Draft Publication of EU and PIC/S GMP Annex 22 Artificial Intelligence
Pharmaceutical Manufacture in the New Era of AI
Do you remember where you were when… the Berlin Wall fell? … Southeast Asia was struck by the tsunami? …a pandemic was declared for COVID-19? And what about when ChatGPT was released to the public? Whilst not as dramatic as the above events, in the short time that artificial intelligence has been available to us all, the societal impact has been vast. It is estimated 66% of people now regularly use AI and it even reaches into highly regulated sectors necessitating the introduction of new regulatory guidance with the draft of GMP Annex 22 Artificial Intelligence.
Annex 22 (AI) Training is now available (Click here).
And if you think there’s a job that AI can’t do, you’re probably wrong.
AI in the Therapeutic Goods Industry
For the science and healthcare industries, the paradigm shift with the introduction of AI has been immeasurable.
- Drug discovery: AI models are designing novel drug candidates in months.
- Medical diagnostics: AI is outperforming human experts in disease detection.
- Climate modelling: Advanced AI systems are delivering highly accurate environmental predictions.
- Autonomous labs: Robotic labs powered by AI are accelerating material and chemical discoveries.
Manufacturing & Quality Control
AI is playing a pivotal role in modernising pharmaceutical manufacturing. It enables predictive maintenance by analysing equipment performance data to anticipate failures before they occur, thereby reducing downtime and improving operational efficiency.
In quality control, AI systems are used to detect anomalies in real time, ensuring that deviations from standard operating parameters are identified and addressed promptly. This contributes to higher batch consistency, reduced waste, and improved product quality. Additionally, AI supports process optimisation by continuously analysing production data to recommend adjustments that enhance yield and reduce variability.
Data Integration & Knowledge Management
Pharmaceutical operations generate vast amounts of data across disparate systems – such as Manufacturing Execution Systems (MES), Laboratory Information Management Systems (LIMS), and historian databases. AI facilitates the integration and harmonisation of these data sources, enabling a unified view of operations.
Natural Language Processing (NLP) tools are particularly valuable in extracting insights from unstructured data, such as shift logs, lab notes, and compliance reports. This capability enhances knowledge management, supports faster decision-making, and enables more effective root cause analysis during investigations or audits.
Regulatory & Compliance
AI is increasingly being leveraged to streamline regulatory processes and ensure compliance with Good Manufacturing Practice (GMP) and other standards. It automates the generation and validation of documentation required for regulatory submissions, reducing the risk of human error and accelerating approval timelines.
Moreover, AI tools assist in continuous monitoring of compliance-related activities, identifying potential risks and deviations in real time. This proactive approach enhances audit readiness and supports a culture of quality and accountability across organisations.
Click here for Annex 22 (AI) Training (online).
Navigating the Future: Annex 22 and the Role of AI in GMP Compliance
A number of draft additions and updates were recently published for EudraLex Volume 4 – Good Manufacturing Practice Guidelines for stakeholder consultation, including:
- Revision of Chapter 4 Documentation
- Revision of Annex 11 Computerised Systems
- New Annex 22 Artificial Intelligence
As AI technologies become increasingly embedded in manufacturing and quality assurance processes, regulatory frameworks must evolve to ensure patient safety, product quality, and data integrity. The recent release of EU GMP Annex 22 Artificial Intelligence reflects recognition of this need by global GMP communities.
This entirely new annex marks a significant milestone, offering structured guidance for the use of AI in GMP-regulated environments. It complements Annex 11: Computerised Systems, focusing specifically on AI models used in critical applications. For pharmaceutical professionals, understanding the scope, principles, and operational requirements of Annex 22 is essential for compliant and responsible AI adoption.
Why Annex 22 Matters
Annex 22 is not just a technical document—it’s a strategic response to the growing use of AI in pharmaceutical operations. From predictive maintenance to automated quality control, AI models are increasingly relied upon to make decisions that can directly affect product outcomes and patient safety.
How Will AI Be Used in GMP
The annex applies to static, deterministic AI models—those trained on data, rather than being explicitly programmed, but not continuously learning or adapting during use. These models will provide the same output when given the same inputs.
It explicitly excludes dynamic models, probabilistic outputs, and generative AI, including Large Language Models (LLMs), from critical GMP applications. Exclusion of dynamic modelling in GMP is reflective of the current technological limitations – in explainability, reproducibility, and risk control – such that identical outputs cannot be guaranteed. For this reason, dynamic learning models should not be used in critical GMP applications.
In scope | Out of scope |
|---|---|
All computerised system used in manufacture | Dynamic models |
Machine learning | Generative AI |
Static models | Large language models (LLM) |
Deterministic output |
DEFINITION: Human in the Loop
“Where a model is used to give an input to a decision made by a human operator (human-in-the-loop), and where the effort to test such model has been diminished, the description of the intended use should include the responsibility of the operator. In this case, the training and consistent performance of the operator should be monitored like any other manual process.”
Note: For use of generative AI/LLM where there is no direct impact on patient safety, product quality or data integrity, personnel with adequate qualification and training should always be responsible for ensuring that the outputs from such models are suitable for the intended use, i.e. a human-in-the-loop (HITL)
Key Changes and Implications for Practice
Key principles of GMP, as always, apply to the management of AI in GMP.
