Undesired process outcomes might manifest in subtle ways, particularly when their underlying causes are not explicitly known or documented. So far, the detection of these causes has been limited, especially if they refer to complex dependencies across multiple data attributes and within trace clusters. Hence, in this paper, we propose the multi-level method PROXEE to detect rules that explain undesired outcomes. The method integrates unstructured textual information such as handbooks and regulatory documents via a retrieval-augmented generation pipeline combined with large language model (LLM) reasoning, as well as structured process data enriched with contextual statistical analysis. We identify contextual trace clusters and automatically generate enhanced features. They are subsequently analyzed in conjunction with the textual knowledge to discover natural language based rules with varying degrees of complexity, ranging from rules found in documents to rules inferred exclusively from data and conceptual understanding. We evaluate PROXEE using two case studies and compare the results with CART decision tree, DeepSeek-R1, OpenAI o1, and OpenAI DataAnalyst. Our findings demonstrate that PROXEE yields concise explanations for undesired outcomes and outperforms existing methods.