In a move that few expected to materialize this quickly, and even fewer anticipated being described in such sweeping terms, the FDA is barreling ahead with an “aggressive, agency-wide” adoption of artificial intelligence. This isn't a tentative exploration. On May 8, 2025, the agency announced the completion of its inaugural AI-assisted scientific review pilot. By June 30, 2025 (yes, next month!), it plans to have generative AI tools embedded throughout every one of its centers.
Make no mistake: these aren’t casual experiments. This is a full-throttle transformation of how the FDA evaluates applications, monitors safety, and interacts with the biopharma industry. The era of AI as a peripheral R&D project is over. The regulator itself is becoming an AI-augmented powerhouse.
"This is a game-changer technology that has enabled me to perform scientific review tasks in minutes that used to take three days,” said Jinzhong (Jin) Liu, Deputy Director, Office of Drug Evaluation Sciences, Office of New Drugs in FDA’s Center for Drug Evaluation and Research (CDER)."
That stark comment encapsulates the magnitude of this shift. What once consumed days of effort can now be accomplished in minutes, fundamentally reshaping timelines and expectations on both sides of the regulatory aisle. But this revolution isn't merely about speed; it’s about a new paradigm, one in which both sponsors and regulators must speak the same AI-infused language.
The FDA’s 2025 AI Marathon: From Zero To Sixty
The velocity of this change is remarkable. In just six months, the FDA has sprinted from releasing its first AI guidance draft to a clear, public commitment to agency-wide integration of generative models. Key milestones paint a picture of deliberate, rapid advancement:
January 6, 2025 – Clear-the-Deck Guidance Published: The agency released draft recommendations on AI for regulatory decisions. This wasn't just a suggestion; it introduced a risk-based credibility assessment framework, outlining seven critical steps for qualifying AI models in specific contexts of use. The core message to industry: sponsors must meticulously “show their work” by detailing model design, data provenance, evaluation metrics, and comprehensive lifecycle maintenance plans.
April 2025 – Rethinking Preclinical Studies: An ambitious roadmap emerged, championing AI-driven computational models and human-derived New Approach Methodologies (NAMs). This signaled a significant shift towards in silico and organoid-based safety assessments, aiming to reduce reliance on traditional animal testing. FDA Commissioner Makary was unequivocal: "By leveraging AI-based computational modeling, human organ model-based lab testing, and real-world human data, we can get safer treatments to patients faster and more reliably, while also reducing R&D costs and drug prices."
May 8, 2025 – Pilot Completion, Proof of Concept: The FDA successfully concluded its first AI-augmented review pilot, validating the approach in a real-world regulatory setting and paving the way for broader implementation.
June 30, 2025 – The Agency-Wide Rollout Deadline: This is the target for embedding generative AI across every center, from New Animal Drug Evaluations to the Digital Health Center of Excellence.
Underpinning these policy and pilot initiatives is a clear structural commitment: the appointment of Jeremy Walsh as the FDA’s first Chief AI Officer. This move cements AI not as an experimental afterthought, but as a central strategic priority for the agency.
Decoding the FDA's Playbook: The Risk-Based Credibility Framework
At the heart of the FDA's January draft guidance lies a crucial seven-step process for any organization using AI to generate data or insights intended to support regulatory decisions. Biopharma sponsors must internalize and build this framework into any AI-driven submission:
Define the Question of Interest: Articulate with precision what decision the AI model is being asked to make. Examples include stratifying patients by adverse-reaction risk or determining vial fill-volume compliance. Clarity here is paramount.
Specify the Context of Use (COU): Explain exactly how the AI model's outputs will be applied within the regulatory decision-making process. Will the model be the sole determinant for patient monitoring, or will it augment, rather than replace, existing sample-based quality checks?
Assess Model Risk: This involves a two-dimensional evaluation:
Model Influence: How significantly does the AI output drive the final decision?
Decision Consequence: What is the potential impact if the AI model makes an incorrect assessment?
