FDA Has Seen 1,000+ AI-Related Submissions. Regulatory Writing Has to Catch Up.

What FDA Said At DIA

At DIA's annual meeting in Philadelphia, FDA officials described a regulatory environment in which AI and real-world evidence are no longer peripheral. Anindita Saha noted that CDER and CBER have received more than 1,000 submissions incorporating AI elements, up from a single submission in 2016. Those AI uses include endpoint development, patient selection, outcome prediction, and trial enrichment, concentrated in oncology and followed by gastroenterology, neurology, and psychiatry.

FDA has also opened a public effort around drug repurposing for unmet medical needs, including chronic and rare diseases, with real-world evidence expected to be part of the conversation. At DIA, officials discussed how RWE may fit into that broader effort. And in the audience Q&A, industry stakeholders asked the agency for more clarity on how AI can be used in drug manufacturing.

None of these developments directly names regulatory writing. All of them create new writing surfaces.

A Thousand Submissions Is Not A Threshold. It Is An Operating Context.

A thousand submissions is not a formal regulatory threshold. But operationally, it matters.

At that volume, AI is no longer an exception category. FDA reviewers are likely moving from occasional case-by-case familiarity toward recurring expectations for how AI-enabled methods should be described. Reviewers who repeatedly see AI-enabled methods will develop expectations for how sponsors describe context of use, training data, validation, performance, bias, limitations, and model governance.

FDA already gave sponsors the vocabulary. The agency's January 2025 draft guidance on the use of AI to support regulatory decision-making provides a risk-based credibility-assessment framework — define the AI model's context of use, assess the model's risk, and establish model credibility proportionate to that risk. The 1,000-submission disclosure suggests that this vocabulary is no longer theoretical. It is becoming review practice.

For regulatory writing teams, that means AI-method descriptions can no longer be improvised in the Module 5 sections where they happen to land. If a sponsor uses AI for patient selection, endpoint prediction, trial enrichment, signal detection, image analysis, or operational decision support, the dossier needs a clear and repeatable way to describe what the model did, where it entered the development or evidence-generation process, what data it was trained and validated on, what performance was observed, what bias or generalisability limits were assessed, what human oversight was used, and how the model output affected the regulatory question.

A submission that describes its AI component in idiosyncratic vocabulary, or that omits sections the reviewer's pattern has come to expect, generates Information Requests that the sponsor pays for in review-cycle time. The writing organisation that pays attention to what is and is not generating clean acceptance — by reading public FDA review documents where available, attending AI-focused regulatory sessions, tracking guidance and advisory-committee materials, and learning from published case studies — is the organisation whose next submission lands in the agency's preferred shape.

RWE For Drug Repurposing Is A Specific Writing Problem

FDA's drug-repurposing effort matters for regulatory writing because repurposing dossiers are evidence-integration exercises.

A sponsor seeking a new use for an already approved drug may need to draw from several evidence layers at once — the original approval package, known safety experience, new clinical evidence in the proposed indication, mechanistic rationale, and real-world data from existing use.

The writing problem is the bridge.

The dossier has to explain what evidence carries forward from the original approval, what does not, what new evidence supports the repurposed use, and what RWE contributes. It cannot assume it will be read as either a clean-sheet NDA or a simple label expansion. The dossier has to teach the reviewer how the evidence should be integrated — what comes from the original approval, what comes from the new indication, and what RWE specifically supports.

That teaching happens in the briefing document, the Module 2.5 clinical overview, and the clinical summaries. If those sections do not make the bridge explicit, the reviewer has to build it themselves — and the dossier becomes harder to defend on its own terms.

When the agency formally opens an RWE-for-repurposing conversation, the immediate implication for sponsors with repurposing programs in their pipeline is that the agency itself is now developing internal expectations for what an adequate repurposing dossier looks like. Writing organisations that have not designed for this dossier shape will be writing into a target the agency is in the middle of defining.

AI In Manufacturing Becomes A Module 3 Question

Industry's question about AI in manufacturing is not just a technology question. If AI is used in manufacturing or quality decision-making, it eventually becomes a Module 3 writing question.

Module 3 already has conventions for process description, process control, validation, analytical methods, continued process verification, computerised-systems documentation, and quality systems. AI-augmented manufacturing does not arrive into a blank space. It adds a new layer on top of an established framework — model role, training data, validation approach, drift monitoring, human override, change control, deviation handling, and relationship to the validated control strategy.

The questions a writing team should be preparing for are practical.

What process step or quality decision does the model support? Is the model advisory or controlling? What data trained and validated the model? How is performance monitored over time? What happens when the model output conflicts with existing controls? How are model updates governed? How does the AI component fit into the validated state of the process?

The use cases are also broader than any single example — AI-supported process monitoring, predictive maintenance, deviation detection, batch-release support, model-assisted process control. As sponsors experiment with AI-supported manufacturing and quality systems, future submissions will need a way to describe those systems without retrofitting the language at the last minute.

That is not just an AI policy question. It is CMC writing architecture.

The Writing Surfaces Sponsors Should Build Now

Sponsors do not need to wait for every future guidance document before building the writing surfaces these submissions will require. The vocabulary FDA has already published — context of use, model risk, model credibility, fit-for-purpose evidence, integrated evidence narratives — is enough to start.

For AI-enabled clinical methods. Context of use, model role, training and validation data, performance metrics, bias assessment, limitations, human oversight, and impact on the regulatory decision. These belong as a reusable structural component in Module 5 sections wherever AI touches the trial.

For RWE-enabled repurposing. Source population, data provenance, fit-for-purpose assessment, endpoint definitions, confounding control, missingness, sensitivity analyses, and the bridge between prior approval and new use. These anchor Module 2.5 and the briefing document and decide whether the repurposing dossier reads as integrated or as stitched-together.

For AI in manufacturing. Model role, process step affected, training data, validation approach, change control, drift monitoring, human override, deviation handling, and relationship to the control strategy. These live in Module 3 quality narratives and need a place in the dossier voice before the first AI-augmented manufacturing decision reaches a submission.

Each of these is a writing surface. Each needs reusable structure before the submission deadline arrives.

The Architecture That Has To Be Built Now

None of the DIA disclosures directly said "regulatory writing." They did not need to.

FDA is already seeing AI-enabled submissions at scale. It is opening a drug-repurposing conversation where RWE will matter. And industry is asking how AI in manufacturing should be handled.

The near-term task for sponsors is not to wait for perfect guidance. It is to build reusable writing architecture now — AI context-of-use sections, model-credibility summaries, RWE integration narratives, and Module 3 language for AI-supported manufacturing.

Sponsors that build those surfaces early will write into FDA's emerging expectations.

Sponsors that do not will discover the expectations through review questions.