Artificial intelligence (AI) is increasingly transforming public administration, yet empirical evidence from developing countries remains limited. This study explores the current use, key challenges, and enabling conditions of AI adoption in Kazakhstan’s public administration system. The study employs an exploratory qualitative design based on semi-structured interviews with 20 experts from government, academia, and related professional domains. The data were analyzed using thematic analysis in ATLAS.ti to identify key themes. The findings show that AI adoption is in a transitional stage, supported by strong government initiatives and shifting from digitalization to its use in decision support and predictive analytics for more proactive public services. While a number of pilot projects and practical applications are already in place, broader adoption remains constrained by interrelated barriers, including data limitations, skills gaps, infrastructural constraints, and regulatory uncertainty. The results also identify a corresponding set of enabling conditions, such as institutional support, human capital development, data governance improvements, and cross-sector collaboration, which can facilitate further progress. By linking systemic barriers with corresponding enabling conditions, the study clarifies how AI adoption unfolds in practice and identifies actionable directions for policy and implementation.

Artificial intelligence is expanding rapidly, intensifying policy concerns that more vibrant AI ecosystems may displace workers and increase unemployment. This study aims to test whether national AI vibrancy is associated with higher unemployment across education groups (advanced, intermediate and basic). Using an unbalanced panel of 34–35 countries from 2017 to 2023, the analysis combines Stanford’s AI Vibrancy Score with World Bank indicators and estimates two-way fixed-and random-effects models, employing Box–Cox/log transformations and dependence-robust inference (including country/time clustering and Driscoll–Kraay standard errors). The results provide little support for the displacement hypothesis. For advanced-education unemployment, AI vibrancy is statistically insignificant in the two-way FE model. It remains insignificant under all robust corrections (ln(AI vibrancy): β=−0.099, country-clustered p=0.494, time-clustered p=0.544, Driscoll–Kraay p=0.468). For basic-education unemployment, AI vibrancy is likewise insignificant in the two-way FE model (p=0.782). It remains insignificant under country clustering (p=0.830), time clustering (p=0.813) and Driscoll–Kraay inference (p=0.819). For intermediate-education unemployment, the AI coefficient remains insignificant under country clustering (p=0.273), time clustering (p=0.310), and Driscoll–Kraay corrections (p=0.226), indicating no robust unemployment-increasing effect across education groups during the observed period.