Regulations and compliance · Practical guide

AI model change control and monitoring under draft EU GMP Annex 22

An unchanged model can perform worse when conditions change. New materials, image acquisition conditions or preprocessing can alter how the application performs. Use monitoring to decide what changed, which previous conclusion it affects and what must be clarified before continued use. This guide treats Annex 22 as a draft.

Table of contents

At a glance

When should a change to a GMP AI system trigger further testing?

Assess changes to the model, software, preprocessing and operational workflow against the intended use already demonstrated. Investigate monitoring signals that suggest a shift in inputs or performance. Draft Annex 22 calls for a reasoned assessment of the need for further testing. Determine the evidence from the effect and risk of the change, including cases where the model weights stay the same.

  • The assessed system includes more than the model file.
  • A monitoring signal triggers an assessment. It does not establish the cause.
  • New evidence and interim limits on use should address the claim affected by the change.

1. Define the controlled version of the static model

Sources checked on 7 October 2026. Annex 22 was a consultation draft. The proposals described here were not in force. For GMP manufacture of human medicinal products, existing applicable requirements continue to apply, including risk-based assessment of changes.

The focus is on static, deterministic models. Releasing a new model version deliberately is a controlled change. A model that learns automatically in operation has a different intended use. The draft does not provide blanket authorisation for dynamic systems.

Identify the assessed combination of model version, preprocessing, software, data sources and intended process. This allows you to trace a change back to the earlier conclusion about fitness for use that it could affect.

Sources: [1] [2] [3] [4]

2. Describe what the change does

A new camera, different lighting or changed image compression can alter inputs while model weights stay the same. A new decision threshold can also change process outcomes. An unchanged model file therefore leaves the impact of other changes open.

The draft covers changes to the model, the system and the process, including physical conditions. Trace each change through the affected input or processing step, the decision that could go wrong, the existing control and the evidence needed.

A small code change can affect a critical calculation. A major infrastructure update may leave certain process functions unaffected. The number of changed files or lines gives no reliable measure of that impact.

Sources: [1] [4]

Further reading:

3. Distinguish a change in inputs from evidence about performance

A change in the input distribution can give an early warning. It does not establish a drop in model performance. Conversely, a stable average distribution does not demonstrate that every relevant subgroup is still classified suitably.

Identify the signals you can observe and what each one shows. Reliable performance metrics need a suitable reference assessment. If reference results become available later, make that delay clear in the assessment.

Use the worksheet to connect a signal with a suitable investigation. Set alert thresholds for your own use.

Connecting an observed signal to a focused investigation
ObservationWhat the observation does not establishNext step
The image brightness distribution shifts That defect detection has deteriorated Check image acquisition conditions and the affected subgroups
Manual overrides increase That every override represents a model error Distinguish the reason for the override from the reference decision
Sensitivity falls in one subgroup That the model alone caused the change Compare data quality, reference assessment, process changes and version
The model checksum changes unexpectedly That the new version is suitable or approved Preserve the version, assess cause and use, and identify the authorised state

Sources: [1]

4. What a change in the monitoring signal actually shows

An image inspection process uses model M-4 and records brightness as a technical signal. In one period, 12 of 400 images fall outside the previously defined input range. In the next period, 36 of 400 do. The share rises from 3% to 9%, an increase of 6 percentage points to three times the previous share.

The calculation shows that the signal changed. It does not establish that the error rate tripled. The team finds that a lighting unit was replaced. It needs suitable reference cases to determine whether the replacement caused relevant wrong decisions.

For this example, the team agreed beforehand that the responsible process expert would decide whether to add an interim manual check of the affected images when this signal could not yet be explained. The control, its capacity and its suitability are assumptions for the example. The draft sets no such general obligation.

Decision worksheet for model M-4
StepObservation or expected resultDecision
Verify the signal 12/400 = 3 %; 36/400 = 9 % Confirm the measurement definition and that the populations are comparable
Identify the change Lighting replaced. Model M-4 unchanged. Identify the affected group of images
Control interim use The example provides for an additional manual check Document effective execution and the remaining limits
Plan the evidence Reference-assessed cases under the old and new acquisition conditions Compare affected performance and relevant subgroups
Close the assessment Results and remaining limitations are available Record the rationale for continued use, restriction or further change

Sources: [1] [4]

5. Select evidence for the performance claim affected

Start by identifying what existing evidence actually covers. If a new material falls outside the assessed use, a startup test is unlikely to establish suitability for the new classification task. A display change demonstrably unrelated to that task may justify a narrower check.

The draft proposes assessing the need for retesting and documenting the rationale where retesting is omitted. It prescribes neither a complete retest after every change nor a general exemption from extensive testing.

Consider the test data and how the model was selected. If known failure cases are used for optimisation, another run on those same cases provides limited evidence. Link the final assessment to the version actually changed.

Sources: [1] [4]

Further reading:

6. Assign the decision that follows a monitoring signal

Monitoring can produce figures automatically. It does not decide whether a batch, process step or test result may continue to be used. Name the responsible process expert, the QA involvement needed and who will provide technical support to investigate the cause.

Agree how to respond outside usual working hours. A signal with no reachable owner can leave a significant gap. Choose response times according to risk and available interim controls. This guide gives no universal number of minutes.

