A few years ago, the question most regulatory affairs teams asked about artificial intelligence was, "Will the FDA accept this?" Today the question is more specific and more useful: "What evidence do we need to show the FDA that this model is credible for what we are using it for?"
That shift is thanks largely to a draft guidance the FDA published in January 2025, and to the joint principles it has since developed with the European Medicines Agency (EMA). For anyone working in, or moving into, regulatory affairs, these documents are becoming essential reading.
The scale of AI in submissions
AI is no longer a future consideration in regulatory submissions. The FDA has reported that since 2016 it has received more than 500 drug and biological product submissions with AI components. These range from models that help identify patients for enrolment, to algorithms that process imaging endpoints, to tools that predict pharmacokinetic behaviour or support manufacturing quality control.
Deloitte's 2025 analysis of pharmaceutical innovation helps explain the appetite. Tracking 20 leading biopharma companies, it projected an average R&D internal rate of return of 5.9% for 2024, up from 4.1% the year before, but with an average cost of around US$2.23 billion to develop an asset and total development timelines exceeding 100 months from Phase I to filing. The report also noted that Phase III cycle times had increased by 12%. When the cost of time is that high, tools that promise to make development faster or more informative attract serious interest.
The FDA draft guidance in brief
The guidance, Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products, was issued as a draft in January 2025. It remained in draft form at the time of writing, so its recommendations are non-binding, but it gives the clearest picture yet of the agency's thinking.
What it covers, and what it does not
The guidance applies when AI is used to produce information or data intended to support regulatory decisions about a product's safety, effectiveness or quality. That includes uses across nonclinical, clinical, post-marketing and manufacturing phases.
It does not cover AI used purely in early drug discovery, such as target identification or lead optimisation, nor AI used for internal operational efficiencies that do not affect patient safety, drug quality or the reliability of study results.
Two ideas at the centre: context of use and model risk
Context of use (COU) describes exactly what role the model plays in answering a specific question. It includes what the model does, what inputs it uses, what outputs it produces, and how those outputs feed into a decision. The same model can have very different risk profiles in different contexts of use.
Model risk is assessed through two factors:
- Model influence: how much the AI output contributes to the decision, compared with other evidence. A model that is the sole basis for a decision has high influence. One whose output is checked against independent evidence has lower influence.
- Decision consequence: how serious the impact would be if the decision based on the model were wrong.
Put together, these determine how much credibility evidence the FDA would expect to see.
The seven-step credibility framework
The core of the guidance is a risk-based framework for establishing and documenting the credibility of an AI model for its context of use:
- Step 1 - Define the question of interest. What specific question, decision or concern is the model helping to address?
- Step 2 - Define the context of use. Describe precisely what the model will do and how its outputs will be used.
- Step 3 - Assess the model risk. Combine model influence and decision consequence.
- Step 4 - Develop a plan to establish credibility. Describe the model, the data used to develop it, the training approach, and how it will be evaluated, proportionate to the risk.
- Step 5 - Execute the plan.
- Step 6 - Document the results and any deviations from the plan.
- Step 7 - Determine the adequacy of the model for the context of use. If credibility is not sufficient, options include reducing the model's influence, adding other evidence, tightening controls, or changing the approach.
A worked example
The following is adapted from one of the illustrations in the guidance itself. Imagine a sponsor running a trial of a drug with a known, rare but serious risk. The sponsor proposes an AI model to stratify participants by risk, so that lower-risk participants can be monitored as outpatients rather than kept in hospital after dosing.
If the model's output alone decides who goes home, model influence is high. If a misclassification could mean a high-risk participant is sent home and suffers a serious reaction without treatment, decision consequence is also high. That places the model in a high-risk category, and the FDA would expect substantial evidence of its credibility: well-characterised training data, independent test data representative of the trial population, clear performance metrics and a plan for monitoring performance during the trial.
