Artificial intelligence has moved from a speculative promise to a working method across pharmaceutical formulation, yet its presence in routine industrial development remains thin relative to the volume of published proof of concept. This review reads that gap directly. It first maps where machine learning now delivers genuine value in formulation science: predicting solubility, dissolution, disintegration and physical stability from composition; navigating the excipient design space faster than sequential experimentation allows; generating candidate formulations by inverse design; and closing the loop between prediction and experiment through autonomous, self-driving laboratories. Representative platforms already predict cyclodextrin complexation and solid dispersion stability with test-set accuracies that would have seemed implausible a decade ago. The review then argues that the constraint on adoption is no longer principally algorithmic. Three coupled barriers dominate. The data problem is structural, because formulation datasets are small, proprietary, fragmented and inconsistently reported, which limits model generalisation more than any shortfall in model architecture. The regulatory position is genuinely evolving rather than settled, and the frameworks now emerging from the United States Food and Drug Administration and the European Medicines Agency ask questions about model lifecycle, credibility and human oversight that most published models were never built to answer. The organisational barrier is the least discussed and possibly the most binding. It reaches across validation under existing quality systems, the scarcity of staff fluent in both formulation and data science, and the difficulty of scaling a pilot into a validated production workflow. The paper concludes that the decisive work now sits downstream of the model, in data infrastructure, benchmarking, regulatory engagement and workforce capability.