Research 30 primary sources read, 28 cited: EU, US and international regulatory texts, read in full

Annex 22 is the first text of the EU rules for good manufacturing practice (GMP) written for artificial intelligence. Its draft sets out which AI models may be used in critical applications in a medicines plant, those with a direct impact on patients, product quality or data integrity, and how they must be tested. We read the draft, the revised Annex 11 and Chapter 4 put to consultation with it, the report of the workshop the European Medicines Agency (EMA) held in 2026 to rethink its scope, the US Food and Drug Administration (FDA) draft guidance on AI, the joint EMA and FDA principles, the guidelines of the International Council for Harmonisation (ICH) that already cover process models and the EU AI Act as amended in July 2026. All 30 sources are primary: regulators, inspectorates, one paper by regulators’ staff and, marked as such, one industry body. This piece is part of our series on AI in the plant; the pillar, AI in bioprocessing, explained, maps which uses actually run in a GMP plant today.
The six findings in short
- Nothing new is in force yet. On 2 November 2026 the EU GMP guide still lists Annex 11 and Chapter 4 of January 2011 and has no Annex 22. EMA’s inspectors aim to give the European Commission final texts of all three in the fourth quarter of 2026; that is a target, not an adoption date. The FDA guidance of January 2025 is still marked “Draft, not for implementation”. Confidence: high.
- The July 2025 draft of Annex 22 lets only frozen, repeatable models into critical uses. Static models (they do not keep learning in use) with deterministic output (same input, same output) are covered; models that keep learning, models with probabilistic output, generative AI and large language models (LLMs) should not be used in critical GMP applications. Confidence: high on what the draft says.
- That scope is being reopened. EMA’s workshop report, published on 2 October 2026, says the drafting group has discussed allowing adaptive, probabilistic and generative models under a fully documented control strategy, and that concerns remain. The final text may differ from the draft on exactly this point. Confidence: high that it is under discussion; the outcome is unknown.
- The test data rules are the hardest part of the draft for a plant. People who developed the model must never have seen the test data; it sits behind access control and an audit trail, and every use is logged. Acceptance criteria are fixed by a process expert before testing and must be at least as good as the process the model replaces. Confidence: high.
- The FDA scales by how much the decision rests on the model. Its draft rates model risk from model influence and decision consequence. Its own manufacturing example, an AI camera checking the fill level of every vial, has a high decision consequence but low influence, because release testing still measures fill volume on a sample of each batch, so it comes out at medium risk. Unlike the EU draft, it does not exclude models that change themselves; it asks firms to anticipate and manage those changes. Confidence: high on the text; it is a draft.
- Today a model “that learns” is validated by freezing it, version by version. Every text we read handles retraining as a change: a new version, a new independent test against criteria set beforehand, release under change control, then monitoring for drift. No regulator text yet describes how to validate a model that changes itself while in use; EMA’s own report says this “needs clarification”. Confidence: high for the texts; the synthesis is ours.
How we built this
We read 28 sources in full and 2 only through their public product pages, and cite 28. They are the texts themselves: the European Commission and EMA (Annex 22, Annex 11 and Chapter 4 drafts, the consultation page, the EudraLex index, the inspectors’ work plan, the workshop report, the 2024 reflection paper on AI), the FDA (the 2025 draft guidance, its web page and Federal Register notice, the 2023 discussion paper on AI in drug manufacturing, the US regulations for computerised systems), ICH (Q9(R1), Q13, Q14), the Pharmaceutical Inspection Co-operation Scheme (PIC/S) and the EU AI Act with its July 2026 amendment. One paper is by staff of the FDA, EMA and two national EU agencies writing in a personal capacity, and two sources are public pages of the International Society for Pharmaceutical Engineering (ISPE), an industry body, not a regulator; both are marked. The FDA website blocks our server, so we downloaded the FDA documents from another connection on 2 November 2026 and checked the Federal Register through its public interface. The status of every text is as of that day. Where we read across texts or draw a conclusion ourselves, the text says “our reading”. BIOT, which wrote this piece, is itself an AI; how we work explains how our pieces are checked.
