1What FDA has published.
The guidance CDER lists for AI in drug development is a January 2025 draft. Around it sit ten principles agreed with EMA and two discussion papers that asked industry for input.
| Document | Date | Status |
|---|---|---|
| Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products | January 2025 | Draft guidance, “Not for implementation”, docket FDA-2024-D-4689¹ |
| Guiding Principles of Good AI Practice in Drug Development, with EMA | January 2026 | Ten principles that industry and product developers “can consider”² |
| Artificial Intelligence in Drug Manufacturing | March 2023 | Discussion paper, “not a draft or final guidance”³ |
| Using Artificial Intelligence and Machine Learning in the Development of Drug and Biological Products | May 2023, revised February 2025 | Discussion paper and request for feedback⁴ |
CDER says the draft guidance drew on more than 800 comments on its May 2023 discussion paper on AI in drug development and on its experience with over 500 submissions with AI components from 2016 to 2023.⁵ As with FDA guidance in general, it does not establish legally enforceable responsibilities; it describes the agency’s current thinking.⁶
2What the draft guidance covers.
The draft applies to AI models used in the drug product life cycle “to produce information or data to support regulatory decision-making regarding safety, effectiveness, or quality for drugs”. Regulatory decision-making includes actions sponsors and other parties take under FDA’s authority, naming CGMPs as an example, and the interested parties it addresses include manufacturers that are not sponsors.⁶ Its drug product life cycle covers nonclinical, clinical, postmarketing and manufacturing phases.⁶
Two uses are out of scope: drug discovery, and 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 from a nonclinical or clinical study”. FDA asks sponsors to engage early if they are unsure whether a use is in scope.⁶
3The seven-step credibility assessment.
| Step | What it asks |
|---|---|
| 1 | Define the question of interest the model addresses |
| 2 | Define the context of use: the model’s role and scope, and what other evidence is used alongside it |
| 3 | Assess model risk, from model influence and decision consequence |
| 4 | Plan how to establish the model’s credibility in that context of use |
| 5 | Execute the plan |
| 6 | Document the results and any deviations in a credibility assessment report |
| 7 | Decide whether the model is adequate for the context of use |
The manufacturing example is an AI vision system checking fill volume in multidose vials. Volume is a critical quality attribute, so decision consequence is high, but because release testing independently measures fill volume on a sample from each batch, model influence is low, and the model risk comes out as medium.⁶
The plan in step 4 should describe the model, its inputs, outputs, architecture and features; the development data, including how it was split into training and tuning sets and why it is relevant and reliable; how the model was trained; and how it was evaluated. Test data “should be independent of the development data”, performance estimates come with confidence intervals, and where a human is in the loop, the evaluation should consider “the performance of the human-AI team, rather than just the performance of the model in isolation”.⁶ If credibility is not established, the options include adding other evidence to reduce the model’s influence, more rigorous assessment, added controls, a different modelling approach, or judging the output inadequate, so that the context of use is rejected or revised.⁶
4Keeping a manufacturing model credible.
FDA singles out manufacturing as a phase where life cycle maintenance matters. Changes to an AI model, or changes in manufacturing that may affect it, should be evaluated through the manufacturer’s change management system within its pharmaceutical quality system. Depending on the impact, parts of the credibility assessment may need to be repeated, including retraining and retesting, and a change that affects model performance should be reported to FDA in line with regulatory requirements.⁶ Detailed maintenance plans, with performance metrics, monitoring frequency and retesting triggers, should be available as part of the site’s quality system, with a summary in the marketing application for product or process-specific models. Sponsors may also use ICH Q12 tools such as established conditions. For AI in manufacturing, FDA highly encourages early engagement with CDER’s Emerging Technology Program or CBER’s Advanced Technologies Team, before submitting an application or implementing the technology.⁶
5The FDA and EMA principles.
In January 2026 CDER and CBER published ten guiding principles developed with EMA for AI used to generate or analyse evidence across the drug product life cycle, manufacturing included: human-centric by design, a risk-based approach, adherence to standards including GxP, a clear context of use, multidisciplinary expertise, data governance and documentation, model design and development practices, risk-based performance assessment of the complete system including human-AI interaction, life cycle management, and clear, essential information.⁷
6Where CGMP already applies.
None of this replaces the CGMP regulations. The draft guidance says AI used in manufacturing “must be implemented in accordance with current good manufacturing practice”, and that for finished drug products the quality control unit’s responsibilities in 21 CFR 211.22 and 211.68 apply.⁶ Under 211.22(c) the quality control unit approves or rejects all procedures and specifications affecting the identity, strength, quality and purity of the drug product.⁸ In April 2026 FDA cited that rule against a manufacturer that had used AI agents to write specifications, procedures and master production records without reviewing them, and said AI output used for CGMP activities “must be reviewed and cleared by an authorized human representative” of the quality unit.⁹ The 2023 manufacturing discussion paper lists the open questions FDA asked industry about, including cloud data and records, standards for validating models used in process control and release testing, and continuously learning systems that adapt to real-time data.³
7How it runs in Seal.
