Multi-Agent AI for Root-Cause and Bottleneck Analysis in Object-Centric Process Mining: A Comparative Study Against a Single-Agent LLM Assistant

Sammanfattning

Background: Object-centric process mining (OCPM) extends classical process mining to event logs in which every event may relate to several typed business objects, as formalised by the Object-Centric Event Log (OCEL) 2.0 standard. This representation is more realistic but raises the analytical complexity of bottleneck and root-cause analysis for non-expert users. Problem and Research Question: Large language model (LLM) assistants offer a way to lower this barrier, but it is unclear whether a multi-agent design with role separation outperforms a single-agent baseline. The thesis asks how a multi-agent LLM assistant compares to a single-agent baseline on object-centric root-cause and bottleneck analysis. Method: The study follows Design Science Research. Two assistants, a planner-analyst-synthesizer multi-agent design and a single-agent baseline, were built on the same 12-tool OCPM toolkit and evaluated on 12 tasks in three categories (A: closed quantitative, B: open root-cause, C: integrated analysis) over 3 runs each on two OCEL 2.0 logs of contrasting complexity, the Business Process Intelligence (BPI) Challenge 2017 and Order Management, yielding 144 runs (72 per log). Outputs were scored with a fixed rubric; a non-expert second rater rated 18 stratified items per log on four structural dimensions, and the two logs' per-task differences were pooled for a secondary task-level analysis. Results: The multi-agent design produced quality gains on open-ended dimensions, larger on the messy BPI 2017 log (mean shifts +0.42 to +0.58 on the 1-10 rubric, on analytical depth and evidence grounding for the root-cause category and on factual alignment and analytical depth for the integrated category) and smaller on the clean Order Management log. Closed-form category-A correctness was at or near ceiling on both logs, leaving little headroom for either architecture; the clearest traceability gain fell on category A (pooled +0.50, driven by +0.75 on the clean Order Management log), whereas open-ended traceability was not consistently improved. A secondary, small-sample pooled task-level analysis found the open-task analytical-depth and factual-alignment advantages significant (p < 0.01), corroborated by a distribution-free sign test. These gains came at a 4-5x latency and 3-5x tool-call cost, and a graceful-degradation pattern was observed in 29 of 36 BPI 2017 multi-agent runs (81%). Contribution: Architecture choice for OCPM assistants should follow task profile and data complexity rather than default to more agents; the artifact and the two-dataset, 144-run harness are released as a reproducible OCEL 2.0 basis.