Why AI is arriving at batch review now
Batch record review is the kind of work AI is good at: reading long documents, comparing them against rules, and spotting what is missing or inconsistent. It is also work that quality teams struggle to staff. A single executed batch record can run to hundreds of pages, and the release clock does not stop while reviewers read.
At the same time, regulators spent 2025 and early 2026 setting out how they expect AI to be used in GMP work:
- January 6, 2025: FDA published draft guidance, Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products. Law firm DLA Piper notes it lists “using AI in pharmaceutical manufacturing” among covered uses and sets out a seven-step, risk-based credibility framework built on the model’s “context of use.”
- July 7 to October 7, 2025: The European Commission consulted on a new EU GMP Annex 22 on artificial intelligence, alongside a revised Annex 11.
- January 14, 2026: EMA and FDA jointly issued 10 guiding principles for good AI practice in drug development, including a risk-based approach and a defined context of use (as reported by RAPS).
- April 2, 2026: FDA issued a warning letter to Purolea Cosmetics Lab that addressed a firm’s use of AI agents for GMP documents directly.
Together these give a clear picture of where AI can and cannot sit in batch review.
For how AI speeds batch record review without weakening data integrity, see AI Batch Record Review: Faster Release Without Weaker Data Integrity.
What draft Annex 22 says
Draft Annex 22 is the most specific regulatory text on AI in GMP so far. Its scope is narrow on purpose.
In scope for critical applications: machine learning models that are static (frozen after deployment) and deterministic (same input, same output), used where there is “direct impact on patient safety, product quality or data integrity.”
Not for critical applications:
- Dynamic models that keep learning in use “should not be used in critical GMP applications.”
- Probabilistic models that can give different outputs for the same input are not covered and should not be used in critical applications.
- The draft states that “the document does not apply to Generative AI and Large Language Models,” and these should not be used in critical GMP applications either.
Non-critical use: where such models are used without direct impact on safety, quality or data integrity, the draft calls for a human-in-the-loop, with qualified people checking the outputs.
For models in scope, the draft sets demanding requirements: test data that “expand the full sample space of the intended use,” acceptance criteria “at least as high as the performance of the process it replaces,” recording of the features that drove each result (explainability), and logging of confidence scores, with low-confidence outputs flagged as “undecided.”
What this means for batch release: an AI system should not be the thing that decides whether a batch is released. That is a critical application by any reading. AI can support the people who decide, provided its output is checked.
The April 2026 warning letter: AI without review
The Purolea letter shows what regulators object to. The firm told FDA it used AI agents to create drug specifications, procedures and master production and control records. FDA’s response was that if AI is used to aid document creation, the firm “must review the AI generated documents” to confirm they are accurate and CGMP compliant, citing 21 CFR 211.22(c).
The letter also records that the firm had not performed process validation and said it did not know this was required because the AI agent “never told you it was required.” Among other findings, batch records were not reviewed before release. FDA said that if the firm resumes production and uses AI for CGMP activities, AI output must be “reviewed and cleared by an authorized human representative” of the quality unit.
The lesson is not that AI is unacceptable. It is that AI does not carry quality unit responsibility. The person who signs remains accountable for what the AI produced.
Where AI fits in batch record review
A practical way to place AI is by the type of task.
| Task | Example | Suitable for AI support? | Human role |
|---|---|---|---|
| Completeness | Find blank fields, missing signatures | Yes | Confirm and correct |
| Rule checks | Values against limits; dates in sequence | Yes (rule engines often better) | Confirm |
| Cross-document checks | Lot numbers in BMR versus CoA and inventory; equipment versus logbooks | Yes, a strong use case | Confirm mismatches |
| Reading free text | Operator comments, deviation descriptions | Yes, to summarize and flag | Judge significance |
| Pattern detection | Same error across batches; recurring deviations | Yes | Decide on CAPA |
| Impact assessment | Does this deviation affect batch quality? | Draft only | Decide |
| Disposition | Release or reject | No | Decide and sign |
The common thread: AI finds and presents; people decide. Every AI finding should link to the source page so a reviewer can check it in seconds. A finding that cannot be traced to evidence is not usable in a GMP review.
