Enveda tested Claude on batch records.

The big change

  • Manufacturing records support a decision about whether a batch meets established requirements. A reviewer needs to understand both the work performed and any departure from the instructions, because an unexplained discrepancy can affect the release decision.
  • A reviewer still needs the underlying evidence and applicable procedures to judge a finding.
  • A task-level evaluation cannot establish how much time a complete development program will take. For a quality team, the next decision is whether a review method is reliable enough for its defined job and how it fits the existing approval process.

What the pilot measured

The records are part of a larger quality process. FDA’s guidance for active pharmaceutical ingredients describes recording the materials, critical process results and signatures associated with each batch. It calls for critical deviations to be investigated, with consideration of other batches that may be affected. The quality unit reviews the relevant records before an active ingredient is released.

A batch record therefore serves more than an inventory of pages. The approved instructions provide the standard for the work; the executed record shows what was documented during production, and an investigation explains a departure. Without that chain, a reviewer cannot assess a reported discrepancy in context or make a well-supported release decision.

Enveda’s technical report describes four executed records from one Phase 2 manufacturing campaign, covering 1,061 scanned pages. Staff and consultants reviewed records; staff checked AI comments on different records. An adjudicator scored comments, and a second person checked actionability labels.

Across runs, Enveda reports nine direct matches to the 11 distinct actionable human findings. Two adjudicated matches were partial: a discrepancy not traced to an unavailable earlier record and a temperature spike not labeled out of limits.

In the Opus 4.7 pass, 33 of 42 distinct comments needed no action. The same pass added seven actionable findings absent from the human reviews, including a handwritten temperature change without documented confirmation. Each of the two reviewed AI runs misread one entry.

Why the quality decision remains

Claude reviews read-only sections with references; the system reconciles findings and presents comments. The advisory tool is not GMP-validated; the quality unit decides batch disposition. The records are proprietary; the evaluation is not independent replication. BIG CHANGE did not examine the records or reproduce the test.

FDA’s separate data integrity guidance explains why preserving records matters. It recommends keeping data with the associated information needed to reconstruct the activity, and describes audit trails that record creation, changes and deletion of electronic records. It also says an electronic copy needs to preserve the original’s content and meaning to serve as a true copy.

Those safeguards give a reviewer a way to trace a result back to its context. A correction or reference elsewhere can change how an entry should be understood.