Now change the context of use. If every participant's risk classification is reviewed by an investigator using standard clinical criteria, the model's influence falls, and so does the expected credibility burden. Same model, different context, different expectations.
Life cycle maintenance
The guidance recognises that AI models can change, or that their performance can drift as the data they encounter changes. Sponsors are expected to have plans for monitoring and maintaining model performance over time, particularly in manufacturing, where models may be used continuously.
Talk to the agency early
The FDA strongly encourages early engagement, and lists several routes depending on the use, including formal meetings and programmes such as the Model-Informed Drug Development (MIDD) paired meeting programme and the Emerging Technology Program.
FDA and EMA: a shared set of principles
On 14 January 2026, the FDA and EMA jointly published Guiding Principles of Good AI Practice in Drug Development, developed by FDA's CDER and CBER together with EMA. The ten principles are:
- Human-centric by design
- Risk-based approach
- Adherence to standards
- Clear context of use
- Multidisciplinary expertise
- Data governance and documentation
- Model design and development practices
- Risk-based performance assessment
- Life cycle management
- Clear, essential information
These principles sit alongside, rather than replace, regional guidance. In Europe, for example, the EMA published its Reflection paper on the use of artificial intelligence in the medicinal product lifecycle in 2024. The consistency between regions is encouraging: sponsors can build one credibility approach that is likely to satisfy both agencies.
What regulatory affairs professionals should do now
Map where AI is used. Many organisations do not have a complete inventory of AI tools that touch regulated activities. That list is the starting point for any credibility assessment.
Learn to write a context of use. A precise, well-bounded COU statement is quickly becoming a core regulatory writing skill, in the same way that indications and endpoint definitions are.
Work across disciplines. Credibility assessments need input from data scientists, statisticians, clinicians and quality specialists. Regulatory professionals often act as the translator between them and the agency.
Keep documentation inspection-ready. Training data sources, model versions, performance results and change history should be traceable, consistent with good data governance under ICH E6(R3) and GxP expectations.
Plan early engagement. For novel or high-risk uses, a pre-submission discussion with the agency can save months later.
Key takeaways
- The FDA has seen more than 500 drug and biologic submissions with AI components since 2016.
- Its January 2025 draft guidance sets out a seven-step, risk-based framework for establishing AI model credibility.
- Context of use, model influence and decision consequence determine how much evidence is needed.
- AI used only for early discovery or internal efficiency is outside the guidance's scope.
- FDA and EMA published ten shared Good AI Practice principles in January 2026.
Developing regulatory expertise for the AI era
Regulatory affairs is becoming more technical, and more interesting. CliniXen Institute's programmes bring together clinical research, data science and machine learning, helping professionals understand both the regulatory framework and the technology it now has to govern.
Advanced PG Diploma in Clinical Research & Pharmacovigilance
References
- US Food and Drug Administration. Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products: Draft Guidance for Industry, January 2025. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/considerations-use-artificial-intelligence-support-regulatory-decision-making-drug-and-biological
- Federal Register. Notice of availability, docket FDA-2024-D-4689, 7 January 2025. https://www.federalregister.gov/documents/2025/01/07/2024-31542/considerations-for-the-use-of-artificial-intelligence-to-support-regulatory-decision-making-for-drug
- US Food and Drug Administration and European Medicines Agency. Guiding Principles of Good AI Practice in Drug Development, 14 January 2026. https://www.fda.gov/about-fda/artificial-intelligence-drug-development/guiding-principles-good-ai-practice-drug-development
- European Medicines Agency. Reflection paper on the use of artificial intelligence (AI) in the medicinal product lifecycle, 2024. https://www.ema.europa.eu
- Deloitte. Measuring the return from pharmaceutical innovation, 2025. https://www.deloitte.com/us/en/industries/life-sciences-health-care/research/measuring-the-return-from-pharmaceutical-innovation.html
Disclaimer: This article is for educational purposes and reflects publicly available regulatory and scientific information at the time of writing. Guidance documents are updated periodically; always consult the latest official version.
