Induced from Binesh Prabhakar’s pharmacovigilance lessons-learned lecture at the KENX AI in GxP East 2026 conference (2026-09-16-1150-ai-validation-pharmacovigilance-it-compliance-lecture), 2026-09-16.
Silence about missing evidence is the dangerous failure
After twenty-four months of validating AI-assisted pharmacovigilance applications, his team hardened guardrail language so the system answers “unable to verify” when evidence is missing. The failure it guards against showed up in the local-LLM investigation too: systems answering without disclosing they only saw part of the data. That is Absence problems are harder to detect than presence problems spoken from the model’s mouth. A confident answer over a partial record is worse than no answer, because it looks like one.
The surrounding discipline
Their six-step GxP process: controlled evidence, fixed questions and prompts, AI analysis, evidence plus rationale, human review, approved output with audit trail. Failure modes they catalogued: severity inflation, duplicate findings, plausible but unevidenced assumptions, citation overreach. Those unevidenced-assumption cases belong in A golden test set must include unhappy paths. Agent personas carry CVs (training scope, versions, validation history, data residency) and get retrained when SOPs change. The human-review step is not optional decoration: Human verification of AI output must be explicit because document creation no longer carries implicit verification, and In the GxP control loop, AI acts, humans decide, and the system remembers is the loop this six-step process implements.
Related
- AI Validation and Acceptance Criteria MOC
- A golden test set must include unhappy paths
- Absence problems are harder to detect than presence problems
- Human verification of AI output must be explicit because document creation no longer carries implicit verification
- In the GxP control loop, AI acts, humans decide, and the system remembers