Experts urge organizations to evaluate intended use, required human oversight, validation needs, revalidation triggers, and whether AI will meaningfully improve quality or save time before adopting a tool.

Edwards Lifesciences demonstrates a human‑in‑the‑loop approach where AI drafts documents and a regulatory affairs professional reviews and finalizes the dossier, preserving human oversight at the end.

Testing strategies vary by risk: low‑risk AI can use unscripted testing, while high‑risk AI may require scripted testing or a mix, with attention to AI changes outside user control.

Validation guidance cites ISO 13485:2016 for QMS, FDA Quality Management System Regulation, and the new FDA computer software assurance guidance, recommending risk‑proportionate testing based on AI use and risk level.

A panel at the RAPS AI in Regulatory Conference advocates a risk‑based, life‑cycle approach to AI in regulatory affairs, stressing validation and ongoing monitoring to catch errors.

Validation should cover data appropriateness, representativeness, quality, task performance, result trust, robustness, security, and end‑to‑end workflow.

Kaspar of TÜV AI.Lab stresses that validation must span the entire lifecycle and be traceable to intended use and risks, focusing on data quality, task performance, trust in results, and end‑to‑end robustness.

IEC 63450, while not AI‑validation‑focused, provides an AI testing method for unknown final answers, aiding understanding of tool behavior.

Human oversight is essential but must be well designed, with trained overseers who understand the AI, plus ongoing monitoring to detect degradation or issues.