Inventory management is a long-standing problem for the manufacturing and distribution industries. Despite attempts to introduce ERP, hire specialized personnel, and conduct various consultations, 64% of companies still suffer from inventory loss. In particular, since the 2000s, the rate of increase in the inventory index has more than doubled, deepening the concerns of small and medium-sized enterprises.
This is because the problems that could not be solved by the old methods are no longer solvable. There are more variables that are difficult to handle with traditional Excel work or ERP systems, such as the increasing complexity of global supply chains, rapid changes in consumer demand, and increased volatility in raw material prices.
However, there is good news. Recently, companies have been making dramatic changes by introducing demand forecasting AI.
In fact, one mid-sized manufacturing company reduced its excess inventory by 70% and saved 1.1 billion won per month after introducing AI. Another company reduced its inventory management work from 15 days to 7 minutes while increasing its prediction accuracy by more than 70%.
At the core of these success stories was a “new level of predictive technology” that went beyond simple automation or data analysis. AI technology that analyzes vast amounts of external data in real time, from 1,700 macroeconomic indicators to 6 million pieces of trend data, is changing the game of demand forecasting.
In this interview, we met with Jung Doo-hee, CEO of Inffective AI, a leading company in the field of demand forecasting AI, and talked about practical insights needed in the field.
This is an interesting question. Generative AI and predictive AI are essentially solving different problems. Generative AI is a tool that supports human creative work, while predictive AI is a tool that supports decision-making directly related to the survival of a company.
What is particularly noteworthy is that the advent of the era of generative AI is making the importance of predictive AI even more prominent. This can be explained in three ways.
First, the pace of market change is accelerating due to the creation of AI. New products and services are being launched faster, and consumer preferences are changing faster, making accurate demand forecasting increasingly important.
As the McKinsey report also points out, corporate decision-making should ultimately be based on quantitative analysis. No matter how great the ideas generated by AI are, their feasibility and business potential must ultimately be verified through demand forecasting AI.
Finally, in an increasingly uncertain business environment, the importance of data-based decision-making is growing. Decision-making based on intuition or experience is no longer sufficient.
First of all, generative AI can be used to improve the accuracy of demand forecasting.
For example, we use generative AI to generate time-series augmented data to supplement insufficient data. When there is a lack of historical sales data for a particular product, we create new scenarios based on data from products with similar patterns.
On the other hand, demand forecasting AI is used to optimize the results of generative AI.
For example, let's say you've created various marketing phrases using generative AI. Our demand forecasting AI helps you establish the most effective marketing strategy by predicting the expected conversion rate and ROI for each phrase.
The most innovative use case is 'scenario-based decision support'. When the generative AI creates various business scenarios, the demand forecasting AI quantitatively analyzes the feasibility and expected performance of each scenario. This enables companies to develop more sophisticated strategies.
This is a very important question in a company's digital transformation strategy. What I want to emphasize is that this is not a question of “choice” but a question of “stage.”
Demand forecasting AI is like strengthening the foundation of a company. It directly affects the company's core operations, such as inventory management, production planning, and raw material procurement. In fact, our customers have seen immediate results, with a 10% improvement in demand forecasting accuracy leading to a 30% reduction in inventory costs.
On the other hand, generative AI is a tool that strengthens a company's innovation capabilities. It can be used to discover new product ideas, create marketing content, and improve customer service. However, in order for this innovation to lead to tangible results, basic operational efficiency must be ensured.
Therefore, we propose a strategy of “demand forecasting AI first, then creation AI.” It is desirable to first strengthen the company's basic strength through demand forecasting AI, and then pursue innovation through creation AI based on this. This is also what we emphasize under our motto of “empowering users.”
Inventory management is a classic problem that has been studied for more than 50 years in the field of industrial engineering, but the reality is that 64% of companies still suffer from inventory loss. What is particularly noteworthy is that the annual growth rate of the producer goods inventory index has more than doubled since the 2000s.
This means that the imbalance between supply and demand in the industry has become more serious. In fact, even companies with annual sales of 2 trillion won have annual inventory costs of 20 billion won.
This shows that the existing inventory management method is no longer able to keep up with the complexity of modern industry.
This is because the complexity of the industry and the nature of data have fundamentally changed. In the past, inventory management assumed a relatively simple demand-supply pattern and created an optimization model.
However, the complexity of the global supply chain has increased dramatically since the 2000s, as evidenced by the fact that the annual growth rate of inventory indices has more than doubled.
What is particularly noteworthy is that the demand patterns of today are so complex that they cannot be explained by simple mathematical models. Consumer behavior has diversified, the impact of global events is reflected in real time, and market volatility has increased significantly.
Traditional industrial engineering approaches have had limitations in capturing these nonlinear and multivariate patterns. However, AI approaches this problem differently.
In other words, it does not use mathematical optimization, but rather learns and predicts patterns from vast amounts of data. Deepflow has become a game changer in the industry because it has been able to comprehensively analyze data at a scale that was unimaginable in the past, ranging from 1,700 macroeconomic indicators to 6 million pieces of trend data.