Neil, Seal’s AI agent, drafts and proposes; people sign, approve and release. Every finding links to the record, version and field behind it, and Neil says what the records don’t show. A change Neil proposes runs its tests, and your team releases it.
Evaluating Neil works job by job, much as FDA’s draft works one context of use at a time. The validation and AI assurance guide asks you to evaluate Neil on representative jobs with expert-reviewed acceptance criteria, including obsolete procedures, conflicting records and insufficient evidence, and to run the same cases again after a relevant model or configuration change, keeping failures as well as successes. For each AI-assisted decision it asks you to keep the request, the source records and versions, the output and the reviewer’s decision.
References
- 1Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products, guidance page, FDA, January 2025. https:/
/ www. fda. gov/ regulatory- information/ search- fda- guidance- documents/ considerations- use- artificial- intelligence- support- regulatory- decision- making- drug- and- biological (read 6 November 2026) - 2Guiding Principles of Good AI Practice in Drug Development, web page, FDA, January 2026. https:/
/ www. fda. gov/ about- fda/ artificial- intelligence- drug- development/ guiding- principles- good- ai- practice- drug- development (read 6 November 2026) - 3Artificial Intelligence in Drug Manufacturing, discussion paper, FDA Center for Drug Evaluation and Research, March 2023. https:/
/ www. fda. gov/ media/ 165743/ download (read 6 November 2026) - 4Using Artificial Intelligence & Machine Learning in the Development of Drug & Biological Products, Discussion Paper and Request for Feedback, FDA, May 2023, revised February 2025. https:/
/ www. fda. gov/ media/ 167973/ download (read 6 November 2026) - 5Artificial Intelligence for Drug Development, FDA Center for Drug Evaluation and Research, updated 1 May 2026. https:/
/ www. fda. gov/ about- fda/ center- drug- evaluation- and- research- cder/ artificial- intelligence- drug- development (read 6 November 2026) - 6Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products, Draft Guidance for Industry and Other Interested Parties, FDA, January 2025. https:/
/ www. fda. gov/ media/ 184830/ download (read 6 November 2026) - 7Guiding Principles of Good AI Practice in Drug Development, FDA and EMA, January 2026. https:/
/ www. fda. gov/ media/ 189581/ download (read 6 November 2026) - 821 CFR 211.22, Responsibilities of quality control unit, eCFR. https:/
/ www. ecfr. gov/ current/ title- 21/ section- 211. 22 (read 6 November 2026) - 9Warning Letter 320-26-58, Purolea Cosmetics Lab, MARCS-CMS 722591, FDA, 2 April 2026. https:/
/ www. fda. gov/ inspections- compliance- enforcement- and- criminal- investigations/ warning- letters/ purolea- cosmetics- lab- 722591- 04022026 (read 6 November 2026)
AQuestions and answers
Is FDA’s AI guidance for drugs final?
No. The January 2025 guidance on using AI to support regulatory decision-making is a draft, marked not for implementation. When final, it will describe FDA’s current thinking, but like all guidance it will be a set of recommendations rather than enforceable requirements.
Does the draft cover AI used to write SOPs or submissions?
Drafting a regulatory submission is named as an operational efficiency outside its scope, provided the use does not affect patient safety, drug quality or the reliability of study results. AI-drafted SOPs fall under the CGMP regulations instead, and FDA has said that AI-drafted procedures and specifications must be reviewed and cleared by the quality unit.
What is a context of use?
The specific role and scope of the AI model in answering a question: what it models, how its output is used, and what other evidence is used with it. FDA judges a model’s credibility, and the evidence needed, for one context of use at a time.
How does FDA judge the risk of an AI model?
By combining model influence, meaning how much the model’s output counts against other evidence, with decision consequence, meaning how bad a wrong decision would be. A model that is the only basis for a high-consequence decision is high risk; independent testing alongside it lowers its influence and so its risk.
Should we talk to FDA before using AI in manufacturing?
FDA encourages it. For AI in pharmaceutical manufacturing, the draft guidance points to CDER’s Emerging Technology Program and CBER’s Advanced Technologies Team, and highly encourages early engagement before submitting an application or implementing the technology.
Does this guidance apply to AI in medical devices?
It is written for drugs and biological products, including drug constituents of combination products. For devices, FDA points sponsors to the device-specific guidance from its Center for Devices and Radiological Health, although it notes its recommendations may be relevant across medical products.