Year: 2026
Type: dissertation
Abstract: Academic workflows in higher-education institutions are rarely as tidy as the systems built to manage them. Exception-prone and deeply contextual, these processes have long resisted the kind of automation that conventional rule-based or low-code tools can offer. This thesis investigates whether agentic AI can do better — and if so, how. To explore this question, a prototype was developed and evaluated that automates a realworld exam-management process, encompassing room allocation, proctor assignment, data consolidation, and document generation, through natural-language instructions. Built within a Design Science Research framework, the system positions a large language model (LLM) as a reasoning core that interprets high-level goals and coordinates a set of deterministic execution modules. This deliberate separation of agentic planning from rule-bound execution forms the architectural backbone of the prototype, complemented by human-in-the-loop checkpoints, audit logs, and traceable decision records designed to keep the system accountable. Evaluation through controlled scenario execution and execution-trace analysis indicates that the prototype can exhibit genuinely agentic behavior while remaining within defined governance boundaries and preserving reproducibility. Beyond the artifact itself, the thesis offers an architectural framework for governed agentic automation and a set of grounded design considerations for institutions considering LLM-based agents in administrative settings. The findings are exploratory and openly so: they are offered not as conclusions, but as a foundation for the research and practice still to come.
Language: English
Subfield: Management Information Systems
Domain: Social Sciences
SDG: Quality education
Open Access status: green
OpenAlex ID: W7171761312