Beyond Logs: AI’s Internal Representations as the New Process Evidence

Abstract

Traditional process mining relies on symbolic event logs that represent activities as discrete labels, often overlooking the rich contextual and semantic nuances found in real-world data such as textual reports, visual records, or sensor outputs. In this paper, we propose a paradigm shift: using the internal representations of AI models—embedding spaces learned from data—as the foundation for process mining. Our framework performs both process discovery and conformance checking directly in these continuous vector spaces, enabling the detection of semantically similar yet lexically divergent events. We evaluate our approach along three dimensions: (i) whether embedding-based discovery maintains or improves accuracy over symbolic baselines, (ii) whether multimodal sources such as video and audio can be processed as unified embeddings for mining purposes, and (iii) whether conformance checking in embedding space enables alignment across noisy or semantically perturbed traces. By treating AI’s internal representations as a novel form of process evidence, we show how process mining can move beyond traditional logs and unlock deeper, semantically enriched interpretations of real-world workflows.