Search

Artificial intelligence is rapidly entering pharmaceutical manufacturing through predictive analytics, process analytical technology, continuous manufacturing platforms, and data-driven quality control. These developments promise earlier detection of process deviation, adaptive optimization of critical process parameters, and more responsive assurance of critical quality attributes. Yet the dominant regulatory and quality vocabulary remains anchored in Quality-by-Design, a paradigm built around pre-defined knowledge, structured risk assessment, and validated control strategies. The central problem is that AI-assisted manufacturing systems do not behave like conventional pharmaceutical processes governed solely through fixed design spaces. Machine learning models may change performance as data distributions shift, as sensors age, as materials vary, or as feedback loops alter the operating environment. This creates a governance gap between the static logic of pre-validation and the dynamic reality of algorithmic decision-making. This critical conceptual review examines whether Quality-by-Design remains sufficient for AI-assisted pharmaceutical manufacturing. It argues that QbD is still necessary but no longer sufficient because it was designed for processes whose boundaries, models, and control logic can be defined before routine operation. AI-assisted systems require an additional governance logic capable of supervising data, models, explanations, adaptation, and accountability throughout the product lifecycle. The review develops the concept of Quality-by-Intelligence as a forward-looking governance framework. QbI does not discard QbD; rather, it extends it by embedding data stewardship, model lifecycle management, explainable decision-making, adaptive risk control, and regulatory translation into the pharmaceutical quality system. Its purpose is to govern not only the manufacturing process, but also the intelligence layer that increasingly mediates quality decisions. Quality-by-Intelligence reframes pharmaceutical quality as an adaptive, evidence-generating, and accountable system rather than a one-time design achievement. The transition will require new regulatory expectations, new validation evidence, new operator competencies, and stronger collaboration between industry, regulators, and academic experts. Without this shift, AI may be deployed into manufacturing faster than the quality systems needed to govern it.

Generative artificial intelligence is becoming increasingly relevant to pharmaceutical formulation because it can propose compositions, excipient combinations, processing conditions, and optimisation trajectories that may not be obvious through conventional experimental design. These capabilities create the possibility of faster development, broader exploration of formulation space, and more systematic use of prior knowledge. Yet the same models that expand formulation creativity often operate through complex latent representations that are difficult to interpret. This creates a trust problem for both scientific and regulatory decision-making. The central problem is that an AI-generated formulation is not only a predicted technical solution but also a claim about product performance, manufacturability, and quality. If the rationale behind that claim cannot be explained, formulation scientists may struggle to convert model outputs into mechanistic understanding. Regulators may likewise find it difficult to assess whether the proposed formulation is supported by transparent evidence. Opaque formulation design therefore risks becoming a translational bottleneck rather than an innovation accelerator. This perspective develops a conceptual framework for dual-purpose explainability in AI-generated pharmaceutical formulation design. The framework is designed to serve two decision contexts simultaneously. Scientific decision-makers require explanations that clarify formulation logic, reveal influential variables, and support hypothesis generation. Regulatory decision-makers require explanations that are auditable, reproducible, uncertainty-aware, and connected to product quality and safety evidence. The article first defines the conceptual gap between existing AI formulation capabilities and explainability expectations. It then describes the logic of AI-generated formulation, identifies distinct scientific and regulatory explanation requirements, and analyses transparency barriers. The proposed framework integrates global model explanations, local formulation-specific explanations, mechanistic interpretation, uncertainty communication, and regulatory evidence packaging. Three tables summarise the gap analysis, explainability requirements, and framework architecture. The article concludes that explainability must be treated as a design requirement rather than a post hoc add-on to pharmaceutical AI. AI-generated formulation designs will become useful only when their rationale can be interrogated, documented, challenged, and connected to established principles of product and process understanding. A dual-purpose explainability framework can help move the field from black-box prediction toward transparent, accountable, and scientifically meaningful formulation intelligence.