Biosimilar Development Covers, Controls for Retrieval-Augmented GenAI in GxP, to be Presented at the 2026 ISPE Annual Meeting & Expo Upcoming Session
18 September, 2026
As pharmaceutical companies seek to leverage generative artificial intelligence within regulated environments, retrieval-augmented generation (RAG) is gaining attention as a practical approach for balancing innovation with compliance. The topic will be explored at the 2026 ISPE Annual Meeting & Expo, where data integrity expert Ulrich Köllisch, Founder of UMK Consulting, will discuss how organizations can validate and govern RAG systems operating in GxP environments.
Unlike standard large language models (LLMs), which generate responses based solely on patterns learned during training, RAG systems combine AI-generated outputs with information retrieved from controlled and validated knowledge repositories. By limiting responses to approved data sources, RAG offers pharmaceutical manufacturers a way to take advantage of generative AI capabilities while reducing concerns related to hallucinations, unsupported conclusions, and inconsistent information. However, the technology also introduces new validation, cybersecurity, data governance, and compliance challenges that traditional computerized system frameworks were not specifically designed to address.
According to Köllisch, existing industry frameworks, including GAMP® 5, ICH Q9 (R1), and established data integrity principles, remain highly relevant for governing RAG systems. Rather than creating entirely new validation methodologies, organizations can extend existing compliance practices to cover the full RAG lifecycle. This includes oversight of data ingestion, preprocessing, embedding, retrieval mechanisms, model versioning, user interactions, audit trails, operational monitoring, and change control. The critical difference is that companies must validate not only the AI model itself but also the quality, integrity, and permissions associated with every data source feeding the system.
One of the primary challenges in validating RAG systems stems from their probabilistic nature. Unlike traditional software applications that consistently produce the same output when provided with identical inputs, AI systems may generate slightly different responses each time. Köllisch notes that this issue is not entirely unfamiliar to pharmaceutical manufacturers. GMP operations such as visual inspection already involve inherent variability and are routinely validated using statistical methods. He suggests that similar risk-based verification approaches can be applied to AI systems, focusing on demonstrating acceptable and controlled performance rather than identical outputs.
The complexity of RAG validation extends beyond model behavior to the entire data pipeline. Organizations must ensure robust governance across every stage of data processing, from source selection and ingestion to retrieval and response generation. Maintaining data integrity throughout this chain is essential because the quality of AI-generated outputs ultimately depends on the quality of the information being supplied to the system. As Köllisch emphasizes, the longstanding principle of “garbage in, garbage out” remains highly relevant in the era of generative AI.
Another key consideration is model lifecycle management. Köllisch recommends maintaining a versioned and approved model registry similar to a controlled software release process. New versions may be required when model performance declines, deficiencies are identified within data pipelines, or improved models become available. However, RAG systems require additional controls beyond those typically associated with conventional software updates because modifications to models, embeddings, preprocessing methods, or knowledge repositories can all influence system behavior. As a result, organizations must implement rigorous approval processes and ongoing monitoring programs to ensure continued compliance.
The article also highlights the limitations of autonomous AI in regulated environments. While agentic AI systems are rapidly evolving, Köllisch notes that fully autonomous decision-making remains largely unacceptable for critical GMP activities. He points to the draft EU GMP Annex 22, published in 2025, which explicitly prohibits probabilistic models from independently making decisions related to critical GMP processes. Human oversight therefore remains a fundamental requirement, with AI serving as an information assistance tool rather than an autonomous quality decision-maker. Users must still review source information and retain accountability for GMP-related decisions.
Cybersecurity presents another significant challenge. Because RAG platforms often pull information from multiple connected databases, they expand the potential attack surface compared to traditional isolated validated systems. Köllisch recommends securing all layers of the architecture, including data ingestion processes, infrastructure components, and application-level controls. Organizations should implement strict identity management policies, role-based access controls, multi-factor authentication, and cloud-provider oversight to mitigate risks associated with interconnected systems.
A major advantage of RAG systems is their ability to support data integrity and traceability. In the information-as-a-service model described by Köllisch, users submit queries through a controlled interface and receive responses accompanied by citations to original source records. Behind the scenes, audit logs capture user identity, timestamps, permissions, source document versions, retrieved content, system configurations, generated outputs, and subsequent actions. This level of documentation supports ALCOA++ principles by ensuring records are attributable, contemporaneous, accurate, complete, and fully traceable to authoritative source data.
Köllisch also sees considerable alignment between regulatory expectations in the United States and Europe. Both the US Food and Drug Administration and European Medicines Agency authorities have expressed support for the responsible use of emerging technologies when applied through scientific, risk-based approaches that enhance patient safety and product quality. While some regional differences remain, particularly regarding access to testing data, he believes companies can develop governance strategies that satisfy both regulatory frameworks.
For organizations beginning their AI journey, Köllisch advises first establishing a shared understanding of the technology among stakeholders and defining an appropriate risk appetite. Successful implementation depends not only on validating the AI model but also on ensuring that the underlying data sources meet the required standards for quality and integrity. These themes are expected to be central to discussions at the 2026 ISPE Annual Meeting & Expo, where industry experts will examine how RAG can enable the benefits of generative AI while maintaining the rigorous controls required in GxP-regulated pharmaceutical operations.
Learn More About the 2026 ISPE Annual Meeting & Expo
About Biosimilar Development
Biosimilar Development is part of the Life Science Connect media group. The publication provides in-depth content from pharmaceutical experts who specialize in the biosimilars segment of the industry.