AI document review was one of the results in an Anthropic post published on October 27, 2026. The company describes early work with six drug development partners, including Bristol Myers Squibb, Generate Biomedicines and Enveda. The post is by Anthropic's own science team, so every figure below is Anthropic's account.
What Anthropic reports
The clearest checking result came from Enveda. Anthropic says Claude read 1,061 pages of manufacturing and regulatory documentation. It caught all 11 issues that human reviewers had already found. It also flagged seven more that the humans had missed.
Anthropic says human review time fell by 89 percent, from more than 25 hours to under three. It also says Claude's first draft of one regulatory document took four hours, against about three months by hand. Human review of another Claude draft still found one minor error.
Other partners reported lab results. Anthropic says one Generate Biomedicines design was about 70 percent shorter than the standard version. It says BigHat Biosciences tested about 1,010 antibody variants over about 10 weeks.
The limits Anthropic states
Anthropic calls the work early and mostly preclinical, so none of the drugs are in human trials. It also says BigHat's team first suspected the standard stability and clumping metrics were misleading. They reanalyzed the data and changed the protocol.
That detail matters. A measurement can look fine and still point the wrong way.
What it means for people who check data
The Enveda result is a useful pattern, but it is one partner's test, reported by the vendor. No independent review of it is cited in the post. Treat it as a claim to test, not a benchmark to copy.
Three habits follow from the post. First, keep a human check on every AI draft, as the one minor error shows. Second, compare the AI's flags against a known set of issues, as Enveda did. Third, question the metric itself before trusting a good score.
Our page on data verification explains how to trace a claim to its source. The trust index shows how we score evidence, and how it works covers the full review flow.