Two layers of scrutiny
The EU AI Act addresses records and human oversight for high-risk AI. Draft EU GMP Annex 22 brings that discussion into GxP operations, where validation, data integrity and accountable decisions already matter.
Annex 22 remains a draft. Organisations should distinguish its proposed expectations from requirements already in force.
A model is not its own validation
For critical uses, the draft focuses on controlled, explainable behaviour and documented validation. The organisation needs to know which model version and inputs supported an output, whether it met its acceptance criteria and when a qualified person intervened.
Evidence belongs in the production path
ALCOA+ principles make the record as important as the result. Capturing the model, input, output, applicable control and any human override as the decision happens is more reliable than reconstructing them for an inspection later.
EU AI Act Articles 12 and 14 · Draft EU GMP Annex 22 (published 7 July 2025) · ALCOA+ data-integrity principles