What GMP Annex 22 Means for Pharma
Artificial Intelligence (AI) is moving from pilot projects to real-world pharmaceutical manufacturing applications—but not without regulatory oversight. With the publication of the draft of Annex 22 to the European Guidelines for Good Manufacturing Practice (GMP), regulators are officially entering the conversation. This new annex, now open for public comment, represents the first structured framework for the application of AI/ML in GMP environments.
Annex 22 is not just regulatory housekeeping—it’s a signal that AI is here to stay. And the message is clear: AI can be used in regulated environments, but it must be transparent, traceable, and well-controlled.
What Annex 22 Covers—and What It Doesn’t
Annex 22 provides additional guidance to Annex 11 for computerized systems, focusing specifically on AI and machine learning (ML) models that are used in critical GMP applications—those impacting patient safety, product quality, or data integrity.
The scope is narrow and deliberate. The annex:
- Applies to:
- AI/ML models with deterministic output (i.e., identical input yields identical output).
- Static models—those that do not continue learning during operation.
- Excludes:
- Generative AI (e.g., ChatGPT, image generators).
- Large Language Models (LLMs).
- Probabilistic or dynamic models that adapt continuously.
- Any model used in non-critical GMP applications.
If you’re exploring AI for tasks like real-time process monitoring, predictive maintenance, or supplier risk classification, the annex is relevant—but only if these systems directly impact critical GMP outcomes.
Themes
- Intended Use and Risk Assessment Are Foundational
AI isn’t plug-and-play. Annex 22 stresses the need for a detailed, documented description of a model’s intended use, input data types, expected output, and risk to the process. Clear documentation ensures traceability and helps QA validate whether AI outputs can be trusted.
- Define subgroups of data (e.g. sites, materials, defect types).
- Identify potential bias or limitations in the data.
- Use subject matter experts to approve intended use before acceptance testing.
2. Test Data Must Be Independent and Representative
A robust AI model is only as good as the data it’s tested on. The annex mandates:
- Independent test datasets not used in training or validation.
- Inclusion of all expected input variations (e.g. across suppliers, shifts, equipment).
- Verified labeling of test data by experts or validated systems.
This avoids “model overfitting” and ensures generalizability, which is crucial in dynamic supply chains with global variability.
3. Explainability and Confidence Are Non-Negotiable
The annex calls for explainability in models used in GMP-critical tasks. QA and supply chain leaders must be able to audit AI decisions and understand when human review is needed. AI outputs should never be blindly accepted in critical decisions. Annex 22 reinforces that responsibility always lies with trained personnel.
- Prominent interpretability tools like SHAP (SHapley Addictive exPlanations) or LIME (Local Interpretable Model-agnostic Explanations) should be used to explain which features contributed to outcomes.
- Confidence scores must be logged, and thresholds set to prevent overconfident predictions.
- Human-in-the-loop (HITL) setups are encouraged. Operator roles must be well-defined and consistently trained.
4. Post-Deployment Monitoring and Change Control
A well-performing model at deployment may deteriorate over time. Supply chains evolve, and so must the oversight of AI tools.
Once in operation, models must be treated like any other GMP-critical system:
- Change control applies to the model, software, and input data sources.
- Performance metrics must be continuously monitored.
- Input data drift must be tracked to ensure ongoing validity.
What This Means for the Industry
Annex 22 is a playbook for adopting AI responsibly. It bridges the gap between innovation and compliance, ensuring that patient safety and product quality remain at the core.
For manufacturers, this is an opportunity to embed AI into GxP processes without ambiguity about what is or isn’t allowed.
✅ Regulators recognize the need for AI-driven innovation, but they demand rigor.
❌ “Black box” models, uncontrolled drift, and unsupported automation will not pass inspection.
Importantly, Annex 22 is currently in draft form and open for stakeholder comments until October 7, 2025. Feedback from industry experts may result in changes to the final version. This is your chance to shape the conversation.
If your organization is considering AI in critical GMP environments, it will soon have a regulatory guidance to do so.
Do you want to learn possible applications of AI in pharma compliance? Listen to the talk of Zuzana Horakova, performed at Sync Space Barcelona 2025 on 17.6.2025
The session covers:
- Real-world use cases of AI in pharma quality and operations
- Predictive compliance tools: where they help, where they can hurt
Responsible AI adoption: why pharma can’t afford to get it wrong