Artificial intelligence is increasingly used to organize formulation variables, predict compatibility and release behavior, prioritize experimental candidates, and connect pharmaceutical data across development stages. Yet much of the current literature remains concentrated on narrowly bounded prediction tasks, whereas drug-delivery development ultimately requires evidence that a formulation can be manufactured reproducibly, generate the intended material behavior, control transport through biological environments, and produce therapeutically relevant exposure. This state-of-the-art review evaluates the changing role of artificial intelligence across that translation arc. The review applies an explicit evidence-selection and critical-appraisal logic to distinguish retrospective benchmark performance, experimentally confirmed capability, mechanistically linked prediction, prospective development utility, and routine pharmaceutical readiness. The resulting synthesis proposes that formulation artificial intelligence should not be treated as a single technological category. Instead, it comprises different evidence classes, including composition screening, compatibility assessment, release modeling, biodistribution prediction, biopharmaceutic and pharmacokinetic integration, manufacturing analytics, and closed-loop experimentation. The central contribution is an evidence-gated interpretation in which predictive models become pharmaceutically consequential only when their outputs remain valid across successive transitions from formulation search to material behavior, release, transport, exposure, and a bounded development decision. Current evidence is strongest for domain-specific prioritization and selected experimentally anchored workflows, whereas generalization across formulations, laboratories, manufacturing scales, biological contexts, and delivery modalities remains limited. The proposed synthesis is conceptual rather than empirically validated and does not constitute a regulatory or clinical framework. Its practical implication is that future pharmaceutical artificial intelligence should be evaluated by the quality of the decisions it supports, the uncertainty it exposes, and the evidence connecting its intermediate predictions to therapeutic exposure.