Novo Nordisk
Novo Nordisk drafts clinical study reports with its NovoScribe AI platform
Novo Nordisk built NovoScribe, a generative AI platform that drafts clinical study reports and other regulatory documents from expert-approved text. Anthropic's case study, which says Claude models drive the system, reports documentation time fell from 10+ weeks to 10 minutes, with a 90% cut in report writing times.
Key figures
Copied exactly from the sources. Copy any figure with its citation.
The problem
Every new medicine needs a large set of regulatory documents. The biggest is the clinical study report, which summarizes a drug trial and can run to 300 pages. Anthropic's case study says producing one was a months-long cycle of writing, review and approval, and staff writers averaged 2.3 reports a year. Anthropic's case study adds that each day of delay can cost pharmaceutical companies up to $15 million in potential revenue.
What they built
NovoScribe is Novo Nordisk's generative AI platform for regulatory documents. Anthropic says Claude models drive it, running on Amazon Bedrock, with MongoDB Atlas as the data store. MongoDB lists Claude 3, Amazon Titan and a private ChatGPT instance. It combines the model with text that domain experts have approved and with variables specific to each study. The team began with clinical study reports. It later added device verification protocols and patient materials.
How they did it
The technique is retrieval-augmented generation: the system looks up approved content and gives it to the model, so drafts rest on vetted text. A development team of 11 people built it, using Claude Code, Anthropic's coding tool. Adapting the setup for patient booklets took little time. Device protocols needed several months to connect legacy systems holding unstructured data. MongoDB's page dates the project start to mid-2023 and says early experiments worked within a few weeks.
Results
Anthropic reports clinical documentation time down from 10+ weeks to 10 minutes, report writing times cut by 90% and device protocol resources down 95%. It says review cycles were cut in half and a study booklet with a patient guide can be produced in under a minute. MongoDB reports 12 weeks to 10 minutes and says NovoScribe compiles about 30% of Novo Nordisk's clinical study reports. Next steps include pilots in medical affairs and automating Common Technical Documents.
What to copy
Pick a document that is costly, repeated and tightly structured, and feed the model only expert-approved content.
All reported figures
| Measure | Figure | Citation |
|---|---|---|
| Time spent producing clinical study documentation | 10+ weeks to 10 minutes1 | |
| Reduction in clinical study report writing times, as stated by the company's digitalization strategy director | 90%1 | |
| Reduction in resources needed to create device verification protocols | 95%1 | |
| Time to create a clinical study report, as stated in MongoDB's case study | 12 weeks to 10 minutes2 | |
| Share of Novo Nordisk clinical study reports compiled by NovoScribe, per MongoDB's case study | around 30%2 | |
| Clinical study reports per year averaged by staff writers before NovoScribe | 2.31 |
Lessons for your project
Start with one document type. Novo Nordisk perfected clinical study reports first, then reused the setup for device protocols and patient booklets.
Ground the model in approved content. NovoScribe works from text approved by domain experts plus case-specific variables, not from open-ended prompting.
Budget for legacy data. Connecting old systems for device protocols took several months, far longer than adapting the setup for patient booklets.
Keep humans in review. The tool aims to get drafts to reviewers faster, and review and approval stay with people.
Caveats
- Both sources are vendor case studies (Anthropic and MongoDB). Neither is independent, and no outside source confirms the figures.
- Neither page is dated, and neither gives a go-live date, so the deployed date shown (June 2023) is only the mid-2023 project start stated by MongoDB, not a go-live date. The tool went live later, at an unstated date; MongoDB says it now compiles around 30% of reports.
- The two pages give different baselines: 10+ weeks (Anthropic) and 12 weeks (MongoDB). Both claim 10 minutes. MongoDB says the model needs just minutes to generate the report and the rest of the time goes to quality checks, so the headline compares machine drafting with the full manual process.
- The MongoDB page lists Claude 3, Amazon Titan and a private ChatGPT instance among the models used, while Anthropic says Claude models drive the system. The exact model mix and versions are not stated.
- Anthropic says regulators gave positive feedback but gives no detail and no regulator named.
Sources
- Novo Nordisk accelerates clinical documentation and drug development with Claude
Anthropic · Undated · Vendor case study · claude.com
- Novo Nordisk and MongoDB Atlas customer case study
MongoDB · Undated · Vendor case study · mongodb.com
Summarized in our own words from the sources above. Spotted an error? See corrections.