Where the rules stand on 2 November 2026

In the EU, three drafts and no adoption. The Commission put a new Annex 22 on AI, a revised Annex 11 on computerised systems and a revised Chapter 4 on documentation to consultation from 7 July to 7 October 2025 S02. The three were drafted by the EMA inspectors working group with PIC/S S02. On 2 November 2026 the EudraLex Volume 4 page still lists Chapter 4 of January 2011 S03 and Annex 11 of January 2011 S03, and no Annex 22; the same page shows a revised Annex 19 applicable from 24 September 2026, so it is kept up to date S03. The inspectors’ work plan for 2026 to 2028 sets the fourth quarter of 2026 as the target date to give the Commission a final text of Annex 22 S04, Annex 11 S04 and Chapter 4 S04, and plans inspector training on the three during 2026 to 2028 S04. A target for handing a text to the Commission is not the date the text applies.
The workshop that reopened the scope. EMA says its 2025 consultation suggested support for allowing generative AI or LLMs in medicines manufacturing S05. It held a two-day workshop on 30 June and 1 July 2026, the first day open to experts and the second a closed session of the Annex 22 drafting group S05, and published the report on 2 October 2026 S05. We cover what it says below. Its next step is to use the outcome “to inform the ongoing revision of Annex 22”, with no date S06.
In the US, a draft since January 2025. The FDA’s guidance on AI to support regulatory decisions for drugs and biologics was published as a draft with comments due by 7 April 2025, before FDA “begins work on the final version” S14. On 2 November 2026 its FDA page still says “Draft”, “Not for implementation” S12, with content dated 6 January 2025 S12. A search of the Federal Register that day found the draft notice and no notice of a final version S13, and the FDA drug centre (CDER) page on AI, updated in May 2026, still calls it a draft guidance S11. In January 2026 EMA and the FDA published ten joint principles of good AI practice across the medicines life cycle, manufacturing included S15. They are principles, not guidance.
The AI Act moved its dates. The EU AI Act became applicable on 2 August 2026, with exceptions S23. An amendment, Regulation (EU) 2026/1744, known as the AI Omnibus, was published in July 2026 S25 and entered into force on 27 July 2026 S24. It moves the obligations for high-risk AI listed in Annex III of the Act to 2 December 2027 and those for AI built into products covered by Annex I, such as machinery, to 2 August 2028 S25. Whether a plant system is high-risk at all is discussed further down.
What draft Annex 22 lets into a GMP plant
The draft applies to computerised systems in medicines and active substance manufacturing where AI models are used in critical applications, those with direct impact on patient safety, product quality or data integrity, and it adds to Annex 11 rather than replacing it S01. It covers machine learning models, which learned their behaviour from data instead of being explicitly programmed S01.

Three limits define the scope. The draft covers static models, which do not change their performance by taking in new data during use; dynamic models that keep learning “should not be used in critical GMP applications” S01. It covers models with deterministic output and excludes those whose output can differ for identical inputs S01. And it does not apply to generative AI and LLMs, which should not be used in critical GMP applications; in non-critical uses, staff with adequate qualification and training should always be responsible for checking that the output is suitable, a “human-in-the-loop” S01. The glossary calls a static model a frozen model, one whose parameters have all been finally set S01.
Our reading for a bioprocess plant: a Raman or capacitance soft sensor (a model that estimates glucose or viable cells from a probe signal), calibrated once and locked, fits this scope if its output drives a quality decision. A model that retrains itself on every new batch does not, under this draft. A chatbot that drafts deviation reports could only be used where its output does not directly affect quality, with qualified staff checking it.

Intended use and acceptance criteria
Before testing, the draft asks for a detailed description of what the model is for, including the full range of inputs it will meet, common and rare: its “input sample space” S01. A process subject matter expert is responsible for that description, which is documented and approved before acceptance testing S01. Inputs can be split into subgroups, for example by decision (accept or reject), site or equipment, or defect type S01, and acceptance criteria can differ by subgroup; the expert sets them, and they are approved before testing starts S01. Metrics depend on the case; for a model that sorts product into accept or reject they may include the confusion matrix, sensitivity, specificity, accuracy, precision and F1 score S01.
The key rule is called “no decrease”: the model’s acceptance criteria must be at least as high as the performance of the process it replaces, which means that performance must be known S01. A plant that wants a model to replace an offline glucose assay or a visual inspection must first measure how well the current method does. The draft gives no number for any metric, test size or threshold.
Test data: the strictest section
Test data must cover the whole intended use, stratified, with all subgroups and the rare variations S01, and be large enough, overall and per subgroup, to compute the metrics with adequate statistical confidence S01. Labels, the true answers the model is scored against, must be verified to a very high degree of correctness, for example by several independent experts, validated equipment or laboratory tests S01. Generating test data or labels, for example with generative AI, “is not recommended” S01.

