GMP Annex 22 readiness: what the public consultation tells us.

EU GMP Volume 4 has accumulated annexes for forty years. Annex 11 (computerised systems) has carried the weight of every regulated computer system since 1992; the 2011 revision still governs at the time of writing. Annex 22 is the first annex written specifically for AI in pharmaceutical manufacturing. The text is short. The implications are not.

The draft is structured in fourteen short sections covering scope, definitions, governance, risk management, data management, models, qualification and validation, operational requirements, change management, monitoring, deviation handling, product recall, decommissioning, and documentation. Each section is a few paragraphs. Each section will be enforced.

/ 01What the draft actually requires.

The architecture is recognisable. It is the lifecycle architecture pharma already uses for any computerised system — Annex 11, 21 CFR Part 11, GAMP 5 — applied with attention to the specific failure modes of machine-learned components. The ten readable obligations of the draft, in operational paraphrase:

• The AI system must have a defined intended use, documented and approved, before deployment
• Risk management under ICH Q9(R1) must be applied at the AI-component level, not just the surrounding system
• Training, validation, and testing data must be documented for provenance, completeness, representativeness, and bias
• Models must be qualified — version-controlled, traceable, signed off by named roles
• Operational performance must be monitored against pre-specified metrics, with thresholds for action
• Deviations involving the AI system must be investigated under the existing GMP deviation framework, with the AI-specific causal layer addressed
• Change control must explicitly address model retraining, data drift response, and infrastructure changes
• Decommissioning must be planned: model retirement, data retention, dependency mapping
• Human oversight must be designed at the workflow level, not assumed at the use level
• Documentation must be inspection-ready by default — not assembled at the time of inspection

Read against Annex 11 and Part 11, the new obligation is not the lifecycle. It is the data and model layers under the lifecycle. The surrounding system has been governed for decades. The AI components — training corpora, model weights, drift detectors, retraining pipelines — are new artefacts that fall under the lifecycle, and pharma manufacturing has not previously had to govern them.

/ 02The data-governance obligation.

Section on data is the most substantive. Provenance for every training data point — where did it come from, was the source GMP-grade, was it sampled in a way that introduces bias, was it labelled by qualified personnel, was the labelling itself audited. For visual inspection AI in vial inspection lines, this means the rejected-vial library used for training is itself a GMP record. For deviation classifiers trained on historical QMS data, the historical QMS data is now training data and its quality flows into model performance.

Most pharma manufacturing sites have not been treating training data as GMP records. They have been treating it as IT artefacts. The first Annex 22 inspection finding will be the absence of training data lineage documentation. The second will be the absence of a labelling audit trail.

/ 03The qualification boundary.

Where IQ/OQ/PQ still applies.

The infrastructure layer — compute, storage, integration with the existing computerised system inventory — is qualified the way it always has been. IQ for installation, OQ for the operational envelope, PQ for sustained performance under production load. The new obligation is that the AI component sits as a tested artefact within that infrastructure, with its own qualification at the model level.

Where it doesn't.

A non-deterministic model does not return identical output to identical input across all model versions. Performance qualification at the model level uses statistical envelopes, reference test sets, and acceptance criteria with confidence bounds. The validation literature for this is borrowed from the FDA AI/ML SaMD pattern, but Annex 22 does not import it directly — the document is framework-level, leaving operational specifics to the manufacturer's QMS. Companies that wait for explicit operational guidance will wait past the inspection.

The annex is short because the discipline is borrowed. The discipline is borrowed because the regulators are confident pharma already has it. The borrowing only works if the discipline is applied.

/ 04The intersection with EU AI Act.

An AI system in pharmaceutical manufacturing that affects product quality is high-risk under the AI Act Annex III(8) reading, regardless of Annex 22. The obligations stack rather than substitute. A manufacturer placing an AI-driven release-decision system on a packaging line must satisfy Article 9 risk management and Annex 22 risk management. The two documents are not redundant — Article 9 is provider-side; Annex 22 is operator-side. Many companies are not yet differentiating the two requirement sets.

Annex 22 also explicitly cross-references ICH Q9(R1) on quality risk management and the existing Annex 11 on computerised systems. The expectation is integration into the existing pharma quality management system — see the governance library — not a parallel AI quality system. Companies building parallel structures are over-engineering.

/ 05What to do before H2 2026.

The realistic readiness sequence: inventory the AI footprint at every manufacturing site (most companies do not have this); classify each system against the draft Annex 22 fourteen-section structure (which gaps does each system fail today); assemble training data lineage where possible and document where not; design the model-qualification protocol for the highest-risk system as a test case; run the protocol; draft the SOP that will support the next system. The expensive way to comply with Annex 22 is to wait for the final text and start cold. The cheap way is to operate as if the draft is already in force.

Final publication is expected in the second half of 2026. Member-state implementation timelines will follow. Any manufacturer waiting to start until publication is missing the fact that the draft is already the inspection reference for AI-driven workflows. PIC/S inspectors are reading it already. So is MHRA.