Towards Explainable Automation of Decision-Making in Tax Consulting: Integrating LLMs with Logical Reasoning for Selection of Transfer Pricing Method

Tax consulting involves interpreting natural-language case descriptions to make logic-based decisions, such as selecting a transfer pricing method for multinational enterprises. While large language models (LLMs) are effective at processing natural language, they lack true logical reasoning capabilities. In contrast, logic programming, e.g., using Prolog, offers explainable and correct reasoning when given a structured fact base. This short paper proposes a hybrid AI approach that combines the strengths of LLMs and Prolog: The LLM extracts structured facts from textual case descriptions using a guided, question-based process, and a Prolog program applies logical rules to perform reasoning. A human remains in the loop to address ambiguities. We present a proof-of-concept implementation and an evaluation of the approach using real-world transfer pricing cases from a tax consulting firm.

Keywords: Hybrid artificial intelligence, logic programming, large language models