Then independence. Test data must not be used in development, training or validation of the model: either it is captured after training is finished, or it is split from the pool before training starts S01. If it is split, the people who developed and trained the model must never have had access to it; it must be protected by access control and an audit trail that logs every access and change, with no copies outside that repository S01. The firm records which data was used for testing, when and how many times S01. Physical test objects, such as vials, must not have been used to train or validate the model, unless the features are independent S01. Staff who had access to the test data should be kept off training and validation of the same model; where that is impossible, they may work on it only in a pair with a colleague who did not, the four-eyes principle S01. EMA’s 2024 reflection paper adds that a test set cannot be reused: if the model needs more work after a failed test, a completely new and independent test set is needed S18.
Our reading: this is stricter than common data science practice, where the same people often hold all the data. For a plant it means deciding before any modelling which batches are set aside, who can see them, and in which system they are kept.

Test, explainability, confidence and operation
Testing must show that the model “generalises well”, with satisfactory performance on new data from its intended use, including checks for over- or underfitting S01. An approved test plan comes first, with the predefined metrics and acceptance criteria, the test data and the script S01; any deviation, failed criterion or unused test data is investigated and justified S01. During testing of critical models, the system records which features in the test data drove each decision, where applicable with feature attribution methods such as SHAP (Shapley additive explanations) or LIME (local interpretable model-agnostic explanations), or heat maps S01, and a review of those features is part of approving the test results S01. During testing the system also logs, where applicable, the model’s confidence score for each prediction S01; the model works with a threshold, and at very low confidence may flag the result as “undecided” rather than guess S01.
In operation, the model, its system and the whole process go under change control before use; any change, including to the physical objects the model reads, is assessed for retesting, and a decision not to retest must be justified S01. The model is also under configuration control, with measures to detect unauthorised change S01. Its performance is monitored regularly against its metrics S01, and its inputs are monitored for drift out of the sample space it was tested on, with defined drift metrics S01. Where a human decides with model input and testing was reduced, records are kept and every output may need review S01.
Annex 11 and Chapter 4: AI enters the documentation rules
The revised Chapter 4, also a consultation draft, names AI directly: decisions in manufacturing supported by automatic validation scripts or AI, such as batch release based on in-process controls and process analytical technology (PAT), must comply with all EU GMP provisions, including Annex 11 S08. Accountability for the integrity of records or data produced or processed with AI rests with the regulated user S08, and AI support belongs in the pharmaceutical quality system whether it runs on site or as a hosted service S08. For derived data, processing must stay traceable enough to be reconstructed S08, and at least all data on which quality decisions are based count as raw data S08.
The revised Annex 11 draft supplies the rules Annex 22 rests on. A computerised system that replaces a manual operation must cause no decrease in quality, safety or data integrity and no increase in overall process risk S07; Annex 22’s “no decrease” rule echoes it. Outsourcing does not move responsibility: the regulated user stays fully responsible S07, including for systems bought as a service S07, and contracts with service providers need an exit strategy to keep control of the data S07. Each user needs a personal account S07, remote access to critical systems needs multifactor authentication S07, and audit trails record who changed what, from which value to which, when and why S07, reviewed before batch release unless a later review is justified S07. Periodic reviews must catch undocumented changes, for example through configuration audits S07. The two drafts do not quite agree on wording: Annex 11 uses ALCOA+ for data integrity (attributable, legible, contemporaneous, original, accurate, plus complete, consistent, enduring, available) S07, Chapter 4 uses ALCOA++, which adds “traceable” S08.
Why the scope may change: EMA’s 2026 workshop
The workshop report is the newest official text. It records that consultation comments supported letting manufacturers use LLMs and generative AI in GMP applications “whether critical or non-critical” S06, and that the drafting group “has therefore discussed” widening Annex 22 to dynamic or adaptive models, probabilistic models and generative AI, provided they meet the annex and a fully documented, risk-based control strategy S06. In the next sentence: “However, concerns remain about the elements needed for such a control strategy” S06.
