Table of Contents
Executive Summary
EMA’s draft Annex 22, the first comprehensive EU regulatory framework dedicated to artificial intelligence in pharmaceutical manufacturing, was published for stakeholder consultation on July 7, 2025, with the consultation window closing October 7, 2025. Finalization has not been dated, and the final text will set when it takes effect. As of May 2026, sponsors are in a critical preparation window: the framework’s contours are clear from the draft, the consultation feedback is informing finalization, and the operational work required to prepare is real but tractable.
This article translates the current state of Annex 22 into operational implications for pharma sponsors. We cover where the timeline stands, what the consultation surfaced, the scope clauses that materially shape the operational impact, how Annex 22 interacts with the EU AI Act and FDA guidance, and the work sponsors should be doing now between May and year-end to be ready when Annex 22 finalizes.
Where the Timeline Stands
The Annex 22 trajectory has moved through several recognizable stages over 2024-2026. The EMA GMDP Inspectors Working Group, in cooperation with PIC/S, drafted the annex. The draft was published for stakeholder consultation by the European Commission on July 7, 2025, alongside companion revisions to Chapter 4 (Documentation) and Annex 11 (Computerised Systems). The consultation, accessible via the European Commission consultation page, closed October 7, 2025.
Following the consultation, the EMA’s working group is reviewing the feedback and finalizing the documents. Finalization is expected during 2026. The European Commission usually sets a date some months after publication when a new or revised GMP annex comes into operation, and the final text will state that date. That means the timing of the implementation window for sponsors depends on when the final annex is published.
The practical implication for pharma sponsors operating in Europe: the window for proactive preparation runs from now through finalization. Sponsors that wait until finalization to begin operational work will find themselves implementing under time pressure, and the Rephine analysis of Annex 22 preparation urged manufacturers to map their AI use and strengthen governance while the draft was still open for consultation.
Two additional considerations affect the timeline. First, EMA’s finalization will need to align with the EU AI Act implementation cadence, which is on its own timeline. Second, the relationship between Annex 22 and PIC/S guidance will need to be articulated, since PIC/S inspectors apply to manufacturers serving multiple jurisdictions. Both of these introduce some uncertainty into the precise finalization date, but neither changes the operational implication: sponsors should be preparing now.
Consultation Outcomes and What Industry Said
The consultation period produced substantive industry feedback. Consultation comments have not been published in full, but public commentary from industry points to some common themes.
Scope of the static/deterministic model restriction. The draft annex restricts critical GMP functions to “static, deterministic” AI models (those that produce consistent outputs for the same inputs) and excludes generative AI and large language models from those critical functions. As described in Continuous Intelligence’s analysis of Annex 22, this scope clause allows only static, deterministic models in critical applications and excludes dynamic models and generative AI from them. A practical question is where the boundary falls, particularly for hybrid systems that combine deterministic and learned components.
Definition of “high impact” AI. The annex’s scope covers AI models used in critical applications with direct impact on patient safety, product quality or data integrity. A key practical question is how to decide which applications are critical (the draft scopes these as having direct impact on patient safety, product quality or data integrity) and how borderline cases should be classified. The clarity of the threshold materially shapes implementation cost.
Validation expectations for AI used in non-critical functions. The draft annex covers only AI in critical GMP applications and sets no validation tier for non-critical uses, but the line between critical and non-critical is consequential. Clearer definitions would help, particularly for use cases that operate next to critical functions (for example, predictive maintenance for critical equipment).
Human-in-the-loop expectations. The annex articulates expectations for human-in-the-loop oversight that vary with the criticality of the process and how much the model has been tested. A key practical question is what counts as evidence of adequate human oversight, particularly for systems where human review is hard to build in (for example, real-time process control).
Lifecycle management and explainability. The annex’s expectations for lifecycle management and explainability are foundational, but the explainability expectations will likely need to be applied differently for different kinds of models. Some AI techniques are inherently less explainable, and the practical operational consequence of the explainability expectation depends on how the expectation is interpreted.
The consultation feedback is not public in aggregate form, but in our view finalization will likely refine the draft’s risk-based structure rather than reverse it.
