Insight Claude Found 11 Drug Manufacturing Issues and Cut Review Time by 89% in Enveda Tests

Enveda's Claude-powered document review system identified all 11 actionable issues found by human reviewers across repeated runs on 1,061 pages of pharmaceutical manufacturing records. A timed run also identified seven additional issues, while reducing the measured review workflow from 1,553 minutes to 166 minutes, an 89% reduction.

The findings were published on 27 October 2026 as part of a collaboration between Enveda and Anthropic. They suggest that AI can help specialists inspect lengthy technical records, but they do not establish that Claude can independently approve pharmaceutical manufacturing documents.

The evaluation involved four records from one clinical manufacturing campaign. Human reviewers remained responsible for assessing the AI's findings, and Enveda's Quality department retained the authority to decide whether a batch could be released.

We examined the company's technical report and the accompanying announcements to distinguish the measured results from broader claims about AI in pharmaceutical development.

Key findings

• 1,061 pages reviewed: The evaluation covered four executed batch records from one Phase 2 manufacturing campaign.

• 11 human-identified issues matched: Across repeated runs, the Claude-powered system raised every actionable issue identified by human reviewers.

• Seven additional issues: A timed run identified further actionable problems that the human reviewers had not flagged.

• 89% shorter measured review workflow: AI processing and human review of its comments took 166 minutes, compared with 1,553 minutes for unaided review.

• 79% of AI comments required no action: In the timed run, 33 of 42 AI-generated comments were judged non-actionable, compared with 79 of 96 human comments, or 82%.

How Claude was used to review pharmaceutical records

Drug manufacturers must document how products are made, tested and controlled. These records help quality specialists determine whether manufacturing processes meet applicable requirements.

Under Good Manufacturing Practice, manufacturers maintain evidence covering production steps, equipment, testing, deviations and other quality-related activities.

An executed batch record documents what happened during a particular manufacturing run. Reviewers compare it with approved instructions and supporting records to identify missing information, inconsistencies and possible departures from required procedures.

These records can contain scanned pages, handwritten entries and technical tables. Reviewing them requires attention to both individual details and the relationships between documents.

Enveda developed a system called CMC Document Processor, or CDP, to assist with this work. CMC refers to Chemistry, Manufacturing and Controls.

The system connects Claude to manufacturing documents stored in Enveda's central compliance environment. It identifies the relevant material and batch, retrieves governing instructions and specifications, and examines records for possible quality issues.

Rather than changing official records, the system produces advisory comments that human reviewers can investigate.

How the document review system works

Enveda's approach involves more than uploading a document to a general-purpose chatbot.

The CDP retrieves supporting material, including approved manufacturing instructions and specifications, so the model can compare an executed record with the requirements that apply to it.

Documents are processed in groups of consecutive pages. The system is designed to report pages it has not evaluated, reducing the risk that missing coverage goes unnoticed.

Claude then identifies possible errors, inconsistencies and questions requiring human attention.

Findings appear as comments associated with the relevant documents. Reviewers inspect the evidence and decide whether each comment requires action.

The model receives read-only document copies. It cannot directly modify the official manufacturing record or authorise batch release.

This separation is important. The AI supports the review process, while the validated document system remains the official record and qualified personnel retain responsibility for decisions.

Claude versus human reviewers: The test results

Enveda evaluated the CDP using four executed batch records from a Phase 2 clinical manufacturing campaign.

The records contained 1,061 scanned pages, including substantial handwritten material.

Enveda compared the AI system's findings with reviews performed by internal specialists and experienced external consultants.

A Quality department member who had not previously reviewed the records assessed the comments, with another department member confirming the scoring.

The most important qualification concerns the issue-detection result.

The system identified all 11 human-found actionable issues across its repeated runs. The published technical report does not establish that every issue was found in a single execution.

The seven additional actionable findings came from a particular timed run. These results should not be combined into a claim that one run consistently detected every known and additional issue.

What the additional findings revealed

The seven additional issues were judged relevant by Enveda's Quality department and addressed with the manufacturer before the batch was released.

Examples included a manufacturing process step that exceeded an agreed time limit and a test marked as failed without a recorded result.

These are the kinds of record-level discrepancies that may require investigation or correction.

The findings suggest that AI-assisted review can complement human attention, particularly when records contain many repetitive checks and cross-document comparisons.

However, additional comments are not automatically useful. An AI system can also flag harmless details, misunderstand handwritten entries or raise questions that do not require action.

In the timed evaluation, 33 of 42 AI comments required no action. Human reviewers also produced a substantial proportion of non-actionable comments, 79 of 96.

These results show why the number of comments alone is a poor measure of review quality.

Why review time fell by 89%

Enveda's internal reviewers spent 1,553 minutes reviewing the records without AI assistance, equivalent to 25.9 hours.

The timed CDP workflow, including model processing and human checking of the generated comments, took 166 minutes, or approximately 2.8 hours.

This represents the reported 89% reduction.

The comparison is meaningful because both workflows addressed the same set of manufacturing records.

