November 19th | 12:00 PM ET | 1 hour
Deploying GenAI and agents for regulated quantitative sciences creates a tough technical challenge. Traditional computer system validation expects predictable, reproducible software, but LLMs don’t work that way. Forcing LLMs into old validation processes produces endless paperwork, while blocking them entirely usually just drives data scientists toward unmonitored shadow AI.
Peter Krusche from Novartis and Matt Tendler from Domino Data Lab will walk through practical solutions. Instead of trying to validate model weights, we focus on reasoning from GxP principles and validating the outputs, having effective review when AI hands over work to humans, and implementing meaningful controls when AI is added to a statistical computing environment. We will also discuss PAIR, a CDISC AI Innovation Challenge Winner for 2026, which implements a human-supervised AI workflow that can draft all TFL programs for a clinical study in one go.
Key ideas:
Shift the approach of using AI from validating models and weights to validating generated programs and outputs - where the 21 CFR Part 11, and EU Annex 11 frameworks work well.
Producing strong evidence that meaningful human supervision has taken place: Spec-based supervision of AI work, stacked pull requests to keep human code reviews fast, thorough, and fully auditable.
PAIR generates study-level TFL programs using metadata and synthetic ADaM datasets inside an isolated sandbox and facilitates the transition into the GxP space by focusing on a strong user experience for planning and review.
