AI and Machine Learning in GMP Drug Manufacturing: FDA and EU Regulatory Expectations

AI and machine learning in GMP drug manufacturing: what FDA and EU regulators expect

Neither FDA nor the EU has a stand-alone binding rule on artificial intelligence (AI) in pharmaceutical manufacturing. Regulators on both sides apply existing GMP principles to AI and add AI-specific expectations through discussion papers, draft guidance and draft GMP text. The positions are already concrete:

• FDA consulted on AI in drug manufacturing in 2023. Its January 2025 draft guidance sets out a risk-based credibility framework that covers manufacturing uses. A 2026 warning letter shows that CGMP quality-unit obligations apply in full to AI-generated content.
• The EU and PIC/S published draft Annex 22 (Artificial Intelligence) on 7 July 2025, together with a revised Annex 11 and a revised Chapter 4. Annex 22 is the most prescriptive text so far on how to validate AI models in GMP. EMA's 2024 reflection paper adds lifecycle-wide principles.

The two regions reach similar conclusions on model validation and data governance. Both require a defined intended use or context of use, test data kept independent of training data, risk-proportionate validation, change control, ongoing performance monitoring, and human accountability for decisions.

FDA

The 2023 discussion paper "Artificial Intelligence in Drug Manufacturing"

FDA announced the discussion paper in a Federal Register notice published 1 March 2023, with a request for information and comments ³¹. The paper set out areas FDA had identified as it evaluated how its existing risk-based regulatory framework applies to AI in pharmaceutical manufacturing. Its focus was drugs marketed under an NDA, ANDA or BLA, and it invited both CDER and CBER stakeholders to respond ¹⁷⁵¹⁷⁶. The notice states that AI may play a significant role in monitoring and controlling advanced manufacturing processes ¹⁷⁷. Comments were initially due 1 May 2023 under docket FDA-2023-N-0487 ¹⁸⁰¹⁸¹, and FDA later reopened the comment period ³³.

Example applications described in the paper ⁵²³⁵²⁴:

• Advanced process control that adjusts manufacturing parameters using real-time data.
• Models that support in-process or final-product testing, real-time release testing, or prediction of in-process quality attributes.
• Trend monitoring of complaints and deviation reports to find clusters of problems and set improvement priorities.

Areas of consideration ⁵²¹⁵²²:

• How cloud applications could affect oversight of manufacturing data and records.
• How the greater volume of AI-generated data could affect data management.
• How AI applications could challenge regulatory oversight.
• AI-model elements in a CGMP environment, validation and maintenance of self-learning models, and appropriate data management.

FDA also noted that there were limited standards and guidance for establishing a model's credibility ¹³⁴¹³⁵.

Stakeholder comments asked for clearer guidance on model validation, maintenance and documentation, and for harmonization with other regulators ¹³⁶¹³⁷. Industry groups asked FDA to separate manufacturing/CGMP uses from other drug-development AI, to make submission and inspection documentation proportionate to risk, and to clarify controls for third-party AI and cloud services ^138139140. These were requests from industry, not FDA requirements.

FDA leadership has spoken positively about adoption. In September 2023, then-CDER director Patrizia Cavazzoni said AI combined with advanced manufacturing could strengthen quality control, reduce defects and potentially help prevent quality-related shortages ¹⁴¹.

The 2025 draft guidance on AI to support regulatory decision-making

On 7 January 2025, FDA issued the draft guidance Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products ⁴⁶³⁴. It is nonbinding and marked "Draft — Not for Implementation" ²²⁸²²⁹.

Scope. The guidance covers AI models used in the drug product lifecycle to produce information or data that support regulatory decisions on safety, effectiveness or quality. Manufacturing is in scope when an AI use serves that purpose. Two uses are excluded: drug discovery, and uses aimed only at operational efficiency that do not affect patient safety, drug quality or the reliability of study results ⁴⁰⁵⁴⁰⁶.

Seven-step credibility framework ^206407408:

1. Define the question of interest.
2. Define the context of use (COU).
3. Assess model risk.
4. Develop a credibility assessment plan.
5. Execute the plan.
6. Document results and deviations.
7. Determine whether the model is adequate for the COU.

Model risk. Model risk combines model influence, meaning how much the AI output contributes relative to other evidence, with decision consequence, meaning how serious an adverse outcome from a wrong decision would be. It describes the risk that a wrong decision follows from the model's output, not risk intrinsic to the model ³⁸⁹.

Manufacturing example. The guidance uses AI-based visual inspection of fill level in every vial of a multidose injectable, where fill volume is a critical quality attribute. Decision consequence is high. Separate fill-volume measurement on a representative sample from each batch lowers model influence, so overall model risk is medium. FDA notes that its examples cover only steps 1 to 3 ^406391383.

