advanced preprocessing techniques that reflect industry characteristics.
For example, in the manufacturing industry, this includes analyzing time differences in production lead times and adjusting data to reflect capacity constraints.
The choice of a demand forecasting model should be made by comprehensively considering the company's data environment and business characteristics. In the manufacturing industry in particular, it is important to find a balance between model complexity and ease of maintenance.
Selection of a demand forecasting model
ARIMA and Prophet, which are basic time series forecasting models, show stable performance despite their relatively simple structure. They are suitable for forecasting data with a certain pattern or seasonality.
In fact, in the manufacturing industry, where production is relatively constant and seasonality is pronounced, these basic models alone can achieve an average forecasting accuracy of over 85%.
However, recently, the need for advanced deep learning-based models has been increasing due to the growing influence of external variables.
LSTM and Transformer-based models are strong in learning complex patterns, but this requires sufficient data and computing resources. Tree-based models such as Random Forest, XGBoost, and LGBM are also used in various ways.
These models also handle nonlinear relationships well and have the advantage of being able to take into account interactions between variables.
The most recommended approach is the hybrid approach. The basic demand pattern is predicted using a traditional time series model, and the impact of special events or external variables is compensated for using a deep learning model. This approach is effective in balancing accuracy and efficiency.
A systematic evaluation and monitoring system is essential for the continuous performance maintenance of the AI demand forecasting system. This is a process that goes beyond simple technical monitoring to continuously check the connection with business performance.
The evaluation of forecasting accuracy uses standard indicators such as MAPE (mean absolute percentage error) and RMSE (root mean square error). However, along with these technical indicators, practical business performance indicators such as improved inventory turnover, reduced lost opportunities, and reduced operating costs should also be managed.
Of particular importance is the system for analyzing the causes of forecast errors when they occur. Forecast errors can occur for a variety of reasons, including problems with data quality, limitations of the model, or the impact of unexpected external variables.
Systematically analyzing and documenting these causes is very important for the continuous improvement of the system.
In order to successfully build an inventory management optimization system, it is essential to design an architecture with scalability and flexibility. In particular, in the manufacturing industry, a flexible structure that can respond to future growth and change is very important.
The basic structure of the system should be designed in a modular fashion. This means that each function, such as inventory level monitoring, order management, and supplier management, should be configured as an independent module so that it can be upgraded or modified individually as needed.
In fact, companies that have adopted a modular structure have been able to reduce system maintenance costs by an average of 35%.
Linking with existing ERP systems requires a particularly careful approach. Many manufacturers already have an ERP system in operation, and the new AI-based inventory management system should be seamlessly integrated with these existing systems. To this end, we recommend a linkage method using standard APIs.
The architecture for real-time data processing is also an important consideration. Changes in inventory levels, order status, and receipt and shipment status should be monitored and processed in real time, and an efficient data pipeline must be established for this.
Algorithms for optimizing inventory management must be carefully selected to suit the characteristics of the company and its operating environment. In particular, in the manufacturing industry, the balance between calculation complexity and feasibility is important.
A hybrid approach that combines a basic inventory optimization algorithm based on the Economic Order Quantity (EOQ) model with modern machine learning techniques is effective. This approach can deliver better performance by combining the stability of traditional inventory theory with the adaptability of AI.
The determination of the safety stock level must take into account both demand uncertainty and supply chain risk. The latest machine learning algorithms can dynamically reflect various variables to suggest the optimal safety stock level.
In a real-life application, this method has reduced inventory holding costs by an average of 25% while keeping the out-of-stock rate below 5%.
It is effective to apply the dynamic threshold method to set the order point. This is a method of adjusting the order timing and quantity in real time by reflecting seasonality, market conditions, and supplier lead times. This can minimize the risk of both excess and shortage of inventory.
For the successful operation of an AI-based inventory management system, it is very important to set the appropriate level of automation. Full automation is not always the best answer, and the optimal level of automation must be found that suits the company's situation and capabilities.
The generally recommended approach is phased automation. In the early stages, the system's suggestions are consulted, but the final decision is made by the person in charge, and the level of automation is gradually increased after the system's reliability has been verified. This approach is effective in increasing the organization's confidence in the system and minimizing operational risks.
The design of the exception handling process is also very important. Even the most sophisticated system can encounter exceptional situations, and a clear process is needed for how to respond to these situations.
In particular, it is important to clearly define the range of exceptions that the system can handle, and to ensure that in other situations, the person in charge is notified immediately.
It is desirable to set the basic timeline for the introduction of AI and the establishment of a system in the manufacturing industry at 12 months. This is the optimal period derived based on the analysis results of 50 AI introduction projects in the domestic manufacturing industry in 2023.
The first to third months are the stage of building the basic data infrastructure. During this period, you should focus on analyzing the current system and improving data quality. It is especially important to establish a data integration plan by closely analyzing the data structure of existing systems such as ERP and MES.
A common mistake made at this stage is spending too much time trying to achieve perfect data.
The fourth to sixth months are the pilot project execution phase. During this period, a single product group with a high contribution to sales is selected to verify the effectiveness of the AI system.
The success of the pilot project is a powerful driving force for change management across the organization. In fact, 85% of successful AI adoption cases have adopted this phased approach.