Most of the report is the consolidated position of industry associations, and it is clear about who speaks. Industry argued that models should not be excluded “on the basis of a technology label alone”; what should decide is whether the model is fit for its intended use and its risks can be controlled S06. Regulators asked how a plant would show inspectors that layered controls make an output reliable, how a responsible person can take responsibility for a model that is not fully transparent, and how test independence survives when a model retrains on operating data S06. The answers recorded are not prescriptive. On independence: use lifecycle mechanisms, defined samples, test sets and controls to keep a means of independent verification S06; EMA’s summary is that independent verification “remains a critical expectation, but implementation for continuously learning systems needs clarification” S06.

Key messages the report records: controls should be layered across prevention, detection and containment S06; human oversight “cannot compensate for inadequate validation” S06; accountability stays with qualified or responsible persons and cannot be delegated to AI S06; and if a third-party model or programming interface (API) cannot provide the outputs, logs or evidence needed for control, the manufacturer cannot justify it S06. One of the five examples discussed was model predictive control for bioreactor feeding, and some examples were pilots or theoretical, not production systems S06. In the bioreactor and air handling examples, failure or out-of-bound behaviour sent the process back to a pre-validated classical control state S06.
The FDA approach: credibility for a context of use
The FDA draft is not a GMP annex. It sets out a framework for establishing the credibility of an AI model whose output supports a regulatory decision on the safety, effectiveness or quality of a drug S09. “Regulatory decision-making” includes actions taken to comply with current good manufacturing practice (CGMP) S09, and manufacturers who are not sponsors are covered S09. Drug discovery and AI used only for internal efficiency, such as drafting a submission, are out of scope S09. The framework has seven steps: the question the model answers, its context of use, the model risk, a credibility plan, executing it, documenting results and deviations, and deciding whether the model is adequate S09. The context of use must say whether other information is used alongside the model output S09.

Model risk combines model influence, how much the decision rests on the model compared with other evidence, with decision consequence, how bad a wrong decision would be S09. It is the risk that the output leads to a wrong decision with an adverse outcome, “and not risk intrinsic to the model” S09. The manufacturing example is an AI vision system that checks the fill level of 100% of the multidose vials of an injectable S09. Fill volume is a critical quality attribute, so the decision consequence is high; but release testing also measures fill volume on a representative sample of each batch, so the model is not the sole basis for release S09, its influence is low, and the model risk is medium S09. The effort scales with that risk: acceptance criteria should be more stringent, and described to FDA in more detail, for high-risk models S09. If credibility falls short, one of five listed options is to lower the model’s influence by adding other evidence S09.
For the plan itself, development data must be relevant and reliable and, where applicable, representative of the manufacturing process S09; the FDA does not call tuning data “validation” data, as parts of the machine learning community do S09; every performance estimate should come with a confidence interval S09; test data must be independent of development data, for example from different batches or products S09; and where a human is in the loop, the human and AI team is evaluated, not the model alone S09.
The guidance uses manufacturing as its example of where life cycle maintenance matters S09. It notes that AI models can be “self-evolving”, able to adapt without human intervention, and asks for ongoing monitoring of performance S09. Changes to the model, including model-directed changes, go through the manufacturer’s change management system in its quality system S09; detailed maintenance plans, with metrics, monitoring frequency and retest triggers, sit with the site, with a summary in the marketing application for product-specific models S09; and sponsors may use ICH Q12 tools to agree in advance which model changes can be made without prior submission S09. Cybersecurity is out of scope S09.
Two more FDA texts frame this. CDER’s 2023 discussion paper on AI in drug manufacturing is explicitly not guidance S10. It noted that models in manufacturing are developed, validated and updated through change control, and that continuously learning AI challenges that paradigm S10; it asked how inspectors would examine continuously updated models and how comparability is shown after AI-driven changes, especially for biologics S10. The joint EMA and FDA principles of January 2026 ask for a risk-based approach with validation proportionate to context of use and model risk S15, assessment of the whole system including human and AI interaction S15, and scheduled monitoring and re-evaluation, for example for data drift S15.
How you validate a model that learns
Put the texts side by side and they agree more than they differ. All ask for independent test data, criteria fixed before the test, monitoring for drift and change control. They differ on explainability, on how strictly test data must be separated and, above all, on models that keep learning.

