What Requirements does the new Annex 22 place regarding Personnel?

On 7 July 2025, the European Commission published the draft Annex 22 'Artificial Intelligence' together with the revised version of Chapter 4 'Documentation' of the EU GMP guidelines.

Annex 22 sets out requirements for the use of AI and machine learning in the manufacture of active substances and medicinal products and contains provisions on the selection, training and validation of AI models. It provides for continuous monitoring of AI systems, including change control, monitoring of model performance and procedures for human review where necessary.

What specific tasks and responsibilities do employees have when dealing with artificial intelligence?

The following requirements are formulated in the 10 chapters.

Chapter 2, 'Principles', Section 2.1 'Personnel'

1.    Employees must understand the intended use and associated risks of the AI model in the GMP environment. This requires close cooperation with other departments in the selection of algorithms, training, validation, testing and commissioning of the model, e.g. with QA and IT staff, consultants and data specialists.
2.    Employees must have appropriate qualifications (training) and clearly defined responsibilities.

Chapter 3, 'Intended Use', Section 3.3. 'Human-in-the-Loop'

• The training and consistent performance of employees who make decisions with AI support must be continuously recorded and documented.

Chapter 6, 'Test Data Independency', Section 6.5. 'Staff independency'

• Staff who had access to test data must not be involved in the training and validation of the same model. If it is not possible to maintain this independence, an employee who had access to test data for a model should only be given access to the training and validation of that model if he works together with a colleague who did not have this access (dual control principle).

Chapter 10, 'Operation', 10.5. 'Human review'

• Records or logs must be created if the output of a model is used as a decision-making aid for an employee and the testing effort for this model has been reduced. Depending on the criticality of the process and the degree of testing of this model, constant human review of this output must be carried out in accordance with a defined procedure.