Data expectations:

• Development data should be relevant to the COU, including sufficiently representative manufacturing-process data, and reliable, meaning accurate, complete and traceable ²³⁰²³¹.
• In proportion to model risk, the credibility plan should describe how data were selected, collected, processed, annotated, stored and controlled, how labels were set, and how data were split into training, tuning and other subsets. These descriptions also help identify algorithmic bias ²⁰³²³².
• Test data should be independent of training and tuning data, for example drawn from a different batch or product. Any overlap must be justified ²⁰⁴.
• Performance metrics should be reported with confidence intervals for all estimates. Sponsors should explain how prediction uncertainty was estimated and describe limitations, including potential biases. Where a human is in the loop, the human–AI team should be evaluated together ²⁰⁴.

Lifecycle maintenance. FDA states that lifecycle maintenance is particularly important in pharmaceutical manufacturing, because model performance can change over time or across deployment environments ⁴⁰⁰. Its expectations:

• Monitor performance on an ongoing basis.
• Evaluate model changes, and manufacturing changes that could affect the model, through the change-management system within the pharmaceutical quality system ⁴⁰⁰.
• Keep risk-proportionate maintenance plans (metrics, monitoring frequency, retesting triggers) available in the site quality system. For product- or process-specific models, include a summary in the marketing application.
• Report model changes that affect performance in line with applicable requirements. The postapproval reporting category can depend on the effect on both model performance and product quality ³⁹⁷⁴⁶⁶.
• Sponsors may use ICH Q12 tools, such as proposing model-related established conditions and a change-management plan ³⁹⁷.

FDA strongly encourages early engagement to agree on credibility activities ⁴⁶⁷.

Enforcement and inspection signals

In January 2026, Tina Kiang of FDA's Office of Pharmaceutical Quality said that AI produces outputs, not decisions, and that the quality control unit remains responsible for how those outputs are used ¹⁰⁸¹⁰⁹. She said manufacturers should:

• Use AI in accordance with CGMP.
• Validate models for their specific manufacturing context of use.
• Keep records available for inspectors.
• Scale the validation effort to model risk, decision consequence and model influence ¹¹⁰¹¹¹.

On 2 April 2026, CDER issued a warning letter to Purolea Cosmetics Lab ¹⁸². The firm had used AI agents to create drug product specifications, procedures and master production or control records ²³⁶. FDA wrote that "if you use AI as an aid in document creation, you must review the AI generated documents to ensure they were accurate and actually compliant with CGMP." It cited the failure to do so as a violation of 21 CFR 211.22(c). FDA also gave an example of overreliance: the firm had not performed process validation under 21 CFR 211.100 because its AI agent "never told you it was required" ²³⁷. FDA stated that any AI output or recommendation used for CGMP activities "must be reviewed and cleared by an authorized human representative of your firm's QU" ²³⁸.

European Union and PIC/S

Draft EU GMP Annex 22 (Artificial Intelligence)

The European Commission proposed Annex 22 to EudraLex Volume 4, together with revisions to Annex 11 and Chapter 4. The consultation ran from 7 July to 7 October ^167171172102. The drafts also appear in the PIC/S publications listing, dated 7 July 2025 and still shown as draft status ^183201186. Annex 22 provides additional guidance to Annex 11 for computerised systems with embedded AI models ²⁹⁸. The Annex 11 currently in force dates from 2011 (in operation since 30 June 2011) ¹¹⁵.

Scope. Annex 22 applies to machine-learning models embedded in computerised systems used to manufacture medicinal products and active substances, where the model is used in a critical application with a direct impact on patient safety, product quality or data integrity ³⁷³. The table below summarises what is covered.

Principles. Process subject-matter experts (SMEs), QA, data scientists, IT and consultants should cooperate throughout model selection, training, validation, testing and operation ³⁷⁶. The regulated user must have access to, and review, the documentation for these activities, including for supplier-provided models ³⁷⁷. All activities follow quality risk management based on risk to patient safety, product quality and data integrity ³⁷⁸.

Model validation requirements:

• Intended use. Describe the model's tasks and characterise its input data, including common and rare variations, limitations and possible biased inputs. A process SME is accountable for this description, which must be approved before acceptance testing ³²⁴³²⁵.
• Acceptance criteria. Use case-appropriate metrics. For classification models, examples are a confusion matrix, sensitivity, specificity, accuracy, precision and F1 score. Criteria may be set per subgroup and must be at least as high as the known performance of the process the model replaces ³²⁶.
• Test data. Use stratified data that cover the full intended-use sample space. The dataset must be large enough for statistically confident metrics, and labels must be verified to a very high degree of correctness. Any pre-processing must be pre-specified, and any exclusions fully justified. Generating test data or labels with generative AI is not recommended ³²⁷³²⁸.
• Test data independence. Use technical and/or procedural controls to keep test data out of development, training and validation. Repositories must be access-controlled and audit-trailed. Record which data were tested and when. Staff with access to test data should not train or validate the same model, except under a four-eyes arrangement where separation is impossible ³²⁹.
• Test execution. Follow a pre-approved test plan that includes metrics, criteria and an executable script. Deviations and failures must be investigated and justified. Test records are retained together with the access-control and audit-trail records ³³⁰³³¹.
• Explainability. For critical applications, record which features drove each decision during testing, using methods such as SHAP, LIME or heat maps where applicable. A risk-based review of those features forms part of approving test results ⁴²⁵.
• Confidence. Log a confidence score for each outcome where applicable. Set thresholds, and consider flagging very low-confidence outcomes as "undecided" ⁴²⁶.

