Cognitive Process Intelligence: Integrating Generative AI and Automated Workflow Discovery

Abstract

Enterprises increasingly rely on complex, cross-functional business processes spanning ERP, CRM, ticketing, and legacy systems, generating massive event logs that conventional process-mining and robotic process automation (RPA) tools struggle to interpret holistically. Rule-based process discovery algorithms such as Heuristic Miner and Inductive Miner produce structurally valid models but cannot explain deviations in natural language, generalize across process variants, or adapt to evolving organizational vocabularies [1], [2]. This paper proposes Cognitive Process Intelligence (CPI), a framework integrating Generative AI with automated workflow discovery to bridge this gap. We introduce ProcessGPT, a 220M-parameter domain-adapted transformer continually pre-trained on enterprise event logs, BPMN repositories, and standard operating procedure (SOP) documentation, and augmented with Retrieval-Augmented Generation (RAG) over an indexed corpus of 1.8 million process artifacts. The system performs four cognitive tasks: event-log-driven process discovery, deviation and anomaly detection, natural-language-to-workflow translation, and automated process documentation generation. Evaluated on the BPI Challenge 2017/2019 logs and a proprietary 42,000-case enterprise dataset, the proposed CPI framework achieves a 0.95 average F1-score in process model discovery, reduces hallucinated process steps from 9.1% to 1.4% through RAG grounding, and operates at 96 MS average latency, substantially outperforming classical miners, sequence-to-sequence deep learning baselines, and general-purpose LLMs. Expert evaluation confirms near-human interpretability of generated workflow narratives, positioning CPI as a practical cognitive layer for enterprise automation.