The biggest limitation of existing ERP systems is that they only use internal data. Actual demand is greatly affected by the market environment, and it is difficult to predict such volatility with internal data alone.
In addition, while some companies' internal IT teams use open source machine learning models, there are limitations to dealing with complex patterns of data.
On the other hand, Deepflow has three key differentiators.
First, it accurately predicts the shipment and order volume for each item for the next six to 12 months using a high-performance AI prediction model. Second, based on the prediction results, it provides intelligent inventory management optimization to automatically calculate the optimal production order volume and production schedule.
Finally, it has implemented explainable AI that can explain the key factors behind the increase in sales by quantitatively analyzing the causes of the prediction results.
This is not a matter of simple task automation, but of an approach that should be taken as an 'evolution of roles'. Looking at the flavoring company's case, the work that a person in charge used to spend 15 minutes on every month was reduced to 7 minutes. However, the key is not to replace the value of the person in charge, but to enable them to create higher value.
Specifically, Deepflow has designed this change in three directions. Based on the various scenarios and cause analyses suggested by AI, the person in charge can focus on making more strategic decisions.
In addition, while the AI processes everyday patterns, the person in charge can focus on making expert judgments and responding to unusual patterns or market changes. Finally, business improvement. The time saved can be used to invest in more valuable tasks, such as improving processes and discovering new opportunities.
What is particularly noteworthy is that it is designed to maximize the synergy between the expertise of the people in charge of the field and AI. This is because the predictions and insights provided by AI are ultimately developed into optimal decision-making through the experience and insights of the experts in the field.
Deepflow is designed as a fully automated AI system. First, internal ERP data and external environmental data are collected and combined.
Then, the automated system automatically performs processes such as data standardization, preprocessing, feature engineering, and model training for AI model training. What is particularly important is that it is designed as a no-coding solution, so even non-AI experts can easily use it.
It also provides predictive insights that can reduce inventory loss and increase operating profit margins from a future-oriented perspective.
It goes beyond simple numerical predictions and serves as a strategic tool to support corporate decision-making. Users can easily view and utilize all of this information through a user-friendly UI.
We have deep technical expertise in a specific area called 'demand forecasting'. We currently operate 224 forecasting models, which is a significant number compared to the 10-20 models typically operated by a typical internal AI team.
In particular, our AI Stacking Ensemble Forecasting Model is an advanced forecasting system that combines various machine learning algorithms hierarchically to overcome the limitations of a single algorithm.
For example, it simultaneously utilizes multiple algorithms, such as LSTM, which is strong in time series analysis, XGBoost, which excels in detecting nonlinear patterns, and Random Forest, which captures interactions between characteristics. It also holds 35 patents (including 5 US patents) and is increasing its prediction accuracy through continuous R&D investment.
Let me explain with a real-life example. In the case of a remote control manufacturer, after introducing our solution, the company reduced inventory shortages by 49% and inventory overages by 70%, achieving a monthly cost reduction of 1.1 billion won.
In another example, a company that manages 1,500 flavors improved the process of calculating the inventory management that previously took 15 days (180 days a year) to 7 minutes after the introduction of AI.
In addition, the accuracy of the forecast has improved by 70-80%. Our solution is characterized by improving work efficiency and accuracy at the same time, beyond simply reducing costs.
Improvement through Deepflow is achieved in three steps. First, there is a technical improvement through the application of advanced predictive AI. This is manifested in improved prediction accuracy. Second, this technical improvement leads to improvements in work processes.
For example, inventory management work time is greatly reduced and work efficiency is increased. Finally, these improvements ultimately lead to financial improvements, which are manifested in reduced inventory costs and improved operating margins.
Importantly, this improvement process is not a one-time event but is continuously expanded and reproduced. As the AI model learns more data, the accuracy of the predictions improves, which in turn creates a virtuous cycle that leads to greater business value.
This is important in that it solves the “black box problem,” which has been the biggest hurdle to AI adoption. It is difficult for corporate decision-makers to make important decisions based solely on the numbers presented by AI. Inventory management, in particular, is a sensitive area that is directly linked to a company's cash flow.
Deepflow quantitatively analyzes and presents the causes of the predicted results. For example, if it predicts that the demand for a certain product will increase by 30%, it shows whether this is due to macroeconomic factors, industry-specific factors, or seasonal factors.
This creates three important business values.
First, the accountability of decision-making is strengthened. Rather than blindly following AI predictions, you can understand the basis for them and make strategic decisions. Second, preemptive response is possible. Knowing the factors that affect you allows you to establish strategies to prepare in advance.
Finally, organizational learning becomes possible. By analyzing the causes of when AI predictions are correct and when they are incorrect, the organization's decision-making capabilities can continue to develop.
This is why we have built a 'predictive risk management system' rather than a simple predictive model. Since prediction inherently involves uncertainty, it is important to not only make accurate predictions but also to respond quickly when predictions are incorrect.
Deepflow has implemented three core mechanisms to achieve this.
In fact, in the case of the remote control manufacturer, the reduction in inventory shortages by 49% and inventory excesses by 70% was not simply due to accurate predictions. It was due to the effective operation of this risk management mechanism.