{"id":35891,"date":"2026-06-23T08:59:00","date_gmt":"2026-06-23T06:59:00","guid":{"rendered":"https:\/\/askme.it\/insights\/ai-pharmacovigilance-the-first-proving-ground-for-the-eu-ai-act-in-pharma\/"},"modified":"2026-05-28T17:55:05","modified_gmt":"2026-05-28T15:55:05","slug":"ai-pharmacovigilance-the-first-proving-ground-for-the-eu-ai-act-in-pharma","status":"publish","type":"insights","link":"https:\/\/askme.it\/en\/insights\/ai-pharmacovigilance-the-first-proving-ground-for-the-eu-ai-act-in-pharma\/","title":{"rendered":"AI pharmacovigilance: the first proving ground for the EU AI Act in pharma"},"content":{"rendered":"
\n

The pharmacovigilance software market was estimated by Precedence Research at USD 2.09 billion in 2025, with a projection of USD 5.06 billion by 2035, a compound growth rate of 9.24%. The broader definition, which includes services and data alongside software, brings the overall value of the pharmacovigilance services market to USD 9.35 billion in 2025, with a projection of USD 16.47 billion by 2030, according to Mordor Intelligence. The market remains structurally concentrated: Oracle Argus holds 19.4% of the PV software share, ArisGlobal LifeSphere 17.2%, for an effective duopoly of around 37%. Veeva Vault Safety, Ennov, IBM Watson, and the legacy of proprietary PV systems from large pharma make up the rest of the market, with increasing consolidation potential in the mid-market.<\/p>\n

Pharmacovigilance is the first pharma use case where the EU AI Act produces concrete operational consequences. Obligations for high-risk AI systems under Annex III apply from August 2, 2026, with an extension to August 2, 2027 for AI systems integrated into products regulated by the Medical Device Regulation or the In Vitro Diagnostic Regulation. The Commission’s classification guidelines are expected in February 2026. The EMA Reflection Paper on AI in the medicinal product lifecycle, finalized in September 2024 and adopted by CHMP and CVMP, explicitly places pharmacovigilance within the regulated perimeter, and requires marketing authorisation holders to carefully monitor AI systems applied to adverse event classification and seriousness assessment.<\/p>\n<\/section>\n

\n

AI in pharmacovigilance workflows<\/h2>\n

AI applied to pharmacovigilance today operates at three distinct levels. The first is the automatic triage of incoming individual case safety reports: the system reads the source \u2014 call center, letter, portal, social \u2014 extracts the relevant entities, maps them to standard ontologies, and routes them to the case processor with a partial completion level. Oracle Argus added automated email intake capabilities in 2024; ArisGlobal LifeSphere integrated LLM-powered triage in 2024-2025, with a record of customer acquisitions in 2024 and global adoption by a major Japanese pharma company announced in 2025. Eversana adopted Argus Cloud in October 2024.<\/p>\n

The second level is MedDRA coding. A 2024 French study published in Drug Safety measured an AUC of approximately 0.97 for transformer-based classification of terms on MedDRA, with seriousness classification accuracy above 80%. PVLens, published in March 2025, reported a recall of 0.983 and a precision of 0.799 for extracting safety information from FDA labels. The numbers are significant but must be read with caution: AUC 0.97 on overall classification does not imply equivalent accuracy on rare terms, where the training dataset is less dense, and seriousness below 80% leaves a margin of error that, in a high-stakes regulatory area, cannot be ignored. The third level is signal detection, where AI works on narrative EHRs, call center transcripts, and social sources to identify emerging signals that traditional statistical methods on spontaneous reporting databases tend to miss or delay.<\/p>\n<\/section>\n

\n

EMA, ICH E2D(R1), and the operational framework<\/h2>\n

The European regulatory framework for PV is structured around three normative pillars that in 2026 enter the phase of operational convergence. ICH E2B(R3), updated for data elements C.1.3 and C.5.4 in EudraVigilance since June 2022, defines the format for the electronic exchange of individual case safety reports. ICH E2D(R1) enters into force in the EU on March 18, 2026 and redefines post-approval safety data management with an updated level of detail on roles and information quality. The EMA Reflection Paper on AI, finalized in September 2024, provides the applicative framework: risk-based approach, model documentation, drift monitoring, alignment with the GVP modules.<\/p>\n

The EU AI Act, from the side of horizontal regulation, classifies as high-risk AI systems used in areas that impact the safety of regulated products or decisions about them. For pharmacovigilance this concretely means: technical documentation accessible to competent authorities, risk management system, quality management of training data, documented human oversight, post-deployment performance monitoring of the model, incident logging. The Sidley Data Matters checklist published in June 2025 \u2014 eight action items for life sciences companies \u2014 has become the de facto reference for pharma compliance officers preparing for the operational transition on August 2, 2026.<\/p>\n<\/section>\n