The Scope Clauses That Matter Most
Three scope clauses in the draft materially shape the operational implications.
Static and deterministic restriction for critical functions. AI used in processes directly impacting product quality or patient safety should be static and deterministic under draft Annex 22. Generative AI and LLMs are excluded from critical functions. This is the single most operationally consequential clause: it limits the AI architectural choices available for the highest-risk pharma applications and shapes vendor selection accordingly. Sponsors planning to deploy LLMs in critical manufacturing functions would need to revisit those deployments under Annex 22.
Risk-based tiering. The risk-based approach triggers the strictest requirements only for high-impact AI. Non-critical uses fall outside the annex, though its principles may be considered where applicable. The draft covers only critical uses and applies its controls based on risk (Section 2.3), which is broadly in line with FDA’s risk-based credibility framework. Sponsors with a tier classification SOP that maps onto the EMA structure are well-positioned; sponsors without one will need to develop one quickly.
Human-in-the-loop oversight. The draft expects a human-in-the-loop when generative AI or LLMs are used in non-critical GMP applications, and when a model gives input to a human decision and its testing has been reduced. The clause is broadly consistent with the FDA’s GMLP and GAIP principles, but the operational specifics matter. As the Merit Solutions analysis of Annex 22 explains, human-in-the-loop review is the expected control for non-critical generative AI uses such as drafting and summarization.
The combination of these clauses produces a recognizable operational picture. Sponsors operating in Europe will need to: (1) inventory all AI in manufacturing systems, (2) classify each use by impact tier, (3) confirm that high-impact uses meet the static/deterministic restriction, (4) demonstrate human-in-the-loop oversight where required, and (5) maintain lifecycle management evidence aligned with the annex’s expectations. None of these are conceptually new disciplines for pharma quality; the work is in extending them to AI.
Interaction With the EU AI Act
Annex 22 does not exist in isolation. AI use cases in pharma manufacturing may also fall under the EU AI Act’s high-risk classification, with corresponding documentation, conformity assessment, and post-market monitoring obligations. The two regimes overlap but are not identical, and managing both efficiently requires deliberate framework design.
The EU AI Act’s high-risk classification applies to AI that is a safety component of a product covered by the EU product laws in Annex I (such as medical devices or machinery) and to the uses listed in Annex III; pharmaceutical manufacturing is not named on either list, so each use case has to be checked. Annex 22 applies specifically to AI in GMP manufacturing of medicinal products and active substances. High-impact AI under Annex 22 will usually not be high-risk under the AI Act, because medicine manufacturing is not on the AI Act’s high-risk lists. Where both do apply, the documentation and conformity assessment requirements differ in important ways.
The practical implication: sponsors operating in Europe need a framework that satisfies both regimes simultaneously. Parallel compliance programs are operationally expensive; an integrated framework is materially more efficient. Sponsors that build their QMS extension with both regimes in mind from the start avoid the more expensive remediation work that parallel compliance produces.
A key area of interaction is post-market monitoring. The EU AI Act requires post-market monitoring for high-risk AI systems; draft Annex 22 calls for lifecycle management and ongoing performance monitoring. The two requirements can be satisfied by a single well-designed monitoring program, but the documentation has to satisfy both. Quality leaders should anticipate this in their monitoring infrastructure design.
Interaction With FDA’s Posture
Annex 22 and the FDA approach share a risk-based outlook, but the two frameworks are not identical. The FDA’s January 2025 draft guidance on AI for regulatory decision-making, available through the FDA guidance landing page, takes a credibility-framework approach that complements Annex 22’s risk-based tiering rather than duplicating it.
For multinational sponsors, the operational implication is that a single AI governance framework can be designed to satisfy both FDA and EMA expectations, with marginal localizations for specific jurisdictional differences. As the BioSlice Blog analysis of the EU consultation notes, outsourcing AI capabilities does not shift regulatory responsibility away from the manufacturer.
Where the two frameworks differ in emphasis: EMA’s Annex 22 is more prescriptive about the static/deterministic restriction for critical functions; FDA’s posture is does not restrict model architecture and instead centers on a risk-based credibility assessment for each context of use. Sponsors operating in both jurisdictions will usually find it practical to meet the more restrictive of the two for critical functions while maintaining the credibility evidence FDA recommends in its January 2025 draft guidance.
