News

Artificial intelligence is no longer a distant promise but has become a tangible component of drug development. The exponential growth of clinical and operational data has created the conditions for algorithms to identify patterns that are imperceptible through traditional methods, reducing research time and costs. The pharmaceutical industry is among the sectors most actively engaged in this transformation: a Minsait/Indra survey indicates that 55% of the pharmaceutical companies surveyed already apply AI to the design of products and services, compared to 39% across the average of all organizations surveyed. The same study, however, reveals the other side of the equation: 45% of these companies identify the lack of a stable regulatory framework as a barrier to adoption, and only 33% report having even a basic understanding of the AI regulatory environment. The diagnosis is clear: technology is advancing more rapidly than the rules that are expected to govern it.

AI is employed across virtually every stage of drug development, from discovery through pharmacovigilance. Nevertheless, the scale of adoption should not be mistaken for regulatory maturity. The Center for Drug Evaluation and Research (CDER), the FDA division responsible for the evaluation of drugs, has recorded more than 500 submissions involving AI components between 2016 and 2023, spanning multiple stages of development. What has yet to be established is the standard by which these submissions will be evaluated: the guidance intended for this purpose remains in draft form, is expressly non-binding, and excludes from its scope both the drug discovery phase and AI applications aimed solely at operational efficiency that do not affect safety, quality, or the reliability of results. The potential is undeniable; what is still lacking is the regulatory framework capable of translating that potential into legal certainty.

The principal reference in this field is the Draft Guidance issued by the FDA in January 2025, which proposes a risk-based credibility assessment framework. Its underlying logic is straightforward: the greater the impact of the model on regulatory decision-making and the greater the risk associated with error, the more rigorous the demonstration that the model is reliable for its intended context of use must be. The document requires a clear definition of both the problem to be addressed and the role of AI, consolidated into a credibility assessment plan. Two limitations of scope deserve particular attention: the guidance does not cover the use of AI in drug discovery or AI applications intended solely to achieve operational efficiencies that do not affect safety, quality, or the reliability of results.

In Brazil, there is still no formal guideline specifically governing the use of AI in drug development. The absence of a specific regulation, however, does not indicate a lack of interest on the part of the Brazilian Health Regulatory Agency (Anvisa). The Agency has already incorporated AI into its own activities, including the analysis of qualification limits for impurities and degradation products in synthetic medicines through Bot Doc Anvisa, which allows users to consult the content of documents through AI-powered chatbots, and the use of machine learning to monitor irregular products marketed on the internet.

Until a specific regulation is enacted, the applicable legal framework consists of the existing healthcare legislation together with the Brazilian General Data Protection Law (LGPD), which classifies health and genetic data as sensitive personal data and requires a specific legal basis for their processing. It is precisely this general framework that makes international references particularly useful: in the absence of explicit national requirements, the credibility standards developed by the FDA and the European Medicines Agency (EMA) provide the best indication of future regulatory expectations.

On the opportunity side, AI offers significant reductions in research time and costs by enabling earlier decisions to discontinue or advance drug candidates during the initial stages of development. It also enhances the accuracy of predictions regarding efficacy, toxicity, and drug interactions. The use of existing data extends to real-world data and real-world evidence. Finally, there is the competitive advantage enjoyed by organizations that begin now to structure their operations in accordance with internationally recognized standards of credibility and governance. The pharmaceutical sector already leads the adoption of AI in the design of products and services (55%, compared to 39% across all sectors) and in new product research (43%, compared to 22%), yet only 18% of companies mention AI in their strategic plans. It is precisely within this gap between adoption and organizational preparedness that the early implementation of these practices, combined with proactive engagement with healthcare authorities, becomes a means of reducing uncertainty and increasing predictability in investment decisions.

On the risk side, everything begins with the data. A model is only as reliable as the datasets on which it is trained, and incomplete or insufficiently representative data may generate results that healthcare authorities simply refuse to accept. From this stem additional concerns, including algorithms that replicate biases inherent in the underlying datasets, results that are difficult to reproduce, and uncertainty regarding liability when a model produces erroneous outcomes. These concerns are compounded by the protection of health data, which the LGPD classifies as sensitive personal data, and by the absence of specific regulation in Brazil. The most challenging issue involves algorithms that continue learning after regulatory approval: if a product changes autonomously over time, the marketing authorization and post-market surveillance systems—designed for stable products—may no longer be adequate.

The responsible adoption of AI in drug development is less a question of whether it should occur than of how it should occur. The reliable use of this technology requires robust and well-documented data governance capable of ensuring the integrity of the datasets used to train AI models, as well as the development of evidence demonstrating credibility that is proportionate to the risk and intended context of use of each application. These measures should be complemented by continuous dialogue with healthcare authorities to align regulatory expectations and by the integrated involvement of multidisciplinary teams from the earliest stages of project development. Finally, given the pace of regulatory developments both in Brazil and internationally, ongoing monitoring of new regulatory requirements becomes an integral component of any strategic approach. Anticipating these developments transforms regulatory uncertainty into a competitive advantage while promoting legal certainty for both patients and innovators.