Abstract
In this review, we examined over 320 records on AI in higher education, of which 102 met the inclusion criteria, and encountered a persistent asymmetry: roughly nine student-focused articles for every one focused on the educator. The instructor who grades 300 assignments, answers routine emails late at night, and still tries to design meaningful assessment remains largely absent from the AI optimisation literature. We considered a meta-analysis but found it unworkable due to heterogeneity of dependent variables, contexts, and effect metrics. A narrative systematic review was the only approach capable of integrating experimental studies, surveys, conceptual frameworks, and regional data. Sources span January 2019 to April 2025, covering Scopus, Web of Science, and Google Scholar, supplemented by nine Kazakhstani sources. PRISMA 2020 version was used, quality appraisal was performed with MMAT and JBI, and dual-expert coding was applied (kappa = 0.84). Next-generation AI agents produce measurable effects: code assessment time drops by 58%, academic risk prediction reaches 80%-95% accuracy ( Ifenthaler and Yau, 2020 ), and GenAI use correlates with improved teaching practice (beta = 0.264; p < 0.01). In Kazakhstan, 87% of instructors know about AI yet only 31% use it – a gap that is methodological, not technological. Five systemic gaps structure the research agenda: persistent student-centredness, absence of longitudinal data, disconnect between engineering and pedagogy, geographic asymmetry, and the unexplored potential of computer science educators as nodal agents of AI diffusion.