Since July 2025, the draft of Annex 22 has been available — the first dedicated GMP framework for artificial intelligence in pharmaceutical manufacturing. It is not yet in force. On May 12, SAP announced a strategic stake in n8n at the Sapphire conference in Orlando — a Berlin-based workflow platform whose solution will be natively integrated into SAP's Joule Studio from Q3 2026. For the pharmaceutical industry, this marks an architectural decision with direct implications for the practical implementation of Annex 22.

What the draft Annex 22 demands from industry

With the draft Annex 22 to the EU GMP Guide, a framework addressing AI and machine learning in drug manufacturing exclusively is available for the first time. It was drafted by the EMA GMDP Inspectors Working Group in cooperation with PIC/S and released for consultation by the European Commission as part of EudraLex Volume 4. The key distinction is drawn between critical and non-critical GMP applications:

  • Model type restriction: Probabilistic models, dynamically learning systems, generative AI, and LLMs should not be used in critical GMP applications — the draft explicitly excludes them from its scope. What it covers are static models with deterministic output, for which it sets out testing and validation requirements.
  • Intended use: Every AI system requires documented purpose definition — scope, data, operating conditions — before going live.
  • Lifecycle validation: Continuous performance monitoring is required. If data or model parameters change, validation must be updated accordingly.
  • Human-in-the-loop: For adequately tested, deterministic models in critical applications, HITL is not mandatory. An explicit HITL requirement applies to the use of generative AI in non-critical applications.
  • Explainability and traceability: Models must be interpretable. Full traceability across input data, model version, output, and release decision must be maintained end-to-end.

Why SAP's investment is regulatorily relevant

n8n combines deterministic automation with agentic AI capabilities. Jan Oberhauser, founder and CEO, describes the approach: "Some work is deterministic — one right outcome, everything else is an error. The rest is non-deterministic."

This distinction maps precisely onto the boundary Annex 22 draws. Three use cases illustrate this:

  • Batch release — a critical GMP application. Static, deterministic models only. AI agents and LLMs are categorically excluded. The process transports data deterministically; an authorized person signs off.
  • Deviation analysis — as a non-critical application, LLM-assisted tools may produce a draft report — under strict HITL. Final responsibility remains with the quality responsible person.
  • Validation documentation — non-critical, provided the draft has no direct influence on release decisions. A machine-generated draft functional spec under HITL is defensible. An automatically approved FS is not.

Three questions for IT, QA, and the business

First: Where does the validated process end — and where does technical integration begin? Business logic belongs in an orchestration layer with its own audit trail. Technical data transport belongs in an encapsulated integration layer. SAP's adoption of n8n confirms this architecture pattern at the enterprise level.

Second: Can you evidence the core requirements of the draft Annex 22 for every AI deployment? In consulting practice, the gaps cluster in lifecycle validation and cross-system traceability.

Third: What vendor dependencies are created — and which are deliberate choices? SAP currently holds 1.3% of n8n. n8n remains independent; self-hosting is possible. That said, what is a partnership today may evolve into full integration over the medium term.

The bottom line

In critical applications: only deterministic, validated models. In non-critical applications: AI may assist — under human oversight and documented accountability.

That is the distinction the draft demands. And it is the distinction that a system architecture must be able to reflect — not as a retrofit governance layer, but as structural foundation. The tools are improving. The question is whether the architectures are ready to deploy them appropriately.