Raising the Bar in GxP Compliance – Part 28: Validating AI Models under Annex 22: Ensuring Trust, Oversight & Performance.
Welcome to the twenty-eighth instalment of Raising the Bar in GxP Compliance, Rephine’s expert-led blog series for QA and regulatory professionals. In this edition, we explore the critical challenge of validating artificial intelligence (AI) in pharmaceutical manufacturing. With the EMA’s draft Annex 22 bringing AI validation into sharp regulatory focus, companies must prove that models are not only reliable and explainable but also governed through robust human oversight. From training data integrity to lifecycle control, discover how Rephine supports organisations in building risk-based AI validation frameworks that safeguard compliance, strengthen quality, and protect patients.
AI is reshaping pharmaceutical manufacturing, but regulators are clear: innovation must not come at the expense of control.
Annex 22 sets out new expectations for how AI models are trained, validated, and governed—demanding transparency, robust oversight, and risk-proportionate lifecycle management.
As AI becomes increasingly integrated into pharmaceutical manufacturing processes, Annex 22 brings into focus how AI models must be validated to ensure they are reliable, explainable, and suitable for their intended use in GxP-regulated environments.
Why AI Validation is Critical in Pharma Manufacturing
Unlike traditional systems, AI models can evolve over time — making their validation more complex and ongoing. Annex 22 defines validation as a risk-based, lifecycle-driven process requiring continuous oversight and control of:
- Model performance and acceptance criteria
- Training and test data quality and representativeness
- Robust human review and intervention procedures
Annex 22: Key Requirements for Validating AI Models
- Intended use must be clearly defined for each AI model.
- Training datasets must be curated, high-quality, and well-documented.
- Test and validation sets should reflect real-life conditions and variation.
- Models must meet pre-established performance thresholds, with justification.
- Version control, change logs, and revalidation must be part of the governance.
- Human-in-the-loop oversight must be implemented where GxP decisions are involved.
Challenges in AI Validation for Pharma and Biotech
- Use of biased or incomplete training data
- Lack of explainability or traceability in AI decision-making
- No formal definition of validation success criteria
- Poor integration of QA and data science teams
- Absence of change control and version tracking in ML environments
How Rephine Helps Companies Achieve Annex 22 AI Compliance
Rephine helps pharma and biotech companies by:
✅ Designing tailored AI validation plans aligned with Annex 22
✅ Assessing data integrity and representativeness of training sets
✅Defining risk-based performance metrics and thresholds
✅ Establishing AI lifecycle management SOPs
✅ Ensuring QA and GxP alignment throughout the AI project
📅 EMA’s consultation period for Annex 22 is open until 7 October 2025. This article is part of Rephine’s educational series on upcoming regulatory changes.
Dr. Eduard Cayón
CSO (Chief Scientific Officer)
About the Author:
Dr. Eduard Cayón is the Chief Scientific Officer (CSO) at Rephine, a global leader in GxP compliance and quality assurance.
We don’t just deliver audits or consultancy services — we partner with clients at every stage of their quality journey, offering end-to-end solutions that empower confidence and compliance.
With over 25 years of experience, Rephine has built an enviable reputation as the gold standard in the industry operating from four primary locations: Stevenage in the UK, Barcelona in Spain, India, and Shanghai in China.
Dr. Cayón, who holds a Ph.D. in Organic Chemistry, is a deeply experienced pharmaceutical industry consultant and auditor.
He is dedicated to supporting pharmaceutical, biotech, and medical device companies in meeting the highest standards of manufacturing and supply chain integrity.