Annex 22: Artificial Intelligence in GMP – What Pharma Needs to Know

Annex 22: Artificial Intelligence in GMP – What Pharma Needs to Know

Artificial Intelligence (AI) is rapidly transforming pharmaceutical manufacturing, but regulatory compliance remains critical. To guide the safe and effective use of AI, the European Commission has introduced Annex 22: Artificial Intelligence under EU GMP (Good Manufacturing Practice). This annex complements Annex 11 (Computerised Systems) and outlines expectations for the use of machine learning (ML) models in regulated pharma environments.

Scope of Annex 22

Applies to AI/ML models that are trained on data, not hard-coded.

Covers only static, deterministic models (those that do not adapt once deployed).

Excludes dynamic learning systems, probabilistic models, Generative AI, and Large Language Models (LLMs) for critical GMP applications.

Such advanced AI tools may only be used in non-critical GMP processes with strong human-in-the-loop (HITL) oversight.

Core Principles for AI in GMP

Cross-functional collaboration: QA, process SMEs, IT, and data scientists must work together.

Quality risk management: Risk to patient safety, product quality, and data integrity drives all activities.

Strong documentation: Model development, validation, and testing records are mandatory.

Qualified personnel: Staff must be trained to understand AI risks and responsibilities.

Intended Use & Acceptance Criteria

The intended use of an AI model must be clearly defined and documented.

Input data, variations, and possible limitations should be characterized.

AI must perform as well as or better than the process it replaces.

Acceptance metrics may include accuracy, sensitivity, specificity, precision, and F1 score.

Test Data & Validation

Test data must be representative, independent, and verified.

Pre-processing, exclusions, or synthetic data use must be justified.

Independence is key: training, validation, and testing datasets must remain separate.

All testing requires a formal plan, documentation, and deviation handling.

Explainability & Confidence in AI Decisions

AI systems must provide explainable results using techniques like SHAP, LIME, or heat maps.

Confidence scores should be logged; low-confidence predictions flagged as “undecided” instead of forcing unreliable decisions.

Ongoing Operations & Monitoring

AI models must be under change control and configuration control.

Continuous performance monitoring is required to detect data drift or system changes.

Human review remains essential for AI-assisted decision-making in GMP.

Why Annex 22 Matters

Annex 22 marks a regulatory milestone in pharma AI adoption. It emphasizes:

Safety first: patient safety and product quality cannot be compromised.

Transparency: AI decisions must be explainable.

Accountability: human oversight and strong governance remain non-negotiable.

For pharmaceutical companies, this guidance provides a clear compliance roadmap for AI implementation. While Generative AI and LLMs are not permitted in critical processes, their use in supportive, non-critical applications is acknowledged — as long as there is qualified human oversight.

👉 In short, Annex 22 bridges innovation and regulation, ensuring that pharma can leverage AI responsibly, without risking GMP compliance.