If you run a pharmacovigilance team, you know the equation well: the volume of Individual Case Safety Reports (ICSRs) grows every year, EMA deadlines are non-negotiable (15 days for serious unexpected adverse reactions, 90 days for periodic ones), and the team has the same size as three years ago.
The result is predictable: chronic overtime, high turnover in specialist profiles that are hard to replace, and the constant pressure of knowing that any deadline breach triggers a regulatory authority procedure.
At Delbion we have over 15 years of experience in IT and cybersecurity processes applied to regulated environments. For the past two years we have been implementing AI agents in pharmaceutical companies that need to scale their pharmacovigilance capacity without scaling their headcount - and without sacrificing the integrity of the regulatory process.
This article explains what part of pharmacovigilance can be automated today, how it works in practice, and what security and compliance guarantees are essential.
Why traditional pharmacovigilance does not scale
Pharmacovigilance is a high-volume, high-precision, high-regulatory-consequence process. The three factors together create a bottleneck that is hard to resolve with traditional approaches.
The volume of ICSRs has grown steadily over the past decade, driven by pharmaceutical market growth, the increase in active pharmacovigilance and greater patient awareness around reporting adverse reactions. At the same time, each ICSR requires medical review and MedDRA coding before entry into EudraVigilance - a process that can take between 30 minutes and several hours per case, depending on complexity.
A mid-size pharmaceutical group with 20-30 products on the European market can receive between 5,000 and 15,000 ICSRs annually. If each case requires an average of 45 minutes of manual work, that is 3,750 to 11,250 hours of specialist pharmacist time - just for case processing, before starting on PSURs and signal monitoring.
What AI agents automate in pharmacovigilance
AI agents do not replace the pharmacist responsible for the causality decision - that remains a human responsibility and should stay that way. What they automate is the extraction, structuring, coding and tracking of the data the pharmacist needs to make that decision.
1. Initial ICSR processing
The agent receives ICSRs from their various input sources (web forms, HCP emails, scientific literature, clinical trial records) and performs initial triage: extracts structured data (suspected product, adverse reaction, patient data, outcome), detects whether the case is serious or non-serious, and identifies the applicable reporting requirements under the EMA timeline. The pharmacist receives a pre-structured case, not a free-text field to interpret.
2. Automatic MedDRA coding
MedDRA coding is one of the most time-intensive processes and the most prone to inconsistencies when done by a large team. The agent automatically codes adverse reaction terms in the correct MedDRA hierarchy (LLT, PT, HLT, HLGT, SOC) with a precision rate above 88% in our implementations. The pharmacist validates and adjusts if needed - rather than coding from scratch.
A laboratory we work with was manually processing between 400 and 600 ICSRs monthly with a team of 6 people. Month-end workload peaks to meet EMA deadlines were generating systematic overtime. After implementing the processing agent, the same team manages the same volume without overtime - and dedicates the freed time to reviewing emerging signals that previously sat in the backlog.
3. Continuous safety signal monitoring
Signal monitoring - identifying emerging risk patterns from multiple data sources - is the most strategic part of pharmacovigilance and, paradoxically, the one that receives the fewest resources because the team is busy processing individual cases. The agent continuously monitors EudraVigilance, scientific literature in PubMed and Embase, specialist social media and patient forums, and generates alerts when it detects signals requiring evaluation. This goes from being a monthly task to a real-time process.
4. PSUR and RMP preparation
Periodic Safety Update Reports (PSURs) are the most time-consuming documents in the pharmacovigilance cycle: they require aggregating and analysing all safety activity for a period, assessing the benefit-risk balance, and preparing a structured document to GVP Module VII. The agent generates the PSUR draft from the period data: exposure statistics, ICSR analysis by reaction type, review of identified signals, and benefit-risk balance update. The specialist pharmacist reviews, adds the clinical judgement analysis, and signs.
Free Assessment
How many ICSRs does your team process manually each month?
In 60 minutes we calculate the automation potential in your pharmacovigilance process: automatable cases, hour reduction, and EMA compliance guarantees for your specific portfolio.
Request Free Assessment →Results in numbers: what changes
| Pharmacovigilance activity | Manual process | With AI agent | Saving |
|---|---|---|---|
| ICSR processing (extraction + structuring) | 25-40 min/case | 5-8 min (review) | ~75% |
| MedDRA coding per case | 15-30 min/case | 2-3 min (validation) | ~85% |
| Seriousness triage and classification | 10-15 min/case | Automatic (immediate alert) | ~90% |
| Signal monitoring (external sources) | Weekly/monthly | Continuous (real-time alerts) | Preventive |
| PSUR draft | 3-5 weeks | 1-2 weeks (review and sign) | ~60% |
| Causality assessment (final decision) | Pharmacist | Pharmacist (cannot be automated) | N/A |
EMA and GVP compliance: what the agent must respect
An AI agent in pharmacovigilance must be designed with the regulatory framework as a technical constraint, not an afterthought. This means every part of the automated process must be traceable, auditable and reproducible before a regulatory authority inspection.