- Personnel of appropriate knowledge and training must be involved
- SOPs on the processes must be available
- Activities should be implemented based on risk to safety, quality and data integrity
Intended Use and Risk Management
A cornerstone of Annex 22 is the requirement for a clear and detailed definition of intended use. This includes:
- The specific task the model automates or assists
- Comprehensive characterisation of input data, including common and rare variations
- Identification of limitations and potential biases
- Confirmation of such definition to be approved by a Subject Matter Expert (SME).
This approach aligns with Quality Risk Management (QRM) principles, ensuring that AI deployment is based on a thorough understanding of process risks and data characteristics.
Acceptance Criteria and Performance Metrics
Before deployment, AI models must undergo rigorous testing against predefined acceptance criteria. These criteria should be:
- Based on relevant test metrics (e.g., accuracy, sensitivity, specificity)
- Defined by subject matter experts (SMEs)
- At least as stringent as the performance of the process being replaced
This requirement ensures that AI does not compromise existing standards and that its benefits are demonstrable and measurable.
Test Data Integrity and Independence
Annex 22 introduces robust controls around test data management, including:
- Documented selection of test data must ensure it is representative, stratified, reflects limitations, complexity and variability
- Ensuring independence of test data e.g. from training and validation datasets
- Implementing access controls and audit trails
- Verifying data-labelling accuracy through expert review or validated methods
These measures are critical to prevent data leakage and biased performance assessments, which could undermine the reliability of AI models in GMP contexts.
Test Execution
- Fit for Intended Use: The model must perform reliably with new, relevant data and avoid overfitting or underfitting.
- Test Plan: A documented and approved plan should outline the intended use, metrics, acceptance criteria, test data, and procedures. An SME should be involved.
- Deviation: Any deviations or failures must be recorded, investigated, and justified.
- Documentation: All test records, data descriptions, and access controls must be retained in line with GMP standards.
Explainability and Confidence
One of the most progressive aspects of Annex 22 is its emphasis on explainability. AI models used in GMP must:
- Record feature attribution
- Justify decisions based on relevant features
- Log confidence scores for predictions or classifications
This ensures that decisions made by AI are transparent and interpretable, allowing for human review and accountability—especially important in regulated environments.
Operational Controls and Monitoring
Once deployed, AI models must be placed under:
- Change control: Any modification to the model, system, or process must be evaluated for re-testing
- Configuration control: Unauthorised changes must be detectable
- Performance monitoring: Regular checks to detect drift or degradation
- Input space monitoring: Ensuring new data remains within the model’s validated scope
These controls mirror traditional GMP expectations but are tailored to the unique challenges of AI systems.
Embracing the Future
The introduction of Annex 22 signals a shift in regulatory thinking—one that embraces innovation while reinforcing the foundational principles of GMP. For professionals across manufacturing, quality assurance, IT, and data science, this means:
- Cross-functional collaboration is essential. AI implementation requires input from SMEs, QA, IT, and data scientists.
- Documentation and traceability must be prioritised. Every step from model training to deployment must be recorded and justified.
- Training and qualification are critical. Personnel must understand AI principles, risks, and operational controls.
- AI governance frameworks should be established. These include policies for model lifecycle management, data integrity, and compliance monitoring.
Looking Ahead: The Role of Generative AI and LLMs
While Annex 22 excludes generative AI and LLMs from critical GMP applications, it leaves the door open for their use in non-critical tasks, provided there is human oversight and suitability checks. Potential uses include:
- Drafting SOPs or documentation
- Supporting training and education
- Assisting in non-critical GMP data analysis
- Project management
Pharmaceutical professionals must remain vigilant. The lack of deterministic outputs, potential for hallucination, and limited explainability of these models pose significant risks if not properly managed.
Conclusion: A Framework for Responsible Innovation
This draft consultation of Annex 22 is a first foray into responsible, compliant and controlled integration of AI into the pharmaceutical industry.
It recognises the promise (and inevitability) of AI with the rigor of GMP, aiming to ensure that innovation does not come at the expense of safety or quality.
For pharmaceutical professionals, the annex provides clarity, structure, and a pathway to harness AI effectively. As technologies evolve, so too will the regulatory landscape. Staying informed and engaged will be key to ensuring compliance within this exciting technological development.
Early adoption and strategic planning will be essential for successful implementation by the 2026 effective date.
PharmOut Services
At PharmOut, we specialise in delivering comprehensive consulting services tailored to the pharmaceutical industry.
If you need assistance determining the impact of regulatory changes on your organisation, there are a number of ways we can help:
- Conduct gap assessments, risk analysis and propose process and documentation updates. Contact us via the website or via email for assistance.
- Provide consultation on the steps that you need to take. To book a one-on-one chat with us, please go to our website: Consultancy Time
- Train you so you have all the knowledge you need, via online Annex 22 AI eLearning or face to face. Visit our website to read about training options.
- Get PIC/S & EU Annex 22 Training (AI & ML in GMP Environments)
Frequently Asked Questions (FAQ)
Annex 22 provides structured guidance for the responsible use of artificial intelligence in GMP-regulated environments, focusing on patient safety, product quality, and data integrity.
Annex 22 applies to static, deterministic AI models that produce consistent outputs. It excludes dynamic models, generative AI, and large language models (LLMs) from critical GMP applications.
Models must be tested for fitness for intended use with a documented test plan, predefined acceptance criteria, and independent test data. Deviations must be investigated and all documentation retained per GMP standards.
AI models must log feature attribution, justify decisions based on relevant features, and record confidence scores to ensure transparency and support human oversight.
Generative AI and LLMs are excluded from critical GMP tasks but may be used in non-critical applications such as documentation or training, provided there is human oversight and suitability checks.
