Can AI be used for GMP batch record review? What regulators expect
By Savi C., Co-Founder · Published Nov 16, 2026
The rules that already apply
An AI review tool is a computerised system in a GxP environment. That brings in 21 CFR Part 11 for electronic records and signatures, EU GMP Annex 11 in its currently effective form, and validation proportionate to risk under your own procedure, typically following GAMP 5.
It also leaves accountability where it was. Under 21 CFR 211.192 the quality control unit reviews and approves production and control records before release. A tool can perform checks and assemble evidence, but the approval remains a human decision in your quality system.
The ten EMA and FDA guiding principles, January 2026
Published jointly to cover AI used across the medicines lifecycle, from research to manufacturing and safety monitoring.
Principles 1 to 5
- Human-centric by design
- Risk-based approach
- Adherence to standards
- Clear context of use
- Multidisciplinary expertise
Principles 6 to 10
- Data governance and documentation
- Model design and development practices
- Risk-based performance assessment
- Life cycle management
- Clear, essential information
Draft Annex 22: what it would change
The European Commission and PIC/S consulted on a draft Annex 22 on artificial intelligence from July 7 to October 7, 2025. It is a draft and is not in force.
As published for consultation, the draft would limit critical GMP applications, those with a direct impact on product quality or patient safety, to static, deterministic models that give the same output for the same input. It would keep dynamic, continuously learning models and generative AI out of critical applications, while allowing them in non-critical uses where a qualified person reviews the output. Manufacturers using AI would remain responsible for it even when it is supplied by a vendor.
Until a final version is published, the practical position is to define each AI use and its risk clearly, keep a qualified person accountable for every decision, and track the draft as it develops.
What to ask an AI review vendor
Fitness for use
- What is the defined context of use for each check, and how is its risk classified?
- How is performance measured, and can it be measured on our own records?
- Does every finding link back to the source page it came from?
Control over time
- How are model, rule and configuration changes controlled and regression tested?
- How is performance monitored for drift after go-live?
- Who approves each disposition, and is the AI ever able to release a batch on its own?
- Where is our data processed, and can it be used to train models for other customers?
Where Litewave stands
Litewave proposes, qualified people approve, and Litewave never dispositions a batch on its own. Every extraction and flag links to its source page. We hold Litewave to the Annex 11 that is currently effective and build toward the drafts rather than claim compliance with standards that are not yet in force.
Questions we get asked
Is EU GMP Annex 22 in force?
No. A draft was published for consultation from July 7 to October 7, 2025. Until a final version is adopted, it is guidance on direction rather than a requirement.
Do we need to validate an AI tool used in batch review?
Yes. It is a computerised system used in a GxP process, so it is validated for its intended use under your own procedure, with effort proportionate to risk, and kept under change control afterwards.
Who is responsible if the AI misses something?
The manufacturer. Using a vendor’s tool does not move accountability, which is why a qualified person approves every disposition and why the tool’s performance should be measured and monitored on your own records.
Sources
- FDA and EMA, Guiding Principles of Good AI Practice in Drug Development
- European Commission, stakeholders consultation on EU GMP Chapter 4, Annex 11 and new Annex 22
- 21 CFR Part 11, Electronic records; electronic signatures (eCFR)
- 21 CFR 211.192, Production record review (eCFR)
- EudraLex Volume 4, Annex 11, Computerised Systems
