Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products (draft)

Context and history

This is FDA's draft guidance on using AI to support regulatory decisions on drugs and biological products. It was published on 7 January 2025 for comment, with comments closing on 7 April 2025. It proposes a risk-based credibility framework. As a draft, it is not yet for implementation, and the FDA-EMA guiding principles of January 2026 are expected to shape later thinking.

The central idea is credibility: how far can you trust the output of an AI model for one specific use? FDA's answer depends on that use, which it calls the context of use. A model trusted for one purpose has not earned trust for another. The guidance is built around a seven-step framework for planning, gathering and documenting that evidence.

Who and what it applies to

Sponsors and other interested parties using AI models to produce information or data supporting regulatory decisions on safety, effectiveness or quality of drugs, across the drug product life cycle, including manufacturing.

Not covered: Two uses are left out: AI in drug discovery, and AI used for operational efficiencies that do not affect patient safety, drug quality or the reliability of study results. FDA gives internal workflows, resource allocation and drafting a regulatory submission as examples. FDA encourages sponsors to talk to it early if they are unsure whether a use is in scope.

What it contains

Step 1: Define the question of interest

State the specific question, decision or concern the model will help address.

Step 2: Define the context of use

Describe the model's role and scope, and whether other evidence is used alongside it.

Step 3: Assess model risk

Model risk combines model influence (how much the output counts against other evidence) with decision consequence (how serious a wrong decision would be). FDA stresses this is the risk that the output leads to a wrong decision, not a risk built into the model.

Step 4: Develop a credibility plan

Describe the model, the data used to develop it, its training, and how it will be evaluated.

Step 5: Execute the plan

Carry out the planned activities.

Step 6: Document results and deviations

Record what happened, including departures from the plan.

Step 7: Determine adequacy

Decide whether the model is adequate for the context of use.

Life cycle maintenance

Managing changes to a model, deliberate or incidental, because performance can shift over time or between deployment settings.

Early engagement

Options for discussing AI use with FDA before relying on it.

Key points

• Credibility is judged for a defined context of use, not in general.
• Rigour, oversight and documentation should match model risk.
• A model that is the sole determinant of a decision will usually carry more influence than one checked by other evidence.
• Model performance can drift, so maintenance planning matters across the life cycle.
• It is a draft and is marked not for implementation.

In practice

General reading, not advice. This is drafted with AI and may not fit your situation. It does not replace the official text, your own quality system or qualified advice, and an inspector may take a different view.

• Write down the question and context of use before choosing or buying a model. Most later decisions depend on them.
• Keep a simple record of influence and consequence for each model, with who decided and why.
• Treat drafting tools and decision-support models differently. The first may fall outside this guidance, and the second may not. Still check GMP expectations for either.
• Use this framework's logic for manufacturing AI too, and read it alongside EU draft Annex 22.

Read from the source document on 1 Nov 2026. This is our own description, not the official text. Check the full dated history in the change log below.