Novo Nordisk built NovoScribe, a documentation platform using Claude models on Amazon Bedrock and MongoDB Atlas, to draft clinical study reports, device verification protocols, and patient materials for human review. Anthropic's case study reports documentation time cut from 10+ weeks to 10 minutes, a 95% reduction in resources for device verification protocols, and review cycles cut in half.
Reported outcomes
Outcomes are reported exactly as stated by the source; we have not independently audited them.
How it works
- RAG combines domain-expert-approved text with case-specific variables.
- Claude drafts the clinical study report (up to 300 pages) or other document.
- Drafts go to human reviewers for approval.
- Expanded from CSRs to device verification protocols and patient booklets.
Risks & compliance
- Regulatory submissions require human sign-off
- Data handling in a GxP context
Healthcare cases cover operational and commercial workflows. Nothing here is medical advice or a claim about a device's or drug's efficacy.
Casefile: MedTech angle
Includes device verification protocols (Novo makes injection pens), relevant to MedTech regulatory and quality teams.
Editor's notes
- Source page is undated; accessed 2026-10-17.
Tools mentioned & how to build it
Disclosure: Some links may earn us a commission; it never affects grades. Tools are listed by fit to this case, never by payment.
Tools mentioned
- Claude & Claude Code (Anthropic)
Anthropic's models, Claude Code and enterprise plans. - Amazon Bedrock (AWS)
AWS's managed model and agent service.
Build a version yourself
Sources
“Claude has helped us cut writing times on CSRs by 90% so we can get documentation directly into human hands for review and approval,”Novo Nordisk accelerates clinical documentation and drug development with Claude
Grade rationale: Named organization with quantified outcomes, published by the model vendor (Anthropic). Automated check on Oct 17, 2026: 1/1 quotes and 2/2 outcome claims matched the source. Compiled by the Agent Casefile research desk (AI-assisted; see methodology). Spot an error? Request a correction.



