EU GMP
EU GMP Annex 22 (Draft) - Artificial Intelligence in the Manufacture of Medicinal Products
Plain-language explanation, inspection context and common violation patterns.
What it says
Draft Annex 22 establishes requirements for the use of artificial intelligence in pharmaceutical manufacturing and quality systems, addressing AI model validation and lifecycle management, human oversight of AI-generated outputs, data governance for AI training and operation, and quality unit review of AI system outputs used in GMP-relevant decisions.
Paraphrased for plain-language clarity. Always verify against the current published regulation text.
What it means in practice
- This draft annex represents the first binding regulatory framework specifically addressing AI use in pharmaceutical manufacturing, following the joint FDA-EMA guiding principles for AI published in January 2026, moving from voluntary principles toward an enforceable EU GMP requirement.
- Human oversight requirements reflect a core theme running through early AI governance frameworks generally: AI-generated outputs used in GMP-relevant decisions, such as specifications, procedures or batch records, cannot bypass quality unit review simply because an AI system rather than a person generated the content.
- This annex is directly relevant to a real, documented enforcement precedent, an April 2026 FDA warning letter citing inappropriate use of AI agents to create drug product specifications, SOPs and master production records without quality unit review, illustrating exactly the failure mode this annex is designed to prevent.
What inspectors look for
- Whether AI systems used in GMP-relevant activities have documented validation appropriate to their intended use and the risk of the decisions they inform.
- Whether quality unit review of AI-generated outputs, specifications, procedures, records, genuinely occurs before those outputs are relied upon, rather than assuming AI-generated content requires less scrutiny than human-generated content.
- Whether data governance for AI training and operation addresses data quality and integrity with the same rigor applied to other GMP data sources.
Most common violation
AI-generated specifications, procedures or records used in GMP-relevant decisions without documented quality unit review, precisely the pattern already cited in the first FDA warning letter to explicitly address inappropriate AI use in pharmaceutical manufacturing.
Want a deeper, interactive breakdown?
GMPify's Regulation Explainer tool gives subscribers an AI-powered analysis of any 21 CFR, EU GMP, USP or ICH reference, with live inspection context.
Try the Regulation Explainer →