Dynamic AI-Enabled Computerized Systems for GMP Usage

This approach may provide benefits in some environments, but it will also introduce risks as models evolve and system behaviors change. The article examines dynamic systems and highlights their potential through a GMP-focused case study.

Background

As increasing numbers of viable AI and ML use cases are emerging in the industry. Regulatory authorities and industry groups have published guidance and even draft regulations on this topic. These authorities aim to establish processes and structures for compliantly applying AI models.

The U.S. Food and Drug Administration (FDA) discussion papers “AI in Drug Manufacturing”1 and “AI in Drug Development” 2 provide use case examples from throughout the pharmaceutical value chain. Also, first-draft guidance is becoming available as per the article “Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products.”3

Similarly, a European Medicines Agency (EMA) reflection paper published in 20244 acknowledges the potential use of AI throughout medicinal products’ life cycles, and the European Union GMP “Annex 22: Artificial Intelligence”5 (in draft at the time of writing) provides further regulatory considerations for manufacturing.

From an industry perspective, ISPE has provided an overview of dedicated use cases, including their respective business potential in the ISPE Baseline® Guide: Pharma 4.0™,6 and has published the ISPE GAMP® Guide: Artificial Intelligence, providing a holistic framework for efficient, effective, and compliant use of AI.7

Technology advancements have enabled AI to learn patterns from historical data and to adapt in real time, improving models automatically. Supporting these advancements, ISPE GAMP^® 5 Guide (Second Edition)8 introduced the term “dynamic systems” in 2022, whereas the AI maturity model9 already anticipated their potential in 2021, with further extensions provided in the ML Risk and Control Framework in 2024.10 Most recently, the ISPE GAMP^® Guide: Artificial Intelligence7 expanded supporting concepts also around dynamic systems.

Despite the available compliance guidance, real-world dynamic systems adoption remains limited. This is mainly because of concerns about automatic model updates without human verification, which could affect product quality or patient safety. This was pronounced in the EMA draft for EU GMP “Annex 22,”5 which states that “the use of dynamic models which continuously and automatically learn and adapt performance during use, [...] should not be used in critical GMP applications.” The ISPE GAMP^® Guide: Artificial Intelligence7 addresses challenges associated with dynamic systems. It suggests applying a control strategy for both the model and its adaptive learning path through the system life cycle, including:

• Robust data management
• Testing activities specifically covering the adaptive learning behavior
• Well-designed human–AI interaction and human oversight
• Clear abort triggers for adaptive learning

Although dynamic AI models introduce complexity, they outperform static models in several key ways. This article illustrates these benefits using a real-world GMP use case and addresses key challenges for adopting these systems. It describes dynamic system designs in more detail, discussing their potential through real and hypothetical use cases before diving into the case study. Challenges to achieving GMP compliance are analyzed based on risks to patient safety, product quality, and data integrity, along with business drivers, modeling choices, and validation approaches. The conclusion provides a forward-looking perspective on the potential of dynamic system designs.

Definitions and Categorization of Static and Dynamic Systems

This article relies on the Organisation for Economic Co-operation and Development (OECD) definition11 of AI, highlighting the two design aspects of autonomy and adaptiveness in focus of our discussion: “An AI system is a machine-based system that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. Different AI systems vary in their levels of autonomy and adaptiveness after deployment.”

Autonomy and adaptiveness relate to the level of human intervention during operation of the system. According to the ISPE GAMP^® Guide: Artificial Intelligence7 (Appendix M10), autonomy is concerned with the level of oversight on the output of AI-enabled computerized systems and their use in the process, divided into stages of autonomy. Similarly, the FDA distinguishes “Assistive AI,” which supports decision-making, and “Autonomous AI,” which makes decisions without human intervention; 12 the European Union AI Act13 and recent guidelines clarify the relation of autonomy to “‘human involvement’ and ‘human intervention’ and thus human-machine interaction.”14

Although higher levels of autonomy have been applied in GxP contexts—such as AI-enabled visual inspection—higher levels of adaptiveness are now seen less frequently. Adaptiveness refers to the “AI subsystem’s ability to automatically perform updates and thereby automatically self-improve”7 (Appendix M10), distinguishing static designs (use of new model versions only after traditional human verification steps) and dynamic system designs (automated adaptation of models). These broad categories of static and dynamic designs are also reflected in the EU GMP “Annex 22” draft.5

Dynamic system designs may be chosen based on an understanding of the temporal development (e.g., wear and tear; see the following case study), alongside sufficient trust in the model and control of data.

