Q&As on "GMP meets AI - How to Use Artificial Intelligence in Quality Assurance and Quality Control"

During the Live Online Training organized by the ECA, titled “GMP meets AI—How to Use Artificial Intelligence in Quality Assurance and Quality Control,” participants asked numerous interesting and practical questions. We have translated a selection of these questions and answers and organized them by topic.

1. Access Control for Password-Protected AI Tools

Question:
How is access to password-protected systems managed when AI tools are involved? Is security ensured?

Answer:
AI and automation solutions are operated within closed enterprise environments. Data are hosted in ringfenced infrastructures and are neither directly nor indirectly accessible to the public. Importantly, company data are not used for training publicly available AI models. Access is strictly role-based. System owners - typically department heads - must regularly review which employees require access to which systems. When new team members join, their required system access is evaluated and granted accordingly. When employees leave the organization or change roles, their access rights are revoked or adjusted. This governance model ensures that only authorized personnel have access to sensitive systems.

2. Data Sources for Predictive Intelligence

Question:
From which systems does Predictive Intelligence extract its data?

Answer:
Predictive Intelligence solutions draw information from several key enterprise systems, often referred to as "North Star systems." These typically include: - deviation management systems, - enterprise resource planning (ERP) systems, - change control systems, - and additional tracking or management applications.
Relevant data are consolidated into an Enterprise Data Backbone (EDB), which serves as the central data architecture. Predictive analytics tools retrieve structured and contextualized information from this backbone.
For major enterprise systems-such as SAP or large e-commerce platforms-dedicated interfaces are established. This approach preserves contextual integrity and ensures appropriate update frequencies. Transferring all data indiscriminately into a single data pool would generate excessive data flows and synchronization challenges. Therefore, dedicated "data highways" are implemented between major systems, while smaller systems continue to feed into the central backbone.

3. Batch Tracking and System Architecture

Question:
How is batch tracking currently organized, and what is the future strategy?

Answer:
Batch tracking is currently supported by front-end solutions that consolidate information from various systems. Legacy elements still include paper-based records as well as SharePoint and Excel-based solutions.
The long-term strategy is to fully integrate batch tracking into a centralized system, such as an SAP-based solution. The objective is to maintain all batch-relevant data within a single, validated environment and eliminate parallel stand-alone tools. This reduces validation effort and prevents data fragmentation.
The current implementation represents an interim solution. Although interim systems may require temporary double investment, the structured data foundation built during this phase is considered essential groundwork for the future target system.

4. AI Usage in Batch Tracking

Question:
Is artificial intelligence currently used in batch tracking?

Answer:
At present, the solution primarily serves as an automated front-end for data aggregation and visualization. Active machine learning or AI-driven decision logic is not yet embedded. Future enhancements may include AI capabilities, but the current focus lies on transparency and process simplification.

5. Review by Exception and Trust in AI

Question:
How can AI be trusted, especially considering known AI errors (e.g., the "six fingers" example)?

Answer:
The principle of "garbage in, garbage out" fully applies. The reliability of AI outputs depends heavily on input data quality. Poor-quality inputs-such as low-resolution screenshots-can lead to incorrect outputs.
Within a review-by-exception framework, each exception must still undergo human quality review. AI may provide suggestions for reducing recurring error classes or improving electronic batch records, but final evaluation and decision-making remain human responsibilities.
AI functions as a decision-support tool rather than an autonomous decision-maker. The "human in the loop" principle remains a core governance element.

In the new edition of the Live Online Training Course “GMP meets AI—How to Use Artificial Intelligence in Quality Assurance and Quality Control” on 17/18 November 2026, you will learn about the latest developments in the use of artificial intelligence in the core GMP areas of quality assurance and quality control.