Every pharmacovigilance event this year is positioning AI as the next layer of compliance infrastructure, automating case triage, coding, and increasingly, elements of medical review itself. The pitch is straightforward: faster case cycle times, lower cost per case, and a workflow that scales without scaling headcount.
Yes, automation can process volume. What it can’t do is exercise clinical judgment on a case that doesn’t fit the pattern, defend a causality assessment to a health authority inspector, or carry a program’s institutional history from one reviewer to the next. Those are not workflow steps. They’re physician functions.
Therefore, the question sponsors should be pricing into a PV partnership isn’t which vendor has the most sophisticated AI, but whether physician oversight in that partnership is structured to be the last thing cut when budget pressure hits, not the first.
What the Physician Layer Actually Does That AI Can’t
Strip away the marketing on both sides of the AI-in-PV debate, and there are three categories of judgment that sit with a safety physician, not a model, no matter how advanced the automation upstream is.
Clinical judgment on ambiguous cases
Signal detection tools and NLP-driven coding are good at flagging patterns and populating fields. They aren’t equipped to make an actual causality assessment on a complex or borderline case, because that requires a clinician who understands the product’s safety profile, and not just the data structure of the report.
Take a case with confounding comorbidities and polypharmacy, where every coded field is complete and internally consistent. To an automated system, that looks like a clean case, with nothing missing and nothing conflicting. To a physician reviewing the same case, that completeness is exactly what raises the question of whether there’s an alternative etiology the coding never surfaced, because it wasn’t built to look for one.
At UBC, physician-led case review isn’t an escalation path reserved for outliers. It’s the standard model. The honest answer to “what percentage of cases need a physician” isn’t a number worth chasing. Every case is reviewed under physician-led oversight; the real question is which case you cannot afford to have missed one on. Framed that way, the absence of a precise percentage isn’t a gap in the argument. It is the argument.
Credibility under inspection
When a health authority inspector asks about a causality determination, they’re asking a question that needs a physician’s name and judgment behind it, not a model’s output. AI-assisted workflows still route to a clinical signature of record, no matter how much of the upstream work is automated.
This isn’t a UBC opinion so much as the regulatory architecture already in place:
| WHAT THE REGULATIONS ACTUALLY REQUIRE |
| • FDA’s December 2025 sponsor guidance (finalizing its 2021 draft) recommends a safety assessment committee that includes at least one physician with relevant clinical expertise to evaluate aggregate safety data and advise on IND safety reporting decisions. • EU pharmacovigilance law requires a qualified person responsible for pharmacovigilance with continuous access to medically qualified review, and GVP Module IX frames signal validation as a judgment of clinical relevance, not a statistical threshold. • CIOMS Working Group XIV’s consensus framework on AI in pharmacovigilance puts human oversight at the center of any AI deployment in the space. |
| An inspector’s questions follow that architecture. They end at a clinician’s name. |
We’ve seen this play out directly in the room. Across recent pharmacovigilance inspection support, including an EMA inspection earlier this year, the questions from inspectors followed a consistent pattern: who made the medical decision, on what basis, and did that person have the qualifications and product knowledge to make it. Every one of those questions ends at a clinician’s name, not a dashboard.
UBC anchor: physician-led safety team retention of 88 to 92 percent over four years, with no leadership turnover on the team. Continuity of medical judgment across a program is itself a compliance asset during inspection, not just an operational nicety.
Institutional memory and continuity
Turnover on a safety physician team means re-litigating product history, prior signal decisions, and sponsor-specific nuance every time a new reviewer rotates in. AI doesn’t solve for this. It can only work with what’s documented, and the judgment calls that matter most are rarely fully captured in the case record.
The cost most sponsors don’t price into a vendor comparison is what it actually takes to rebuild physician-level product familiarity after a vendor’s team turns over mid-program. That cost shows up months later, usually in exactly the ambiguous case or inspection moment where it’s most expensive to pay.
Where This Leaves AI: Combination, Not Replacement
The AI-first narrative frames this as substitution: more automation should mean less need for expensive clinical headcount. The more accurate framing is division of labor.
AI is genuinely good at triage, structured data capture, duplicate detection, initial coding suggestions, and surfacing patterns across a volume of cases that no physician reviewing one case at a time would catch. None of that is in dispute.
