- Generative AI consultation directly in the KakaoTalk chat window without web redirection... conversation history fully preserved

- Understands complex rules such as card issuer registration conditions and daily fares... performs pro-rata calculations and conditional reasoning

A generative artificial intelligence (AI) chatbot service that can be used directly through the KakaoTalk chat window has been introduced to Korea’s leading public transportation payment platform. This is a full-scale application of generative AI to the public service sector, where regulations are complex and answers change depending on each individual’s situation. The so-called “native” approach, in which generative AI responses are exchanged directly within the KakaoTalk chat window, is the first of its kind in the transportation, aviation, or mobility industries.


Makebot, an AI chatbot specialist led by CEO Kim Jiwoong, announced on July 29 that it has applied LLM (Large Language Model)-based generative AI consultation functions to the KakaoTalk chatbot for the country's largest transportation payment platform. The latest upgrade also incorporates Makebot's ‘HybridRAG’ technology, which was presented at SIGIR 2025, the world’s most prestigious AI academic conference, held in Padua, Italy last year.


No web redirection with “KakaoTalk Native” generative AI... uninterrupted conversation

While the adoption of generative AI and LLM-based chatbots is spreading across industries, most services until now have required users to move from KakaoTalk to a separate web browser or app. When the conversation moves to a web window, the chat disappears as soon as you close the window, making it impossible to revisit responses and leading to user drop-off.


This new chatbot allows users to ask questions and receive answers directly in their familiar KakaoTalk chat rooms. The conversation remains in the chat thread, so users can check previous answers days later and ask follow-up questions, such as “So, how does postpaid work?” which rely on earlier messages, supporting multi-turn responses. Messenger chatbots generally face restrictions in response time and message structure, making the direct use of LLMs—which generate long sentences in real time—challenging. Makebot addressed this limitation using its HybridRAG method.


Not just searching, but “reasoning”… LLM directly calculates complex fares

The most notable feature of this generative AI chatbot is its reasoning capability: it not only finds information but also makes judgments and performs calculations based on each user’s situation. For example, if a user asks, “If I register a postpaid Climate Companion Card on July 15, how much will I be charged for the first month?” the chatbot recognizes that July has 31 days and sequentially applies the standard monthly burden limit and the pro-rated reduction for the number of days the card was not registered, then calculates and provides the first month’s charge. Similarly, if asked, “If I use the Hangang bus twice, how much will be added to my monthly burden limit?” the chatbot informs the user that the base monthly maximum of 62,000 won is increased by an additional 5,000 won for bus usage, for a total of 67,000 won (based on 30 days).


The chatbot can also distinguish between regions and conditions. For questions like “Can I use it in Busan?” it explains the difference in eligible regions for prepaid and postpaid Climate Companion Cards. For questions without a clear, definitive answer, it refrains from giving a forced response, and for inquiries about specific error codes, it breaks down the possible causes by service. Thanks to the incorporated hallucination-elimination technology, it accurately recognizes abbreviations such as “Kidongka.”

Photo: Response screen of the KakaoTalk native generative AI chatbot built by Makebot. It informs the prorated monthly burden limit for the first month of the postpaid Climate Companion Card (left) and calculates and presents the monthly burden limit changes based on the use of the Hangang bus (right). (Photo by Makebot)

Photo: Response screen of the KakaoTalk native generative AI chatbot built by Makebot. It informs the prorated monthly burden limit for the first month of the postpaid Climate Companion Card (left) and calculates and presents the monthly burden limit changes based on the use of the Hangang bus (right). (Photo by Makebot)

View original image

NLP + generative AI hybrid... LLM calls reduced to 30%, token costs saved

Another core feature is a hybrid structure that does not assign every question to generative AI. Questions included in the completed NLP (Natural Language Processing) intent database are answered instantly without calling the LLM, and only questions outside of the intent list are handled by generative AI. In those cases, the chatbot displays a message saying, “Generating an answer,” to minimize user wait time.


The effect was evident from data on the first day of operation: of 974 inquiries processed, 685 (70.3%) were instantly answered by NLP intents, while 289 (29.7%) were handled by generative AI. Compared to processing all inquiries via LLM, token costs were cut to about one-third. Since generating each answer via generative AI incurs a token cost, operational expenses increase as user numbers rise. This has been an obstacle to adoption, but the hybrid structure allows for repetitive and common queries to be absorbed by NLP. As a result, annual operating budgets can be estimated even at the introduction stage, which is especially significant for public institutions facing strict budget limitations.


The HybridRAG technology, jointly developed by CEO Kim Jiwoong with Hanyang University, is a combination of LLM and RAG and achieved a verified 26.6% accuracy improvement and 6.2% faster response time compared to conventional RAG. It is also evaluated as optimized for the transportation payment domain because it can accurately interpret documents containing tables or charts.


974 inquiries processed in half a day... replaced 64.9 hours of human consultation work

The generative AI chatbot was officially launched at 1:42 p.m. on July 28, and within only half a day, it processed 974 inquiries. This output is equivalent to approximately 64.9 hours of staff consulting work, and when overtime and holiday premium pay are factored in (in accordance with Article 56 of the Labor Standards Act), the estimated labor cost saving is about 1.65 million won.


Makebot has integrated and operated customer centers for six services—including transportation cards, mobile top-up, and easy payment—into a single KakaoTalk chatbot on this platform, used by hundreds of thousands of users monthly. With the introduction of generative AI, service gaps during nights and weekends can now be filled without increasing staff.


This development is expected to accelerate the adoption of generative AI in other industries such as telecommunications and finance, which also handle complex rate plans and terms.


“The core of civil complaints is, ‘What does this mean for me?’... now intelligent generative AI can answer”

Kim Jiwoong, CEO of Makebot, commented, “Questions like whether someone qualifies for a local tax reduction, how many days remain before the application deadline, or the amount of support under individual conditions cannot be solved by just reading regulations. This project proves, in a large-scale public transportation service, that LLM can actually understand business documents and perform calculations and reasoning.” He added, “Previously, public institutions hesitated to adopt generative AI not because of performance, but because of operational costs. The same approach can be applied to any area with complex regulations and individual-specific cases, like municipal inquiries, public fees, or qualification screening.”



Meanwhile, Makebot has built chatbot services for dozens of clients, including Korean Air, Korea Airports Corporation, Kakao T, SRT, various municipalities such as Cheongju-si and Bucheon-si, Korea Tourism Organization, Korea Real Estate Board, Seoul Housing and Communities Corporation, and other public institutions. It has been selected as a supplier for the government's 'hyper-scale AI ecosystem construction' project and has filed dozens of LLM and RAG-related patents.


This content was produced with the assistance of AI translation services.

© The Asia Business Daily. All rights reserved. Unauthorized AI training and use prohibited.

Today’s Briefing