Can AI review batch records?

1Who must review the record.

In the US, 21 CFR 211.192 requires that all drug product production and control records, including those for packaging and labelling, “shall be reviewed and approved by the quality control unit to determine compliance with all established, approved written procedures before a batch is released or distributed.” Any unexplained discrepancy, or a failure to meet specifications, must be thoroughly investigated, the investigation extended to other batches that may be associated with it, and a written record made with its conclusions and follow-up.¹ Section 211.22(a) gives the quality control unit the authority to review production records “to assure that no errors have occurred or, if errors have occurred, that they have been fully investigated”, and the responsibility to approve or reject drug products.²

In the EU, medicinal products are not sold or supplied before a Qualified Person has certified each production batch.³ Annex 16 lists what the QP must be assured of before certifying, including that “all records are complete and endorsed by appropriate personnel” and that investigations pertaining to the batch have been completed to a level that supports certification. Those checks “may be delegated to appropriately trained personnel or third parties”; the QP relies on the pharmaceutical quality system and “should have on-going assurance that this reliance is well founded”.⁴

2What a review tool may already do.

EU GMP already accepts records and reviews that concentrate on exceptions. For the batch processing record, Chapter 4 says that where a validated process is continuously monitored and controlled, “automatically generated reports may be limited to compliance summaries and exception/ out-of-specification (OOS) data reports”.⁵ EMA’s GMP questions and answers permit review by exception “when scientifically justified”, and require the automated system and its batch exception report to be tested and validated.⁶

FDA’s data integrity guidance treats audit trail review as part of record review. The people responsible for reviewing a record review the audit trails that capture changes to its data, and because 211.22 requires data review before batch release, the audit trail is reviewed on the same schedule.⁷ Annex 11 sets the condition for any computerised system that takes over a manual task: “there should be no resultant decrease in product quality, process control or quality assurance.”⁸

3Where the draft Annex 22 draws the line.

The draft Annex 22 on AI, published for consultation in July 2025 and still under revision in 2026, applies to AI models in “critical applications with direct impact on patient safety, product quality or data integrity”. It covers static models with deterministic output in those applications, tested against acceptance criteria that are “at least as high as the performance of the process it replaces”. Generative AI and large language models “should not be used in critical GMP applications”.⁹¹⁰

Where a model only gives an input to a decision a person makes, the draft calls it human-in-the-loop. If the model’s testing has been reduced on that basis, the operator’s responsibility belongs in the intended use, their training and consistent performance are monitored like any manual process, and records of the review are kept. Depending on criticality and testing, that “may imply a consistent review and/or test of every output from the model”.⁹ When a model that classifies data is tested, its confidence score for each outcome should be logged, and where confidence is very low the draft says to consider flagging the outcome as “undecided”.⁹

4What FDA has said about AI and release.

FDA’s January 2025 draft guidance on AI for regulatory decision-making uses a release example. An AI visual-inspection system checks the fill level of every vial, while release testing still measures fill volume on a representative sample of each batch. Because of that independent check, “the AI-based model will not be the sole determinant for the release of product”, which lowers the model’s influence and so its risk. The guidance adds that the quality control unit responsibilities in 211.22 and 211.68 apply to AI used in manufacturing.¹¹

In April 2026 FDA’s warning letter to Purolea Cosmetics Lab cited a quality unit that had failed to ensure batch records were reviewed before release. It also told the firm that if it resumed production and used AI for CGMP work, any output or recommendation from an AI agent “must be reviewed and cleared by an authorized human representative of your firm’s QU”.¹²

5Which parts of a review AI can take on.

6How it runs in Seal.

In Seal, Neil can prepare a batch for review. It reads the batch execution record and your review requirements, the linked QC results and deviation records, and the recorded signatures and approval status, then drafts a review brief with an evidence checklist, open items and the source behind each finding. QA verifies the evidence and decides disposition; in the EU a QP certifies the batch. Neil’s summary does not release it or replace signatures.

With electronic batch records, Seal evaluates completeness and exceptions while the batch runs, so review begins before it ends. Values outside limits, corrections and overrides, alarms and open samples or deviations come to the reviewer with every instruction, value and signature still available, and QA can disposition the batch only when the effective release requirements are resolved. Neil drafts and proposes; it signs, approves and releases nothing.