AI in drug manufacturing: where FDA stands and what to build for
By QLogix · Published · 4 min read
FDA has no final guidance written only for AI in drug manufacturing. It has a 2023 discussion paper naming five open areas and a 2025 draft framework that covers manufacturing models. Together they show what a process model has to prove.
- Areas of consideration in FDA's 2023 discussion paper on AI in drug manufacturing.
- 5
- Year of FDA's PAT guidance, the starting point for model-based control.
- 2004
- Draft credibility guidance, which covers manufacturing models.
- Jan 2025
Key takeaways
- FDA's drug center published the discussion paper Artificial Intelligence in Drug Manufacturing on 1 March 2023. It asks questions; it sets no requirements.
- The January 2025 draft guidance on AI credibility applies to models that support decisions about drug quality, which includes models used in manufacturing.
- A model that advises an operator carries less regulatory weight than one that controls the process or replaces a release test.
- Existing CGMP rules already apply: equipment and automated systems under 21 CFR 211.68, in-process controls under 211.110, and change control for any model update.
Does FDA have guidance on AI in drug manufacturing?
Not a final guidance written only for that subject. What exists is a line of documents that build on one another. The 2004 guidance on process analytical technology (PAT) accepted real-time measurement and model-based control as part of pharmaceutical quality. ICH Q13, published by FDA as final guidance in March 2023, set expectations for continuous manufacturing, where process models and automated control do most of the work.
In March 2023 FDA's Center for Drug Evaluation and Research published a discussion paper on AI in drug manufacturing and asked for comment, followed by a public workshop with PQRI in September 2023. In January 2025 the agency published its draft guidance on establishing the credibility of AI models used to support regulatory decisions for drugs and biological products. That draft applies across the product life cycle, manufacturing included.
FDA documents that shape AI in drug manufacturing
Sep 2004Final
PAT framework guidance
Real-time measurement and model-based control accepted as part of quality assurance.
Mar 2023Final
ICH Q13, continuous manufacturing
Expectations for process models, control strategy and batch definition in continuous processes.
Mar 2023Discussion paper
Discussion paper: AI in drug manufacturing
Five areas of consideration, published for comment.
Sep 2023Workshop
FDA and PQRI workshop
Public workshop on the regulatory framework for AI in manufacturing.
Jan 2025Draft
Draft guidance: AI to support regulatory decision-making
Seven-step credibility framework, covering manufacturing models.
What are the five areas in FDA's discussion paper?
The paper covers drug products marketed under an NDA, ANDA or BLA. It names five areas where the current framework raises questions, and states that the list is not exhaustive.
- Cloud applications may affect oversight of pharmaceutical manufacturing data and records.
- The Internet of Things may increase the amount of data generated during manufacturing, which affects existing data management practices.
- Applicants may need clarity about whether and how the application of AI in manufacturing is subject to regulatory oversight.
- Applicants may need standards for developing and validating AI models used for process control and to support release testing.
- Continuously learning AI systems that adapt to real-time data may challenge regulatory assessment and oversight.
Where does AI sit in a manufacturing control architecture?
The regulatory weight of a model depends on where its output goes. The same soft sensor can be a monitoring aid, an input to a control loop or the basis for releasing a batch without an end-product test. Each step up moves decisions from a person to the model, and raises what the model has to prove.
This is the same idea as model influence in the 2025 draft guidance. A design that keeps an independent measurement and a qualified person in the release decision keeps the model's influence low.
From sensor to release decision
- People decideBatch dispositionQuality unit reviews the record and releases or rejects the batch.
- People decideReview of alarms and excursionsOperators and QA assess model alerts, with the reason recorded.
- AIProcess modelSoft sensor, multivariate model or digital twin predicts a quality attribute and flags drift.
- DataData platformContextualized, time-aligned process and laboratory data with audit trails.
- DataSensors and analyzersPAT instruments, equipment signals, in-process tests.
What does a manufacturing model have to show?
- A stated context of use: which attribute it predicts, on which product and equipment, and what the output is used for.
- Development data that covers the process range the model will see, including the variation between lots, materials and seasons.
- Performance on data that played no part in building it, against acceptance criteria set in advance.
- Behaviour outside its range: what happens when an input drifts or a sensor fails, and how the system falls back to a conventional control.
- Monitoring in routine use, with triggers for review and retraining.
- Change control for every model update. FDA's paper notes that deployed models are updated as needed through the change control process, which is the practical limit on continuous learning today.
How should a manufacturing science team start?
- Pick a use where the model advises and a person decides, such as predicting a drift before it reaches an action limit.
- Get the data foundation right first. A model trained on data without context or audit trails cannot be defended in an inspection.
- Write the context of use before building, and size the validation to the influence the model will have.
- Talk to FDA early for uses that touch the control strategy or release testing. The Emerging Technology Program exists for that conversation.
Common questions
Is there a final FDA guidance on AI in drug manufacturing?
No. As of October 2026 FDA has a 2023 discussion paper on AI in drug manufacturing and a January 2025 draft guidance on AI credibility that covers manufacturing models. Existing CGMP regulations and the PAT and ICH Q13 guidances continue to apply.
Can a continuously learning model be used in GMP manufacturing?
Not as a model that changes by itself in production. FDA's discussion paper lists continuously learning systems as an open challenge, and in practice models are locked and updated through change control with revalidation.
Does a digital twin need to be validated?
Yes, when its output supports a GMP decision. The depth of validation follows its use: a twin used to explore scenarios offline needs less than one that feeds a control loop or supports batch release.
Sources
- FDA CDER discussion paper: Artificial Intelligence in Drug Manufacturing (March 2023)
- FDA draft guidance: Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products (January 2025)
- FDA guidance: Q13 Continuous Manufacturing of Drug Substances and Drug Products (March 2023)
- FDA guidance: PAT, A Framework for Innovative Pharmaceutical Development, Manufacturing, and Quality Assurance (September 2004)
This article is general information for quality and regulatory professionals. It is not regulatory or legal advice. Check the source documents for their current status before acting on them.
More from the blog
- AI in quality · 12 October 2026FDA's seven-step credibility framework for AI, in plain termsFDA's draft guidance asks sponsors to show that an AI model is credible for a specific context of use. The effort scales with model risk: how much the model's output drives the decision, and how serious a wrong decision would be.
- Medical devices · 12 October 2026AI-enabled medical devices: what FDA's list and guidance showFDA's public list counted 1,614 AI-enabled medical devices through June 2026, about three in four of them in radiology. Two documents shape how they are built and changed: the final guidance on predetermined change control plans and the draft guidance on lifecycle management.
- Manufacturing science · 12 October 2026Robotics in aseptic manufacturing: what Annex 1 expectsThe revised EU GMP Annex 1 names robotic systems, with isolators and RABS, as technologies to consider for protecting sterile products. Taking the operator out of the critical zone removes the largest contamination source and moves the quality work to design, software and data.