Uploaded October 2026 | Updated November 2026, 4 hours ago
As AI takes on more of the work in clinical trials, how do you make sure a human stays accountable? A talk from the Validated AI for Pharma Summit.
Speaker:
Peter Krusche, Novartis (AI and Data Engineering, Development Advanced Quantitative Sciences)
Peter starts from two GxP basics: a qualified human is accountable for every analysis, and the analysis runs in a validated system. AI sits between the two, and it can’t be validated or held accountable itself. His answer is to validate what the AI produces, the same way teams already handle people writing different code from the same spec.
What he talks about:
- Why the AI should draft and challenge, while a qualified human checks the evidence and approves the result
- Why agents shouldn’t get the same access to validated systems as the people using them
- Oversight that scales: agree the spec first, then check the work against the spec instead of reading thousands of lines of code
- Specs clinical trials already have, like USDM, data standards and controlled terminology, and how agents can use them
- Mock data for testing agents, and keeping double programming independent when AI writes both versions
- Q&A on generating TFLs with agents, working with quality teams, and building mock data from real metadata
His main advice: start by writing down the context of use, meaning what exactly you want AI to do with the data. Everything else builds on that.
Recorded on 24 September 2026 at the Validated AI for Pharma Summit, hosted by Appsilon.
All summit sessions on demand: events.ringcentral.com/events/validated-ai-for-pharma-summit?utm_source=youtube&utm_medium=video&utm_campaign=validated_ai_summit_2026&utm_content=ai_and_accountability
#ValidatedAI #GxP #ClinicalTrials #PharmaAI
As AI takes on more of the work in clinical trials, how do you make sure a human stays accountable? A talk from the Validated AI for Pharma Summit.
Speaker:
Peter Krusche, Novartis (AI and Data Engineering, Development Advanced Quantitative Sciences)
Peter starts from two GxP basics: a qualified human is accountable for every analysis, and the analysis runs in a validated system. AI sits between the two, and it can’t be validated or held accountable itself. His answer is to validate what the AI produces, the same way teams already handle people writing different code from the same spec.
What he talks about:
- Why the AI should draft and challenge, while a qualified human checks the evidence and approves the result
- Why agents shouldn’t get the same access to validated systems as the people using them
- Oversight that scales: agree the spec first, then check the work against the spec instead of reading thousands of lines of code
- Specs clinical trials already have, like USDM, data standards and controlled terminology, and how agents can use them
- Mock data for testing agents, and keeping double programming independent when AI writes both versions
- Q&A on generating TFLs with agents, working with quality teams, and building mock data from real metadata
His main advice: start by writing down the context of use, meaning what exactly you want AI to do with the data. Everything else builds on that.
Recorded on 24 September 2026 at the Validated AI for Pharma Summit, hosted by Appsilon.
All summit sessions on demand: events.ringcentral.com/events/validated-ai-for-pharma-summit?utm_source=youtube&utm_medium=video&utm_campaign=validated_ai_summit_2026&utm_content=ai_and_accountability
#ValidatedAI #GxP #ClinicalTrials #PharmaAI