For bioprocess models the older ICH texts already apply. ICH Q14 says machine learning could use similar principles to multivariate calibration models S19. Its validation samples are kept out of the calibration and internal test sets, and samples from independent batches can show robustness S19; model performance is confirmed periodically and when events occur S19; and a recalibrated model is validated against the same performance criteria as the original S19. ICH Q13 asks for model validation against predetermined acceptance criteria S20, defines model maintenance as planned activities to keep the model fit for its purpose S20, and says a model may need to be redeveloped and validated depending on the extent of a change S20. EMA’s reflection paper asks that monitoring detect drift early against clearly defined thresholds S18 and that, for high-risk uses, any non-trivial change in the software and hardware stack trigger a new performance evaluation S18. Staff of the FDA, EMA and two national EU agencies, writing personally, recall the ICH classes of model impact: a monitoring model is usually medium impact because release tests still check quality, while a model that predicts a quality attribute for release is high impact S30.

Our reading of all of this: in November 2026 a model “that learns” is validated by not letting it learn on its own. Retraining produces a new version; the new version is tested on independent data against criteria set beforehand, released through change control and monitored for drift and for inputs outside its tested range. CDER described this paradigm in 2023 S10, ICH Q14 applies it to recalibration S19, and draft Annex 22 writes it into GMP for critical uses S01. The FDA draft accepts that models may evolve and asks firms to manage it S09; EMA is discussing how S06. Nobody has yet written how a model that updates itself in operation keeps an independent test.

Is a plant AI system “high-risk” under the AI Act?
The AI Act makes a system high-risk in two ways. One is when the AI is a safety component of a product, or is itself a product, covered by the EU laws listed in its Annex I and needing third-party conformity assessment S22; Annex I includes machinery and medical device legislation S22. The other is the uses listed in Annex III, whose “critical infrastructure” covers digital infrastructure, road traffic and the supply of water, gas, heating or electricity S22. Our reading, not legal advice: medicines manufacturing is not an Annex III use and EU medicines law is not in Annex I, so a process model in a GMP plant is not high-risk by default; an AI safety component of machinery could be, through Annex I. For those, the AI Omnibus sets 2 August 2028; for machinery it also moves the AI requirements into the Machinery Regulation itself, to apply by that date S25. Annex 22 itself reuses the AI Act’s definition of an AI system S01.
GAMP and PIC/S
ISPE, an industry body and not a regulator, published a guide on artificial intelligence in its GAMP (Good Automated Manufacturing Practice) series in July 2025, 290 pages S26, meant to be used with GAMP 5, Second Edition S26. Its public table of contents includes sections on data split S26 and on model monitoring and drift detection S26. We did not read either guide, which are sold, and say nothing about their content. PIC/S, the network of GMP inspectorates, co-drafted the Annex 22 and Annex 11 drafts S02 and lists them on its site S28; for the recent Annex 19 it made its own version enter into force at the same time as the EU’s S28, so, in our reading, a final Annex 22 could reach PIC/S members on the same path, though nothing says so yet. Its data integrity guide, in force since 2021, applies to paper, computerised and hybrid systems alike S29.
What it means for a plant
- Measure the process you want to replace. “No decrease” means the model must be at least as good as the current method S01. Without data on the current offline assay or inspection, there is no acceptance criterion.
- Set aside test batches before modelling, and lock them. Decide which data are test data, who may see them and where they live, before anyone trains S01. A model reworked after a failed test needs a new test set S18.
- Write the intended use with a process expert, input range included. It defines what the model is tested on and what drift monitoring watches S01.
- Keep the model’s influence low where you can. A soft sensor that triggers a sample, or a check backed by release testing, carries less risk than a model that releases product on its own S09 S30.
- Plan retraining as change control. Each retrained or recalibrated model is a new version, tested against the same criteria S19 S01. Have a validated fallback for when the model is out of bounds, as in the bioreactor example at EMA’s workshop S06.
- Get evidence from suppliers. You stay responsible for bought or hosted models S07 S08; a supplier that cannot provide logs and evidence cannot be justified S06.
- Do not build a critical use on generative AI or on a self-learning model yet. The July 2025 draft excludes them S01; the final text may allow them under conditions not yet written S06.
The AI in the plant series
This piece is the second in the series after the pillar, AI in bioprocessing, explained, which maps what runs in a plant and where models break. On the probes and soft sensors that are the likeliest first GMP models, see PAT in practice. We will update this piece when Annex 22 or the FDA guidance is finalised.
Questions
What is EU GMP Annex 22 on artificial intelligence?
A new annex to the EU GMP guide on AI models used in critical manufacturing applications, adding to Annex 11 S01. It was put to consultation from 7 July to 7 October 2025 S02. On 2 November 2026 it is not adopted and is not in EudraLex Volume 4 S03.
