The European Medicines Agency has published a detailed update showing that artificial intelligence is moving from experimental pharmacovigilance projects into operational regulatory workflows. The presentation, published on 30 October 2026, describes deployed AI tools, measurable pilot results, a network-wide framework for sharing AI tools and a five-year European project on AI-powered signal detection.
The update was presented during the 21st EMA Industry Stakeholder Platform on the operation of EU pharmacovigilance legislation, held on 20 October 2026. EMA published the meeting highlights on 3 November 2026.
The development is important because it provides evidence of how a medicines regulator is beginning to use AI in real pharmacovigilance operations rather than discussing AI only at the level of principles or future guidance.
Why Pharmacovigilance Is a Strong Candidate for AI
EMA identifies several characteristics that make pharmacovigilance potentially suitable for artificial intelligence:
- large and continuously growing volumes of safety data;
- repetitive and standardised workflows;
- extensive use of narratives and scientific literature;
- established terminologies such as MedDRA;
- existing human quality-control and review stages.
Potential AI applications exist across the complete safety workflow:
- case intake;
- duplicate detection;
- literature screening for individual case safety reports;
- generation of case-series summaries;
- literature screening for emerging signals;
- case adjudication;
- signal prioritisation;
- analysis of historical safety assessments;
- drafting of assessment reports.
However, EMA also identifies several controls that remain essential:
- human oversight and accountability;
- validation against curated case sets;
- particular attention to false negatives;
- auditability and inspection transparency;
- protection of patient information.
The message is therefore not that pharmacovigilance should be automated without restriction.
It is that pharmacovigilance may be suitable for carefully controlled human-AI workflows.
EurEKA Structures Adverse-Reaction Information from Product Information
One of the deployed tools is:
EurEKA – EU SmPC Entity Extraction and Knowledge Acquisition.
EurEKA extracts adverse-reaction information from Summaries of Product Characteristics and converts it into searchable structured information.
EMA reports that the tool has mapped approximately:
155,000 adverse drug reactions across approximately 1,500 products.
The system combines regular-expression logic with a pre-trained machine-learning model and LLM-enabled functionality.
Its intended benefits include:
- faster checking of whether an adverse reaction is already listed;
- structured comparison of adverse-reaction information across products;
- improved access to product-information data during signal evaluation.
This illustrates an important use of AI in regulated environments.
The value does not come only from generating text.
It comes from converting large volumes of controlled but unstructured regulatory information into structured, searchable knowledge.
ERATO Reduces Manual Literature Screening
The second deployed tool is:
ERATO – Enhanced Review of Abstracts with Transformer Models.
ERATO uses a Generative AI / Large Language Model approach to screen scientific literature for possible emerging safety signals.
EMA reports an approximate:
70% reduction in literature abstracts requiring manual review.
This does not mean that the AI independently determines whether a valid safety signal exists.
Instead, the tool reduces the number of abstracts that human assessors need to review manually.
The operating model can therefore be expressed as:
Scientific Literature
↓
AI Screening and Prioritisation
↓
Human Pharmacovigilance Review
↓
Signal Assessment
This is a useful example of risk-proportionate automation.
AI reduces repetitive workload while the consequential regulatory decision remains with qualified assessors.
False Negatives May Be More Important Than Average Accuracy
Literature screening also demonstrates why a single overall accuracy result may be insufficient.
A false positive creates additional human work.
A false negative may result in relevant safety information not being presented for review.
The two errors therefore have different consequences.
For pharmacovigilance AI, performance evaluation should consider:
- sensitivity and recall;
- false-negative rate;
- types of safety information most frequently missed;
- performance for rare or unusual events;
- human correction and override rates;
- performance after changes in literature or terminology.
This is a broader lesson for pharmaceutical AI:
performance criteria should reflect the consequence of the error, not only aggregate model accuracy.
Several Other AI Applications Remain Experimental
EMA’s pharmacovigilance workflow also includes several AI applications that remain in experimentation.
These include:
- literature screening for individual case safety reports;
- duplicate detection in EudraVigilance;
- AI-enhanced case adjudication;
- generation of insights from historical signal reviews.
This distinction between deployed and experimental tools is important.
It demonstrates that EMA is not treating every technically promising AI use case as ready for routine regulatory operation.
Different applications may require different levels of:
- validation;
- human oversight;
- data-quality assurance;
- operational monitoring;
- regulatory confidence.
EMA Is Creating a Catalogue for Sharing AI Tools Across the Regulatory Network
EMA also described a framework for sharing AI tools across the European Medicines Regulatory Network.
The framework is intended to:
- support collaboration and reuse of AI tools;
- provide guidance for collecting and maintaining information about AI tools;
- provide transparent and trustworthy information supporting adoption;
- reduce duplication across regulatory organisations.