\n

The Italian constraint: AIFA RNF and the EudraVigilance flow<\/h2>\n

The Italian Pharmacovigilance Network (Rete Nazionale di Farmacovigilanza, RNF), managed by AIFA, has been active since 2001 and, since June 9, 2022, uses an online platform that replaced the previous Vigifarmaco system. Reports of suspected adverse reactions acquired by the RNF are automatically transmitted to EudraVigilance, the European system managed by EMA. AIFA has published quarterly updates of the RAM database since 2002. For Italian pharmaceutical companies, the operational flow is therefore articulated on three levels \u2014 internal corporate, national RNF, European EudraVigilance \u2014 with deadlines, formats, and responsibilities precisely codified by European Good Vigilance Practice.<\/p>\n

The introduction of AI systems into this chain produces a specific constraint: every step in which an AI model makes an automated decision that impacts the flow \u2014 seriousness classification, MedDRA coding, causality assessment, intake triage \u2014 must be documented, validated, monitored for drift, and must have an active human oversight mechanism. The “human-in-the-loop” model is the default required by the EMA Reflection Paper for decisions that directly impact regulatory assessment; the “human-on-the-loop” model is acceptable for support tasks such as initial triage. The design of the Italian PV-AI process in 2026 requires explicitly stating, for each step, which of the two models applies.<\/p>\n<\/section>\n

\n

Automation bias and the deskilling problem<\/h2>\n

The risks documented in the 2024-2025 academic literature \u2014 the review published in PMC in September 2025 is one of the most cited references \u2014 are not technical but behavioral. The first is automation bias: PV analysts tend to trust AI system decisions even when the model has limited confidence, with a risk of under-detection of rare signals on which the training set is underdimensioned. The second is the lack of explainability of LLMs: the decision on a single case is not always reconstructible in a transparent way, and the regulatory requirement of an audit trail becomes more difficult to satisfy. The third is the progressive deskilling of PV operators if validation governance is weak \u2014 young operators entering a workflow where AI does 70% of the work accumulate less decisional experience than their predecessors, and the capacity for manual intervention in non-standard cases erodes.<\/p>\n

The operational response is structured on three points. A periodic validation mechanism of the model on manually coded gold standards, with defined cadence and documented responsibility. Random sampling of automatically processed cases for human verification, with sampling rates that allow statistical detection of drift. A continuous training plan for PV operators that maintains manual competence on cases the system does not handle. For Italian pharmaceutical companies, 2026 is the year in which these three points stop being best practice and become regulatory specification. Pharmacovigilance is not the most visible AI pharma use case, but it is the first in which the regulatory framework is closed and enforcement becomes operational.<\/p>\n<\/section>\n","protected":false},"excerpt":{"rendered":"

Oracle Argus at 19.4% of the market, ArisGlobal LifeSphere at 17.2%. NLP that codes MedDRA with AUC 0.97. The EMA Reflection Paper of September 2024 explicitly places PV within the regulated post-authorization scope. ICH E2D(R1) enters into force in the EU on March 18, 2026. Pharmacovigilance is the first pharma use case classified as high-risk.<\/p>\n","protected":false},"featured_media":35893,"menu_order":0,"template":"","insights_category":[968],"insights_tags":[925,715,923,924,976],"class_list":["post-35891","insights","type-insights","status-publish","has-post-thumbnail","hentry","insights_category-ai-and-pharma","insights_tags-aifa-rnf","insights_tags-eu-ai-act-en","insights_tags-meddra","insights_tags-oracle-argus","insights_tags-pharmacovigilance"],"acf":[],"_links":{"self":[{"href":"https:\/\/askme.it\/en\/wp-json\/wp\/v2\/insights\/35891","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/askme.it\/en\/wp-json\/wp\/v2\/insights"}],"about":[{"href":"https:\/\/askme.it\/en\/wp-json\/wp\/v2\/types\/insights"}],"version-history":[{"count":1,"href":"https:\/\/askme.it\/en\/wp-json\/wp\/v2\/insights\/35891\/revisions"}],"predecessor-version":[{"id":35892,"href":"https:\/\/askme.it\/en\/wp-json\/wp\/v2\/insights\/35891\/revisions\/35892"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/askme.it\/en\/wp-json\/wp\/v2\/media\/35893"}],"wp:attachment":[{"href":"https:\/\/askme.it\/en\/wp-json\/wp\/v2\/media?parent=35891"}],"wp:term":[{"taxonomy":"insights_category","embeddable":true,"href":"https:\/\/askme.it\/en\/wp-json\/wp\/v2\/insights_category?post=35891"},{"taxonomy":"insights_tags","embeddable":true,"href":"https:\/\/askme.it\/en\/wp-json\/wp\/v2\/insights_tags?post=35891"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}