However, the assisted workflow began with AI-generated comments for reviewers to inspect, while the unaided reviewers had to examine the documents and identify potential issues themselves.

The measurement also does not establish the time needed for every subsequent quality activity. Enveda's Quality department continued to conduct its final review before authorising batch release.

We would therefore describe the result as a reduction in the measured document-review workflow, not proof that the entire pharmaceutical quality process became 89% faster.

Errors and limitations in Claude's review

Enveda reported that its evaluation did not detect hallucinated findings, meaning the CDP did not cite content absent from the reviewed records.

That is encouraging, but it does not establish that the system is incapable of hallucinating or misinterpreting documents.

In the timed run, Claude misread a struck-through handwritten correction. The system marked the interpretation as requiring verification rather than presenting it as certain.

This example illustrates the importance of communicating uncertainty.

Scanned manufacturing records can contain handwriting, corrections, poor image quality and unusual formatting. These features may be difficult for automated systems to interpret consistently.

The evaluation was also limited to four records from one campaign. It does not demonstrate the same detection performance across different manufacturers, document formats or regulatory environments.

Furthermore, Enveda developed and evaluated the system for its own operations. The findings have not been independently reproduced across a broad range of pharmaceutical organisations.

Claude also helped draft regulatory documents

The collaboration extended beyond manufacturing record review.

Enveda used Claude to prepare substantial portions of Module 2 of an Investigational New Drug application.

The workflow extracted structured information from source documents covering areas such as toxicology, pharmacokinetics and bioanalysis.

These facts were organised into a shared data layer before Claude generated draft regulatory content.

This approach helped the team maintain consistency between sections and trace statements back to their underlying sources.

Enveda reported that Claude generated a first draft in four hours, compared with approximately three months for a comparable manually prepared draft.

Neither figure includes the subsequent expert review process.

The company also used Claude to draft an Investigator's Brochure. Reviewers identified one minor unit-conversion error, which was corrected before the document was submitted as part of the regulatory application.

These examples demonstrate potential time savings in document preparation, but they do not show that regulatory submissions can safely proceed without specialist review.

Why the system retained human oversight

Pharmaceutical manufacturing decisions can affect product quality and patient safety.

For this reason, Enveda designed the CDP as an advisory system rather than an autonomous decision-maker.

The AI could identify possible problems, but qualified staff decided whether the comments were valid and what action was required.

The system could not write directly to regulated records, approve deviations or release manufacturing batches.

Enveda also stated that the CDP had not been validated as a Good Manufacturing Practice computerised system.

This limits the role it can play within regulated operations.

For organisations considering similar tools, the important distinction is between using AI to assist a reviewer and allowing AI to perform a regulated decision independently.

The published results support investigation of the former approach. They do not establish the safety or compliance of the latter.

What this means for AI document processing

The findings are relevant beyond pharmaceutical manufacturing because many organisations review lengthy, structured records against established requirements.

Potential applications include technical quality assurance, internal compliance checks and document consistency reviews.

However, Enveda's results should not be treated as direct performance evidence for those other industries.

The system benefited from access to governing documents, defined review criteria and specialists who could judge the resulting comments.

Those conditions are important when designing similar workflows.

A general-purpose chatbot reading an isolated file would not necessarily reproduce the CDP's performance.

Teams also need to consider data security, document permissions, record retention, audit trails and the reliability of citations linking findings to original material.

What this means for you

We recommend treating this study as an example of a carefully constrained AI-assisted review workflow, not as evidence that AI can replace professional quality reviewers.

Organisations evaluating document-review tools should follow a structured process.

1. Choose a narrow task. Start with documents governed by clear rules and review criteria.

2. Establish a human baseline. Record review time, actionable findings and errors before introducing AI.

3. Provide supporting documents. Give the system access to the approved instructions and reference material required for each review.

4. Check document coverage. Verify that every relevant page was processed and that omissions are clearly reported.

5. Require source references. Each finding should point reviewers to the evidence supporting it.

6. Measure individual runs. Evaluate consistency rather than combining findings from repeated attempts.

7. Track false alerts. Distinguish actionable issues from comments that require no correction.

8. Protect official records. Keep AI-generated comments separate from authorised changes and approvals.

9. Retain human decisions. Assign qualified staff responsibility for accepting findings and approving outcomes.

10. Test unfamiliar documents. Evaluate performance on new records, formats and handwriting styles before wider deployment.

For developers, the strongest technical lesson is the value of combining document retrieval, source references, explicit coverage checks and human adjudication.

For business teams, the main question is whether the system reduces review effort while preserving accuracy and accountability.

Conclusion

Enveda's Claude-powered CDP produced promising results in a limited pharmaceutical document-review evaluation.

Across repeated runs, it identified all 11 actionable issues found by human reviewers. A timed run also identified seven additional issues, while the measured review workflow required 89% less time than unaided review.

However, the study involved four records from one campaign, the system made a documented reading error, and human specialists remained responsible for every consequential decision.

We recommend viewing these results as evidence that AI-assisted document review deserves further testing. Organisations should verify detection consistency, review quality, data protection and total workflow performance before relying on similar systems in production.