Retour au blog
Cas d'usage 11 min

AI and Pharmaceutical Manufacturing: GMP, Quality and Annex 22 in 2026

Belgium bills itself as a « Biopharma Valley ». In 2024, its biopharmaceutical exports reached 79.0 billion euros, roughly 216 million a day (pharma.be). In one of the most regulated sectors on earth, AI is leaving the lab for the production floor: the McKinsey Global Institute puts the potential value of generative AI for pharma and medtech at 60 to 110 billion dollars a year. The decisive question remains: how do you manufacture a medicine with AI without losing GMP compliance? The answer is being written right now, with the new European Annex 22, the FDA guidance and the EMA reflection paper.

Article generated by AI. Content written with the help of an artificial intelligence model and reviewed by a human before publication. The figures cited point to their sources, listed at the end of the article.

The Sector in Numbers

$60-110B
Potential annual economic value of generative AI for pharma and medtech (McKinsey Global Institute)
€79B
Belgian biopharmaceutical exports in 2024, that is 216 million euros a day (pharma.be)
Annex 22
New EU GMP framework for AI: consultation closed October 2025, EMA workshop July 2026

Pharma did not wait until 2026 to use models: AI-assisted drug discovery is already common. What changes is the shift of AI towards manufacturing and quality control, two areas where every decision is traced, validated and auditable. The McKinsey Global Institute, which analysed 63 generative AI use cases in life sciences, values the potential at 60 to 110 billion dollars a year for pharma and medtech, of which 18 to 30 billion for commercial functions alone.

Belgium is on the front line. Its 79.0 billion euros in biopharmaceutical exports in 2024 make the sector the country's largest exporter, but that figure is down 1.4% and employment down 0.5%, for the first time in years (pharma.be). Amid cost pressure and heavy dependence on the US market (nearly 24% of exports), automating production and quality is no longer a luxury: it is a competitiveness lever.

The challenge: AI meets "GMP by design"

In pharmaceutical manufacturing, a model is judged not on its average accuracy but on its ability to be validated, explained and replayed. An algorithm that releases a batch or drives a critical process must produce reproducible, documented decisions. That is exactly the friction with generative AI: the draft Annex 22 held that dynamic, probabilistic models should not touch critical GMP applications. The whole question of 2026 is which guardrails could relax that line.

What AI Changes Concretely in Production

Beyond discovery, four uses are now mature or being deployed in pharmaceutical plants.

The 2025-2027 Framework: Annex 22, FDA, EMA

Three texts now structure every AI project in pharma. The system you design in 2026 will be judged by these standards.

Annex 22: the GMP of AI (EU)

A new EudraLex Volume 4 annex dedicated to AI and machine learning in manufacturing. It sets requirements for model selection, training and validation, data quality and continuous monitoring. Public consultation from 7 July to 7 October 2025, with a parallel revision of Annex 11 (computerised systems).

EMA: the 2026 workshop

The draft barred generative and probabilistic models from critical GMP applications. After a consultation that favoured relaxing this, EMA convened an expert workshop on 30 June and 1 July 2026 to build a risk-based approach: guardrails, human oversight, data governance.

FDA: the January 2025 guidance

On 7 January 2025 the FDA published its first draft guidance on AI to support regulatory decision-making (nonclinical, clinical, post-market, manufacturing). A risk-based credibility assessment framework. The FDA says it has received over 500 submissions with AI components since 2016.

EMA: the reflection paper

Adopted on 9 September 2024 by the CHMP and CVMP, it covers the entire medicine lifecycle, from discovery to post-authorisation surveillance, advocating a risk-based approach to the development, deployment and performance monitoring of AI tools.

The Method: Industrialise AI Without Losing the Inspector

The risk is not adopting AI, but adopting it without validatable governance. Inspectors do not ban models: they demand to be able to audit them. Four steps structure a defensible rollout.

1

Define the intended use and risk level

A model that assists non-critical visual inspection is not validated like one that releases a batch. Classify each use case by patient impact and detectability, in the spirit of ICH Q9 on quality risk management.

2

Start with non-critical uses

Quality vision, document prioritisation, drafting assistance: quick wins, low regulatory risk, that build data quality and team trust before you approach critical processes.

3

Document the model end to end

Training and test data, performance metrics, explainability, robustness tests, change control and a human review procedure. That is the file a GMP inspector, Belgium's FAMHP or the FDA will want to see.

4

Monitor drift continuously

A model validated once is not validated forever. Track production performance, detect drift, trigger retraining under change control and report to the quality system.

