Lee Jun-ki’s Big Data
big data
Two years have passed since ChatGPT was announced. There have been many changes in the meantime. Last week, during a big data class, students were asked whether they had a monthly subscription to generative AI (artificial intelligence), and 29 out of 30 students were subscribed. There were 15 students who subscribed to two or more subscriptions. Companies are also competitively introducing generative AI. According to Forrester’s global survey released last June, 80% of executives expected generative AI to play a major role as a strategic asset for their companies within a year.
Attention as an attempt to demonstrate generative AI performance
Individuals within companies are using generative AI for report writing, drafting, translation, simple programming, and information search, and at the corporate level, it is also being used for customer consultation and providing guides based on corporate manuals. However, there are some limitations for companies to use the current large-scale language model (LLM) as is. First, the trained model has general knowledge, but does not contain company-specific information. Second, generative AI is still not completely free from the hallucination phenomenon of providing false information. Third, the latest external information, such as from the web, is not included in the training. To solve this problem, companies use N-shot learning through prompts (a technique that provides examples of answers and induces answers in an illustrative format) and RAG (Retrieval Augmented Generation: a method that presents the context of a specific document and induces answers within it). , Fine Tuning (a method of training a new LLM to suit the purpose), etc. are used. However, these attempts are still mainly used to improve individual productivity. There are major barriers to connecting this to corporate productivity. According to the global survey mentioned above, contrary to expectations, the proportion of respondents who responded that they are actually achieving results through generative AI was only slightly more than 20%.
Recently, research and attempts that can show great results through generative AI in terms of corporate strategy are attracting attention. One of them is an AI business solution through a multi-agent system. Simply put, a multi-agent system is a method of creating multiple computer agents and giving each one a different purpose and orientation. For example, instead of asking the generative AI, “Create a travel schedule to Hokkaido, Japan for two weeks in November with a budget of 8 million won,” say, “My name is the head of a travel agency. Please make a flight schedule that suits him.” “For Jiyeon, search Instagram to find travel destinations that women in their 30s would visit near Hokkaido,” “For Donghu, search hotel sites and recommend accommodations that fit their budget.” roles are given specifically.
Individuals can easily assign multiple roles to generative AI, but the key here is that each agent communicates with each other and autonomously coordinates toward a single goal. For example, choosing a travel destination and choosing accommodation are closely related. An individual can review the travel itinerary created by an agent and pass that information back to the generative AI that is using it to complete the travel plan. However, in a multi-agent system, each agent performs tasks independently. Each agent plays a role such as passing on newly generated information to other agents while performing tasks, conducting additional analysis, or determining how the supervisory agent interprets different information. Each agent uses different tools such as web search, data analysis, and PDF document search, and has its own information access rights, regulations, and memory.
In fact, the multi-agent concept is not new. However, with the development of artificial intelligence, a new era has opened with the emergence of a system that allows each agent to make decisions independently. For example, in supply chain management, one agent may determine current inventory, another agent may analyze hourly sales, and yet another agent may negotiate with suppliers to determine inventory replenishment and purchase prices. The advantage of this model is that artificial intelligence can make analysis-based decisions for various product groups in real time. Even today, many companies can automatically handle these processes with rule-based systems. However, unlike rule-based processes, multi-agent systems can actively respond to environmental changes and can also design processes (escalation) that make the final decision with human intervention in ambiguous cases.
Although these multi-agent AI business solutions are still in their early stages, systems implementing them are appearing on the market and showing visible results in a short period of time. Let’s take a look at a hospital case in New York that was implemented this year. In Korea, medical insurance is run by the government, and the state pays the remaining hospital expenses after treatment to the hospital in the form of medical insurance. In the United States, insurance companies intervene instead of the state in this process. American hospitals bill insurance companies for the remaining amount after treatment, but there is a conflict of interest between hospitals and insurance companies. It is essential for hospitals to preserve medical expenses, but insurance companies sometimes refuse to pay for various reasons. Recently, it is known that about 22% of billed medical expenses are rejected, which causes significant losses to hospitals. Hospitals must resubmit claims for rejected claims. This is a very complex process that combines medical knowledge and regulations, and cannot be resolved with simple administrative knowledge alone, requiring the time and effort of a specialist.
Linkage between AIs may produce unexpected results
This hospital introduced a multi-agent AI solution to solve the problem. One agent extracted reasons from the insurance company’s denial letter, while another took patient records and checked the eligibility rules against the denial reasons. Another agent was responsible for extracting compliant patient records. If additional patient records are needed, the artificial intelligence agent requests them from the doctor. Finally, another agent gathers this information to write an appeal letter to the insurance company, and sends feedback to the supervising agent for final approval or re-extraction of the patient record if any deficiencies are found. As a result, the hospital wrote and sent more than 2,500 appeal letters in just 6 months, and considering the time and manpower invested in the retrial process, it achieved cost savings of approximately 1.5 billion won in addition to the medical fees refunded through the successful retrial.
As in the case of the New York hospital, the multi-agent AI system does not use only generative AI separately. It is a comprehensive system that combines the company’s database system, current system, machine learning analysis, and document and regulatory information. One of the advantages of this system is that it stores the company’s past business flows in the company’s memory. At this hospital, not only can doctors refer to the existing reasons for denial and the reconsideration process when making future insurance claims, but they can also operate a simulation system to predict the insurance company’s response.
Multi-agent AI systems are still in their early stages, so there are many challenges to be solved. In order to encourage smooth collaboration between agents and build a successful system, high-level business knowledge, AI knowledge, and database knowledge must be harmoniously combined. In addition, unlike before, we must prepare for unexpected results that may occur as decisions are made through interaction between artificial intelligence. Therefore, it is important to increase the transparency of the system and design the monitoring process together. In addition to the initial construction cost, it is necessary to consider the token cost (charge) incurred when using commercial LLM.
In order to effectively implement this system, companies must think deeply about where in their processes the system should be applied to achieve maximum benefit. The key is not to simply replace humans, but to design how artificial intelligence agents and humans will cooperate as new decision-making entities. Despite these expected difficulties, the potential of multi-agent AI systems poses too great a risk for companies to ignore them at this point.
Junki Lee is a professor at the Graduate School of Information at Yonsei University. After graduating from Seoul National University with a degree in computational statistics, he received a master’s degree in social psychology from Carnegie Mellon University and a doctorate in business administration from the University of Southern California. He serves as co-chairman of the National Public Data Strategy Committee along with the Prime Minister and has published books such as 『Manage with AI』, 『Open Collaboration』, and 『Web 2.0 and Business Strategy』.
