Intelligent Automation of Invoice Processing and Validation Using Computer Vision and AI Assistants for Financial Control in Public Institutions

This work presents a modular, open-source framework that fuses YOLOv8 for field localization, EasyOCR for text extraction, and a GPT-4-based conversational assistant within a BPM-guided redesign of public-sector invoice workflows. Evaluated on 900 real, anonymized invoices from Peruvian public projects, the system achieves [email protected] = 91.4% for detection, 93.2% field-level OCR accuracy, 88.5% end-To-end invoice accuracy, and reduces processing time from ~18 to 1.98 min per invoice (-88.9%). In a 14-day simulated deployment, the assistant answered 96% of 200 task-oriented queries, enabling auditable, near-real-Time oversight. Descriptive benchmarking indicates competitive accuracy versus ABBYY FlexiCapture and UiPath Document Understanding while substantially lowering licensing and infrastructure barriers. Our contributions are: (i) a rigorously evaluated, replication-ready pipeline tailored to public-sector constraints; and (ii) evidence that combining computer vision, OCR, and conversational AI within a BPM framework yields measurable gains in efficiency, accuracy, and transparency for financial control.