Examples of such dynamic systems include:

1. Automated model retraining and instantiation after elapse of time
2. Automated model retraining and instantiation after a certain number of data points
3. Automated model retraining and instantiation as controlled by key performance indicators (KPIs) and thresholds from ongoing monitoring
4. Automated updates of the model with each and every data point

AI subsystem in autonomy and adaptiveness stages can be com-bined in a maturity model, which may also be used as a starting point for a risk-based approach as presented in the ISPE GAMP^® Guide: Artificial Intelligence7 (Section 2.2.4 Science-Based Quality Risk Management, detailed chapter 5 and Appendix M3), in line with ISPE GAMP^® 5 (Second Edition) Appendix D11 on AI and ML (see,8 Appendix D11). When companies consider adopting an AI subsystem with a higher degree of maturity, those companies must carefully consider risk impact in terms of product quality, data integrity, and patient safety in GxP use cases (see also,7 Appendix M10).

To better understand dynamic systems’ practical relevance, the following section highlights how they’re used in life sciences.

Use of Dynamic Systems in Life Sciences

Dynamic systems in the life sciences industries can achieve a higher performance or enable new use cases that static approaches cannot manage. For instance, dynamic models create new opportunities in which the ratio between input and output data changes over short intervals. They also offer potential for improvement in use cases in which large volumes of data are generated and can be continuously used for training.

Specific advantages include:

• Lower risk of overfitting: The model continuously incorporates new data. As a result, it is less likely to overfit to a specific data set.
• Increased training efficiency: Dynamic learning approaches mean smaller amounts of data can be used per cycle, which can reduce training times, depending on the model architecture.
• Accuracy improvement: By using incremental learning approaches to regularly or continually incorporating new data, models may incorporate changes in the environment or system behavior more quickly to maintain or boost operational performance and accuracy.

These advantages illustrate dynamic systems’ potential through a higher degree of digitization and automatization, which could improve overall process quality. However, to seize this potential, companies need robust governance and control strategies. The advantage of continually incorporating new data, for example, may prompt the system to shift unexpectedly and overemphasize more recent events.

Use Cases of Dynamic Systems

Speaking technically, dynamic system designs are generally feasible only when no external information is needed for retraining, enabling closed feedback and data loops. From a process perspective, companies must consider a thorough control strategy to satisfy quality expectations and a clear business case for a dynamic design.

However, some life sciences regulatory guidance, such as the EMA reflection paper,4 discourage using dynamic systems in areas such as late-stage clinical trials, as does the draft EU GMP “Annex 22”5 for critical use in GMP. Moreover, the model’s architecture must be tailored to fit a specific dynamic design, capable of adapting with changing inputs and conditions.

A system may be initially adopted following a dynamic design or starting at a lower level of AI maturity with the goal of achieving a fully adaptive design. The following examples illustrate cases in which a dynamic system design could benefit companies that use them:

• Advanced process monitoring and control of key process parameters (described in the following case study)
• Recommendations of corrective and preventive actions based on historical data with direct operator feedback loop
• Deviation description and regulatory submission report generation with direct operator feedback loop
• Yield optimization based on chromatography data with dynamically adjusted separation based on recent process parameters

However, these benefits come with important risks, including unintended model drift, reduced transparency, and validation challenges. We’ll discuss these aspects further in the next section.

Challenges and Risks of Dynamic System

Although dynamic systems exhibit promising benefits, they also present distinct challenges and risks that companies must manage. These activities focus on the operational governance and control of the update cycles of models, specific design considerations, and human interaction and change management.

Operational Governance and Controls

Controlling model quality and performance is critical for keeping dynamic systems validated. The model quality can be controlled, for example, by defining a controlled space with quantitative boundary conditions and rules under which it may evolve. The corresponding KPIs capturing input data quality and model performance must be defined and monitored during operational use to verify consistency with the controlled space.

Performance monitoring requires closed data loops. For example, data or process outputs need to be labeled for supervised learning. For unsupervised learning absent a target variable, other performance control strategies need to be designed. These may include controls of incremental shifts of hypothetical model outcomes compared with previous versions and use of the model. In either case, periodic review of the model’s evolutionary path and periodic retesting should be performed; see also ISPE GAMP® Guide: Artificial Intelligence7 (Appendix P3). These activities should focus on changes over multiple versions of the model that together may cause the model to shift or misbehave.

