Annex 22 Is Not A Brake On AI. It Is The Framework Clinical Supply Has Been Waiting For

By Chris Yuen, Director, Tenthpin Management Consultants

The first dedicated EU GMP draft guidance for artificial intelligence does not give AI a free pass. It gives clinical supply and IT leaders something more valuable: a regulated path to use AI with discipline, confidence, and audit readiness.

For years, the question around AI in clinical supply has been stuck in a holding pattern: “Can we even use it in a GMP environment?” Annex 22 changes that conversation. With the European Commission’s draft of EU GMP Annex 22 “Artificial Intelligence,” published on July 7, 2025, alongside revised drafts of Annex 11 and Chapter 4, the regulatory direction is becoming clearer: AI is not outside GMP. It must be brought under GMP control.

That shift matters. It moves the discussion from vague innovation language to practical implementation questions: What is the model intended to do? What GMP processes or decisions does it influence? What data supports it? How is performance verified and how will people remain accountable when the model is in use?

What Annex 22 Is, And What It Is Not

Annex 22 is the proposed new AI annex to the EU GMP Guide. It is still draft guidance, not a finalized standard. It should therefore be read as a clear regulatory direction of travel rather than as a completed compliance checklist.

It also does not replace Annex 11 with computerized systems. It builds on it. Annex 11 remains the foundation for validated computerized systems in GMP. Annex 22 addresses the additional questions raised when those systems contain AI or machine learning models.

There is one provision R&D and supply chain leaders must not read past. Annex 22 applies specifically to AI in critical GMP applications, meaning models with a direct impact on patient safety, product quality, or data integrity. Within that scope, the current draft states that dynamic, adaptive, and probabilistic models, including generative AI and large language models, should not be used for GMP applications. This is the most consequential line in the draft, and it does not close the door on clinical supply AI; it defines where the door is. It determines which use cases sit inside critical classification and must be controlled accordingly, which can be re-architected to keep a human as the accountable decision maker, and which fall outside critical scope altogether.

In practical terms, Annex 22 asks a simple but demanding set of questions:

• Intended use: What exactly is the model meant to do, and which GMP process or decision does it support?
• Data: What training, validation, and test data are used, and is that data relevant, representative, and controlled?
• Performance: Which acceptance criteria and metrics prove that the model performs well enough for its intended use?
• Explainability and oversight: Can users understand the output, challenge it where needed, and retain meaningful human accountability?
• Life cycle control: How are model changes, drift, monitoring, deviations, and retirement managed over time?

That is why Annex 22 matters. It does not say “AI is approved.” It says: if AI affects GMP, it must be specified, validated, controlled, monitored, and governed across its life cycle.

Why This Matters For Clinical Supply

Clinical supply is one of the best places to begin because the operational problem is real and measurable. Studies still carry avoidable overage, expiry risk, emergency shipments, manual reconciliation, depot complexity, and resupply decisions made under uncertainty. These are not abstract AI opportunities. They are everyday clinical supply constraints that translate directly into cost, waste, patient risk, and avoidable work.

The work is data-rich, forecast-driven, and full of decisions that reward pattern recognition over static assumptions. That is exactly where well-governed AI can help, provided the use case is narrow enough to validate and important enough to matter.

Where AI Can Move The Needle First

• Demand forecasting and resupply: AI can combine enrollment trends, screen-failure rates, site behavior, and protocol assumptions to detect demand signals earlier than static planning rules.
• Inventory and depot optimization: Smarter allocation across depots, countries, and sites can reduce waste, expiry write-offs, and urgent shipments.
• Comparator and label complexity: AI can support sourcing risk assessment, booklet label logic, and country-specific supply scenarios that are difficult to manage manually at scale.
• Cold chain and logistics risk: Anomaly detection can identify patterns in lane, carrier, depot, or site performance before they become patient-impacting deviations.

The prize is not “AI transformation” as a slogan. It is less overage, less waste, fewer stockouts, fewer avoidable escalations, faster decisions, and a clinical supply organization that spends less time compensating for uncertainty manually.

Annex 22 Gives Leaders The Language To Say Yes Responsibly

What makes Annex 22 useful is not that it removes regulatory burden. It makes the burden more explicit. It gives clinical supply, quality, CSV, IT, and data teams a shared language: intended use, risk classification, acceptance criteria, test-data independence, explainability, human oversight, change control, and ongoing monitoring.

That language is powerful because most high-value clinical supply AI use cases do not need uncontrolled black-box autonomy. They can often be framed as decision-support capabilities with defined inputs, measurable outputs, controlled use, human review, and clear escalation paths.

This is the practical opening. The organizations that learn to express their AI opportunities in Annex 22 language will be able to validate faster, govern better, and discuss AI with inspectors in terms the quality system already understands.

Now We Are At Dawn Of Regulated AI In GMP

The consultation on the draft package closed on October 7, 2025. EMA is continuing to gather expert input, including through a multistakeholder workshop on June 30–July 1, 2026, to inform the development of Annex 22 guidance. The details are still being shaped.. The 2025 consultation surfaced support for enabling the very technologies the draft would restrict, including generative AI and large language models, and the workshop was convened in part to revisit whether probabilistic and adaptive models can have a place in critical applications. The scope of permissible AI is being decided now, not merely awaited, which is precisely why early movers can help shape the questions rather than inherit the answers.

That creates a rare moment. R&D and supply chain leaders do not need to wait passively for the final text. They can begin preparing now by mapping use cases, defining intended use, assessing data readiness, identifying validation evidence, and deciding where human oversight must remain explicit.

Clinical supply is a natural proving ground because the value is operational, the risks are visible, and the improvement levers are concrete. The companies that begin now will not be scrambling to interpret AI governance later. They will already have use cases, evidence, controls, and lessons learned. The next move is to understand the clinical supply process deeply enough to define intended use, acceptance criteria, test data, controls, and inspection-ready evidence. That is how organizations can turn Annex 22 from a compliance challenge into a practical framework for safer and more confident AI adoption.

Sources:

1. European Commission stakeholder consultation on EudraLex Volume 4 Chapter 4, Annex 11, and Annex 22 (published July 7, 2025; consultation deadline October 7, 2025)
2. Draft EU GMP Annex 22 “Artificial Intelligence”
3. PIC/S joint stakeholder consultation note
4. EMA multistakeholder workshop on Annex 22, June 30–July 1, 2026

About The Author:

Chris Yuen is a Director at Tenthpin Management Consultants with over 20 years in clinical supplies, GxP regulatory and quality systems, and cutting-edge technology integration across life sciences. He has led complex system integration programs at SAP AG, Novartis, Roche, Sandoz, and others, reconciling MES, LIMS, and SAP landscapes under strict GxP compliance for batch release, quality certification, and audit readiness. His work now extends into AI-assisted forecasting and planning within regulated environments, applying deep regulatory rigor to next-generation clinical supply technology and digital transformation initiatives for global pharmaceutical and biotech clients. A Dutch national with Asian heritage rooted in Hong Kong and South Korea, Chris brings a genuinely international background to his work, fluent in Cantonese alongside English and Dutch. Chris holds an M.Sc. in international business studies from Maastricht University and is based in the Netherlands.