Implementation of an LLM-Based Agentic Solution for Finance using Microsoft Foundry

This master's thesis focuses on the design, implementation, and verification of a functional prototype (MVP) of the conversational analytical tool DataChat, developed for the financial reporting needs within the BU Data & AI division at Seyfor, a.s. The main objective is to create a tool capable of reliably translating natural language user queries into SQL queries, interpreting the retrieved data, and supporting specific reporting scenarios. The thesis aims to answer the research question of how to design a secure and auditable Text-To-SQL solution in a corporate environment that minimizes the risk of Large Language Model (LLM) hallucinations and ensures full control over data outputs. In terms of methodology, the text combines a literature review of the theoretical framework for deploying artificial intelligence in finance, an analysis of business requirements, practical implementation, and subsequent experimental testing combined with quantitative evaluation. The practical part describes the development of the solution primarily on the Microsoft Foundry platform and its migration to the LangGraph framework. The author's main contribution is the end-to-end implementation of the initial multi-agent orchestration, the design of visualization logic, and the development of automated reporting. The result of the thesis is a tested solution that demonstrates the real possibilities and limitations of using agentic systems in corporate financial controlling.