EU pharma regulators publish annual AI observatory report

The EU's medicines regulatory network is moving decisively from treating artificial intelligence (AI) as an emerging priority to building it into practical expectations for drug development, according to the recently published second annual AI observatory report by the Network Data Steering Group (NDSG), a joint body of the European Medicines Agency (EMA) and the Heads of Medicines Agencies (HMA).

The 2025 report, which was published in June this year and covers both human and veterinary medicines, monitors AI activities and trends across the European medicines regulatory network (EMRN). Where the 2024 edition mapped AI and its growing and future potential, the 2025 report is framed around a clear transition from adoption to practical implementation. For pharmaceutical and biotech companies, it is a useful guide both to where regulators are directing their attention and to the directions in which EU and international regulatory frameworks are heading.

The policy environment for AI in medicines development became considerably more structured in 2025. The EU AI Act, which entered into force in August 2024, moved into practical application in phases: prohibitions on certain AI uses and AI literacy obligations – which bear specific consequences for life sciences – under articles 4 and 5 of the Act became applicable in February 2025, with obligations for general-purpose AI models and governance structures following in August 2025.

Guidance roadmaps

The most significant new guidance for trial sponsors and marketing authorisation (MA) applicants is the EMA and US Food and Drug Administration (FDA) joint guiding principles on good AI practice in drug development. Agreed in 2025 and published in early 2026, the ten principles cover the use of AI in evidence generation and monitoring across the full medicines development lifecycle. They span all phases from early research and clinical trials through to manufacturing and post-marketing safety monitoring. While not legally binding, the principles represent a common transatlantic reference point and will increasingly inform what EMA and FDA expect from pharmaceutical businesses that rely on AI tools in their submissions.

At the EU level, the EMRN completed a public consultation in 2025 on guidelines for AI in product manufacturing, covering good manufacturing practice (GMP) requirements, predictive models and digital twins. The NDSG 2026-2028 workplan also foresees guidance development on AI in clinical development and AI in pharmacovigilance, as well as a coordinated roadmap for future AI guidance aligned with the proposed European Biotech Act. The Council for International Organisations of Medical Sciences' working group focused on AI in pharmacovigilance also published a specific report on the subject in 2025, contributing to international alignment on how AI tools should be governed in the post-marketing setting.

Broad applications

The report sets out the AI applications that EMA and national competent authorities discussed with industry stakeholders in 2025. 

The early regulatory interaction mechanisms provided by EU pharmaceutical laws are the principal pathway through which competent authorities such as the EMA signal the likely regulatory acceptability of different AI tools. The range of applications discussed in 2025 is notably broad.

Pre-clinical development

In early development, discussions covered AI-assisted drug discovery and evidence generation, including inference of disease-relevant molecular pathways, mechanisms of action, toxicity prediction and biomarker identification. In silico toxicology tools, including those already used to assess mutagenic impurities under the International Council for Harmonisation's M7 guideline, are being coupled with machine-learning algorithms to reduce reliance on animal studies.

Clinical development

In trials and studies, regulators engaged in 2025 on a wide range of applications: clinical outcome prediction, clinical trial and patient selection, medical imaging, endpoint assessment, digital endpoints, patient-reported outcome measures, generative AI-based assistants, safety biomarker identification and AI-based in silico trials. 

A qualification opinion in the medical imaging category was published in 2025. Generative AI was also discussed as a new use case, in this instance to assist with drafting regulatory submissions and technical documentation, and to generate answers to regulatory queries.

GMP and post-marketing

In manufacturing, highlights included predictive stability modelling, pharmaceutical process models and digital twins. These developments sit within the broader public consultation on manufacturing guidance (including Volume 4 of the EudraLex GMP guidelines and their new draft annex 22 on artificial intelligence), with the public consultation closing in October 2025. Post-marketing discussions focused on AI for real-world evidence generation and individual case safety report management, both of which have direct implications for pharmacovigilance operations.

Regulators' use of AI 

The report provides a degree of transparency on how EMA and national competent authorities are deploying AI within their own operations. Regulators are currently using AI primarily for knowledge mining, personal productivity and process automation. The Scientific Explorer tool, for example, now allows EMA staff to search both scientific advice procedures and, since 2025, initial MA applications.

In 2025, the NDSG adopted a Network AI Tools framework and catalogue to support sharing and reuse of AI tools across the EMRN. From 61 AI use cases collected from national competent authorities, the NDSG is developing validated prompt libraries for drafting, summarisation and quality assurance. It has further planned to launch an EMRN prompt community pilot to standardise generative AI practices across the network.

EU-funded research

Several EU-funded initiatives continued or were launched in 2025 across the medicines development lifecycle and applied AI to specific research questions. Research spans in silico toxicology, disease modelling and AI-assisted drug repurposing in the pre-clinical domain; digital pathology, patient stratification, large-scale data integration and clinical decision support in oncology and neurology in the clinical domain; AI-enabled decentralised production and cell-based therapy manufacturing; and AI-powered signal detection for pharmacovigilance.

The report is nonetheless candid about where gaps remain. Tools for regulatory assessment are underdeveloped relative to the pace of AI adoption. These tools include model validation, explainability evaluation, audit frameworks and what the report describes as "regulatory-grade acceptability".

The NDSG responded by adopting a set of Network AI Research Priorities in late 2025, identifying seven domains for further investigation: research integrity and intellectual property; accuracy and reliability of AI tools; data governance, confidentiality and consent; regulation and oversight; ethics, fairness and bias prevention; resources and support for AI use; and impacts on jobs and skills. The European platform for regulatory science research, launched in March 2025, brings together academia and regulators to address these and related questions.

Osborne Clarke comment

The 2025 AI Observatory report captures an EU pharmaceutical regulatory system that has moved past treating AI as a novelty and started building it into day-to-day expectations for drug development. The EMA/FDA guiding principles are a clear expression of that shift: although not legally binding, they are likely to become the reference point against which EMA assessors judge whether AI used in evidence generation or safety monitoring is fit for regulatory purpose. Companies relying on AI anywhere in their development pipeline, not only in clinical trials, should treat the principles as a practical checklist rather than a policy statement to note and move past.

Regulators and pharma are already engaging with AI well beyond early research. In addition to trial site and patient selection, digital endpoints, manufacturing process models and pharmacovigilance signal detection all feature in the discussions EMA and national authorities held with industry in 2025. Some of these applications may fall within the scope of the EU AI Act, either as high-risk AI systems or in lower-risk categories. Companies that assume AI oversight is confined to a single function or development phase could be caught out by how broadly regulators are now looking.

Perhaps the most useful signal in the report is where it admits the system is still incomplete. Validation, explainability and audit standards for AI have not kept pace with adoption, and the NDSG's own research priorities acknowledge as much. For life sciences companies generating real-world evidence through AI or exploring AI-based in silico trials, this is a live opportunity: the guidance that will eventually govern these tools is still being shaped and early, well-documented engagement with competent authorities now is more likely to influence where the bar ultimately lands than reacting to it once it is set.