Lastly, companies should assess specific risk scenarios in the context of use to plan appropriate quality assurance measures, including respective monitoring. As the model moves out of specification, a reasonable control is to switch to a stable model without dynamic elements, until reinitiated. Further measures include root cause analysis and remediation strategies with particular focus on input data and shifts of model behavior as well as human assessment of plausible and implausible model results.

Design Considerations

Besides general considerations from an AI governance perspective, further system design considerations have to be established during the dynamic system’s development.

A key dynamic system design decision is version release, that is, when a new version is created, depending on the amount of new training data, elapse of time, or KPIs with the goal to incorporate changes between input and output via (immediate) feedback. The system must ensure full traceability. It should record which model version was active and which decisions that version prompted. Also, interfaces are needed to relate data input, the model, and the data output as used in the process; see also ISPE GAMP^® Guide: Artificial Intelligence7 (e.g., Section 4.4.8).

This is especially relevant for systems that change with every data point. From a model quality assurance perspective for these special cases, setting interim versions is highly recommended as is defining a controlled space for proving that output data goes out within the boundaries. If required, companies may partially rebuild historic models for investigation. Alternatively, version release plans let companies demonstrate model control by showing the compliant behavior of surrogating model versions. This may matter because, based on system architecture, it may be impossible to fully trace a defined version’s behavior to a given time.

Human Interaction and Oversight

Human interaction and oversight are important when designing risk-and-control strategies. Users should get information for understanding implications and risks to keep the model within boundaries and behaving plausibly, at the same time limiting changes by humans outside the process’s validated parameter ranges. Explainable AI techniques may provide insights not only into what drives model results but also about changes between model versions (e.g., higher relevance of certain data characteristics in an updated model version); see also “Road to Explainable AI.”15 For this evaluation, the autonomy and the impact to patient safety, product quality, and data integrity are crucial. Because the degree of explainability can vary among models, the required level of information to perform their oversight role should be evaluated and included during the model selection process in the concept phase of the system.

Organizational Change Management and Upskilling of Staff

Using existing complex AI models may already be challenging for some organizations; dynamic system designs can pose further challenges for managing organizational change. Nevertheless, companies must manage these challenges to foster an environment for safely and effectively establishing and adopting these systems. Because the degree of explainability can vary among models, companies need to evaluate how much information is necessary for system oversight and keep it in mind when selecting models and designing the system.

As described in the preceding section, human oversight is important for ensuring systems stay validated, and developing evolutionary models of dynamic system elements. Training and upskilling are required to apply effective oversight. This includes learning new concepts to govern the human and AI interactions so users can cope with and control such systems’ dynamically changing character and ensure trust in model results. Upskilling will also let people explain the model, its changing behavior, and its controls during regulatory inspections. These topics should be integrated into staff training through structured learning paths that show how model behavior affects individual results. This may include the usual behavior of evolving models, but also situations in which models bump up against control-strategy-imposed boundaries. This should let end users distinguish between expected dynamic learning behavior and unusual or suspicious evolutionary trajectories of the model and changes of its results; a knowledge base may collect further insights during operation.

Risk Strategies and Regulations

Patient safety, product quality, and data integrity are at the core of all life sciences regulations. The notion of impact is critical to consider, as included in the EU GMP Annex 225 draft, which discourages the use of probabilistic models or dynamic systems that directly affect patient safety. A robust, well-reasoned control strategy is important to establish, leveraging critical thinking and an understanding of risks to patient safety, product quality, and data integrity. Suitable risk management approaches depend on the use case and the system design.

In most cases, in areas with higher risk, AI is used only as one subcomponent or module of a larger computerized system. A layered examination of the system and its components may determine that AI subsystems, such as for process monitoring purposes or adding an additional layer of control, do not directly affect patient safety or product quality. Therefore, the AI subsystem would potentially not be seen as a critical application with direct impact under consideration of EU GMP “Annex 22” draft.5

As further assurance, inspired by the pharmaceutical product control strategy, boundaries of critical process parameters (CPPs) and input variables of raw materials—the design space—can be defined, within which the AI subsystem would operate; see also the international Conference on Harmonisation of Technical Requirements for Registration of Pharmaceuticals for Human Use, Pharmaceutical Development Q8.16 By applying the control strategy, and verifying its effectiveness (see earlier section), AI subsystem operation would be monitored and human intervention triggered if a limit is breached and the dynamic learning would be stopped. Additionally, rule-based algorithms may take the role of the AI subsystem in the process. Such cases would be subject to incident management, with an assessment and root cause analysis, as applicable. Also, human intervention would be controlled by respective standard operating procedures, and operating personnel should be adequately qualified to execute the process. For additional control, a critical quality attribute (CQA) monitoring strategy would also be in place to ensure product quality.

In summary, multiple methods can be used to manage dynamic systems’ risk to patient safety and product quality. The best risk strategy to control the system for regulatory inspections depends on its use and settings. The following case study illustrates how these concepts can be applied in a real GMP environment.

Case Study for Primary Packaging Container

The subsequent case study will deal with high-volume production of prefilled syringes made of high-performance plastics cycloolefin copolymer (COC) or cycloolefinpolymer (COP) for administering drugs, enabling a long shelf life due to significantly better barrier protection against gases (moisture and oxygen), higher chemical inertness, and tighter dimensional tolerances for administering drugs. The production line consists of a high-volume injection molding machine, subsequent processing stations, and several fully automated inspection stations along with packaging practices for nesting, tub insertion, and tub sealing. The material flow is defined by a workpiece carrier system throughout the machine and simplified as depicted in Figure 1.

Figure 1: Simplified manufacturing process flow of the syringe production line use case.17

Image

In dynamic learning, an incremental learning system for a full automated and validated production line was considered to systematically collect feedback and evaluate results. The learning system would have the required data integrity controls in the automation to prevent alteration of filed parameters outside of the validated ranges. It would also have alarms to alert the operator to any CPP/CQA-related excursions.

The optimization system aims to increase operational efficiency during multishift 24/7 operations running with varying input factors such as raw material properties, wear on molds and machines, and human-based experiences. This helps identify suitable counteractions at any given time.

The main motivations for this particular case study include business drivers and technical influencing factors. On the one hand, many influencing factors change over a machine’s lifetime. A dynamic control strategy helps address these changes more effectively. This can increase operational efficiency without needing to define manual root causes and countermeasures.

On the other hand, raw material properties are a key factor. Depending on supplier and batch, the specific properties change within an allowed range. There is also often a primary batch material mixed into raw materials to generate certain properties that can change slightly with each batch and supplier over several years or decades. Machine wear occurs continuously throughout the production machine’s lifetime. Therefore, the machines’ behavior in month 1 is slightly different than in month 60, also considering exchange of parts, resulting in a “reset of wear to 0” for dedicated components but not for the whole machine. So, interaction among machine components with different wear states changes over time.

The Dynamic Learning Model: Applied Methods

For this case study, a set of model pipelines, each containing several ML models, is applied to generate situational recommendations. These recommendations contain set parameter changes to react to the current process, at the same time learning and evolving from it.

For further reference, the following publications illustrate the current technical state of the art of models being used: We distinguish between 1) explicit probabilistic modeling approaches, such as Bayesian networks,18 which define a structured probabilistic dependency model, and 2) implicit data-driven approaches, such as state machine induction19 or temporal behavior graph models, which infer process structure directly from observational data,20 which we mainly use. For a general overview, please see.21

New data, from running machines for mainly two pipelines in parallel, lets end users consider continuous influencing factors:

1. As input data of the most recent qualified model to generate an output recommendation
2. As input to retrain the algorithm stack to align with recently observed behavior; this is a new facet typically not done after design-freeze for currently running AI models in GMP environment

A key element of the technology is learning the process model implicitly from hundreds of continuously changing, high-frequency time-series inputs. This lets the system generate counteractions for given situations and capture previously seen situations and new effects that occur over time. Static models cannot fully anticipate these changes because they rely on historical data with fixed cutoff points.

Figure 2: Conceptual overview describing the functional setup of the pipeline. 20

Image

The following conceptual overview describes the functional setup of the pipeline (see Figure 2): Based on high-frequency raw data from the machine, a behavioral state model is implicitly trained and optimized, and an alternative set of controllable process parameters is derived. If it complies with allowed boundary conditions based on previously validated and filed/registered ranges and/or within the design space validated process windows, it is either:

• Sent to a person providing “human-in-the-loop” oversight (orange arrow) for acceptance or denial
• In a “closed-loop” mode written on the machine controller to be immediately active

The retraining scheduler running in the second point varies: At the beginning, new contexts are expected to occur more often, for example, by introducing and qualifying each new material supplier coming in for each validated product produced. Model updates can be done on a shorter timeline to reflect all new behaviors introduced.

GxP Considerations

The production line has been operating without any additional optimization tool before this case study started. As such, varying high scrap rates during 24/7 operations and efficiency losses at certain points in time have been observed. The optimization software is, on the one hand, a non–mission-critical additional service on top of the existing validated production procedures. It is also affecting the production process and its CPP/CQA, which need to be analyzed and have its risks assessed and mitigated (e.g., by adopting safe-fail modes).

