Can Celonis Predict AI-Agent Failure Before It Happens? An Empirical Process-Mining Study of WildClawBench Execution Traces

Abstract

This independent empirical study examines whether process mining can support early detection of AI-agent execution errors using Celonis and WildClawBench execution traces. The investigation covers 720 executions, 35,771 events, 12 AI models, and six task categories. The study demonstrates process reconstruction, execution-error analysis, cross-model comparisons, and fixed-horizon evaluation of early behavioral indicators. The exploratory predictive assessment did not establish reliable discrimination of subsequently flagged execution errors, but provides a reproducible analytical framework for studying AI-agent observability and identifying future opportunities for process intelligence beyond conventional business workflows.