Artificial intelligence (AI) in reproductive medicine is shifting from brittle, task-specific models toward versatile foundation models, yet translation into clinical practice remains uneven due to data heterogeneity and domain shift. This review bridges the gap between model metrics and clinical utility by defining an ‘IVF Foundation Model Stack’ and establishing engineering principles for safe deployment. We discuss emerging in vitro fertilization (IVF) AI scenarios from an engineering perspective, proposing a layered architecture spanning data, pretraining, adaptation, deployment and monitoring. Seven core engineering principles are identified, including defining action boundaries prior to prediction, ensuring provenance-first data contracts and implementing privacy-preserving federated learning; these are incorporated into a scenario-specific checklist to guide clinical implementation. IVF is a decision-dense, safety-critical environment, and clinically successful foundation models will be defined not just by scale but also by system design that prioritises provenance, human–AI teaming and continuous lifecycle governance.
Liu et al. (2026) studied this question.