Neuro-symbolic reasoning engine for tax optimisation

Automating tax calculations and optimisation is challenging because modern AI systems such as large language models operate probabilistically, while legal and financial reasoning requires deterministic compliance with statutory rules. This research presents a new approach to automate taxes in India through the Neuro-Symbolic Tax Optimizing Engine (NTOL) that combines the advantages of large language models (LLMs) and a deterministic symbolic reasoning component based on a structured knowledge graph of tax statutes. This research builds on prior work in neural semantic parsing, combining neural semantic parsing with formalized symbolic rules that are immutable ensuring that output complies with the law, producing reproducible calculations and being independently auditable. The system was evaluated using a curated benchmark dataset consisting of 20 tax computation scenarios derived from provisions of the Indian Income Tax Act, 1961, including business income adjustments, deductions under Sections 80C and 80D, and presumptive taxation under Section 44AD. The proposed approach was compared against two benchmarks, an LLM only system and with a simple Retrieval-Augmented Generation (RAG) system. Evaluation results indicate that the proposed approach outperformed both the benchmarks across multiple performance characteristics including improved accuracy, compliance with legal statutes, robustness to complexities within legal statutes and support for explainable results. In particular, in contrast to the LLM-only system, which achieved a 75% accuracy, and the simple RAG system, which only achieved a 60% accuracy; the proposed system achieved an accuracy of 80%. Most of the inaccuracies were due to incomplete knowledge graph coverage rather than hallucinated reasoning, whereby the symbolic layer limits non-valid computation paths and does not produce unsupported output when the system lacks an output. In addition, the system's performance was analyzed using qualitative methods such as determining semantic completeness and reasoning consistency. These qualitative methods provide an assessment that demonstrates the system generates computations that are consistently maintainable and auditable, therefore reducing inconsistencies, such as artificial intelligence hallucinations; this provides tremendous value to stakeholders through the provision of transparent computation trails that facilitate audit readiness, reduce the amount of time required for compliance verification, and help to mitigate the risk of litigation due to incorrect tax calculations.