Artificial intelligence in the pharmaceutical sector has moved beyond the phase of generic enthusiasm. It is no longer merely a conference promise, nor a scenario-planning exercise entrusted to innovation departments. It is entering the processes that support the medicinal product lifecycle: preclinical development, clinical trials, regulatory documentation, manufacturing, pharmacovigilance, real-world evidence, quality control and post-marketing surveillance.
This is the most relevant finding emerging from the 2025 AI Observatory report published by EMA and HMA. The document is not an operational guide and does not claim to provide a systematic mapping of all AI applications in the sector. It does, however, offer a significant snapshot of what is happening within and around the European regulatory network: AI has moved “from concept to practice”, with applications being discussed both by companies with regulators and by the authorities themselves to increase efficiency, analytical capacity and the quality of decision-making processes.
The new question: not whether to use AI, but how to govern it
For years, the debate focused on the simplest question: can artificial intelligence be useful in pharma? Today, that question appears to have been overtaken. The issue is no longer to establish whether AI will have a role, but to understand how to make it acceptable, controllable and documentable in a sector where every decision must be explainable, traceable and linked to clear accountability.
The EMA AI Observatory 2025 report places this transformation within a much broader regulatory framework. The European AI Act entered into force on 1 August 2024 and its application is proceeding in phases: from 2 February 2025, rules on prohibited practices and AI literacy obligations have applied, among others, while from 2 August 2025 the rules on general-purpose AI models have come into application. Full application of the Regulation is expected from 2 August 2026, with some exceptions and specific transitional periods.
For the pharmaceutical sector, this means that AI cannot be considered only as a technology to be adopted, but must be understood as a new object of governance. It is not enough to choose a tool, train a model or automate an activity. Companies need to establish who is responsible for its use, which data enter the system, how outputs are verified, which limits are acceptable, and how errors, bias, updates, cybersecurity and human oversight are managed.
This is a profound shift in perspective: AI is not entering pharma as a shortcut, but as a new layer of complexity to be integrated into quality systems, regulatory processes and professional culture.
Where AI is already entering the medicinal product lifecycle
One of the most interesting aspects of the report is the variety of areas in which AI applications have already been discussed with companies in the context of regulatory interactions. These areas do not concern only early research or hypothesis generation, but span the entire medicinal product lifecycle.
In the preclinical phase, AI is being explored to identify drug candidates, support drug discovery and generate evidence. In clinical development, applications are emerging in patient selection, outcome prediction, medical imaging, digital endpoints, patient-reported measures, clinical event assessment, dosing and even in silico trials. In manufacturing, the report refers to pharmaceutical process models, digital twins, process digitalisation and cell analytics. Across functions, generative AI is being explored to support the drafting of technical and regulatory documentation or the generation of responses to requests from authorities.
This map says a great deal. AI is not confined to computational research, nor can it be left solely to data science specialists. It touches different functions, each with its own language and responsibilities: clinical development, regulatory affairs, quality, manufacturing, safety, medical writing, CMC, market access, IT, legal and compliance.
For companies, the risk is that each function moves independently, adopting local solutions that may be effective in the short term but become difficult to defend when they enter a dossier, an audit or a discussion with the authority. The issue is not to slow innovation, but to prevent AI from becoming a collection of isolated experiments without a common framework.
Regulators are using AI: a shift that should not be underestimated
The report does not look only at AI used by industry. A significant section concerns the adoption of AI tools by the European regulatory network. EMA identifies applications mainly focused on knowledge mining, personal productivity, process and system automation, document analysis, validation and quality assurance.
This is a crucial step. If regulators begin using AI tools to navigate large volumes of documents, retrieve regulatory precedents, support analytical activities and improve internal efficiency, the way industry thinks about its own documentation also changes.
This does not mean that dossiers should be “written for machines”. Rather, it means that consistency, structure, traceability, data quality and clarity of argument become even more important. A dossier that is poorly structured, redundant, inconsistent or built on weak documentary logic risks becoming more fragile in a context where authorities also have more powerful tools to compare, query and analyse information.
EMA has also extended the functionalities of Scientific Explorer, an AI-enabled knowledge mining tool, to support EMA and national competent authorities in searching for information related to initial marketing authorisation applications, including public assessment reports. The extension has been available since March 2026 and allows assessors to search regulatory precedents by medicinal product name, disease area or relevant keywords.
For industry, this points to a clear direction: pharmaceutical regulation is becoming increasingly data-driven, comparative and knowledge-based. Content quality remains central, but the environment in which that content is read, verified and discussed is changing.
Generative AI and the regulatory documentation challenge
Among the most sensitive applications is the use of generative AI in the drafting of technical and regulatory documentation. It is one of the most promising areas, because documentary production in pharma is vast, repetitive in some parts, highly structured and often subject to tight timelines. But it is also one of the riskiest.