Considering the risk-based validation approach of ISPE GAMP^® 5 (Second Edition),8 the main impact in this case study relates to the following:

• Direct impact on data integrity as risk category: As a data-intense and live-running integrated information technology system, it must be ensured that all other data flows responsible for process data documentation (e.g., track and trace application running in parallel) are not compromised
• No direct impact on product quality as risk category (not direct because 100% inline inspection per the current validated process control strategy will remain unchanged and in place after the technology introduction): The ML pipeline itself can be systematically qualified but not on a frozen data set, because no local (on the historic data relying) qualification is possible because of continuously changing data sets and counteractions

The validation standpoint is aligned with the AI maturity model as provided in the ISPE GAMP^® Guide: Artificial Intelligence 7 as a general guideline: The particular incremental learning technology was divided into three subsequent stages (1 to 3) and an outlook stage (4). Note that the high-frequency data connectivity as precondition is also qualified with a conventional risk-based approach based on ISPE GAMP^® 5 (Second Edition) (see,8 appendix M3). It is necessary for assessing the impact and potential risks, particularly on data integrity, when reading a significant amount of additional data from all involved machine controllers (PLCs) and data interfaces. It’s also necessary when creating additional network load.

In summary, the life cycle as per ISPE GAMP^® 5 (Second Edition)8 (Appendix D11) was considered as suitable for this dynamic learning approach. It took into account further considerations as outlined in the ISPE GAMP^® Guide: Artificial Intelligence7(see Figure 3) and concepts enabling scalable life-cycle activities and risk-based approaches.

Considerations on the Validation and Control Strategy

The following graphic (see Figure 3) illustrates the subsequent staged validation as an evolution path within the framework. The individual stages 1 to 3 are introduced below. For a detailed classification with background information on the framework, we recommend the ISPE AI Maturity Model publication,9 including the idea of an incremental approach to achieve higher levels of AI maturity, and the updated conceptual basis of AI maturity in line with the ISPE GAMP® Guide: Artificial Intelligence7 (Appendix M10). Generally, the human oversight and critical thinking element are crucial (i.e., the assurance that the control strategy is operationally adopted and effective).

This case study focuses on the classification of an incremental learning optimization tool.

In the following, we describe key validation-related considerations; note that only dynamic learning specific content is depicted. Nevertheless, qualification and management of the IT infrastructure and validation of the non-AI computerized system must be addressed.

Stage 1

Stage 1 is the “nonbinding AI glimpse”: The incremental learning tool is introduced as an “in parallel running” tool; process optimization and troubleshooting in case of scrap being produced is officially, at this stage, still a manual process. The responsible process expert can assess the AI-enabled computerized system’s access parameter recommendation but must apply critical thinking when doing so. Already by installation and beginning training, the AI subsystem adaptiveness stage 3 was chosen. During this first phase, all involved people can individually evaluate system functionalities and situational recommendations without investing in full validation. From a business perspective, this allows teams to forecast economic benefits that the following next steps, including significant validation, could yield.

From a technical AI perspective, it is also crucial to compare both reactions for each situation: Is the process expert considering partly/fully the AI-generated recommendation to understand human-based decisions? After this, context-information and additional experiences from the past should be considered, or both systems should act in the same manner, meaning humans follow the generated recommendations with a high probability.

Figure 3: Iterative system design shown similar to the ISPE GAMP^® Guide: Artificial Intelligence.7

Image

Stage 2

Stage 2 is “human-in-the-loop AI”: Each automatically created recommendation is suggested to an experienced and responsible person who decides about acceptance. This leads to the AI subsystem autonomy stage 2 system, and part of the AI maturity level IV, which highlights the following aspects on top of subsequent level III aspects.

Monitoring of model quality in operation: In this case study, the monitoring framework relies on the following. The first is an explainable AI approach (see Figure 4) to ensure that skilled technical people can understand, in a graphical qualitative and quantitative way, how models changed and evolved between two iterations. For instance, Shapley additive explanations (SHAP) values can be used to highlight that parameters that were once important for the model are no longer important in the next generation of a trained model, and that the first five most important features are still the same but also in slightly different order of importance. For an overview survey of typical methods for explainable AI and their field of applications, see.22

Second is controlling quality KPIs and notification process: Besides those qualitative investigations on model input-output relations, there is a full set of KPIs for the currently running model behavior. There are also send notifications about the performance profile and KPI trends to defined user groups in charge of operations of the AI-enabled computerized system.

