This paper describes a method for automatically detecting key performance indicator (KPI) thresholds by dividing and aggregating process instances on the basis of differences in process models. The thresholds can be used as an analysis axis of data exploration to investigate process models that are discovered from huge logs. The proposed method enables users to minimize the time needed to detect KPI thresholds through trial and error. We applied the method to real-life logs and experiment results showed that thresholds were detected for two types of KPIs. Although one type did not correlate with process patterns, the other highly correlated with them. Such findings are usually obtained from the domain knowledge of business users and analysis results acquired by data analysts with technical expertise. However, with our approach the thresholds can be detected automatically and this helps to expand process analysis for end users.