How MyRealTrip's Finance Director Built an AI Agent System

Lee Seung-gil, Director of the Finance Management Office at MyRealTrip and a 12-year veteran of accounting and finance, built his own AI staff using Claude Code. Rather than writing code directly, he used a vibe coding approach—designing the system through conversation—focusing on automating repetitive accounting, tax, and legal-reference tasks.

On the first day, he created 8 skills such as YouTube summarization, Slack summarization, and company wiki search, and added a meeting-minutes auto-analysis system along with contract review, accounting advisory, and inspection agents, securing 7 finance-specialist agents in a single day. The next day, he integrated KIFRS to make standards, paragraphs, and interpretive rulings searchable, and connected DART OpenAPI for disclosure monitoring, bringing the total to 9 agents and 2 APIs.

On the third day, he built a settlement-analyst agent to analyze settlement structures for products like flights, hotels, and tours, and connected the company's Confluence wiki and a tax-law ruling search system. At this point, agents reached 19 and APIs reached 3.

On the fourth day, he discarded existing templates and completely rewrote 21 agents to fit the context of MyRealTrip's Finance Management Office. Accounting advisory incorporated KIFRS and samili, tax advisory incorporated the Corporate Tax Act, Value-Added Tax Act, and Income Tax Act, and internal controls incorporated K-ICFR and COSO, with every answer required to carry a source tag. At the same time, he built a regulation-alert skill to detect changes from Samil.com, the K-ICFR website, and FSS press releases on a weekly basis.

On the fifth day, he rebuilt fintech-engineer, risk-manager, and business-analyst—chosen from the marketplace—customized for MyRealTrip, and added a Google Sheets-based task inspection dashboard and a tech-advisor agent to assist with technical decision-making. At this stage, the build was evaluated at grade A.

On the sixth day and in a final push, he integrated the National Law Information Center's statute API to directly look up provisions such as the Corporate Tax Act, Commercial Act, and Civil Act, and extended statute-reference capability to 6 agents. This completed a structure that cross-verifies accounting standards, tax rulings, legal provisions, and DART disclosures, and with an organizational chart centered on a multi-agent-coordinator and a Weekly Brief skill finalized, he received a final grade of A+.

The build results are summarized as 26 sub-agents, 21 custom skills, 5 in-house APIs, and 4 automation pipelines. Real-time numerical analysis and FP&A integration remain as tasks for the future, but Director Lee Seung-gil sums it up: "The key to AI Native transformation isn't development skill—it's the ability to define problems."

• He structured repetitive tasks to be handled by AI
• He required sources for answers that need evidence
• He wove accounting, tax, law, IR, IT, and business analysis together like a single command staff.