FDA Rules for AI in Drug Development. What CMC Teams Need to Know

Written By Simon Letarte

In January 2025, FDA published a draft guidance titled Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products. It covers CDER, CBER, CDRH, and CVM, essentially the full regulatory apparatus. If you work in CMC or analytical development and you haven't read it, you should. Not because it changes everything you're doing right now, but because it draws a line that every team deploying AI needs to understand.

The guidance establishes a risk-based credibility assessment framework for AI models used to produce information or data that supports regulatory decision-making. That's the scope. And the guidance is explicit about what it does not cover: AI used for "operational efficiencies (e.g., internal workflows, resource allocation, drafting/writing a regulatory submission) that do not impact patient safety, drug quality, or the reliability of results."

Drafting a regulatory submission is explicitly out of scope. This is actually good news for the growing number of CMC teams using LLMs to generate first drafts of Quality Overall Summaries, method validation reports, or regulatory justifications. Amgen has published on an LLM-based tool that maps Module 3 content to the appropriate QOS sections and produces a structured first draft in under an hour compared to roughly two weeks by a human writer.

What FDA is asking about is more consequential: the AI models that actually generate the data and conclusions that go into those submissions.

What Falls Under This Guidance

Think about the AI applications that are increasingly common in analytical development and manufacturing:

  • An AI-based visual inspection system performing 100% automated fill-volume assessment for release testing

  • A machine learning model predicting shelf-life or stability outcomes from early time-point data

  • An AI model used to set process parameters or flag out-of-trend results in real-time manufacturing

  • A predictive model informing specification ranges based on historical characterization data

These applications do impact drug quality and the reliability of regulatory data. They fall squarely within the scope of this guidance. And if you're building or deploying any of them, FDA has now told you exactly how they want you to think about it.

The Framework: Context of Use and Model Risk

The centerpiece of the guidance is a seven-step risk-based credibility assessment framework. The two concepts that will drive most of the practical work are the Context of Use (COU) and model risk.

The COU defines the specific role and scope of the AI model: what question it addresses, what its output is, and whether other evidence will be used alongside it to answer that question. In the guidance's manufacturing example, an AI-based fill-volume inspection tool is used alongside independent release testing of a representative sample. That orthogonal verification reduces the model's influence on the final decision, which in turn lowers model risk. If instead the AI model were the sole determinant of batch release, model risk would be much higher.

Model risk, as FDA defines it, is a function of two factors: model influence (how much the AI's output drives the decision) and decision consequence (how bad it is if that decision is wrong). A model that flags anomalies for human review has lower influence than one making autonomous decisions. A model informing a release decision for a drug with a narrow therapeutic index has higher decision consequence than one used to flag early out-of-trend signals in a development-stage program. The combination determines how much credibility evidence you need and how much FDA scrutiny your AI model will receive.

This is genuinely useful framing. It gives CMC teams a principled way to prioritize and a vocabulary for conversations with FDA.

Life Cycle Maintenance: The Part Everyone Will Underestimate

FDA makes clear that AI model performance can change over time as inputs change. What the guidance calls data drift. In pharmaceutical manufacturing, this means that as your process evolves, your raw material suppliers change, or your equipment ages, an AI model trained on historical data may degrade in ways that are not immediately obvious. FDA expects sponsors to have a plan for monitoring model performance on an ongoing basis and for managing changes through the pharmaceutical quality system.

For teams that have been treating AI models like validated analytical methods (qualify once, use forever) this is a wake-up call. AI models are not static instruments. They are data-dependent systems, and their performance is only valid within the distribution of data they were trained on. Building life cycle maintenance into the deployment plan is not optional if your model is supporting regulatory decisions. FDA is recommending it be included in the marketing application for product- or process-specific AI models.

What This Means Practically Right Now

What hasn't changed: using AI for document drafting, literature review, validation report generation, batch record data extraction and other operational tasks is still fair game and still outside FDA's formal framework. Do it. The efficiency gains are real.

What has changed: if you are deploying AI to generate analytical data or inform manufacturing decisions that go into regulatory submissions, you now have an FDA-endorsed framework for how to document and justify that use. Ignoring that framework is not going to make your submission cleaner, it is going to create questions. Adopting it gives you a structured approach to a conversation that FDA is ready to have.

FDA strongly encourages early engagement when you're uncertain about whether and how this guidance applies. For CDER/CBER biologics programs, CBER's Advanced Technologies Team (CATT) is the right contact for AI in pharmaceutical manufacturing. A pre-IND meeting that addresses your AI deployment plan is far less expensive than a deficiency letter that challenges your data package on credibility grounds.

Four Takeaways for CMC and Analytical Teams

1. Map your AI tools to the COU framework now. For each AI application in your program, ask: What question does it answer? Is it the sole determinant of a regulatory decision, or one input among several? What happens if it's wrong? This exercise will tell you where your credibility documentation gap is.

2. Operational AI tools (QOS drafters, report generators, literature search) are outside this guidance. Deploy them. The regulatory overhead is low and the productivity gains are real — as the Amgen QOS experience demonstrates.

3. For in-scope models, calibrate your documentation effort to your model risk. A low-risk, low-influence model used alongside orthogonal evidence needs less credibility documentation than a model making autonomous release decisions. FDA is not asking for the same package across the board. Match the rigor to the risk.

4. Plan for life cycle maintenance. If your AI model is being deployed in manufacturing or supporting ongoing regulatory decisions, build in performance monitoring, drift detection, and a change management pathway tied to your quality system. Retrofit is much harder than building it in from the start.

If you're thinking through where AI fits in your CMC or analytical development program and how to engage FDA on it, I'd be glad to talk. Reach out here.

Sources:

Simon Letarte
Next

Mass Spectrometry for ADC Characterization: What You Really Need to Know for IND Filing