FDA Agentic AI - Implications for IVD/CDx Programs of Regulating Wellness and Medical Wearables in an Age of Blurring Boundaries

The FDA has deployed “Elsa”, an enterprise-generated large language model (LLM) platform. The agency has described the current state as “exploratory”, using AI agents in a limited internal capacity. In this sense, the agency’s effort is analogous to that of millions of individuals, small businesses, and large organizations in adopting tools like Claude Code, Codex, and Gemini.

The first wave of LLM tools from 2022 to 2024 was centered on chat-based answer engines that could engage in conversations, analyzing and synthesizing billions of pieces of text and other datasets.

Use Cases for Agentic AI

Agentic AI, rapidly maturing since 2025, is a significant advancement of artificial intelligence capability, enabling autonomous workflows that span multiple complex steps, such as sorting, scheduling, purchasing, prioritizing, communicating with multiple parties, and even writing computer programs, in a manner analogous to a team of human assistants. Agentic systems are well-suited to common tasks in FDA review, such as:

Collecting structured data from unstructured submission documents

Comparing submitted information against guidance documents, prior approvals, and regulatory standards

Identifying missing elements and deviations from expected formats

Writing draft review questions, deficiency letters, and internal memos

Cross-referencing clinical protocols, statistical analysis plans, and final reports for consistency

What We Know and Next Steps

In this sense, Agentic AI is an obvious candidate for operational domains like pre-market review, post-market surveillance, administration, inspection, and compliance. In pilots across multiple industries, it is increasingly common to hear professionals rave that agentic AI enables tasks that previously required days of effort to be completed in minutes. This is exactly what the Center for Drug Evaluation and Research (CDER) has indicated in pilot programs. For Clinical Research Organizations (CROs), this accelerated velocity means submissions will need to be structured for AI-triaged interfaces, not just sequential human reading.

The FDA has repeatedly emphasized that Elsa is in the early stages of maturity. One obvious example is that core technical architectural changes are still in process. The agency has not yet published specific workflow protocols, operational guidance, or format requirements related to 510(k), De Novo, and Pre-Market Application (PMA) submission reviews. It is, however, safe to assume that this is an inflection point at which device makers and CROs need to anticipate a future in which Agentic AI is deeply embedded in clinical trial and regulatory review processes.

IVD/CDx Risks in an AI Workflow

The agency has already flagged cross-document inconsistency checks as an obvious use, particularly for in Vitro Diagnostics (IVD) and Companion Diagnostics (CDx) submission packages that often exceed a million words. This can amplify the risk of false positives if language describing the same item is used differently across document types. There are already known cases of AI detecting meaningless patterns that required further human intervention to resolve.

This risk is particularly pronounced for CDx submissions, which require coherence across device and drug/biological domains. Biomarker-drug, pair-specific evidence for CDx tests must align with the drug sponsor's New Drug Application (NDA) or Biologics License Application (BLA) submission. AI systems are trained on historical models. An AI system trained on historical 510(k) clearance models may not be validated for the joint drug-diagnostic evidence architecture inherent to PMA CDx submissions.

CDx manufacturers, and CROs must anticipate how the AI and human reviewers might assess that a biomarker claim is scoped to the co-approved drug’s label. This is especially important considering the FDA’s November 2025 proposed reclassification of certain oncology CDx devices from Class III to Class II, which require test claims to be anchored to drug risks and benefits.

Preparing for an AI-Powered Regulatory Future

 Based on the known validation gaps and regulatory pitfalls, IVD/CDx manufacturers, clinical trial sponsors, and CROs can make specific compliance adjustments. With the introduction of LLM-based cross-document scans, consistent language is essential across all intended use statements, endpoint language, claims, and labeling. This rigor also applies to biomarker definitions, performance claims, and statistical conclusions, which should use a tightly controlled and well-documented vocabulary.

For CDx submissions, it will be essential to create a drug-IVD label alignment narrative as an explicit traceability document. This should summarize how biomarker claim language in the IVD submission aligns with the co-approved drug label. This significantly reduces the risk that Elsa flags cross-document inconsistencies. If the submission includes Next Generation Sequencing (NGS) panels, it is also critical to include a bioinformatics glossary that defines variant nomenclature to minimize potential misinterpretation of terms that are scientifically equivalent.

Regulatory engagement also provides an opportunity to learn more about the use of Agentic AI and to collaborate on risk minimization and avoidance. In Q-Submissions and pre-Investigational Device Exception (IDE) meetings, sponsors and CROs should ask the FDA to clarify whether and how Elsa will be used in the review of the specific submission, including which model version will be used. Through regulatory processes, be sure to document FDA interactions contemporaneously, noting any indications that AI-influenced workflows may have influenced deficiency letters or information requests.

Conclusion

Agentic AI is quickly becoming the new normal for digital knowledge work in a vast array of global industries and public/private organizations. The FDA’s initial pilot with Elsa provides IVD/CDx leaders with an opportunity to plan and phase in adjustments to their product development, clinical research, and regulatory affairs management in a way that anticipates potential challenges with proactive solutions. These are short-term growing pains, and there is every reason to believe that, in the long run, a mature AI-human hybrid process will offer dramatic improvements in speed and productivity, bringing innovative new therapies and devices to market for a fraction of the current time and at a fraction of the current cost in human resources.

Questions on how Agentic AI impacts your IVD/CDx strategy? Let's discuss! Landrich Group combines deep expertise and a forward-looking approach to support our clinical research partners!