Hello Celonis Community,
I recently completed an independent empirical study exploring a question that I believe could be relevant to the future of Process Intelligence:
Can the principles of process mining be extended beyond conventional enterprise workflows to analyze the execution behavior of autonomous AI agents?
Using Celonis and execution traces from WildClawBench V4, I investigated how agent activity sequences, tool interactions, and explicitly flagged execution errors could be reconstructed and analyzed through a process-mining approach.
Research scope
-
720 AI-agent executions
-
35,771 recorded events
-
12 AI models
-
6 task categories
-
Process reconstruction, model-level comparisons, and exploratory early-warning evaluation
What did the research find?
1. Process observability: AI-agent execution traces could be represented and investigated through a process-mining framework, making execution paths, tool interactions, and flagged errors more interpretable.
2. Behavioral differences: The analysis identified variations in execution behavior across models and task categories.
3. Predictive limitations: A fixed-horizon evaluation considered 452 eligible executions, including only six positive future-error cases. The exploratory heuristic achieved an apparent ROC AUC of 0.467 and did not establish reliable early-warning performance.
An important distinction is that explicitly flagged execution errors are not equivalent to overall benchmark task failures.
Why this matters
The results do not establish that Celonis can reliably predict AI-agent failures. Instead, they suggest a potentially useful research direction: applying Process Intelligence to the observability, analysis, and eventual operational improvement of agentic AI systems.
Open research questions for the community
-
How would you model AI-agent tool calls, retries, and reasoning transitions using Celonis?
-
Could object-centric process mining provide a stronger representation than a case-centric execution model?
-
What additional event features or evaluation strategies would be necessary to investigate reliable early-warning signals?
I would welcome feedback from Celonis practitioners, process-mining researchers, and AI-agent engineers.
Open-access preprint (Zenodo):
https://doi.org/10.5281/zenodo.23410265
This is independent exploratory research and is not an official Celonis study or product-performance validation.
#Celonis #ProcessMining #ProcessIntelligence #AgenticAI #AIObservability
