Answers are grounded in this paper and cite their sources.
Information classification is critical for organizational security and regulatory compliance; however, current practices remain constrained by limited automation and subjective judgment. To address these limitations, we leverage large language models (LLMs) and propose an agentic retrieval-augmented generation (RAG) framework for classifying organizational information assets. Unlike prior approaches that treat information classification merely as a static labeling task, we reconceptualize it as a compliance-driven decision process, naturally formulated as a Markov Decision Process (MDP). Our architecture comprises three key components: (i) a hybrid retrieval module combining vector search with knowledge graph traversal grounded in the ISO 27000 standards and EU regulations, forming the system’s reasoning state; (ii) an agentic orchestration, where a hierarchical policy coordinates the agent actions, including query rewriting, context optimization, classification, reward, and reflection, along the decision trajectory; and (iii) regulatory adjudication that reviews the proposed classification against ISO controls, threats, and impacts before the final decision. To the best of our knowledge, this is the first work to model information classification as an MDP-based agentic workflow explicitly grounded in regulatory knowledge. We evaluated the proposed system on an automotive enterprise dataset (n=200) and a generic public dataset (n=400) over five seeds with 95% confidence intervals, using classification metrics and RAGAS for retrieval quality and reasoning fidelity. On the enterprise corpus, where sensitivity is largely implicit, agentic RAG achieves 0.86 accuracy and improves recall for Confidential and Restricted information to 0.91 and 0.92, respectively. Faithfulness increases by 59% from the retrieval baselines to agentic RAG. On the public corpus, where sensitivity cues are more explicit, MDP-based agentic RAG provides greater value for evidence grounding than for classification accuracy. An ablation study further reveals that MDP actions wire together as one decision trajectory but contribute to different aspects of the outcome: some influence the predicted label ˆ 𝑦 , whereas others primarily strengthen the quality and evidence grounding of the reasoning 𝑟 . More notably, the classifier’s self-reported confidence is an unreliable indicator of correctness (AUC 0.44/0.38), whereas deterministic adjudication based on ISO controls, threats, and impacts provides a stronger signal for decision correction (AUC 0.72/0.75). This suggests a useful balance for compliance-driven AI: LLMs provide flexible reasoning, while deterministic rules offer a more reliable basis for decision review. The MDP also records decision trajectories and rewards, providing a basis for future policy learning.
No takes yet. Share an insight, caveat, or question.
Zhang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: