The European Medicines Agency (EMA) will hold a multistakeholder workshop on 30 June focused on the incorporation of artificial intelligence into GMP environments, with the objective of collecting technical and regulatory input for the development of the future EU GMP Annex 22, currently under review.
Although the event has not yet taken place, its announcement already represents an important signal: the EMA recognizes the need to review and further develop how the use of artificial intelligence should be regulated within GxP processes and systems.
A draft that has generated debate
The initial draft of Annex 22 introduced a particularly conservative position regarding the use of artificial intelligence in critical GMP processes, significantly limiting the use of dynamic, adaptive, and probabilistic models.
This approach has generated debate within the sector, and the organization of this workshop suggests that the EMA is open to reconsidering part of this approach based on technical evidence and practical experience provided by experts from industry, academia, and other stakeholders.
According to the agenda published by the EMA, the workshop will address several aspects that will likely be decisive for the final version of Annex 22.
Principles for acceptable AI in GMP environments
One of the key points will be the definition of the requirements that an AI must meet to be considered suitable in a GMP environment, including aspects such as robustness, traceability, explainability, and reproducibility.
This approach aims to transfer classical validation and quality assurance principles to systems based on algorithmic models.
Risk, technical reliability, and compliance boundaries
The workshop will analyze aspects related to risk management, compliance boundaries, and the technical reliability of AI systems, as well as the response to incidents.
In this context, typical risks associated with AI systems (for example, non-deterministic behavior or model degradation) may be considered within the framework of system reliability assessment and control.
Likewise, the EMA has indicated interest in exploring control and mitigation measures, introducing the concept of guardrails, understood as mechanisms intended to limit, monitor, and contain model behavior within a risk-based approach.
Data governance
Data management will be another pillar of the discussion, with particular focus on: dataset quality, traceability, versioning, segregation of GMP and non-GMP data, and data integrity under ALCOA+ principles.
New validation models
The EMA will also open the discussion on how to validate AI beyond the classical IQ/OQ/PQ models, which could imply the incorporation of concepts such as challenge testing, boundary testing, continuous monitoring, etc.
Human oversight
The principle of Human-in-the-loop will foreseeably be consolidated as an essential requirement for any AI application in critical GMP processes.
Cybersecurity and outsourced activities
The workshop will address cybersecurity and outsourced activities in AI systems applied to GMP as a critical risk area and will evaluate the risks when the infrastructure is outside the direct control of the manufacturer and how to preserve control.
What it could mean for Annex 22
Although it is still too early to anticipate conclusions, this workshop could have a significant influence on the final drafting of Annex 22.
Among the possible changes that could arise after the discussion process, the following may be highlighted:
• An evolution from prohibition towards a risk-based approach;
• The formal incorporation of guardrails as a regulatory requirement;
• The adoption of a specific lifecycle for AI models;
• And a reinterpretation of the concept of validated state in non-deterministic systems.
An event that may shape the regulatory direction
Beyond its technical content, this workshop may become a turning point for the pharmaceutical sector and for all regulated GxP environments.
The way in which the EMA decides to integrate artificial intelligence into the GMP framework will have a direct impact on the implementation of AI in regulated environments, as well as on validation strategies, supplier qualification, and operational oversight in the coming years.
Therefore, this event will not only be relevant for what is discussed, but also for what it may anticipate about the future European regulatory framework for artificial intelligence in regulated environments; and the event will be available to follow live…