Full traceability of every agent decision
The agent records every action: what data it processed, what coding it proposed, what signals it identified and when. If a regulatory authority asks why a particular ICSR was classified in a specific way, there is an auditable log to justify it. The agent does not make decisions in a black box - every output has a recorded reasoning chain.
Human-in-the-loop at every regulatorily significant decision
The agent automates the process, but does not sign the ICSR or the PSUR. Validation and signature by a qualified pharmacist is a non-negotiable requirement under the GVP guidelines - and the agent design reflects this. The workflow requires human approval before any submission to EudraVigilance or any official notification.
Clinical and personal data security
ICSRs contain identifiable patient health data - the most sensitive category under GDPR. With 15 years of cybersecurity experience in regulated sectors, at Delbion we design agents with anonymisation at the processing layer, encryption of data in transit and at rest, and granular access controls. Data never leaves the European environment without appropriate contractual safeguards.
Periodic agent performance validation
The agent is not a static system. MedDRA terminology updates, products evolve, and the regulatory context changes. We include a quarterly performance validation process - reviewing coding precision rates, the quality of signals identified, and the correct application of seriousness classification criteria.
The EU AI Regulation classifies systems that assist in the evaluation of medicinal product safety risks as high-risk applications. This is not an impediment to using them - it is a framework of transparency, traceability and human oversight requirements that must be met. At Delbion we design pharmacovigilance agents to comply with the AI Act from the outset, with all the technical documentation needed for an eventual regulatory inspection.
The implementation process: from zero to production
Implementing a pharmacovigilance agent requires more care than in other sectors because errors have direct regulatory consequences. Our approach starts in parallel with the existing process - the agent learns from historical data before touching new cases.
Weeks 1-2: Assessment and agent design
Analysis of your product portfolio, ICSR volume and typology, case input sources, and pharmacovigilance management system (Argus, ARISg, Ennov, or whichever you use). Definition of applicable GVP modules, products in scope, and EudraVigilance integration requirements. Identification of the case types with the greatest automation potential in your specific context.
Weeks 3-5: Training with historical data
The agent trains on resolved historical ICSRs - cases where the correct output is known. This allows validation of MedDRA coding quality, seriousness classification and signal identification before the agent touches new cases. In this phase we measure and adjust precision rates for your specific portfolio.
Weeks 6-8: Parallel production run
The agent goes live but the manual process continues in parallel for 2-3 weeks. We compare agent outputs with those of the human team, identify discrepancies and adjust. At the end of this phase, the team has confidence in the agent and the results have documented cross-validation - useful for an eventual inspection.
Week 9+: Handover and ongoing support
The manual process is retired and the agent takes over the full processing flow, with the pharmacist in the role of reviewer and validator rather than processor. Delivery of the agent technical documentation (required for the pharmaceutical quality system and GxP inspections). Start of monthly support with periodic performance validations.
What pharmacovigilance directors ask
Will an EMA inspection accept cases processed by AI?
Yes, provided the GVP requirements are met: process traceability, human validation of the final decision, and documentation of the system used. European regulatory agencies have published reflections on the use of AI in pharmacovigilance that explicitly acknowledge its potential - with the appropriate safeguards. The agent we implement generates all the documentation needed to demonstrate the process meets GVP standards.
What if the agent incorrectly classifies a serious case?
The agent is designed to flag cases with high uncertainty as "priority review required" and escalate them to the pharmacist before definitive processing. Seriousness classification is one of the points where the human-in-the-loop is most critical - and the workflow reflects this. In our implementations, serious cases are always reviewed by a pharmacist before any notification.
Can the agent handle ICSRs in multiple languages?
Yes. ICSRs arrive in multiple languages, especially if your company operates in several European markets. The agent processes cases in the main European languages (English, Spanish, French, German, Italian, Portuguese) with integrated translation and structured data extraction. MedDRA coding is always done in English, the standard language of EudraVigilance.
The pharmacovigilance directors we work with describe the change like this: their teams go from being case processors to being signal analysts. The high-value work - interpreting patterns, evaluating emerging risks, preparing the scientific argument for the benefit-risk balance - was what professionals wanted to do but did not have time for. Now they have time.
Next step
If your pharmacovigilance team is managing current volumes with overtime or a growing backlog, the first step is a 60-minute Assessment with our team. We analyse your ICSR volume, your product portfolio, your current pharmacovigilance system, and show you exactly what part of the process is automatable and what impact it would have on hours and deadline compliance.
No commitment, no generic sales pitch. Just specific analysis for your organisation and your portfolio.