Despite the exponential growth of artificial intelligence (AI) solutions designed to assist radiologists in clinical practice, their actual impact on radiology departments remains below initial expectations. Daily workflows have not been profoundly transformed. The actual clinical benefit of these tools is often limited to single, modality-specific "narrow AI" tasks, and their return on investment is unclear. The purpose of this article was to analyze: ( i ), the human and perceptual challenges that shape attitudes toward AI among radiologists, referring clinicians, and patients; ( ii ), the technical and clinical limitations of current AI models, including the mismatches between target tasks and real-world needs, and between published versus real-life performances; ( iii ), the lack of objective return on investment quantification and the paucity of medicoeconomic studies in a context of constrained hospital budgets; ( iv ), the limitations of the current " assistive models " of human-AI interaction in radiology; ( v ), the technical and organizational difficulties that information and technology departments face in integrating, maintaining, and securing a growing number of AI applications across specialties within complex hospital information systems; ( vi ), the ethical and patient safety concerns related to bias, transparency, data protection, and regulatory compliance with respect to data protection officers and the European General Data Protection Regulation; and ( vii ), the underexplored environmental and energy implications of large-scale AI deployment. Finally, potential solutions relating to AI governance, national data infrastructures, user education, and the design of randomized clinical trials and cost-effectiveness studies, are discussed to promote the responsible, evidence-based integration of AI into radiology practice.