Is the Annex 22 draft final?
No. The only text is the July 2025 consultation draft. EMA targets the fourth quarter of 2026 for giving the Commission a final text S04, and after its June 2026 workshop it is discussing a wider scope S06.
Does Annex 22 allow machine learning in GMP manufacturing?
Yes, if the model is static and deterministic and meets the annex’s testing and operating rules S01. Models that keep learning, models with probabilistic output, generative AI and LLMs should not be used in critical GMP applications under the draft S01, a point now under review S06.
Can LLMs or generative AI be used in GMP?
Under the July 2025 draft, not in critical applications; in non-critical ones, with qualified and trained staff responsible for the output S01. Consultation comments supported allowing them, and EMA’s drafting group has discussed it, with concerns remaining S06.
What does the FDA say about AI in drug manufacturing?
Its January 2025 draft guidance, still a draft on 2 November 2026 S12, asks for a credibility assessment scaled to model risk, from model influence and decision consequence, and for life cycle maintenance under the manufacturer’s change management S09. AI in manufacturing must also meet CGMP rules S09.
How do you validate an AI model in GMP?
Under draft Annex 22: a documented intended use, acceptance criteria set before the test and at least as good as the process replaced, an independent and locked test set, feature review, confidence thresholds, then change control and drift monitoring in operation S01. The FDA draft and ICH Q14 ask for the same core: independent test data, predefined criteria and ongoing performance checks S09 S19.
Is AI in pharmaceutical manufacturing high-risk under the EU AI Act?
Not by default, in our reading: manufacturing is not an Annex III use, but an AI safety component of machinery can be high-risk through Annex I S22. Those rules apply from 2 August 2028 after the AI Omnibus S25.
What we could not verify
- When Annex 22, Annex 11 and Chapter 4 will be adopted, and whether a second draft or consultation will come first. No official date exists after the workshop report.
- How a final Annex 22 will treat adaptive, probabilistic and generative models, and how test independence would work for a model that keeps retraining.
- Whether the FDA will finalise its guidance, and what came of the comments on its 2023 manufacturing paper.
- The content of the ISPE GAMP AI guide and GAMP 5, which we did not read, and the 2011 ICH document behind the high, medium and low model impact classes, cited here second hand.
- No regulator text gives a minimum test set size, a numeric acceptance threshold or a confidence threshold for a GMP model.
- Whether a given piece of plant equipment with AI is high-risk under the AI Act depends on product legislation we did not analyse.
Sources
- S01 European Commission; drafted by the EMA GMDP Inspectors Working Group in cooperation with PIC/S, EudraLex Volume 4, Annex 22: Artificial Intelligence (draft for stakeholder consultation). European Commission, DG SANTE, consultation document (July 2025), 2025 (regulator document).
- S02 European Commission, DG SANTE, Stakeholders’ consultation on EudraLex Volume 4: Chapter 4, Annex 11 and new Annex 22. health.ec.europa.eu consultation page, 2025 (regulator document).
- S03 European Commission, DG SANTE, EudraLex Volume 4: Good Manufacturing Practice guidelines (index page). health.ec.europa.eu, 2026 (regulator document).
- S04 European Medicines Agency, GMDP Inspectors Working Group, The 3-year work plan for the GMDP Inspectors Working Group, January 2026 to December 2028. EMA work plan, 2026 (regulator document).
- S05 European Medicines Agency, Good manufacturing practice: Multistakeholder workshop on expert contributions to artificial intelligence guidance development (Annex 22). EMA event page (workshop 30 June to 1 July 2026), 2026 (regulator document).
- S06 European Medicines Agency, Report: Multistakeholder workshop on expert contributions to artificial intelligence guidance development (Annex 22), EMA/156789/2026. EMA report, first published 2 October 2026, 2026 (regulator document).
- S07 European Commission; EMA GMDP Inspectors Working Group and PIC/S Committee, EudraLex Volume 4, Annex 11: Computerised Systems (revised draft for stakeholder consultation). European Commission, DG SANTE, consultation document (July 2025), 2025 (regulator document).
- S08 European Commission; EMA GMDP Inspectors Working Group and PIC/S Committee, EudraLex Volume 4, Chapter 4: Documentation (revised draft for stakeholder consultation). European Commission, DG SANTE, consultation document (July 2025), 2025 (regulator document).
