Delegation Is Not Assistance: Agentic AI and Assessment Governance in Higher Education - A Comparative Documentary Analysis of 12 Universities

Abstract

Generative AI (GenAI) assessment policies in higher education have centred on originality, disclosure, and permission to use AI. Agentic AI, which can plan, use tools, access platforms, and act with limited human intervention, challenges this logic because it can take over the whole action chain of an assessed task rather than only generate content. This study examines how assessment governance changes from GenAI to agentic AI. Using exploratory comparative documentary analysis, public central guidance from 12 universities was theoretically sampled: six agentic-aware cases and six GenAI-dominant comparators, with a retrieval cut-off of 3 October 2026. Documents were coded on six dimensions using a directed framework informed by Bacchi’s “What’s the Problem Represented to Be?” approach. All agentic-aware cases distinguished agentic AI from GenAI, and most set explicit delegation and access boundaries, whereas no comparator explicitly addressed agents operating institutional systems on a student’s behalf. The comparators, however, were strong on evidence of learning and task-level transparency, and programme-level assurance was sparse in both groups. The study proposes the Delegation-Evidence-Access (DEA) framework, which moves governance from tool permission towards responsibility allocation, and derives four assessment contexts and a task-, course-, and programme-level implementation model. Implications for Malaysian higher education are discussed. Findings are descriptive, based on selected public documents, and not representative; no inter-rater reliability was assessed.