What stays with the physician is the judgment call automation surfaces but can’t close: medical credibility with regulators and sponsor medical directors, and the accountability that must sit with a named clinician.
CIOMS XIV’s language for this is useful: human-in-the-loop (HITL) versus human-on-the-loop (HOTL). For causality assessment, seriousness and expectedness adjudication on complex cases, and signal validation decisions, UBC operates human-in-the-loop by design. AI supports triage and pattern-surfacing; it doesn’t get a vote on the determination.
The counter to the AI-first pitch isn’t “AI is a risk,” it’s that AI without a stable physician layer underneath it doesn’t close the compliance gap—it just moves it.
Questions Sponsors Should Ask About Medical Oversight in an AI-Augmented PV Program
These questions apply regardless of which vendor you’re evaluating.
- Who signs? For a complex case, who makes the causality and seriousness determination: a named, medically qualified reviewer, or a workflow step? Can the vendor show you that person’s qualifications and their role in the audit trail?
- What happens when the model and the physician disagree? Is there a documented process for physician override of AI-generated assessments, and is the override rate monitored?
- Who will be in the room at inspection? If a health authority inspects this program in year three, will the physicians who made the signal decisions still be on the team to defend them? What is the vendor’s safety physician retention rate?
- What is the AI validated to do? Which PV tasks does it perform, under which oversight model (human-in-the-loop vs. human-on-the-loop), and how is that documented in the PV system master file and vendor qualification?
What This Looks Like in Practice
The differentiator here isn’t a claim about having physicians on staff, because most vendors can say that. It’s retention and continuity, sustained over years, on a team that hasn’t seen leadership turnover.
- 30 regulatory inspections, no critical findings
- 88 to 92 percent safety physician team retention over four years
- No leadership turnover on the physician team across that period
- Physician-led case review as the standard model, not an escalation tier
- A full-service, technology-enabled PV model where AI supports the case team rather than replacing the clinical layer
The Layer That Doesn’t Get Automated Away
Budget pressure hits every PV program eventually. What tends to get cut first is whatever is most visible on a line item. What should be last to cut is the layer holding clinical judgment and regulatory credibility together. In a well-structured program, that is exactly what occurs.
This isn’t a case against AI in pharmacovigilance, but an invitation to ask any vendor, how the two are actually structured together, and not just claiming to coexist.
AI can be a powerful support for physicians, and when appropriately implemented, it can significantly improve productivity by helping teams process information faster, surface relevant patterns, and focus clinical time where it matters most. But that productivity gain does not replace the need for physician-led review and evaluation. In the end, the medical judgment, accountability, and regulatory defensibility of the assessment still need to remain with qualified physicians.
If you’d like to talk through how UBC structures physician oversight alongside its technology stack for your specific program, we’d welcome the conversation.
About UBC
United BioSource LLC (UBC) is the leading provider of evidence development solutions with expertise in uniting evidence and access. UBC helps biopharma mitigate risk, address product hurdles, and demonstrate safety, efficacy, and value under real-world conditions. UBC leads the market in providing integrated, comprehensive clinical, safety, and commercialization services and is uniquely positioned to seamlessly integrate best-in-class services throughout the lifecycle of a product.
About the Author
Alexis Pinçon, MD, Director, Global Medical Services & Medical Writing
Dr. Alexis Pinçon serves as Director, Global Head Medical Services & Medical Writing. In this role, Dr. Pinçon is responsible for supervising the medical aspects of all safety and clinical operations activities at UBC, including medical support to the case processing group and the integrated safety services team. In addition, Dr. Pinçon is also responsible for leading the safety writing and safety physician group.
Following clinical practice as anesthesiologist, Dr. Pinçon earned more than 13 years of industry experience, in drug safety and risk management at Sanofi Pasteur MSD, Sanofi Pasteur and UBC. He has a broad pharmacovigilance experience for a wide range of drugs and vaccines, both in clinical development and post-marketing including medical monitoring activities, signal management activities, aggregate safety report and risk management writing
Alexis Pinçon holds an MD from the University of Besançon (France) and is a specialist in public health. He also obtained a master’s degree in epidemiology, biostatistics, and clinical research.