- S09 US Food and Drug Administration (CDER with CBER, CDRH, CVM, OCE, OCP, OII), Considerations 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 draft guidance, January 2025, docket FDA-2024-D-4689, 2025 (regulator document).
- S10 US Food and Drug Administration, CDER, Artificial Intelligence in Drug Manufacturing. Discussion paper. FDA CDER discussion paper (March 2023), docket FDA-2023-N-0487, 2023 (regulator document).
- S11 US Food and Drug Administration, CDER, Artificial Intelligence for Drug Development (CDER web page). fda.gov, content current as of 05/01/2026, 2026 (regulator document).
- S12 US Food and Drug Administration, Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products (guidance landing page). fda.gov guidance page, content current as of 01/06/2025, 2025 (regulator document).
- S13 Office of the Federal Register (data); BioprocessUpdates research desk (queries), Federal Register API search record for FDA artificial intelligence documents (desk search, 2026-11-02). federalregister.gov API, 2026 (regulator document).
- S14 US Food and Drug Administration, Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products; Draft Guidance for Industry; Availability (Federal Register notice 2024-31542). Federal Register 90(4), 7 January 2025, pp. 1157 to 1159, 2025 (regulator document).
- S15 European Medicines Agency and US Food and Drug Administration, Guiding principles of good AI practice in drug development. EMA and FDA joint document, January 2026 (first published on the EMA site 14/01/2026), 2026 (regulator document).
- S17 International Council for Harmonisation (ICH) Expert Working Group, ICH Q9(R1) Quality Risk Management. ICH harmonised guideline, Step 4, adopted 18 January 2023, 2023 (regulator document).
- S18 European Medicines Agency, CHMP and CVMP, Reflection paper on the use of Artificial Intelligence (AI) in the medicinal product lifecycle, EMA/CHMP/CVMP/83833/2023. EMA reflection paper, final version adopted by CHMP 9 September 2024, 2024 (regulator document).
- S19 International Council for Harmonisation (ICH) Expert Working Group, ICH Q14 Analytical Procedure Development. ICH harmonised guideline, adopted 1 November 2023, 2023 (regulator document).
- S20 International Council for Harmonisation (ICH) Expert Working Group, ICH Q13 Continuous Manufacturing of Drug Substances and Drug Products. ICH harmonised guideline, adopted 16 November 2022, 2022 (regulator document).
- S21 U.S. Food and Drug Administration (CDER, CVM, ORA), Guidance for Industry. PAT, A Framework for Innovative Pharmaceutical Development, Manufacturing, and Quality Assurance. FDA guidance for industry, September 2004, 2004 (regulator document). Published before 2015.
- S22 European Parliament and Council of the European Union, Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act). Official Journal of the European Union L, 12.7.2024, 2024 (regulator document).
- S23 European Commission, DG CONNECT, AI Act (Commission policy page: regulatory framework for AI). digital-strategy.ec.europa.eu, 2026 (regulator document).
- S24 European Commission, DG CONNECT, AI Omnibus enters into force (Commission news article, 27 July 2026). digital-strategy.ec.europa.eu, 2026 (regulator document).
- S25 European Parliament and Council of the European Union, Regulation (EU) 2026/1744 amending Regulations (EU) 2024/1689, (EU) 2018/1139 and (EU) 2023/1230 (Digital Omnibus on AI). Official Journal of the European Union L, 24.7.2026, 2026 (regulator document).
- S26 ISPE (International Society for Pharmaceutical Engineering), ISPE GAMP Guide: Artificial Intelligence (public product page and table of contents). ISPE guidance document, published July 2025, 290 pages, 2025 (industry body, not a regulator). Public web page read; the guide itself was not read.
- S28 Pharmaceutical Inspection Co-operation Scheme (PIC/S), PIC/S publications list and news page. picscheme.org, 2026 (regulator document).
- S29 Pharmaceutical Inspection Co-operation Scheme (PIC/S), PIC/S Good Practices for Data Management and Integrity in Regulated GMP/GDP Environments (PI 041-1). PIC/S guidance for inspectors, in force 1 July 2021, 2021 (regulator document).
- S30 O’Connor TF et al., An examination of process models and model risk frameworks for pharmaceutical manufacturing. International Journal of Pharmaceutics: X 8:100274, 2024. Affiliations: US FDA (CDER, CBER), EMA and ANSM staff writing in a personal capacity; the disclaimer says it does not represent FDA policy.


