%Aigaion2 BibTeX export from HES SO Valais Publications
%Sunday 20 September 2026 10:07:30 AM

@INPROCEEDINGS{,
     author = {Brandas, Claudiu and Plesas, Mihai and Gard, S{\'{e}}bastien and Reychav, Iris and McHaney, Roger and Sharma, Madhav},
   keywords = {Agentic AI, Auditability and provenance, Data pipeline transparency, ETL, Explainable AI, Low-code workflow orchestration},
      title = {Transparency and Auditability in AI-Driven Data Pipelines: Toward Explainable Agentic ETL Systems},
  booktitle = {The 13th International Conference on Information Technology and Quantitative Management},
     series = {Procedia Computer Science},
     volume = {287},
       year = {2026},
      pages = {1093-1103},
  publisher = {Procedia Computer Science, Elsevier B.V},
        url = {https://www.sciencedirect.com/science/article/pii/S1877050926027882},
        doi = {https://doi.org/10.1016/j.procs.2026.08.152},
   abstract = {The growing adoption of large language model (LLM)-driven agentic systems is reshaping how organizations automate data workflows in management and analytics contexts. However, their increasing complexity introduces critical challenges in transparency, explainability, and operational oversight, particularly within extract-transform-load (ETL) pipelines where routing decisions, data interpretation, and cleaning actions may be difficult to trace or audit. This study proposes and evaluates an auditable multi-agent ETL architecture implemented on a low-code automation platform. The system employs a central orchestrator coordinating four specialized agents—Extraction, Validation and Quality Assurance, Cleaning and Transformation, and Analytics—supported by OpenAI GPT-5-mini. The pipeline integrates data from Google Sheets, generates visualizations via QuickChart, and embeds auditability mechanisms including transformation logs, QA reports, and non-destructive corrections. Using a design science methodology, the system was iteratively developed and empirically evaluated for transparency, performance, and failure modes. Experimental runs completed in 5–7 minutes, demonstrating effective anomaly detection across formatting, completeness, and domain rules, while applying conservative fixes and escalating uncertain cases to human oversight.

The findings highlight the importance of observability, verification, and tamper-evident audit structures in agentic ETL systems, and underscore the need for scalable, transparent AI-driven data pipelines.}
}