Automated Power BI report generation using an event-driven agentic workflow
Kangas, Juho (2026)
2026
Tietotekniikan DI-ohjelma - Master's Programme in Information Technology
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
Hyväksymispäivämäärä
2026-06-16
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202606046933
https://urn.fi/URN:NBN:fi:tuni-202606046933
Tiivistelmä
Large language model enabled AI agents plan and act through external tools rather than emitting a single completion. This capability lets a language model orchestrate a multi-step workflow, validate its outputs against external systems, and recover from errors without further human input. Enterprise business intelligence has a workflow that fits this pattern. A small group of developers authors Power BI reports against shared semantic models through repetitive graphical-editor interactions whose cost scales with the number of visuals on a page rather than with the difficulty of the underlying request.
This thesis designs, implements, and evaluates an event-driven cloud-based system that converts a natural-language report request, filed as an Azure DevOps work item, into a deployable Power BI report. Prior automation in this domain targets templating or single-shot generation rather than a tool-grounded agent that compiles, validates, and self-corrects against a stored semantic model. The work in this thesis follows the Design Science Research Methodology. The analysis covers the system architecture, the deterministic compiler that emits the Power BI Report Format, and the report-generation agent that drives the compiler through the Model Context Protocol. Verification combines a 30-run benchmark, an end-to-end walkthrough, and an architectural inspection. Each result is verified against the requirements that motivated the system.
The deployed system reaches structural benchmark success on 29 of the 30 runs. Every run terminates within the budgeted reasoning turns, and the agent recovers from any validation errors observed during the experiment. A short-lived agent grounded by typed Model Context Protocol tools and a deterministic compiler can therefore author deployable Power BI reports from natural-language requests without human intervention. The result supports the use of event-driven agentic workflows for comparable artifact-generation tasks in enterprise business intelligence use cases.
This thesis designs, implements, and evaluates an event-driven cloud-based system that converts a natural-language report request, filed as an Azure DevOps work item, into a deployable Power BI report. Prior automation in this domain targets templating or single-shot generation rather than a tool-grounded agent that compiles, validates, and self-corrects against a stored semantic model. The work in this thesis follows the Design Science Research Methodology. The analysis covers the system architecture, the deterministic compiler that emits the Power BI Report Format, and the report-generation agent that drives the compiler through the Model Context Protocol. Verification combines a 30-run benchmark, an end-to-end walkthrough, and an architectural inspection. Each result is verified against the requirements that motivated the system.
The deployed system reaches structural benchmark success on 29 of the 30 runs. Every run terminates within the budgeted reasoning turns, and the agent recovers from any validation errors observed during the experiment. A short-lived agent grounded by typed Model Context Protocol tools and a deterministic compiler can therefore author deployable Power BI reports from natural-language requests without human intervention. The result supports the use of event-driven agentic workflows for comparable artifact-generation tasks in enterprise business intelligence use cases.