Key figures
- 89% Less review time in one timed run, from almost 26 hours to just under three, including human adjudication. 1,061 pages of batch records. (Enveda release, 27 Oct 2026)
- 11 of 11 Actionable issues from the human review also identified by the tool, across its runs. (Enveda release, 27 Oct 2026)
- 7 Additional issues relevant to an auditor, raised in one timed run. (Enveda release, 27 Oct 2026)
Announced by Enveda about its collaboration with Anthropic, the AI vendor. Internal evaluations, not independently verified.
Before a batch of an experimental drug can go to patients, someone has to read the paperwork. All of it. Manufacturing records run to hundreds of pages, every figure has to trace back to its source, and a missed line can cost a batch or delay a trial. It is careful, slow, essential work.
Enveda, a Boulder, Colorado biotech that looks for new medicines in natural chemistry, has been testing whether AI can carry some of that load. On October 27, 2026 it shared the results of two internal projects built with Anthropic's Claude.
The batch-record review
Enveda built a tool it calls the CMC Document Processor, which "flags possible issues for the company's reviewers, who decide what to do about each one." To test it, the company used "1,061 pages of batch records from a Phase 2 manufacturing campaign." Four internal reviewers and a team of outside consultants went through the same records, and a member of the Quality team ruled on every comment.
Across its runs, the tool "identified all 11 actionable issues found during the human review." One timed run "also raised seven additional issues relevant to an auditor," which the Quality team resolved with the manufacturer before the batch was released. Among them: "a process step that exceeded its agreed limit and a test marked 'FAIL' without a recorded result."
Thoroughness and freshness are different skills. The human review was thorough, and it found all 11 actionable issues. A reader that is as alert on page 800 as on page one adds something else: it notices what is simply sitting there. The test marked "FAIL" is a good example. The word was on the page, with no result beside it. The tool raised it, and people decided what it meant, as they did for every finding.
On time, the same timed run "reduced total review time from almost 26 hours for Enveda's four reviewers to just under three hours, including human adjudication, representing an 89% reduction for this use case."
It was not flawless, and its one slip will be familiar to anyone who has squinted at someone else's handwriting. "The system made one error after misreading a handwritten note," the company says, but flagged it as needing verification rather than stating it as fact.
The regulatory draft
The second project tackled part of an Investigational New Drug application, the filing a company sends the FDA before testing a drug in people. Claude first pulled structured facts out of toxicology, bioanalytical and pharmacokinetic reports, then drafted from that shared set of facts.
Enveda compared the result with a section its team had already written by hand. "Claude generated the first draft in four hours, compared with approximately three months to write a comparable first draft manually at Enveda. Neither figure includes time for expert review." The company's Drug Development and Quality teams judged the draft "comparable to work produced by professional medical writers."
Who stays in charge
Both projects kept people in the loop. "The system never wrote directly to a regulated GxP record. Enveda personnel reviewed and approved every finding and draft," and the Quality team "still carries out its own final review of every record before any batch is released."
In Enveda's words, Claude "supported Enveda's experts rather than replacing established quality processes." The same reviewers got through the same pile far faster and ended up with a longer list of real problems.
How to read this
These are Enveda's own internal evaluations, announced in a press release about a collaboration with the AI vendor. They have not been independently verified. The 89% figure comes from a single timed run on one set of records, and the four-hour draft excludes the expert review time that any real filing needs. Enveda has published a preprint describing the review tool for anyone who wants the method.
Still, this is skilled people in Colorado spending less time on page-turning and more on the medicine itself. The question for the next filing is whether, once expert review is counted, a three-month first draft really can become an afternoon.
