Document Type
Article
Publication Date
5-9-2026
Abstract
Enterprises are increasingly adopting Large Language Models (LLMs) to enhance automation in document driven workflows. However, their probabilistic behavior raises challenges related to governance, auditability, and regulatory compliance. This work proposes an agentic automation framework that integrates LLM-based cognitive services with Robotic Process Automation (RPA) to enable intelligent document processing while preserving determinism and traceability. The architecture separates probabilistic interpretation from deterministic execution, using LLMs for bounded information extraction and RPA for rule-based orchestration, validation, and audit logging. To mitigate risks associated with nondeterministic AI behavior, the framework incorporates guardrails such as structured prompting, confidence thresholds, human-in-the-loop review, and comprehensive audit trails. The approach is demonstrated through an enterprise use case involving automated case creation for unbilled invoices, showing reduced manual effort, improved processing accuracy, and end-to-end auditability. This work provides practical architectural guidance for deploying agentic AI systems responsibly within regulated environments