RAJARAJAN GANESAN

DOI: https://doi.org/10.5281/zenodo.21606742

Enterprise Resource Planning (ERP) systems remain, for most users, an exercise in navigating deep menu hierarchies, memorized transaction codes, and rigid form-based screens. This paper proposes and evaluates a reference architecture for real-time, voice-driven interaction with ERP systems built around large language model (LLM) function calling. The architecture couples a streaming automatic speech recognition (ASR) front end with an LLM-based orchestrator that maps natural spoken utterances to strongly typed function calls against a governed tool registry, interposes a guardrail and policy validation layer before any ERP-side effect is committed, and returns results through low-latency text-to-speech (TTS) synthesis. We describe the design of the function-calling schema, the security and audit boundary between the LLM and enterprise systems of record, and a prototype implementation spanning finance, inventory, human resources, and customer relationship management (CRM) modules. Evaluation on a corpus of 1,240 synthetic ERP task utterances shows the proposed pipeline achieving 97.2% single-function call accuracy and 91.5% accuracy on multi-step chained tasks, compared to 78.4% and 54.2% respectively for zero-shot prompting baselines, while sustaining a median end-to-end turnaround of approximately 840 milliseconds. The results suggest that schema-constrained, guardrail-mediated function calling is a practical and auditable path toward voice-native ERP interaction at enterprise scale.

large language models; function calling; enterprise resource planning; conversational AI; automatic speech recognition; voice user interface; tool use; enterprise integration; guardrails; retrieval-augmented generation

This paper presented a reference architecture for real-time, voice-driven ERP interaction built around LLM function calling, structured around four layers: voice capture and streaming ASR, an LLM-based function-calling orchestrator with scoped tool retrieval, a deterministic guardrail and policy validation layer, and an enterprise integration layer with full audit logging. A prototype evaluation across finance, inventory, HR, and CRM modules showed that schema-constrained, guardrail-mediated function calling substantially outperforms simpler prompting strategies, reaching 97.2% accuracy on single-function tasks and 91.5% on multi-step chained tasks, while sustaining a median end-to-end latency near 840 milliseconds. These results support the broader claim that the binding constraints on production voice-driven enterprise AI are architectural (scoping, validation, and auditability) rather than purely a function of underlying model capability.
Future work includes field evaluation with production ERP user populations across multiple industries; extension of the tool registry and guardrail rule set to manufacturing, logistics, and regulatory-reporting modules; formal risk-scoring for the destructive-action classifier in place of manually assigned thresholds; and investigation of proactive clarifying-turn strategies to reduce error propagation in multi-step chained tasks, which the results in Section 7 identify as the largest remaining source of error.

[1] Anonymous, "Function Calling in Large Language Models: Industrial Practices, Challenges, and Future Directions," OpenReview preprint, 2025.
[2] T. Schick, J. Dwivedi-Yu, R. Dessì, R. Raileanu, M. Lomeli, L. Zettlemoyer, N. Cancedda, and T. Scialom, "Toolformer: Language Models Can Teach Themselves to Use Tools," Advances in Neural Information Processing Systems (NeurIPS), 2023.
[3] M. Fowler, "Function Calling Using LLMs," martinfowler.com, 2024. [Online]. Available: https://martinfowler.com/articles/function-call-LLM.html
[4] S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. Narasimhan, and Y. Cao, "ReAct: Synergizing Reasoning and Acting in Language Models," International Conference on Learning Representations (ICLR), 2023.
[5] Floworks Research Team, "Benchmarking Floworks against OpenAI and Anthropic: A Novel Architecture for Enhanced LLM Function Calling," arXiv:2410.17950, 2024.
[6] G. Zeng, W. Ding, B. Xu, C. Zhang, W. Han, G. Li, J. Mo, P. Qiu, X. Tao, W. Tao, and H. Hu, "Adaptable and Precise: Enterprise-Scenario LLM Function-Calling Capability Training Pipeline," arXiv:2412.15660, 2024.
[7] Databricks, "Beyond the Leaderboard: Unpacking Function Calling Evaluation," Databricks Engineering Blog, 2024.
[8] Anonymous, "Voice-Activated AI Assistants in ERP and CRM Applications," ResearchGate preprint, 2025.
[9] Anonymous, "Routine: A Structural Planning Framework for LLM Agent System in Enterprise," arXiv:2507.14447, 2025.
[10] Anonymous, "Voice Assistants: Future of Interaction," ResearchGate publication, 2022.
[11] Anonymous, "Transforming Enterprise Resource Planning with Voice and Text-Based AI Assistants," ResearchGate preprint, 2025.
[12] A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, "Robust Speech Recognition via Large-Scale Weak Supervision," Proceedings of the 40th International Conference on Machine Learning (ICML), 2023.
[13] M. Zuluaga-Gomez et al., "ATCO2 Corpus: A Large-Scale Dataset for Research on Automatic Speech Recognition and Natural Language Understanding of Air Traffic Control Communications," arXiv:2211.04054, 2022.
[15] Anonymous, "Augmented Intelligence with Voice Assistance and Automated Machine Learning in Industry 5.0," PMC / National Library of Medicine, 2024.

RAJARAJAN GANESAN (2026). Real-Time Voice-Driven ERP Interaction: An LLM Function-Calling Architecture for Enterprise Resource Planning Systems. International Journal of Computer Techniques, 13(4), 108–121. ISSN: 2394-2231. DOI: https://doi.org/10.5281/zenodo.21606742