Combining these factors allows for a risk rating of Low, Medium, or High for the model in its specified COU.
Draft a Credibility Assessment Plan: Tailor your validation activities directly to the assessed model risk level. This plan must comprehensively cover:
Model Description: Detail the architecture, parameters, features, and the rationale for selecting that particular approach.
Data Characterization: Explain how training, tuning, and testing datasets were collected, annotated, and partitioned. Crucially, justify why these datasets are "fit for use" for the intended COU.
Performance Metrics: Specify metrics such as ROC AUC, sensitivity, specificity, positive/negative predictive values, calibration curves, and uncertainty quantification (complete with confidence intervals).
Software Quality Controls: Outline the measures taken to ensure the reliability and robustness of the AI software.
Execute the Plan: Carry out the planned testing meticulously, documenting every step and any deviations from the original plan.
Document Outcomes and Deviations: Compile a comprehensive Credibility Assessment Report. This report can be part of your regulatory submission or maintained within your quality-system dossier for potential FDA audit.
Determine Model Adequacy: Based on the evidence, assess if the model's credibility has been sufficiently proven for its intended COU. If not, iteration is required. This may involve adjusting the COU, incorporating additional evidence sources, retraining or redesigning the model, or applying further controls until the risk is deemed acceptable.
What This Means for Biopharma
Regulatory review is no longer a human-only endeavor. The agency you engage with today is an AI-augmented powerhouse with access to more data than any single company. Every biopharma organization must forge a unified, end-to-end AI and data strategy that permeates R&D, clinical operations, regulatory affairs, and commercial functions. Your operating models, governance frameworks, and data channels must be designed to converse fluently with an AI-driven regulator. Key components include:
Robust AI Governance: Establish a unified framework defining roles, responsibilities, validation standards, and risk-based controls for all AI tools. This ensures consistent oversight and allows you to demonstrate reproducibility, traceability, and explainability for every model, whether it's for dose simulations or post-market surveillance.
"AI Packages" as Standard in Submissions: Prepare to include comprehensive model artifacts with your INDs, NDAs, and other regulatory filings. These "AI packages" should encompass detailed training data descriptions, performance metrics, boundary conditions, and the full Credibility Assessment Report as outlined by the FDA.
Ethics, Transparency, and Traceability: Black boxes are unacceptable. Embed explainability features directly into your AI solutions. Document every critical decision-making step (data provenance, feature selection, model retraining logs) so that both your teams and regulators can audit the AI’s reasoning and ensure ethical application.
Real-World Evidence (RWE) at Scale: The ability to generate high-quality RWE efficiently will be paramount. AI-powered platforms that can leverage vast, anonymized patient records will provide the critical fuel for swift, confident, and compelling regulatory interactions.
The FDA's timeline speaks volumes: all centers operating on a common, secure generative AI system integrated with the FDA's internal data platforms by June 30. This isn't a distant future state - it's next month. The agency hasn't just dipped its toe in the AI waters; it's diving in headfirst with a clear message to industry: adapt now or get left behind.
Why Endpoint?
We partner with forward-looking biopharma teams to deploy advanced AI agents that unlock the power of Real-World Evidence (RWE) at scale. Our platform is engineered to generate limitless, high-quality RWE, the essential fuel for next-generation R&D, clinical development, regulatory strategy, and commercial success. With access to multimodal data from over 25 million patients, Endpoint delivers the foundational insights you need to make critical decisions today, not tomorrow.
The capabilities the FDA is now championing, a robust data infrastructure, AI-driven insights, continuous learning, are no longer aspirational. They are non-negotiable for building competitive and compliant regulatory, safety, R&D, and medical functions. Mirroring the FDA's accelerating pace, Endpoint provides the AI-powered RWE platform to help you meet these new standards.
If you'd like to explore how we can help you stay ahead of the regulatory curve, get in touch today, and let’s discuss whether a research preview is right for you.