The Agentic Controller: Designing a Self-Correcting Multi-Agent System for Corporate Finance

This thesis addresses the design, implementation, and evaluation of a self-correcting multi-agent system for corporate finance, developed in collaboration with KPMG Czech Republic for a mid-sized retail company. The motivation stems from the challenge of data democratization — the gap between enterprise data availability and the ability of senior executives to leverage it without IT mediation. The system enables natural language querying of financial data through a conversational interface, replacing request-response workflows with direct self-service analytics. The architecture comprises six orchestrated components, achieving 90 % accuracy on a held-out test set with zero hallucinations detected across all test queries — confirming that deterministic verification mechanisms effectively eliminate a critical failure mode of LLM-based systems. The economic analysis evaluates three deployment variants, all yielding positive net present value, discounted return on investment exceeding the required threshold, and payback periods significantly shorter than typical enterprise IT projects, confirming the investment viability of agentic systems for mid-sized enterprises.