Rethinking the Architecture of Adverse Event Processing

By: Lucinda Smith, Chief Safety Product Officer, ArisGlobal

Contract service providers handling adverse event case processing for pharmaceutical sponsors are well placed to understand how the underlying workflow has resisted fundamental change. Over more than two decades of frontline pharmacovigilance (PV) — before moving to the technology side — I watched the industry digitize the paperwork, accelerate individual tasks, and expand the capacity of outsourced teams. The architecture itself was a constant: sequential, dependent on human review at every stage, and built around periodic data snapshots rather than continuous monitoring. But what happens when that architecture is no longer the only available option?

AI’s potential in PV to date has been framed mostly around operational gains — e.g. faster MedDRA coding, streamlined intake, reduced clerical errors. While these improvements matter, they address symptoms of the underlying constraint rather than the constraint itself. The more consequential opportunity lies in making signal management genuinely autonomous at the processing stage: deploying AI agents to handle data collection, coding, medical review and quality checking so that structured, validated case data is available to signal detection teams within hours of receipt rather than days later. That shift changes everythingdownstream.

A Regulatory Direction With No Operational Counterpart

Regulators have been explicit about where they want PV to go. The European Medicines Agency has stated for some years that, by 2030, PV for key new medicines should support real-time regulatory decision-making — transforming the discipline from a reactive data collection activity into proactive, continuous safety monitoring.¹ GVP Module IX on signal management reinforces this, identifying early detection and prompt evaluation as central objectives.² The FDA’s Sentinel Initiative is working toward comparable goals through active surveillance infrastructure.³

Yet the dominant operational discussion in the drug safety space remains focused on process optimization within the existing sequential model. The ambition articulated by regulators will not be delivered by optimization. Closing the distance between intent and reality requires a structurally different approach to how cases are handled from the moment of receipt.

If cases could be processed automatically — extracted, coded, medically reviewed, quality-reviewed and entered into a safety database within hours — signal detection teams would no longer work from snapshots. They would have access to real-time, structured data as a matter of course. The monthly line listing cycle would become redundant. The seven-day Suspected Unexpected Serious Adverse Reaction (SUSAR) reporting timeline,⁴ calibrated to the realities of manual processing when first established, would begin to look more like a historical baseline than a performance target. Near-instantaneous case processing would make reporting within one or two days entirely feasible, with direct benefits for the speed of patient protection.

The Data Environment Has Outpaced the Workflow

Greater data availability has not produced sharper signal intelligence. Across the past decade, the proliferation of sources has made the extraction of meaningful safety insights harder, not easier. Internal datasets — clinical trial records, exposure data, safety databases — are supplemented by real-world evidence, electronic health records, published literature and regulatory repositories such as FAERS and EudraVigilance. Sequential, manual workflows were not designed to cross-analyze such a range of sources rapidly or coherently. The result is more noise rather than better signals.

ICSR duplication compounds the problem. Research conducted by TransCelerate BioPharma across seven major pharmaceutical companies found a mean of three submissions per case version across 2.5 million case versions, with a meaningful proportion reaching ten or more health authority recipients.⁵ External factors add further distortion: heightened media attention on a drug class, or public concern about a newly authorized product, can generate reporting surges with no underlying change in product risk. The documented spike in GLP-1 receptor agonist adverse event reports as public and media interest in that drug class grew is an illustrative recent example. Genuine signals risk being buried in volumes that no human-curated workflow can filter efficiently at scale.

AI-powered signal evaluation offers a workable route through this, not by replacing expert analysis but by making it substantially more effective. Agentic AI is well suited to this because agents can observe, reason and recommend — performing the data collection, analysis and cross-triangulation of multiple sources, then surfacing cases most likely to represent true safety signals and directing expert attention accordingly. The safety scientist’s job then becomes one of assessing those recommendations and making decisions, rather than conducting the search from scratch.

Reconsidering What Human Involvement Is Actually For

Anxiety about reducing human involvement in case processing tends to assume that current manual performance is consistent, reliable, and essentially irreplaceable. That assumption does not hold up especially well. Routine inspection findings relating to ICSR quality exist precisely because human-dependent workflows are difficult to standardize at scale. A quality control process that samples five or ten percent of case volume is not a guarantee of systematic quality, and the effects of fatigue and interpretive inconsistency become more pronounced as processing volumes rise.