A generative model can help summarise, compare versions, prepare drafts, extract information, produce preliminary responses or accelerate medical and regulatory writing activities. However, it cannot replace expert control. The issue is not only the possible “hallucination” of the model, but also the possibility of introducing imprecise nuances, unproven causal links, incorrect references or formulations that appear plausible but are weak from a regulatory perspective.
EMA and HMA have already published principles for the safe and responsible use of large language models by the regulatory network, highlighting aspects such as the security of input data, critical thinking, output verification, continuous training and the need to know whom to contact in case of doubts.
It is reasonable to assume that similar principles should also become part of companies’ operational practice. It is not enough to state that a text “has been reviewed by a human”. Companies need to define what review means, who performs it, with which competencies, on which version, according to which procedure and with what level of traceability.
AI, manufacturing and quality: the most sensitive ground
In pharmaceutical manufacturing, artificial intelligence opens up very concrete opportunities: predictive process models, digital twins, trend analysis, predictive maintenance, advanced control, early detection of deviations, optimisation of manufacturing parameters and quality support.
But this is precisely where a characteristic tension of the sector emerges. Pharmaceutical manufacturing cannot adopt AI with the same logic used to introduce a business intelligence tool. If a model supports a GMP-relevant decision, it inevitably enters the perimeter of validation, change control, data integrity, cybersecurity, audit trails, deviation management and decision-making accountability.
The fact that the EMA report identifies manufacturing as one of the areas where further reflections and guidance activities are being prepared is particularly important. AI applied to the factory is not only about production efficiency: it concerns the ability to demonstrate that a process remains under control even when certain analytical or predictive activities are supported by complex, probabilistic and potentially adaptive systems.
A decisive part of industrial maturity will be shaped here. Companies that are able to connect AI, process understanding, quality risk management and lifecycle validation will gain not only a technological advantage, but also a regulatory one.
Ten common EMA-FDA principles: towards international convergence
Another element that deserves careful attention is international convergence. EMA and FDA have identified ten principles of good AI practice across the medicinal product lifecycle, with the aim of guiding the use of AI in evidence generation and monitoring, from early research to clinical trials, manufacturing and safety.
This is an important signal because AI, by its nature, goes beyond national borders. Models can be developed in one country, trained on international datasets, integrated into global platforms and used in dossiers intended for multiple authorities. Without convergence at least on basic principles, the risk would be regulatory fragmentation that is difficult to manage.
For global companies, this means that AI governance cannot be built only to respond to a single jurisdiction. It must be designed in a way that is consistent, documentable and adaptable to different regulatory contexts. Local compliance remains necessary, but the credibility of the approach will increasingly depend on the solidity of the overall system.
Human expertise becomes more important, not less
One of the most common misconceptions about AI is that it can reduce the importance of expertise. In pharma, the opposite is true. The more capable the tools become, the greater the need for professionals who can ask the right questions, assess outputs, interpret uncertainty, recognise subtle errors and translate technical results into regulatory or industrial decisions.
The AI literacy required by the AI Act should therefore not be interpreted as a generic training obligation to be fulfilled through an introductory course. In the pharmaceutical sector, it should become a distributed and differentiated capability: different for those working in regulatory affairs, quality, manufacturing, pharmacovigilance, clinical development or medical writing.
The question is not whether everyone should become a data scientist. The question is whether each function understands AI well enough to use it without being driven by it. This means knowing its limits, risks, conditions of use, documentary implications and responsibilities.
The message for industry: build the operating model now
The EMA AI Observatory report does not impose a new procedure on companies. But it sends a very clear message: AI is entering the ordinary language of pharmaceutical regulation. Those who continue to treat it as a collection of local experiments risk being unprepared when those experiments become part of a dossier, a clinical strategy, a manufacturing decision or an inspection.
For industry, the necessary step is to build a real operating model for AI. Not only general policies, but practical criteria to classify use cases, assess risk, define roles and responsibilities, control data, verify outputs, document decisions, train people and integrate all of this into existing quality systems.
AI does not replace the pillars of pharmaceutical regulation. It puts them to the test. And precisely for this reason, it can become a powerful tool: not when it promises to eliminate complexity, but when it helps to read, organise and govern it more effectively.
The future of AI in pharma will not be decided only by the quality of algorithms. It will be decided by the quality of the questions, data, processes and responsibilities that companies are able to build around those algorithms.
And perhaps this is the most concrete message of the EMA report: artificial intelligence is now part of the medicinal product lifecycle. It must now become a mature part of the way the pharmaceutical sector generates evidence, makes decisions and preserves trust.
