IKEA: Intelligent Knowledge Extraction with Reasonable and Agentic AI Agents

Administrative document analysis represents a critical organizational challenge: organizations generate vast quantities of documents containing valuable strategic insights, yet traditional analysis approaches remain manual, opaque, and fragmented across systems, leaving organizations vulnerable to missed opportunities and delayed decision making. Conventional document management systems relying on keyword search and isolated machine-learning pipelines lack intelligent reasoning capabilities and autonomous coordination, leaving critical insights buried in unstructured documents. Organizations relying on traditional document processing tools and lacking explainable AI remain blindsided by complex patterns and correlations that could be anticipated and leveraged for strategic advantage. We introduce IKEA, an intelligent document analysis framework that integrates Reasonable AI (explainable reasoning agents) with Agentic AI (autonomous intelligent agents) to provide transparent, autonomous knowledge extraction from administrative documents. Our system employs GPT-3.5-turbo powered reasoning agents that autonomously extract knowledge, generate transparent reasoning chains, and provide conversational AI assistance, all while maintaining end-to-end explainability through step-by-step reasoning decomposition, confidence breakdown analysis, evidence extraction, and assumption identification. Evaluated on a primary corpus of 50 administrative documents (performance reports, incident reports, meeting minutes, policy documents) with 200+ performance metrics and 100+ AI-generated insights, the Reasonable AI agents achieve 87% explanation accuracy aligned with expert assessments, while the Agentic AI agents demonstrate 92% knowledge extraction accuracy with 85% reasoning quality score. Against state-of-the-art baselines under matched conditions—including GraphRAG, DocAgent (text-adapted), dense RAG, and ReAct—IKEA attains the highest entity F1 and a large margin on explanation accuracy and auditability. Cross-dataset tests on Kleister (Charity and NDA), a DocVQA text-only subset, and an external 25-document administrative corpus confirm that these gains transfer beyond the primary corpus, with only a modest drop relative to in-domain performance. On the primary corpus, the integrated framework also delivers a 40% reduction in document analysis time and a 35% improvement in insight discovery rate compared with traditional manual analysis.