The Inspection Model Passed Its Test. The Retest Decision Is Waiting on QA.

The EU and PIC/S draft Annex 22 on Artificial Intelligence went out for public consultation from July 7 to October 7, 2025. It is still a draft and not in force. It covers AI models used in critical applications with direct impact on patient safety, product quality or data integrity, and its example is a model that classifies products as "accept" or "reject."

Section 10.1 is about operation. A tested model, its system and the process it supports "should be put under change control before it is deployed in operation." Any change to the model, the system or the process "should be documented and evaluated to determine if the model needs to be retested. Any decision not to conduct such retest should be fully justified." Section 4.2 says a process subject matter expert should be responsible for defining the acceptance criteria.

In early October 2026, EMA published the report of its June 30, 2026 workshop on the revision of Annex 22. It records that the drafting group has discussed widening the annex to adaptive, probabilistic and generative models, provided they are "supported by a fully documented, robust, risk-based control strategy." One of the key messages recorded from the discussion reads: "Accountability remains with qualified/responsible persons and cannot be delegated to AI."

Hypothetical: a fill-finish site uses a tested vision model to flag vials for rejection. Engineering replaces a lighting fixture on the inspection line. The change record is complete: the old and new specifications, a side-by-side image check and the engineer's assessment that the input conditions are unchanged.

Under the site's procedure, only the head of QA can decide whether the model needs a full retest, and the head of QA is at a supplier audit until next week. The line stays on manual inspection while the record waits.

The file is ready. The person who can release it is not.

Manual inspection slows the line and adds its own variability. If the retest decision is finally made in a hurry, the justification is thin, which is exactly what the draft annex says it should not be.

The change arrives as one decision packet: the change record, the intended use, the acceptance criteria, the input monitoring data and a drafted justification for or against a retest.

A named QA approver with written delegated authority decides within a day. The justification is recorded at the moment of decision, the process expert's input is attached, and the head of QA reviews the decision on return. Accountability stays with a qualified person, and the line moves.

WisdomTwin.ai is building a sovereign judgment layer for regulated enterprises. It is designed to give each named approver a role-specific judgment twin that prepares the decision packet the way that person reviews, with the evidence attached, so the decision moves faster and leaves an auditable record. It is designed to run on-premises, on-device or in your private cloud, and the named person still makes the call.

We don't replace judgment. We remove the wait.

Examples in this issue are hypothetical, and nothing here describes a customer result.

Book a 20-minute call: https://calendly.com/romanbodnarchuk/20min

Roman Bodnarchuk

Co-Founder and CEO, WisdomTwin.ai