Operation:

• Change and configuration control. The model, its implementing system and the process it supports must be under change and configuration control before deployment. Any decision not to retest after a change must be fully justified ⁴²⁷.
• Monitoring. Performance metrics must be monitored regularly, and input-data drift must be monitored against the model's sample space ⁴²⁸⁴²⁹.
• Human review. Where testing effort was reduced because a human reviews outputs, records of that human-in-the-loop review must be kept ⁴³⁰.

Draft revised Annex 11 and Chapter 4: the data governance foundation

Draft revised Annex 11 sets the general controls for computerised systems ²⁰¹:

• Audit trails must capture who, what, when and why, remain enabled and protected from editing, and be reviewed in a risk-based way, normally before batch release ³⁰⁵³⁰⁶.
• Security expectations include an information security management system, least privilege, segregation of duties, MFA for remote access to critical systems, and network segmentation ^{307308309310311}.
• Service providers. Relying on a vendor or service provider does not remove the regulated user's responsibility. Oversight is expected through audits, SLAs, KPIs and contracts ³¹².
• Periodic review must be documented and its frequency justified by risk to product quality, patient safety and data integrity ³⁰¹³¹³.

Draft revised Chapter 4 extends data integrity to an ALCOA++ standard for paper, electronic and hybrid records ²⁶³²⁶⁴. It requires:

• A documented, regularly reviewed data governance system within the pharmaceutical quality system, with assigned data ownership and controls proportionate to data criticality ^265266267.
• Governance across the full data lifecycle, from creation through processing, review, decisions, retention and destruction ^270271272.
• Retention of relevant metadata as part of complete records ²⁷³²⁷⁴.
• Validation of hybrid systems, including their manual-to-computerised interfaces ²⁷⁵²⁷⁶.

EMA reflection paper on AI in the medicinal product lifecycle

EMA published the final reflection paper on 30 September 2024 ⁷⁸, following a July 2023 draft ⁷⁹. Its manufacturing section expects AI/ML use in process design, scale-up, optimisation, in-process control and batch release to increase. It states that model development, performance assessment and lifecycle management should follow quality risk management principles, and that for human medicines ICH Q8, Q9 and Q10 should be considered "awaiting revision of current regulatory requirements and GMP standards" ⁵³⁹. The paper describes itself as conveying EMA's current experience and says it should be read alongside applicable legal requirements ⁵⁴¹.

Its general expectations on validation and data overlap with Annex 22:

• Data. Training data should be sufficiently large and balanced. Data sourcing and processing should be traceable and consistent with GxP. For high-impact or high-risk uses, GxP data and algorithm governance SOPs are expected ³⁵⁵³⁵⁶.
• Test sets. Final performance should be assessed once on a hold-out test set. A model that is redeveloped needs a new, independent test set. Data should be split before normalisation, and high-risk models should be prospectively tested on newly acquired data ³⁵⁷.
• Performance. Metrics should be insensitive to class imbalance, performance thresholds should be predefined, and performance should be monitored after deployment ^358359360.
• Explainability. Transparent models are preferred. Black-box models must be justified, and SHAP or LIME should be used where possible ³⁶¹.
• Changes. In high-risk, high-impact uses, non-trivial software or hardware changes require re-evaluation of performance ³⁵³³⁵⁴.

Where FDA and EU expectations converge

Practical implications for manufacturers

• Classify each AI use by criticality first. Under draft Annex 22, that classification decides whether the model is allowed at all in a critical GMP application ³⁷³³⁷⁴. Under FDA's framework, it decides how much credibility evidence is needed ³⁸⁹.
• Build test-data independence into data architecture early. Both regions expect demonstrable separation of test data, and the EU draft expects access controls and audit trails to enforce it ²⁰⁴³²⁹.
• Do not treat generative AI as a substitute for quality-unit judgment. The Purolea letter shows FDA will cite AI-generated CGMP documents under existing 21 CFR Part 211 obligations when they are not adequately reviewed ²³⁷²³⁸. Draft Annex 22 keeps generative AI out of critical applications ³⁷⁵.
• Plan lifecycle controls before deployment. That means monitoring metrics, drift triggers and retesting criteria, documented within the pharmaceutical quality system ^400397427428.