What FDA's draft AI guidance asks for, and what it leaves out

Alyse Gonthier, PhDHead of Content

Published September 23, 20264 min read

In brief

FDA's draft guidance Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products, issued in January 2025 and still a draft as of September 2026, sets out a seven-step, risk-based framework for establishing that an AI model is credible for a defined context of use. It applies to AI that produces information supporting regulatory decisions about a drug's safety, effectiveness, or quality. It excludes AI used in drug discovery and AI used for operational efficiencies that do not affect patient safety.

Key takeaways

FDA published the draft guidance in January 2025, with comments due April 7, 2025, and as of September 2026 the guidance page still labels it a draft.
The framework has seven steps, from defining the question of interest and the context of use to deciding whether the model is adequate for that use.
Model risk is assessed on two factors: how much the model's output influences the decision, and how serious the consequence of a wrong decision would be.
The guidance excludes AI used in drug discovery and AI used for operational efficiencies, such as internal workflows, that do not affect patient safety.
In January 2026, EMA and FDA published ten shared Guiding Principles of Good AI Practice in Drug Development, which explicitly include manufacturing.

What the guidance is

The guidance is FDA's first cross-center draft on how sponsors should show that an AI model is trustworthy when its output is used to support a regulatory decision. It was issued jointly by CDER, CBER, CDRH, CVM and other FDA offices and published in the Federal Register on January 7, 2025, under docket FDA-2024-D-4689.

It describes the kinds of uses it has in mind. Its examples include reducing the number of animal-based pharmacokinetic, pharmacodynamic and toxicology studies, predictive modeling for clinical pharmacokinetics, integrating data from many sources to understand disease, processing real-world and digital health data for endpoints, handling postmarketing adverse event information, and facilitating the selection of manufacturing conditions.

The seven-step credibility framework

The core of the guidance is a risk-based process for establishing and documenting that an AI model is credible for its intended use.

Define the question of interest the AI model will address.
Define the context of use: the specific role and scope of the model in answering that question.
Assess the AI model risk.
Develop a plan to establish the credibility of the model's output within the context of use.
Execute the plan.
Document the results of the credibility assessment and discuss any deviations from the plan.
Determine whether the model is adequate for the context of use.

The context of use carries most of the weight. The same model can be low risk in one use and high risk in another, depending on what the output is used to decide.

How model risk is assessed

Step 3 combines two factors in a matrix.

The two factors FDA's draft guidance uses to assess AI model risk.
Factor	The Question It Asks
Model influence	How much does the AI output contribute to the decision, relative to other evidence?
Decision consequence	How serious would the outcome be if the decision were wrong?

A model whose output is one of several lines of evidence behind a low-consequence decision is lower risk than a model whose output alone determines a decision that affects patients. Higher risk calls for more extensive credibility evidence and documentation.

What the guidance leaves out

The scope section is easy to skip and important to read. The guidance does not cover AI used in drug discovery, and it does not cover AI used for operational efficiencies, such as internal workflows, resource allocation or drafting a regulatory submission, that do not affect patient safety, drug quality or the reliability of results from a study.

That has a practical consequence for manufacturing and quality teams. An AI system that assembles evidence for an investigation or drafts a record for expert review is not governed by this draft simply because it uses AI. It is governed by the same CGMP framework as any other system in the operation: the predicate rules in 21 CFR Part 211, the electronic records requirements in Part 11 and the operation's own validation approach. Where AI output does feed a regulatory decision, the credibility framework becomes relevant.

What’s happened since January 2025

The guidance is still a draft. As of September 2026 FDA's guidance page still labels it a draft.
EMA and FDA aligned on principles. On January 14, 2026, the two agencies published ten Guiding Principles of Good AI Practice in Drug Development, covering the lifecycle and explicitly including manufacturing.
Europe proposed rules for AI in GMP. The EU GMP Annex 22 draft, published for consultation in July 2025, keeps generative AI and large language models out of critical GMP applications and requires a qualified person to check outputs in non-critical ones. It was still a draft as of September 2026.

The direction is consistent across all three: define the intended use, scale the evidence to the risk, keep a qualified person responsible and document how the result was produced. The guide [Page on this site: Agentic AI in GxP environments: a practical guide] covers what that means for operational AI.

What to do with it now

For each AI use, write down the question of interest and the context of use, even where the guidance does not formally apply. It is a useful discipline for any AI system.
Decide whether the output feeds a regulatory decision. If it does, plan credibility evidence in proportion to model influence and decision consequence.
For operational uses, apply the operation's existing validation, data integrity and change control practices.
Talk to FDA early for novel uses. The guidance encourages sponsors to engage early on AI they intend to rely on.

Questions

Is FDA's AI guidance for drug development final?: No. FDA issued it as a draft in January 2025, and as of September 2026 the guidance page still labels it a draft. Sponsors can comment on drafts, and FDA may revise it before finalizing.
Does the guidance apply to AI used in manufacturing?: It applies where AI produces information used to support a regulatory decision, and it names facilitating the selection of manufacturing conditions as an example. It excludes AI used for operational efficiencies that do not affect patient safety, drug quality or study reliability.
What is a context of use?: In FDA's draft, the context of use defines the specific role and scope of an AI model in addressing a question of interest. It determines how much credibility evidence is needed.
How is this different from the EU's Annex 22 draft?: FDA's draft is about establishing the credibility of AI used to support regulatory decisions. The EU GMP Annex 22 draft sets requirements for AI models used in GMP manufacturing and quality operations, and it restricts generative AI in critical applications. Both were still drafts in September 2026.

Sources

Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products. Draft guidance., U.S. FDA, January 1, 2025
EMA and FDA set common principles for AI in medicine development. , EMA, January 14, 2026
Stakeholders' consultation on EudraLex Volume 4, Chapter 4, Annex 11 and new Annex 22., European Commission, July 1, 2025

September 21, 2026Why unstructured records slow regulated operations
In pharmaceutical operations, much of the evidence behind a batch sits in unstructured records: paper batch records, supplier certificates of analysis, lab reports and deviation records that no system can query. The records are complete and controlled, but answering a question across them means finding, reading, and retyping them by hand. That slows investigations, supplier qualification, and batch review, and it keeps experts on assembly work. Turning those records into structured, source-linked data removes that step without replacing the systems that hold them.