Generative AI systems are governed by a set of structural regularities, what we term the Physics of AI, that define how they behave in practice. These regularities are not optional design choices but unavoidable constraints. Through three case studies, a retrieval-augmented manual assistant, a semantic document-matching service, and a test-data generation tool, we examine how these physical constraints manifest in real projects and identify ten recurring challenges. By interpreting these findings through the lens of the Physics of AI, we propose five actionable practices that make AI engineering more reproducible, interpretable, and accountable. While unavoidable, we believe that allowing for the Physics of AI changes engineering methodology and practices, not necessarily the outcome.