Batch release AI: a safe operating model
If you are planning to use AI in batch review or release, build these controls in from the start:
- Define the context of use. Write down what the AI does and does not do. Example: “Identify potential gaps in executed batch records for QA reviewer confirmation. Does not approve, reject or close any record.” This matches the FDA draft guidance’s emphasis on context of use.
- Classify criticality. If the AI output can directly change a GMP decision without human review, it is critical, and draft Annex 22 limits which models can be used. Keeping a human reviewer between the AI and the decision is what keeps most review-support uses non-critical.
- Require evidence links. Each finding must cite the record, page and rule it relates to.
- Measure performance against your current process. Run AI-assisted and standard review side by side and compare what each finds. Draft Annex 22 asks for performance at least as high as the process replaced; it is a sensible benchmark even outside its scope.
- Do not let AI narrow the review silently. If reviewers start reading only what the AI flags, you have created review by exception without the validation that RBE requires.
- Record the human decision. The reviewer’s acceptance or rejection of each AI finding should be recorded, with their name and the date.
- Control changes. Model versions, prompts and rule sets belong under change control, like any other GMP configuration.
What to ask an AI batch review vendor
- Is the model static, or does it change with use? How are version changes controlled?
- Is it deterministic? If not, how do you make sure the review result is reproducible?
- Does every finding cite its source?
- What happens when the model is unsure? Is low confidence shown?
- How was performance measured, on what data, and against what baseline?
- Is any customer data used for training?
- Which Part 11 and Annex 11 controls apply to the outputs?
AI batch review readiness checklist
- Context of use written and approved
- Criticality classified; AI kept out of the disposition decision
- Every AI finding links to source evidence
- Human confirmation of findings is recorded
- Side-by-side performance comparison completed
- Review scope is not reduced on the basis of AI output alone
- Model, prompt and rule changes under change control
- Training for reviewers on AI limits and failure modes
- SOP updated to describe AI use in batch review
Try evidence-linked AI review with AskDona
AskDona is built for the support role described above. It reads batch records, CoAs, deviation reports and SOPs, checks them against 21 CFR 211, EU GMP and your procedures, and returns findings linked to the exact page and requirement. Reviewers confirm or reject each one; AskDona does not release or reject batches. Run a sample batch record review on a batch you have already reviewed and compare the results. AskDona shows the evidence; your quality unit decides.
FAQ
Can AI review batch records? AI can support review by finding gaps, mismatches and patterns, but the quality unit must review and approve the records under 211.192.
Can AI release a batch? No. Release authority sits with the quality control unit (US) or the QP (EU). Draft Annex 22 also keeps generative AI and probabilistic models out of critical GMP applications.
Does EU GMP allow large language models in batch review? Draft Annex 22 does not cover them and says they should not be used in critical applications. In non-critical, supporting roles, it expects a human in the loop.
What did FDA say about AI in the April 2026 warning letter? That AI-generated GMP documents must be reviewed for accuracy and compliance, and that AI output must be reviewed and cleared by an authorized person in the quality unit.
How should AI performance be measured? Against your current review process, on representative records, with findings verified by experienced reviewers.
References
All pages opened and checked by October 14, 2026.
| Source | Type | Used for |
|---|---|---|
| Draft EU GMP Annex 22: Artificial Intelligence (2025 consultation) | EU draft guideline | Scope, critical applications, performance and explainability |
| EC consultation page | EU consultation page | Consultation dates |
| FDA warning letter, Purolea Cosmetics Lab (April 2, 2026) | FDA warning letter | Statements on AI-generated documents |
| DLA Piper, FDA releases draft guidance on use of AI | Law firm analysis | FDA draft AI guidance, context of use, manufacturing scope |
| RAPS, EMA and FDA issue joint AI guiding principles | Professional association news | January 2026 principles |
| 21 CFR Part 211 (eCFR) | US regulation | 211.22 and 211.192 |
| EU GMP Annex 16 | EU GMP guideline (copy hosted by ECA) | QP certification |