The Reviewer That Reads First

When Anthropic described its drug-discovery partnerships on Oct. 27, its most concrete result came from Enveda, one of six partners. Over 1,061 pages of manufacturing documents, the company wrote, Claude “caught all 11 issues that human reviewers had spotted,” flagged seven more that they had missed and “reduced human review time by 89%.” Enveda published the technical report behind those claims the same day. It describes something narrower than a machine that outperforms people, and in some ways more useful: a review in which people start from a machine’s list.

The documents are executed batch records, the scanned paperwork showing how each batch of an investigational drug was actually made. Enveda’s tool, which it calls CDP, reads copies of them, attaches comments and proposes an advisory outcome; the report says it does nothing beyond that. It never edits the originals, and classifying deviations and deciding a batch’s fate stay with the company’s quality unit. That boundary predates the tool. U.S. drug-manufacturing rules require that production and control records “shall be reviewed and approved” by a quality control unit before a batch is released (21 CFR 211.192). CDP was built to sit in front of that signature.

The trial was small, and Enveda calls its analysis “exploratory.” It covered four batch records from a single Phase 2 clinical manufacturing campaign. Two internal teams, each pairing a manufacturing-chemistry reviewer with a quality reviewer, reviewed two records by hand and checked the tool’s comments on the other two. Each team thus never saw the same record twice, which keeps memory of a first reading from speeding the second.

Anthropic’s summary reads as a single result; the report describes several runs. Human reviewers raised 11 distinct issues that needed action, and the report says the tool raised each of them “in at least one of its runs.” The seven additional issues come from a run on Claude Opus 4.7 and were confirmed by one quality adjudicator who had not reviewed the records. A later run on Claude Opus 5.5 found “no missed human findings” and six actionable issues no person had raised, four of them new; its review time was not recorded. The report gives no run-by-run table showing which run caught which of the 11, so the summary alone does not tell a reader how its figures line up with particular runs.

The 89% follows from the reported totals: 166 minutes for the assisted review against 1,553 minutes for unaided review, about a ninth of the time. The 166 minutes include the tool’s processing as well as the reviewers’ check of its comments, so the figure describes the assisted review as a whole rather than human effort alone. That is a difference in what was measured, and it does not inflate the result: whatever share of the 166 minutes was machine time, the human share was no larger.

One ratio in the report got no mention in Anthropic’s summary. Of the human reviewers’ 96 distinct comments, 79, or 82%, were judged to need no action. Of the tool’s 42 comments in the timed run, 33, or 79%, were judged the same way. The trial did not ask whether those comments were worth making—a request for clarification can matter in a quality review even when nothing has to be fixed. It does show that sorting through comments that lead to no action is part of the work for people and machine alike, and that the assisted reviewers did that sorting for the tool’s comments within the 166 minutes. What changed is who drafts the first list.

That change is also where the trial runs out. The report notes that its measure “does not capture misreads that produce no comment,” and its reviewers saw the tool’s comments “presented in advance.” The design cannot show whether people who begin from a machine’s list look as closely at what the list leaves out, and an issue that both the people and the tool missed would leave no mark in these numbers. That is my reading of the design, not a finding of the report. It is the question a quality unit would most want answered before leaning on such a tool across many campaigns instead of four records from one.

Regulators have been working out where AI belongs in drug manufacturing. In 2023 the Food and Drug Administration’s drug center issued a discussion paper on the subject as a request for comments. Three years later the detailed numbers for this particular use come from the developer’s own exploratory study, with a generic companion implementation released as open source, the production prompts and rule sets kept proprietary, and supplementary data that the report says will accompany the analysis.

Anthropic is plain about the larger program—“All of this research is still at an early stage,” it writes, most of it preclinical. In Enveda’s paperwork the machine already reads every page first and writes in the margins, while the decision on each batch stays where the rule puts it. The next trial should measure what people still find on the pages where the machine wrote nothing.