Hana Bank has developed a search service using an in-house finance-specialised artificial intelligence (AI) model, cutting its search failure rate from about 20 percent to within 5 percent. It is now testing the service for use in the Hana1Q app, and plans to use AI-UX and a knowledge graph to strengthen its ability to handle complex financial queries.

At the 'Digital Insight 2025' conference hosted by DigitalToday at COEX in Seoul's Samseong-dong on the 29th, Woo-seob Lee (이우섭), head of Hana Bank's Digital Strategy Business Department, introduced the bank's AI buildout cases and future strategy under the theme, 'Financial AX, attempts and lessons.'

Hana Bank is pursuing a parallel approach of using external commercial AI while running its own AI on an internal network, considering the financial sector's regulatory environment such as network separation and personal data protection. It said building its own system allows it to use internal data without exporting it outside, but comes with a heavy burden for infrastructure buildout, operations and maintenance.

To that end, Hana Financial Group developed a finance-focused foundation model, HNFM, based on an open-weight model. It used about 100 GB of data for training to improve understanding of Korean-language financial knowledge and complex financial queries.

In the Q&A session, Lee said, "Personal information was not included in the training data, and we used data related to laws and regulations, internal rules, terms and conditions, and products and services."

It is also testing a search service using HNFM. Hana Bank said its existing keyword-based search failed to find correct answers for some inquiries, pushing the overall search failure rate to about 20 percent, but that fell to within 5 percent after it implemented its own service. The service is not yet fully applied to the Hana1Q app and is under testing.

Hana Bank said it sees limits to handling financial tasks with retrieval-augmented generation (RAG) alone. It said ambiguous questions such as "sending money to another bank" can be interpreted as multiple tasks, including interbank transfers, overseas remittances and automatic transfers, and that complex tasks such as loans require consideration of multiple criteria, including laws, internal rules and terms and conditions.

Hana Bank plans to raise AI use by applying 'AI-UX' that supplements missing user context through follow-up questions, and a knowledge graph that structures priorities and relationships in financial knowledge such as laws, internal rules and terms and conditions.

Hana Bank said trust and control are important even if AI use expands in the financial sector. Considering the regulatory environment such as network separation, personal data protection and AI guidelines for the financial sector, it is pursuing a direction that combines external AI use with in-house development on internal networks.

Lee said, "Even if AI use expands in the financial sector, the core lies in trust and control," and added, "We will view financial regulation positively and continue to consider what we will do and how we will proceed within that regulatory environment."