Imagine a sponsor evaluating an AI tool that analyzes clinical images. The demonstration is compelling, and the vendor reports strong accuracy. Before those results inform a development decision, however, the team needs to know what was tested, which patients were represented, and whether the tool can perform reliably in the proposed study. A promising demonstration begins the discussion; the supporting evidence determines how far the sponsor can take it.
That distinction is central to credible AI use in drug development. Sponsors need a defensible reason to rely on an output for a particular purpose. The following questions offer a practical starting point for evaluating a proposed application, whether developed internally or supplied by a partner.
In January 2026, FDA and EMA published ten joint principles for good AI practice across the medicines lifecycle. They address how AI is developed, assessed, used, and maintained, including its role in generating evidence. EMA describes the principles as a foundation for future guidance. They do not provide a universal testing protocol or guarantee acceptance of a particular application.
FDA’s January 2025 draft guidance offers a more detailed credibility assessment framework for AI supporting regulatory decisions. It remains draft guidance, labeled not for implementation. Its definition of credibility ties confidence in model performance to a specific context of use.
A useful first exercise is to complete this sentence: “We will use this system to produce ___, which will inform ___, under these conditions: ___.” If the team cannot agree on the blanks, it is too early to judge whether a vendor’s evidence answers the sponsor’s needs.
For the hypothetical imaging tool, consider the difference between helping researchers explore patterns and producing measurements used in a trial endpoint. The sponsor’s assessment should account for how much the decision depends on the output and the consequences of an error. Both are components of model risk in FDA’s draft framework.
Questions to ask
Document the agreed use before procurement or development progresses. This gives the scientific team and vendor a common target and helps prevent a successful pilot from quietly expanding into an application it was never designed to support.
EMA’s 2024 AI reflection paper emphasizes representative training data, potential bias, and traceable data processing. It also calls for remaining data limitations to be documented, with alternatives for cases where the model is unsuitable.
Returning to the imaging example, suppose the vendor’s demonstration relies on scans collected at one specialist center, while the proposed study will involve several hospitals using different equipment. The sponsor should investigate whether that difference matters before agreeing to use the tool. A larger dataset from the same center would not, by itself, answer that question.
Questions to ask
Ask the vendor to walk through a sample record from collection to model input. Use that discussion to identify unresolved assumptions and decide what additional information the team needs.
FDA’s draft recommends evaluating the trained model using test data independent of development data, with metrics suited to the proposed use and performance estimates accompanied by confidence intervals. These details help distinguish a reported score from evidence supporting a decision.
For the imaging tool, a sponsor might ask the vendor to show difficult cases alongside successful ones. Does performance change when image quality is lower? What happens when two expert readers disagree? The answers can help the sponsor decide whether the proposed study needs additional testing or a narrower application.
Questions to ask
EMA additionally highlights prospective testing on newly acquired, representative data for high patient risk or high regulatory impact settings. Sponsors should therefore avoid treating every validation claim as equivalent; the relevant question is what the testing establishes for their proposed use.
The joint principles call for multidisciplinary expertise and assessment of the complete system, including human–AI interactions. That brings the user’s role into the performance discussion.
A practical oversight plan for the imaging example might specify who reviews uncertain measurements, how disagreements are resolved, and when analysis pauses for investigation. The reviewer would need enough time and information to challenge the result. Naming a clinician as the final approver leaves these workflow questions unanswered.
Questions to ask
Before launch, rehearse a difficult case with the people who will use the system. This can expose ambiguity in responsibilities that a technical demonstration would miss.
The January 2026 principles emphasize verifiable documentation of data sources, processing, and analytical decisions, alongside scheduled monitoring and periodic reevaluation. EMA’s reflection paper also addresses detecting performance deterioration and planning how to suspend a model when necessary.
For a sponsor, a useful operational test is whether another team member could explain how a reported result was produced. If the vendor updates its software midway through the program, the team should be able to identify the affected outputs and determine what review is needed.
Questions to ask
Set these expectations with the vendor early. Documentation that becomes available only after a concern arises may arrive too late for the sponsor’s decision.
The imaging demonstration becomes more useful when the sponsor can explain the decision it supports, the evidence behind it, and the circumstances that would prompt reconsideration. Those explanations should guide the development plan from the beginning.
Where AI will support regulatory decisions, FDA’s draft encourages early discussion of the credibility assessment plan. Bring specific questions and a clear proposed use to that conversation. The resulting feedback can inform what evidence to gather before the program commits to relying on the system.
Planning to use AI-generated evidence in your development program? Connect with The Bracken Group to discuss the scientific and regulatory questions your strategy needs to address.