An automated system with continuous, full-coverage quality monitoring — with a governance layer that detects drift and enables systematic updating — offers a tangible improvement on this. Crucially, when quality problems do arise, teams can more reliably instruct agentic agents to reference updated materials than redirect hundreds of case processors across multiple sites. AI-powered processing is also inherently more auditable: every decision can be logged and every deviation rendered traceable in ways that manual processes cannot replicate.

The industry has navigated comparable transitions before. When PV functions began outsourcing case processing to contract organizations in the late 1990s and early 2000s, the concerns raised closely mirror those being voiced today about AI: how would sponsors maintain oversight? How would quality be assured? Those questions were resolved, governance frameworks were established, and that transition is now so thoroughly embedded that these concerns barely feature in the conversation.⁶ Replacing outsourced human processing with AI-managed processing is structurally the same journey.

The objective being described here is not autonomous PV, but rather augmented PV: AI handling the high-volume, labor-intensive processing tasks to which it is well suited, with human expertise reserved for interpretation, regulatory dialogue, benefit-risk assessment and communication with healthcare professionals. For CROs and contract service providers, this distinction matters operationally — it defines what their teams will actually be doing, and where the value of that expertise is concentrated.

What the Next Few Years Could Reasonably Deliver

The technology components to enable a step-change in PV workflow are now in active development or already available. Automated ICSR processing pipelines, agentic AI coding tools, continuous quality review layers, cross-domain data integration platforms, and AI-powered signal evaluation capabilities together represent a substantial and increasingly deployable capability set. What is needed alongside the technology is organizational readiness: the willingness to deploy these tools at scale, underpinned by governance frameworks robust enough to withstand regulatory scrutiny of the outputs.

The case is particularly acute for products approved after exposure in relatively small patient populations. Products reaching market having been studied in only a few hundred patients — increasingly common in areas of significant unmet medical need — present a demanding signal management challenge. The safety database at launch may be limited, making early post-authorization signal detection correspondingly more critical. Faster case processing means earlier access to emerging signals, and earlier intervention where intervention is warranted.

Organizations that commit to this transition could, within a couple of years, be operating with near-instantaneous case processing as a baseline. That step change is not a prerequisite for improved signal management — much can be done today with a signals agent to accelerate the workflow — but it means signal detection will by then be running on real-time data as a matter of course. Safety scientists would spend the majority of their time on analysis and decision-making rather than data management. The seven-day SUSAR timeline — currently a regulatory floor rather than an ambition — would become a ceiling that most organizations could operate comfortably beneath.

EMA’s 2030 vision for PV sets a clear expectation around real-time decision-making. Delivering against that expectation will require infrastructure investment and organizational commitment, not further optimization of the existing model. For contract service providers positioned at the center of the industry’s adverse event processing workflows, that represents a significant strategic inflection point — and a concrete opportunity to lead rather than follow.

References

1. How will pharmacovigilance look in 2030? European Medicines Agency, March 2020. https://www.ema.europa.eu/en/news/how-will-pharmacovigilance-look-2030
2. Guideline on Good Pharmacovigilance Practices (GVP) Module IX — Signal Management (Rev. 1). European Medicines Agency, November 2017. https://www.ema.europa.eu/en/human-regulatory-overview/post-authorisation/pharmacovigilance-post-authorisation/signal-management
3. Leveraging real-world data for safety signal detection and risk management in pre- and post-market settings. Frontiers in Drug Safety and Regulation, September 2025. https://www.frontiersin.org/journals/drug-safety-and-regulation/articles/10.3389/fdsfr.2025.1626822/full
4. ICH Guideline E2A: Clinical Safety Data Management — Definitions and Standards for Expedited Reporting. International Council for Harmonisation; in force since 1994. https://www.ich.org/page/safety-guidelines
5. Individual Case Safety Report Replication: An Analysis of Case Reporting Transmission Networks. Drug Safety, January 2023. https://pmc.ncbi.nlm.nih.gov/articles/PMC9870831/
6. Might We Come Together on a Paradigm Shift to Manage ICSRs with a Decentralized Data Model? Drug Safety, April 2025. https://doi.org/10.1007/s40264-025-01539-4

About the Author

Lucinda Smith is ArisGlobal’s chief safety product officer. Before joining ArisGlobal, she spent more than two decades working in frontline scientific and strategic pharmacovigilance and drug safety roles at a major pharma brand.