Lee Young-soo, head of the AI Research Institute at Shinhan Bank, said on the 29th, “Agentic AI is not just a simple chatbot. It is an autonomous system that thinks, plans, and executes on its own.” However, he emphasized that sufficient learning must be a prerequisite for Agentic AI to function properly.
During a lecture at the ChosunBiz ‘2025 Future Finance Forum’ held at the Westin Chosun Hotel in Jung-gu, Seoul, the director introduced actual application cases of Shinhan Bank and the flow of technological development under the theme of ‘AI, big data, and future finance.’
First, he explained that while there are various definitions of Agentic AI, there are common essential requirements.
He said, “These days, even simple prompt-based large language models (LLM) are referred to as ‘agents,’ but true Agentic AI must execute tasks based on reasoning and planning and should be able to learn the process to achieve personalization.”
He also explained that recent global financial companies like Mastercard are gradually introducing Agentic AI. In particular, Mastercard is implementing ChatGPT-based payment and reservation services, while Goldman Sachs is applying it to internal control and transaction monitoring systems.
The director noted that Shinhan Bank is also actively working to adopt Agentic AI. Currently, Shinhan Bank is considering the introduction of Agentic AI in areas such as ▲ AI branch ▲ AI private banker (PB) ▲ internal control and risk management.
First, Shinhan Bank's AI branch is designed to handle the entire customer journey on its own. According to the director, the AI branch can currently complete simple tasks, such as card issuance, in under one minute.
In the asset management sector, the AI PB project is underway. The director emphasized that a multi-agent system composed of market experts and investment strategists is key to the AI PB.
The multi-agent structure is designed to allow multiple agents performing a single task to be organically connected, enabling comprehensive asset management tasks to be carried out independently.
Third, the director introduced the potential application of Agentic AI in internal control and risk management. He stated, “According to the five stages of AI development proposed by OpenAI, we have currently reached the third stage, where ‘planning and execution-based reasoning’ is possible.”
However, the director pointed out that there are also significant challenges to be addressed, as much as there are expectations for the technology.
He explained, “While there are expectations that everything will be automated with the introduction of Agentic AI, in practice, the realization of a ‘data hell’ occurs due to a lack of data and unstructured work knowledge.” He noted that ultimately, AI is similar to a ‘new employee,’ as without systematically learning work processes and regulations, practical application becomes difficult.
He also pointed out that in complex and unstructured tasks like internal control, it is difficult for LLM to ensure consistency and accuracy.
He added, “There is a risk of making repeated poor judgments, and for Agentic AI to deliver practical results, it is essential to refine the internal context and domain knowledge of corporations so that the AI can understand it.”