What Does the FDA’s Risk-Based Approach to AI Guidance Mean for Sponsors?

With all the data being collected in today’s clinical trials, it is becoming increasingly difficult to oversee studies. It’s no surprise, then, that AI has begun to feature more prominently in discussions surrounding clinical trial efficiencies, with sponsors and technology vendors alike asking how AI can make oversight more scalable. The lingering issue with AI, however, is that it introduces a new layer of uncertainty. If a model is helping identify a potential quality or safety signal, prioritize a site for review, or support another clinical decision, sponsors need to understand how much they can rely on that output and what controls are appropriate if that output is wrong. 

That is where FDA’s emerging guidance becomes useful. In January 2025 the FDA published, “Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products,” draft guidance for AI use in clinical trials. In it, they introduced a risk-based credibility assessment framework for AI models used to produce information or data that support regulatory decision-making around drug safety, effectiveness, or quality. The framework was designed to help sponsors determine how much evidence and oversight an AI model needs, based on how it will actually be used.  

Below, I break down the first four steps of that framework and what they mean for sponsors looking to leverage AI in their clinical trials. 

1. Defining the Specific Question the AI model Will Address 

The first step the FDA recommends is defining exactly what you are asking the AI model to help you determine. In a clinical trial, for example, that could be which participants are at increased risk of developing a serious adverse event, where protocol deviations are most likely to occur, or which sites are showing data quality issues that could affect study reliability.  

Defining a specific question is incredibly important. Without a clearly defined question the AI model might solve the wrong problem or study teams could underestimate risks that the model poses to the trial—there’s also the risk that your team won’t be able to explain your results to the FDA if asked.  

2. Defining the Context of Use 

Once the question is clear, the FDA asks sponsors to define the AI model’s context of use, or COU. 

The COU describes the specific role and scope of the AI model will play in helping answer the question: what will it do, how will its output be used, etc. The COU is important to consider in support of the next step which involves AI risk. 

For example, an AI capability that drafts a risk statement for a user to review has a very different COU from an AI model that identifies a potential patient-safety concern for a medical monitor. Both may support operations of a trial, but the impact the model has on the trial is very different. 

3. Assessing the AI Model Risk 

The FDA’s recommended third step is to assess model risk based on two factors: model influence and decision consequence. 

Model influence is the contribution of the AI-derived evidence relative to the other evidence used to inform a decision. Decision consequence is the significance of the adverse outcome that could result from an incorrect decision. In practical terms: How much can the AI impact our outcomes, and what could happen if it is wrong? 

4. Developing a Plan to Establish AI Model Credibility 

If steps 1-3 are asking what the AI is doing and to what extent it will impact your final decisions, step 4 is asking: What evidence will you generate to show the model can be trusted for its purported use?  

The purpose is to determine what evidence will establish confidence in the AI model. Depending on the model and its risk, that can include information about the model itself, the data used to develop it, how it was trained, and how its performance will be evaluated.  

Importantly, the FDA does not prescribe the same plan for every model. For lower-risk models, minimal information may be appropriate; higher-risk applications may require considerably more detail and more stringent performance criteria.  

That gives sponsors room to innovate without treating responsible AI adoption as a one-size-fits-all validation exercise. 

How Can You Ensure Your Technology Adheres to the FDA’s AI Guidance? 

If sponsors are going to apply the FDA’s AI guidance, there are certain foundational features that must exist in their tech stack.  

• Clinical Data Repository: AI is only as useful as the data it can draw from. If the use case involves identifying site-level risk, supporting patient-safety review, or surfacing data quality concerns, AI needs access to the relevant clinical and operational data in context. A clinical data repository brings sources together and gives AI access to data that is appropriate for the question being asked. 

• Automated Data Ingestion: Without automated data ingestion, teams risk asking AI to make recommendations based on information that is already incomplete or out of date. In a clinical trial, where new data is arriving continuously from multiple sources, sponsors should not have to manually collect, clean, and prepare that data every time they want to use AI. Automated data ingestion helps keep the data feeding the model up to date while reducing the manual effort required to get it there.  

• Transparency and Human Oversight: Sponsors need visibility into the data being used and the role AI plays in the workflow. Teams should be able to understand what information is feeding the AI model, review its outputs, and retain appropriate control over the decisions that follow. 

The FDA is clear that a risk-based approach to AI should be tailored to the context of use and commensurate with an AI model’s influence and consequences. For clinical operations teams looking to deploy AI into their workflows: Look for technology that can bring together the right data, align with sponsor needs (context of use), preserve appropriate human oversight, and provide the transparency needed to ensure AI use adheres to FDA guidelines.  

To discover how FDA guidelines can be applied to AI features in elluminate, you can watch my full webinar here.  

Download the white paper, Risk-Based Data Management: Bridging RBQM Strategy and Data Execution. The white paper lays out an RBDM Maturity Model across planning, execution and closeout, and oversight, along with a step-by-step guide for putting RBDM into practice. You can find the white paper here.