The planned output is an AI-tool catalogue containing information required to understand and potentially reuse a tool.
Such a catalogue could include information on:
- intended purpose;
- development history;
- performance;
- support and maintenance;
- limitations;
- availability to other regulatory authorities.
This is an important governance development.
Reusing an AI tool safely requires more than transferring software.
The receiving organisation needs sufficient information to understand:
what the tool does → how it was evaluated → where it may fail → what operational support it requires.
AI Tool Reuse Requires Controlled Context Transfer
A tool that performs acceptably in one regulatory organisation may not automatically perform equally well in another.
Differences may include:
- data sources;
- languages;
- case populations;
- regulatory workflows;
- technical infrastructure;
- user expertise.
A network catalogue should therefore support informed assessment rather than automatic acceptance.
Before reuse, the receiving authority may need to confirm that:
- the intended use remains appropriate;
- local data are sufficiently comparable;
- performance remains acceptable;
- human-review arrangements are defined;
- local change and incident management are available.
EMA Joins a Five-Year AI Signal-Detection Project
EMA has also joined a five-year Innovative Health Initiative public-private project on AI-powered signal detection.
The project is scheduled to begin in:
January 2027.
Participants are expected to include:
- regulators;
- pharmaceutical industry;
- academic organisations;
- international organisations.
Expected project outputs include:
- signal-detection models;
- evaluated data sources;
- risk-prediction models;
- implementation recommendations;
- recommendations for human oversight;
- templates and tools;
- training and educational materials.
This is significant because it combines technical model development with implementation science and governance.
The project is not limited to asking whether AI can detect a signal.
It also aims to address:
how AI signal detection should be implemented, governed and supervised.
Guidance on AI in Pharmacovigilance Is Being Drafted
EMA also confirmed that a dedicated questions-and-answers document on AI in pharmacovigilance is under development.
Its scope was informed by a stakeholder survey involving:
176 respondents from industry, regulatory assessment and inspection communities.
Topics of particular interest include:
- human oversight;
- documentation;
- risk-based control;
- inspection expectations.
EMA’s current timetable indicates:
- scoping in Q3 2026;
- drafting in Q4 2026;
- European regulatory-network consultation in Q1 2027;
- public consultation in Q2 2027.
This will be an important document to monitor because it may translate general AI principles into practical expectations for pharmacovigilance systems.
The Regulatory Direction Is Becoming Clearer
EMA’s update points toward a multi-layered model for AI in pharmacovigilance:
Good AI Practice Principles
↓
Defined Pharmacovigilance Use Case
↓
Curated Evaluation Data
↓
Human Oversight
↓
Auditability and Data Protection
↓
Operational Performance Monitoring
↓
Lifecycle Governance
This is substantially broader than one-time model testing.
Implications for Marketing Authorisation Holders
Pharmaceutical companies should not assume that EMA’s internal tools create immediate new obligations for Marketing Authorisation Holders.
However, the programme provides a strong indication of the areas regulators consider important.
Companies using AI in pharmacovigilance should be prepared to explain:
- the intended use of each AI tool;
- which pharmacovigilance decision it influences;
- which datasets were used for evaluation;
- how false negatives are controlled;
- what human review is required;
- how the activity is documented and audited;
- how patient information is protected;
- how performance and model changes are monitored.
The forthcoming AI-in-pharmacovigilance Q&A may provide more detailed expectations.
Why This Development Matters
The most important aspect of EMA’s update is that it combines:
- deployed tools;
- measured operational outcomes;
- network-level governance;
- future regulatory guidance;
- long-term research collaboration.
AI in pharmacovigilance is therefore moving beyond isolated experimentation.
The emerging regulatory model is:
pilot carefully → measure performance → retain human accountability → build shared governance → develop formal guidance.
This is likely to be highly influential for pharmaceutical companies developing their own AI-enabled safety operations.
Regulatory Status Note
The EMA presentation and meeting highlights describe regulatory-network projects, pilot deployments and planned guidance.
They do not establish new legally binding pharmacovigilance requirements.
The reported pilot metrics should not be interpreted as universal validation thresholds for industry systems.
The dedicated AI-in-pharmacovigilance Q&A remains under development and is expected to enter public consultation in the second quarter of 2027.
Sources
EMA – Highlights from the 21st Industry Stakeholder Platform on the operation of EU pharmacovigilance legislation, published 3 November 2026:
https://www.ema.europa.eu/en/media/75901
EMA – AI in pharmacovigilance update, published 30 October 2026:
https://www.ema.europa.eu/en/media/75842
EMA – 21st Industry Stakeholder Platform on the operation of EU pharmacovigilance: