SSI Strategy

Artificial intelligence has been piloted across pharmacovigilance (PV) for several years, but adoption has remained cautious due to regulatory uncertainty. Companies have been exploring how to apply automation, natural language processing, machine learning, or generative AI to enable PV efficiencies and quality. Early pilots around 2021 produced some success and companies adopted robotic process automation for activities like case intake. Subsequent pilots explored more advanced applications in case processing and signal detection, but these efforts did not progress to broader implementation.

Although PV teams continue to be cautious about adopting AI, clinical and commercial groups have quickly embraced these technologies for tasks like patient recruitment, patient engagement, healthcare provider engagement, and marketing insights. These activities outside of PV often introduce new and uncontrolled sources of adverse events (AEs), requiring PV teams to react faster in monitoring, managing and controlling AI systems. Additional drivers come from the regulators themselves. The FDA has recently begun implementing advanced AI tools within its own operations. Beyond simply adopting AI, the agency is also taking steps to increase data transparency. For example, the FDA’s shift to real-time FAERS (FDA Adverse Event Reporting System) publishing means that information about AEs is now made available much quicker. This heightened transparency raises the urgency for the industry to respond, as regulators’ use of AI and instant data access set new expectations around the ability to monitor and manage safety information.

To assist the industry, regulators and industry groups have finally begun to offer guidance regarding AI frameworks and their validation approach. Examples include the CIOMS guidance from 2025, which advocates for a risk-based approach, emphasizes human oversight and underscores the importance of transparency and governance in applying AI to PV. The FDA Draft Guidance issued in January 2025 introduces a credibility assessment framework that requires a clear definition of the context of use, rigorous validation procedures, and ongoing monitoring of lifecycle performance. Similarly, the EMA Reflection Paper published in September 2024 promotes risk-based, human-centric principles throughout the medicinal product lifecycle, with scrutiny given to applications with high regulatory impact.

For PV teams, the main takeaway from these guidances is the need to ensure that use of AI follows risk-based credibility standards. This means clearly defining how and where AI tools will be used, how human oversight will be maintained, and then validating systems according to the potential impact on patient safety and regulatory outcomes.

The guidances highlight that expert judgment and final safety decisions must always remain with qualified professionals. Additionally, AI models must continuously be evaluated using real-world data and performance be reassessed whenever circumstances or data sets change. Thorough documentation of model logic, limitations, and performance is required to meet regulatory expectations.

"AI adoption accelerates across company functions, PV teams find themselves at an inflection point. This is a pivotal moment where the complexity and interconnectedness of AI systems require PV to remain vigilant and responsive, ensuring that patient safety is not compromised by innovation."

So, as non-PV functions deploy AI tools which may gather or surface AE information, PV teams must maintain proactive oversight. PV needs to assess whether the AI’s capabilities could reasonably result in the collection, inference, or transformation of data that meets AE reportability criteria. PV must demonstrate retained responsibility for key activities such as AE identification, validity assessment, timely reporting, and inspection readiness.

To effectively maintain control, PV must partner across all company functions to ensure they receive early notification of any AI system that interacts with patients, healthcare professionals, or consumers. PV must also be informed of any AI system that handles unstructured data such as text, voice, chat, emails, social media, and documents. For each system, PV should systematically document what types of data are processed and which populations are involved.

It is important to understand whether AI can extract or summarize medical events, if it generates new content and whether any human review happens before actions are taken and information is stored.

Further, to address the requirement to establish risk frameworks for AI oversight, PV should focus on classifying AI systems according to their level of exposure to AE information. This may range from no exposure to passive exposure, to active detection of AE information.

A documented risk-based approach clearly identifying when a human must remain in the loop will enable PV to tailor governance and escalation pathways ensuring the organization remains inspection-ready and compliant.

In summary, as AI adoption accelerates across company functions, PV teams find themselves at an inflection point. This is a pivotal moment where the complexity and interconnectedness of AI systems require PV to remain vigilant and responsive, ensuring that patient safety is not compromised by innovation. Fortunately, with regulators now offering specific guidance and frameworks, PV teams have clearer direction on how to evaluate, validate, and govern AI systems. This shift in regulatory guidance empowers PV to address AI challenges with greater confidence and consistency, while prioritizing patient protection.