Validating AI in development software

AI is showing up everywhere in the tools software teams use to build medical devices, from code assistants to design review copilots to AI-powered features inside the products themselves. Auditors are starting to ask about all of it, and most teams do not have a clear answer ready.

AI models are non-deterministic: the same input can produce a different output each time, and the underlying model can change without notice. That makes them a poor fit for a validation approach built for deterministic software. This session breaks down verifying an AI feature versus validating it. Jesseca Lyons and Will Sanders give quality and regulatory affairs professionals a practical, CSA-aligned framework for deciding what belongs in a validation package, what should be verified and documented separately, and how to explain that distinction to an auditor.

You will learn:

• Why AI features are typically verified but not formally validated, and what that means for your documentation
• How to apply FDA Computer Software Assurance (CSA) principles to AI-powered functionality
• What to document so human review of AI outputs holds up under audit scrutiny
• How to talk to auditors and prospects about AI validation without overpromising or underpreparing