Regulatory News

The QPPV in an AI-native PV system: oversight redefined, accountability unchanged

There is a quiet conversation happening in every pharmacovigilance department evaluating AI. It rarely shows up in the RFP. It shows up later, in the meeting after the demo, when someone asks the question everyone has been thinking.

If the system triages cases, codes MedDRA terms, and drafts narratives, what exactly is the QPPV signing off on?

It is a fair question, and the honest answer matters. Because the QPPV is not a workflow step that can be optimized away. Under Article 8(3) of Directive 2001/83/EC and GVP Module I, the Qualified Person for Pharmacovigilance is personally and legally accountable for the marketing authorization holder's pharmacovigilance system. That accountability does not soften when a model enters the room. It sharpens.

So the right framing is not "does AI replace the QPPV." It is: what does oversight look like when the work underneath it has changed shape?

What does not change

Three things stay exactly where they were.

The QPPV remains the single named individual accountable to the EMA and national competent authorities. The PSMF still has to describe the pharmacovigilance system accurately, including any automated components. And the obligation to detect, evaluate, and report safety signals on time is unchanged, regardless of who or what produced the underlying case data.

If a vendor tells you their AI removes the need for QPPV review, walk away. They have misread the regulation.

What does change, and meaningfully

The substance of oversight shifts. In a traditional setup, the QPPV reviews a sample of cases, signs off on aggregate reports, and trusts the SOP to have caught the rest. The unit of review is the case. The implicit assumption is that human case processors make idiosyncratic errors that random sampling will surface.

In an AI-native setup, that assumption breaks. Models do not make idiosyncratic errors. They make systematic ones. A model that misclassifies "shortness of breath" in elderly patients will misclassify it the same way every time, in every case, until someone notices.

This changes what the QPPV needs to look at. The unit of review is no longer just the case. It is also the system that produced the case. That means QPPV oversight now extends to:

  • Model performance monitoring. Precision, recall, and drift on the tasks the system performs, reviewed at a defined cadence and documented.
  • Configuration changes. Every change to prompts, models, or decision thresholds is a change to the PV system and belongs in the PSMF.
  • Exception and deviation logs. What did the system flag as low confidence, what was overridden, what was missed and caught downstream.
  • Audit trail integrity. Who or what made each decision, on what input, with what version of the system.

This is more demanding than the old model, not less. It is also, done properly, far better for patient safety. The QPPV moves from sampling cases to governing the system that processes all of them.

What to demand from a vendor

Three things, non-negotiable.

  • A QPPV-facing oversight layer, not just an operator dashboard. The QPPV needs a view designed for their accountability: model performance over time, configuration history, exception trends, and a clear line from any output back to its inputs and the system version that produced it.
  • Validation documentation that meets GAMP 5 and Annex 11 expectations for computerized systems, adapted for the non-deterministic behavior of AI components. If a vendor cannot tell you how they handle model retraining as a change-control event, they are not ready for production PV.
  • Human-in-the-loop and human-on-the-loop modes that the QPPV controls, not the vendor. The level of autonomy granted to the system on any given task should be a configuration the QPPV can see, change, and justify in an inspection.

The role is not shrinking. It is concentrating.

The QPPV of 2030 will spend less time reading case narratives and more time governing the system that reads them. That is not a diminishment. It is the role finally being asked to do what the regulation always intended: take responsibility for the pharmacovigilance system as a system, not as a sum of cases.

AI does not let the QPPV off the hook. It hands them a better hook.

To learn more about how Vigintake designs its platform around QPPV oversight, visit our Intake Management and Case Processing pages.