The right reflex

Do not put generative AI on the critical process first. Start with a low-risk, high-return use such as visual inspection or document triage: a return on investment within months, without touching the validated GMP baseline, and the foundations (data, governance, quality culture) to tackle sensitive applications next.

What It Changes for the Belgian Company

The subject reaches beyond the big production sites. Belgium's Biopharma Valley is also subcontractors (CDMOs), contract manufacturers, control laboratories, distributors and equipment makers, often SMEs, all subject to the same GMP. For them, Annex 22 is not a distant obligation: it is the common language a client will demand as soon as a model touches a batch or a quality record.

Concretely: a filling site can deploy AI visual inspection with a validation file aligned to Annex 22; a quality control lab can automate chromatogram reading while keeping human review; a CDMO is well advised to formalise its model governance now, to reassure clients and anticipate inspections. The logic matches our sector-by-sector analysis of the EU AI Act: map the uses, classify the risk, document, and reserve humans for the decisions that require them.

Frequently Asked Questions

Can generative AI be used in GMP medicine manufacturing?

With caution and under conditions. The July 2025 draft Annex 22 held that dynamic, adaptive and probabilistic models, such as generative AI or LLMs, should not be used in critical GMP applications. After the 2025 consultation, EMA is reassessing this position and convened an expert workshop on 30 June and 1 July 2026 to define a risk-based approach, with guardrails and human oversight.

What is Annex 22 of the European GMP?

The new EudraLex Volume 4 annex dedicated to AI and machine learning in the manufacture of active substances and medicines. It sets requirements for model selection, training, validation, data quality, continuous monitoring and human review. The draft was in public consultation from 7 July to 7 October 2025, alongside Annex 11 and Chapter 4.

What does the FDA say about AI in drug development?

On 7 January 2025 the FDA published its first draft guidance on the use of AI to support regulatory decision-making for drugs and biological products. It proposes a risk-based credibility assessment framework covering the nonclinical, clinical, post-market and manufacturing phases. The FDA says it has received over 500 submissions with AI components since 2016.

Is AI in pharmaceutical production high risk under the EU AI Act?

AI used in medicine manufacturing is first governed by sectoral pharmaceutical law: GMP, Annex 22, Annex 11 and the EMA reflection paper. The EU AI Act articulates with that harmonised legislation rather than mechanically reclassifying every production use as high risk. GDPR, model governance, explainability and traceability remain required.

Sources

  1. McKinsey & Company, Generative AI in the pharmaceutical industry: moving from hype to reality ($60 to $110 billion potential annual value for pharma and medtech; 63 use cases analysed; $18 to $30 billion for commercial functions). mckinsey.com
  2. European Commission, DG SANTE, Stakeholders' consultation on EudraLex Volume 4, new Annex 22 (Artificial Intelligence) (opened 7 July 2025, closed 7 October 2025; parallel revision of Annex 11 and Chapter 4). health.ec.europa.eu
  3. EMA, Multistakeholder workshop on AI guidance development (Annex 22), 30 June and 1 July 2026 (the draft held that dynamic, adaptive and probabilistic models, GenAI and LLMs, should not be used in critical GMP applications; seeking guardrails and a risk-based approach). ema.europa.eu
  4. FDA, Artificial Intelligence for Drug Development / Draft guidance on the use of AI to support regulatory decision-making, 7 January 2025 (risk-based credibility assessment framework; more than 500 submissions with AI components received since 2016). fda.gov
  5. EMA, Reflection paper on the use of artificial intelligence in the lifecycle of medicines (adopted 9 September 2024 by the CHMP and CVMP; risk-based approach, from discovery to post-authorisation surveillance). ema.europa.eu
  6. pharma.be, New pharma.be figures, 2 April 2025 (79.0 billion euros in biopharmaceutical exports in 2024, that is 216 million a day, down 1.4%; employment down 0.5%; United States nearly 24% of exports). pharma.be
  7. EMA, 2024 Annual Report on EudraVigilance for the European Parliament, the Council and the Commission (about 1.76 million individual case safety reports, ICSRs, managed in 2024). ema.europa.eu

Is your pharma production AI project ready for Annex 22?

Molderez Consult helps production sites, control laboratories and CDMOs in Belgium map their AI use cases, classify GMP risk, frame model governance and validation, and prepare for inspections in the Annex 22 era.

Assess my project
Article generated by AI. Content written with the help of an artificial intelligence model and reviewed by a human before publication. The figures cited point to their sources, listed at the end of the article.
Partager