Therefore, internal model quality criteria and model application rules are defined. These rules and criteria are initially validated as mutually exclusive and collectively exhaustive (MECE guideline) to control the dedicated AI pipeline for the intended use. This depicts an automated quality assurance procedure comparable to a ML operations approach.

Validation of the human factors related to overrides, qualifications, and AI system acceptance: The person in charge must be able to understand and judge, based on their own critical thinking approach, the situational recommendation. This person must consider existing control strategy elements when deciding to accept or decline the proposed action. This is linked to the question of user group definition, and a clear user role model must be defined and ensured. Also, a control strategy to continually verify the effectiveness of human oversight and intervention should be in place (e.g., by ex post facto evaluation of a representative set of scenarios).

Stage 3

Stage 3 is “adaptive design within well-defined boundaries”: As part of the next stage (AI maturity level V), an automated model update procedure is established on top of the predefined KPIs and explainability features defined in the last stage. The observation of the control strategy is key to placing the system in a continuous updating mode. The consistency and value range of all input data are actively observed (as part of level III validation). Data is also evaluated by each model iteration, the potential output data-range, and the quality KPIs of the recently updated model. This data is then compared with past models. Only during in-specification operating range and a quality KPI improvement is the updated model considered for the live system. This prevents people from reacting to random noise and model oscillations that may make the system unstable.

Figure 4: Explainable AI methods applied in the context of the use case.20

Image

Stage 4

Stage 4 is “AI maturity level VI”: After collecting experiences with adaptiveness stage 3, authors approach an increase of adaptiveness from stage 3 to stage 4 by setting up a fully closed-loop target system for retraining models. This will be part of further investigations.

Business Impact

Two main effects are highlighted as main business impacts. First, for the 24-hour running syringe production line, an output increase of 5% was observed consistently. A reduction of 30% of variation of output was observed, comparing all shifts because of situational reaction immediately on slight changes of models. In conclusion, using the modeling approach saves scrap and cuts manual reaction time, which could boost production uptime and make the process layout more robust.

Second, reducing process disturbances (e.g., manual intervention in the process) benefits validation—a lower number of events happening means less operational manual documentation must be created and reviewed. The observed number of short stops (less-than-10-minute production machine stops to manually change machine panel parameters) was reduced during the case study. In summary, the case study has proved that dynamic learning models can help reduce the number of operational disturbances throughout multishift operations.

Conclusion

As this article has highlighted, dynamic systems are no longer just a possibility—real-world use cases are starting to emerge in which AI models are adapting to patterns in new real-time process data. In some instances, ruling out dynamic model approaches may itself introduce risk. Apparently less complex, static models may develop drift from the real-world relationships they are intended to represent. For certain use cases, this divergence undermines the model’s utility and may compromise process integrity. Exploring dynamic AI within governance boundaries offers a way to reinforce control strategies and perhaps inform static systems’ designs, underpinning and augmenting foundational frameworks developed in the GAMP® body of knowledge. This includes explainable AI,13 ML-specific risk management,10 and maturity models9 and further considerations consolidated in the ISPE GAMP^® Guide: Artificial Intelligence.7

Dynamic AI use must be grounded in process understanding and justified by clear business benefits. The higher levels of complexity should be justified to warrant adoption; it should be matched by robust quality risk management that accounts for model behavior and learning dynamics. Effective human oversight remains essential—not just as a formality but as a functional safeguard. The effectiveness of human–AI interaction should be demonstrable, and system controls should ensure the model remains controlled. Monitoring the model’s evolution, particularly through adaptive retraining strategies, is critical for detecting and responding to shifts in input data or model performance.

As dynamic systems move from concept to application, their coexistence with traditional control strategies becomes increasingly feasible, and complimentary. Use cases that operate alongside validated processes provide practical insights but must be supported by forthcoming regulatory frameworks. Such AI adoptions contribute to the broader discourse on AI applicability in GMP environments. We hope this article contributes to continuing discussions on the use of AI in the current EMA work plan23 and the revision of the EMA GMP Annex 11,24 the new Annex 22,5 and future US FDA regulatory initiatives.3

Acknowledgments

The authors would like to thank the ISPE GAMP^® Global Software Automation and Artificial Intelligence Special Interest Group (SIG) for its support in this article’s creation. The authors would like to thank Brandi Stockton, Sebastian Pfaff, and Veit Mengling for their review.