What to Do Now: A May-to-Year-End Plan
For sponsors operating in Europe, the May 2026 to year-end window is the right time to move Annex 22 preparation from planning into action. A workable plan:
| Window | Workstream | Deliverable |
|---|---|---|
| May – June | AI inventory in manufacturing systems | Comprehensive inventory of all AI use in manufacturing, including vendor-embedded features |
| June – July | Tier classification SOP aligned with Annex 22 structure | Published SOP with risk-based classification criteria |
| July – August | Apply classification to inventory | Classified inventory with high-impact use cases identified |
| August – September | Gap assessment for high-impact uses | Documented gaps against Annex 22 requirements with remediation plans |
| September – October | Remediation work for highest-priority gaps | Resolved gaps for the use cases with the largest regulatory exposure |
| October – November | Human-in-the-loop oversight design | Validated oversight workflows for high-impact AI |
| November – December | Lifecycle management and monitoring infrastructure | Operational monitoring for high-impact AI with defined response procedures |
By year-end 2026, a sponsor following this plan would have a published tier classification SOP, an inventory, a gap-and-remediation register for high-impact uses, validated oversight workflows, and operational monitoring infrastructure. This is a defensible posture going into the implementation window.
Operational Implications for Sponsors
The operational implications of Annex 22 are significant but tractable for sponsors that begin work now.
Vendor relationships will need to be revisited. Vendor contracts for AI-enabled manufacturing systems will need to address model architecture (static/deterministic versus dynamic), model version pinning, change notification, and validation cooperation. Sponsors with contracts negotiated before Annex 22 should anticipate amendments or renegotiations.
Some current AI deployments will need to be reconsidered. AI use in critical functions that does not meet the static/deterministic restriction will need to be reconsidered. This is most acute for generative AI and LLM deployments in critical manufacturing contexts, although as discussed in the case-pattern article, most pharma LLM deployments have not been in critical manufacturing.
Documentation burden will increase. Draft Annex 22’s lifecycle management and explainability expectations call for documentation (such as feature attribution records and performance monitoring metrics) that, in our experience, traditional CSV programs do not typically produce. Sponsors will need to extend their documentation discipline rather than build parallel AI-specific documentation.
Cross-functional coordination becomes more important. Manufacturing AI use cases typically involve IT, Engineering, Quality, and Regulatory. The draft Annex 22 calls for close cooperation among all relevant parties, including process experts, QA, data scientists and IT, during algorithm selection and model training, validation, testing and operation. Sponsors with weak cross-functional coordination will find Annex 22 implementation slower than sponsors with strong governance.
Inspector training will lag. EMA inspectors will be applying Annex 22 starting from implementation, but their depth of AI expertise will vary. Sponsors should anticipate that early inspections under Annex 22 may be inconsistent and should prepare documentation that is intelligible to inspectors with varying AI backgrounds. The 21 CFR Part 11-style architecture that pharma quality leaders are familiar with is the right reference for documentation design.
The broader strategic implication is that Annex 22 will not be the last EU regulatory action on AI in pharma. Sponsors that build AI governance frameworks expecting ongoing regulatory evolution, rather than treating Annex 22 as a one-time compliance event, will be materially better positioned for the post-2027 regulatory environment than sponsors who treat it as a discrete project. The framework being built now should be designed to accommodate evolution, not to solve a single regulatory release.
How the Annex 22 work interacts with Chapter 4 and Annex 11 revisions
A practical implementation point that often gets underweighted: Annex 22 was published in consultation alongside revisions to Chapter 4 (Documentation) and Annex 11 (Computerised Systems). The three documents are designed to work together, and sponsors that focus exclusively on Annex 22 risk missing the operational changes the Chapter 4 and Annex 11 revisions will require. As noted in Epista’s preparation analysis for the GMP revisions, the three documents collectively introduce material changes to documentation expectations, computerized system validation, and AI-specific controls.
