AI in CMC: what regulators allow

The short version

• AI can help to prepare, draft, rank, and find work in a CMC program. A named person must make the decision.
• If AI output supports a regulatory decision about safety, efficacy, or quality, the FDA expects a credibility case for that specific use.
• Under the EU's draft Annex 22, generative AI should stay out of critical GMP decisions, such as batch release. This position is not final yet.

What the FDA draft guidance asks for

The FDA draft guidance of January 2025 asks sponsors to define the question that the AI answers. It also asks them to define the exact role of the AI. The guidance calls this role the "context of use." Then the sponsor finds the model risk from two factors: model influence and decision consequence. Model influence is how much weight the AI output has in the decision, and decision consequence is how serious that decision is. When the risk is higher, the FDA expects more evidence of credibility.

The guidance does not address AI that a company uses only for operational efficiency. This exclusion applies only when the AI output does not affect patient safety, drug quality, or the reliability of study results. Several tools on this site stay in that space by design. But a tool is not exempt only because it is project-management software. The intended use decides, and the key question is whether the output could affect quality. The guidance is still a draft.

What the EU's draft Annex 22 says

The EU published the draft of EU GMP Annex 22 on 7 July 2025. The draft names three types of AI that companies should not use in critical GMP applications. The first type is a model that changes while people use it. The second type is a model that can give different answers to the same input. The third type is generative AI. In practice, the draft keeps these models out of batch disposition, release decisions, and process control.

The consultation closed on 7 October 2025 with about 1,300 comments. The drafting group now reviews whether regulators could allow these models under risk-based safeguards. The target for the final text is late 2026. Until then, the draft gives the clearest signal of the direction that regulators will take.

What this means for AI in a CMC program

How the tools on this site stay inside the limits

• A named person makes each decision. The tools prepare, rank, and draft work, but they never approve it.
• Each claim shows its source document and page, so a reviewer can check it.
• Each tool states its context of use and risk level on its own page.
• By design, none of my tools has a part in batch release, disposition, or process control.

The next change to watch: ICH M4Q(R2)

ICH M4Q(R2) changes the structure of the quality part of the submission. Module 3 becomes a data repository on a standard template (description, manufacture, control, storage) with explicit links to the Module 2.3 summary. The target for final adoption (Step 4) is June 2027, and the expected start of implementation is 2028. The difficult task for teams will be to map existing Module 3 content into the new structure without broken cross-references. The Module 3 tool on this site focuses on that problem.

See the Module 3 tool →

Sources

FDA draft guidance

Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products

Draft, January 2025. It gives a risk-based framework with 7 steps to assess credibility.

EU GMP Annex 22

Artificial Intelligence

Draft published 7 July 2025 · consultation closed 7 October 2025 (about 1,300 comments) · final text targeted Q4 2026 · not yet in force

EMA reflection paper

Use of Artificial Intelligence (AI) in the medicinal product lifecycle

Adopted by CHMP and CVMP, 30 September 2024. Risk-based, human-centred approach

ICH M4Q(R2)

Common Technical Document: Quality (Modules 2.3 and 3 restructure)

Step 2 draft May 2025 · FDA draft January 2026 · comments closed March 2026 · Step 4 target June 2027

See how each tool applies this