Record the observation, interim action, investigation and final decision separately. A closed IT ticket alone cannot establish that the boundary for continued use has been assessed.

Sources: [1] [4]

7. Prepare for the next model or process change

An SME can focus on a small set of interpretable signals and defined responses. Missing reference data, unclear version links and unavailable process decision makers are common sources of effort.

  • Document the assessed system configuration and intended operating range.
  • Track changes to the model, software, preprocessing and physical conditions.
  • Define signals, denominators and subgroups.
  • Identify a suitable reference assessment and when it will be available.
  • Agree interim control options with those responsible.
  • Document the need for retesting and the rationale for exceptions.
  • Link the final decision to permitted use and remaining limitations.

Sources: [1] [4]

8. Verify the cause before closing the investigation

Timing can suggest a cause without establishing it. If inspection findings increase after a lighting change, also examine the material mix, reference assessment and other concurrent changes. Correcting an apparently obvious cause may otherwise leave the real problem unresolved.

A return to the previous average is insufficient for closure. A subgroup may still show poor performance. Record the process question the investigation answered and how the results support the chosen limits of use.

Sources: [1] [4]

Project experience and collaboration

Trace an unexpected monitoring signal to its cause

When results change, model version is one possible cause. Different lighting, new inputs or changed preprocessing can produce the same effect. We link the signal to the affected use and existing test results. A limited environmental change may need a targeted retest. A retrained model requires a broader assessment of whether earlier conclusions still hold. We distinguish immediate actions from permanent changes so a temporary fallback does not quietly become routine operation.

Frequently asked questions about AI model changes and monitoring

Must we assess an unchanged model after replacing the camera?

Yes. Assess the impact on the intended use. A camera replacement can alter inputs while the model weights remain unchanged. The impact and available evidence determine the tests needed.

Sources: [1] [4]

Does a drift signal establish a model error?

No. It first identifies a change in observed inputs or metrics. Investigate performance with suitable references and consider other causes. A technical signal and a wrong process decision are different findings.

Sources: [1]

Why is a rise from 3% to 9% more than a 6% increase?

The absolute difference is 6 percentage points. Relative to the starting value, the share rises by 200% to three times its original value. In this example, the share counts images outside the input range. It does not establish the model's error rate.

Sources: [1]

Does every change require full revalidation?

No. Section 10.1 of the draft proposes documenting changes to the model, system and process and assessing the need for retesting. A decision to omit retesting needs a full rationale. The effect of the change may justify a focused additional test or a more extensive reassessment.

Sources: [1] [4]

Can a model continue to learn automatically during operation?

A model that continues to learn automatically in operation falls outside the draft's proposed scope for critical GMP applications. Deliberately retraining a static model and verifying the new version under controlled conditions is a different case. Section 10.1 of the draft calls for documenting the change and assessing the need for retesting. Annex 22 remains a draft. Assess currently applicable requirements separately.

Sources: [1]

Is a new version number sufficient evidence of a change?

No. It identifies a version but cannot establish its impact or suitability. Link it to the reason for change, the functions affected, the evidence and permitted use. Changes outside the model file can also matter.

Sources: [1] [4]

Who decides whether to introduce an interim manual control?

That depends on the process and assigned responsibilities. The responsible process expert assesses whether the control is suitable and workable. Involve QA and IT according to their roles. An additional manual check is not a universal solution.

Sources: [1] [4]

What should the final decision record after an unusual monitoring result?

Record the signal investigated, affected data and versions, assessment of the cause, actions taken and the rationale for continued use. Identify any remaining limitations. A closed technical ticket without a process decision does not provide this evidence.

Sources: [1] [4]

Sources and review date

Sources were checked on 7 October 2026. Annex 22 was a consultation draft. The scope is static, deterministic AI in GMP for human medicinal products. The case numbers, signal assessment and interim action are assumptions for the example.

  1. European Commission, draft Annex 22 – Artificial IntelligenceThe 2025 consultation draft was not an annex in force on 7 October 2026. Sections 1, 3–7 and 10 address static, deterministic models in critical GMP applications.
  2. European Commission, consultation on Chapter 4, Annex 11 and Annex 22Official consultation from 7 July to 7 October 2025, closed by the source check on 7 October 2026. Closure of the consultation does not bring the draft into force.
  3. European Commission, EudraLex Volume 4 official document indexChecked on 7 October 2026. The index listed the January 2011 Annex 11 revision. The veterinary section is separate. Regulations 2025/2091 and 2025/2154 have applied there since 16 July 2026.
  4. European Commission, EU GMP Annex 11 – Computerised SystemsJanuary 2011 revision, applicable from 30 June 2011. Sections 1, 4, 6–7, 9–13 and 16–17. This is published GMP guidance for human medicinal products.

Sources checked: .

Daniel Herrmann, Daniel Herrmann Consulting

Daniel Herrmann Consulting

Link monitoring signals to controlled change decisions

We help your team identify useful signals, assign responsibility and select evidence for changes to models and processes. The aim is a documented decision about continued use of the specific system.

  • AI system validation →

    Bring your AI system into GxP use with traceable evidence and keep it under control in operation. Once the system and intended GxP use are defined, we validate according to risk, addressing data, supplier, tests, release and change control.

Send a project enquiry Customer testimonials →