The Chapter 4 revisions are expected to bring records produced or processed with AI under GMP documentation rules, with accountability for their integrity resting with the regulated user and cross-references to Annex 11 and Annex 22. The draft Annex 11 revision does not mention AI. It sets general lifecycle, validation and data integrity rules for all computerized systems. AI-specific guidance is in a separate new draft, Annex 22, which adds to Annex 11 for computerized systems in which AI models are embedded. The Annex 22 is the dedicated AI annex but operates in the context that the Chapter 4 and Annex 11 revisions establish.
Quality leaders preparing for the implementation window should be reading all three documents as a coherent set, not Annex 22 in isolation. Documentation programs designed to satisfy Annex 22 alone will miss the broader expectations that the companion revisions establish, and remediation under inspection pressure later is materially more expensive than preparation now.
Inspector preparation and the early enforcement environment
An underappreciated dimension of the implementation window is how the inspector cadre will be prepared to enforce Annex 22. EMA and national competent authority inspectors have varying degrees of AI expertise; even after Annex 22 finalizes, the operational interpretation of the framework will evolve through actual inspection experience over the first year or two of enforcement. Sponsors anticipating the implementation window should expect inconsistent interpretation across inspections during this learning period, and should design their documentation to be intelligible to inspectors with varying AI backgrounds.
The 21 CFR Part 11-style architecture that pharma quality leaders already know is the right reference for documentation design. Inspectors familiar with computerized system validation will navigate AI documentation built on the same scaffolding more readily than documentation built on AI-specific architecture. The familiar scaffolding is itself a compliance posture: it reduces the cognitive load on inspectors and produces faster, smoother inspections.
The risks of not acting now
A final strategic dimension worth understanding: the downside of waiting until Annex 22 finalizes before beginning operational work. Sponsors that defer preparation face several real problems.
First, the work compresses into the grace period after publication, which is typically too short for the full operational build. Programs that compressed analogous preparation into grace periods for other regulatory actions have consistently delivered lower-quality compliance posture under time pressure.
Second, vendor contract renegotiations cannot reasonably be compressed. Vendor contracts for AI-enabled manufacturing systems often run on multi-year cycles; sponsors that defer preparation will find themselves needing to renegotiate during the grace period, when bargaining power is weakest because the deadline is fixed.
Third, organizational learning takes time. Building QA capability for AI requires repeated exposure to the work, not just training. Sponsors that start the operational work now will have time for organizational learning before the implementation window; sponsors that defer will be starting that learning under deadline pressure.
Fourth, regulatory engagement opportunities are time-sensitive. Pre-submission meetings, scientific advice procedures, and similar regulator engagements that can clarify operational expectations are most useful when conducted ahead of the implementation deadline. Sponsors that defer preparation typically have fewer opportunities to engage productively with regulators during the implementation window itself.
The cumulative effect is that waiting carries materially more risk than acting now. Quality leaders making the business case for proactive preparation should be explicit about these consequences, because the conservative-feeling decision to wait often produces the more expensive outcome.
References & Sources
For Further Reading
References & Sources
- Stakeholders’ Consultation on EudraLex Volume 4: Chapter 4, Annex 11 and New Annex 22 — European Commission Public Health. Official consultation page documenting the July 7 to October 7, 2025 consultation window for Annex 22 and companion revisions.
- Multistakeholder workshop on expert contributions to AI guidance development (Annex 22) — European Medicines Agency. Reference for the EMA’s stakeholder engagement process during Annex 22 development.
- Annex 22: EMA’s AI Regulation for Pharma Manufacturing — Continuous Intelligence. Industry analysis of the static/deterministic restriction and other key clauses in the draft Annex 22.
- How to Prepare for Annex 22 — Rephine. Practitioner-grade guidance on the implementation timeline and the operational work sponsors should do during the grace period.
- EU GMP Annex 22: The New AI Regulatory Standard — Merit Solutions. Industry analysis of Annex 22’s human-in-the-loop expectations and broader operational implications.
- EU Consults on New GMP Rules for AI in Pharma Manufacturing — BioSlice Blog. Documentation of the consultation process and the EMA-FDA coordination on AI in regulated manufacturing.








Your perspective matters—join the conversation.