Towards a Conversational Invoice Issuance LLM-Based Agent

The traditional invoice issuance process within tax administration is labor-intensive and prone to errors, necessitating a shift towards digitalization. Despite the advent of digital invoicing systems that streamline invoice generation and automate rule-based audits, integration with existing financial accounting systems remains a challenge. Particularly in the hospitality and bookkeeping sectors, the adoption of these systems is hindered by the lack of standardized software, high costs, and the absence of technical expertise among small and micro enterprises. The integration of digital invoicing systems with diverse financial software presents significant barriers to uniform adaptation. Furthermore, the complexity of tax regulations and the dynamic nature of tax categories require advanced understanding beyond the capabilities of standard Large Language Models (LLMs). The need for a specialized system that can comprehend finance and tax contexts, securely handle sensitive information, and adapt to user interactions is paramount. This paper introduces an autonomous agent based on a finance and tax-specific Large Language Model (LLM) designed to address the aforementioned challenges. The system includes a Specialized Training Framework to enhance domain comprehension, a Hierarchical Memory Architecture for dynamic user interaction, and a Tax Domain Security Module to ensure compliance with tax regulations. The proposed agent aims to improve the efficiency and accuracy of the invoice issuance process, providing a robust solution for tax administration in the digital era.