The integration of artificial intelligence (AI) into Agile IT project management has emerged as a transformative strategy for enhancing risk assessment and decision-making processes. Despite the increasing interest, a fragmented understanding persists regarding the application, effectiveness, and limitations of AI-driven tools in Agile environments. This thematic review systematically analyses 20 peer-reviewed studies published between 2010 and 2025 to synthesise current knowledge and identify research gaps. The literature reveals that machine learning, predictive analytics, natural language processing, and reinforcement learning offer substantial benefits, including improved sprint planning, early risk detection, optimised resource allocation, and greater responsiveness to dynamic project demands. The study categorises findings into four core themes: AI integration in Agile workflows, AI-driven risk assessment methodologies, AI-enhanced decision support systems, and challenges to adoption. Barriers such as data quality issues, model interpretability, integration complexities, organisational resistance, and high implementation costs are critically examined. A conceptual framework, the Dynamic structuration framework (DSF), which leverages the adaptive structuration theory and Dynamic capability theory, is proposed to guide the structured integration of AI within Agile project lifecycles, emphasising the need for recursive co-evolution, iterative feedback loops, augmented capabilities as well as a governable human–AI collaboration. The review highlights the superiority of AI-driven approaches over traditional methods in terms of predictive accuracy and operational agility, while also calling for further empirical validation and development of explainable AI models. The findings offer valuable insights for researchers and practitioners seeking to advance AI-enabled Agile project management.

Language model development prioritizes scaling to ever-larger sizes despite massive computational and environmental costs, yet whether size alone determines reliability remains unclear. Through systematic evaluation of seven models from one to seven billion parameters, this study reveals that non-scale factors—encompassing architectural design, training data quality, and alignment methodology—account for nearly one-quarter of performance variation, sufficient for modern three-billion-parameter models to match seven-billion-parameter predecessors. This advantage at moderate scale equals gains from doubling model size while requiring only half the parameters and associated computational resources. However, fundamental limits persist: one-quarter of carefully designed prompts defeat all tested models regardless of scale or architecture, with questions requiring multi-step reasoning and temporal tracking proving universally problematic. Efficiency analysis shows returns diminish severely beyond three billion parameters, with marginal gains collapsing up to sixteenfold across this heterogeneous cross-family model set. These findings challenge the bigger-is-better paradigm, showing that architectural optimization at moderate scales offers superior cost-effectiveness, yet fundamental training limitations remain resistant to both approaches.