Expert comment: AI in Pharma – considerations of Annex 22

Artificial Intelligence (AI) is increasingly being adopted in pharmaceutical industry, offering new opportunities to save time and enhance quality assurance. To ensure these technologies are used responsibly, the European Commission has drafted the new Annex 22 to the EU GMP Guidelines – providing a regulatory framework for AI applications used in the manufacture of active pharmaceutical ingredients (APIs) and medicinal products.

Annex 22 at a Glance
Annex 22 applies to static AI models – those that do not adapt or learn after deployment. They provide deterministic outputs (same input always gives the same output) to ensure reproducibility and traceability. This is especially critical for applications with direct impact on patient safety, product quality, and data integrity, such as batch release decisions, deviation assessments, and quality trend analysis. Dynamic AI models, which continuously and automatically learn and adapt performance during use, is not in scope for the draft annex 22, and it is stated that these models should not be used for critical GMP applications. However, it can be used for non-critical GMP application if personnel with appropriate training and qualification should ensure the output from such models are suitable for intended use (human-in-the-loop).

Key Considerations
Annex 22 presents several essential requirements including:

• Validation for Intended Use: AI systems must be validated for their specific application, with documented input/output behaviour and acceptable variability.
• Representative Test Data: Data used to test AI models must accurately reflect the intended use environment, ensuring the model performs reliably under real-world conditions.
• Explainability and Auditability: Outputs must be interpretable and traceable, especially when influencing decisions that affect product quality.
• Data Integrity Controls: Training and test datasets must be version-controlled, independent, and protected against manipulation.

Also worth mentioning is that personnel are described as one of the key principles. Process SMEs are critical in understanding the intended use and associated risks when applying AI models in a GMP environment. It also requires a close cooperation with QA, data scientists and IT.    

Why It Matters
AI can significantly improve efficiency and consistency in GMP processes. However, when used in areas that directly affect medicinal product quality or patient safety, the risks must be carefully managed. The purpose of Annex 22 is ensuring that AI is deployed with robust safeguards, reinforcing GMP principles and ultimately protecting public health.