The seventh to ninth months are the system implementation and expansion phase. The company builds an enterprise-wide system based on the experience gained from the pilot project. The important thing in this process is to continuously reflect the feedback from the business departments. Improving the usability and practicality of the system should be prioritized over technical perfection.
The 10th to 12th month is the stabilization and advancement stage. During this period, internal capabilities are strengthened for the stable operation of the system, and a continuous improvement system is established. In particular, sufficient time should be spent on training operational personnel and maintaining manuals.
The risks that may arise during the process of building an AI system should be managed by categorizing them into three types: technical, organizational, and operational risks. Specific countermeasures for each risk should be prepared in advance, which is an essential element for the success of the project.
The most important aspect of managing technological risks is data quality control. Clear quality standards must be established from the data collection stage, and the reliability of the data must be secured through regular monitoring. An emergency response plan for system failures and performance degradation is also essential.
Organizational risks are mainly related to resistance to change. To overcome this, it is important to secure the active participation of the business departments, along with the firm support of the management.
In fact, according to successful cases, companies that have established a channel for employees to participate in and express their opinions on projects from the early stages have a success rate that is more than three times higher than companies that have not.
The performance of an AI system should be managed in a balanced manner, with quantitative and qualitative indicators. Quantitative indicators include prediction accuracy, inventory turnover, and delivery compliance rate, while qualitative indicators include user satisfaction, decision-making speed, and work efficiency.
Performance measurement should be done systematically on a monthly, quarterly, and annual basis, and the measurement results should be used as basic data for system improvement. In particular, continuous monitoring of the return on investment (ROI) is important, as it serves as the basis for decision-making on future system expansion or enhancement.
In fact, an analysis of successful cases shows that, on average, companies achieve a 30% reduction in inventory costs, a 25% improvement in forecast accuracy, and a 40% improvement in operational efficiency within one year of introducing an AI system.
However, these results may vary depending on the situation of the company and the scope of the introduction, so each company should set realistic targets that take into account its own characteristics.
For Company A, accurate inventory management was a key factor in its competitiveness due to the nature of its industry, which involves producing a wide variety of products in small quantities.
The existing inventory management method was carried out manually by the person in charge using Excel, which took an average of 15 days per month. This resulted in serious inefficiency of up to 180 days per year and also led to a decline in data accuracy due to manual work.
Company A introduced Deepflow, an AI-based inventory management system, to solve these problems. After the system was introduced, the time spent on inventory management was dramatically reduced from 15 days a month to 7 minutes, and the accuracy of forecasting improved from 70 to 80 percent.
This change has led to a qualitative improvement in strategic decision-making, beyond simply streamlining work.
In particular, the Deepflow system has enabled the company to determine the optimal inventory level by comprehensively considering the characteristics of each flavor, expiration date, and substitutability. This has enabled Seoul Flavor to simultaneously ensure the accuracy and efficiency of inventory management, which is drawing attention as a successful example of digital transformation in the flavor industry.
Company B was facing supply chain inefficiency due to the “whiplash effect” that commonly occurs in the electronics manufacturing industry. It was difficult to solve the problem of inaccurate demand forecasting and the resulting excessive inventory costs with the existing ERP system alone. This was a major factor that hindered the company's profitability.
To solve these problems, Company B introduced the Deepflow system. After the introduction of the system, the company saw a 49 percent reduction in inventory shortages and a 70 percent reduction in inventory overstock, which led to a reduction in inventory costs of 1.1 billion won per month.
The system's greatest feature is that it can derive an optimal inventory management strategy by taking into account the variability of lead times of component suppliers, the seasonality of market demand, and even new product launch plans. This has improved visibility across the supply chain, which has led to a qualitative improvement in strategic decision-making.
The introduction of AI in the manufacturing industry is no longer an option, but a necessity. This is because data-based decision-making has become a key factor in the survival of companies due to intensified global competition and increased market uncertainty.
However, this change can be a significant burden for mid-sized manufacturing companies. This is because there are realistic constraints such as limited budgets, insufficient professional manpower, and inadequate data infrastructure. The phased approach strategy presented in this report is a feasible solution that takes these realistic constraints into account.
In particular, the introduction of AI in the areas of demand forecasting and inventory management can be expected to have a high investment return with relatively low risk. As confirmed by the analysis of actual cases, it was possible to achieve tangible results within a year through a systematic approach.
In the future, AI technology is expected to develop further, lowering the cost of adoption and expanding the scope of application.
In particular, the development of cloud-based services is expected to significantly lower the barriers to AI adoption for companies. From 2024, the government's support for the digital transformation of small and medium-sized enterprises is also expected to expand, which will further improve the environment for AI adoption.
However, the introduction of AI is more than just the implementation of a simple technology. It is a process that entails a fundamental change in the way a company operates and its decision-making system. Therefore, the firm commitment of the management and company-wide participation are essential.
Manufacturing companies should recognize that now is the time to take the first step toward introducing AI. Even if it is not a perfect preparation and large-scale investment, a strategy that starts with a small-scale pilot project and gradually expands will be effective.
We hope that the framework and implementation strategies presented in this report will serve as practical guidelines for manufacturing companies to successfully adopt AI. While the adoption of AI is no longer an issue that can be put off, it will be a fully achievable goal with a systematic approach and clear strategy.