Novo Nordisk's NovoScribe: how a RAG pipeline cut clinical study report production from 12 weeks to 10 minutes

Last Updated: October 12, 2026

The bottleneck: a 300-page document that took three months

Pharmaceutical development is, past a certain point, a documentation business. Anthropic's case study on the deployment frames the problem bluntly: each new treatment arrives with its own mountain of required paperwork, and producing the documentation for a clinical trial consumed so much time and resource — even with external agencies — that it slowed the organisation's ability to bring medicines to market.

The clinical study report sits at the summit of that mountain. It is the comprehensive account of a trial's methodology, execution, results and analyses, and it is what regulators, clinicians, researchers and legal teams read to judge whether a product is safe and effective. Louise Lind Skov, who leads Content Digitalisation at Novo Nordisk, described the pre-automation cycle to MongoDB.

A CSR usually takes around 12 weeks to compile, involving a multidisciplinary team of statisticians, scientists, and technical authors.

— Louise Lind Skov, Head of Content Digitalisation, Novo Nordisk, MongoDB case study

The mechanics were as laborious as they sound. Statisticians analysed field data into tables and figures. Technical authors extracted those outputs and merged them into regulatory templates. Then came extensive quality assurance to confirm that every number across a 100-plus-page document was consistent, complete and compliant. Anthropic's write-up adds a detail that quantifies the throughput ceiling better than any cycle-time figure: staff writers averaged only 2.3 CSRs per year, and the manual process stayed prone to errors and subjective interpretation.

What was actually built

NovoScribe started in mid-2023 as an experiment by Skov's team: could a CSR be compiled dynamically by prompting large language models with the trial's own statistical outputs plus vector embeddings of the approved report templates? Within a few weeks the experiments were producing drafts faster, more accurately and with fewer people than the manual route, and the project moved toward production.

The production stack is unglamorous in the best way. Foundation models — Claude 3 and Titan — run on Amazon Bedrock, alongside the company's own private instance of ChatGPT. LangChain sits in front as the orchestration layer so the team can change models without touching application code. MongoDB Atlas Vector Search holds the embeddings and the associated text snippets and metadata. Tobias Kröpelin, NovoScribe's tech lead and a statistical programming specialist at the company, explained why that optionality was designed in from the start.

Each foundation model has its own strengths and weaknesses, so we typically experiment with a variety of different embedding and generation models for each report we compile.

— Tobias Kröpelin, NovoScribe Tech Lead, Novo Nordisk, MongoDB case study

The generation logic is deliberately narrow. NovoScribe produces validated text against defined content rules and statistical output; Atlas Vector Search scores how closely each candidate snippet matches the relevant statistics; the LLM assembles the draft from that evidence. Critically, full lineage of all sources is presented, so authors verify provenance rather than re-reading from scratch. That is the design decision that turns an LLM from a liability into a filing-grade tool in a regulated workflow.

Why they left the relational model behind

One of the more useful engineering details in the public record is a failure. Kröpelin's team began on the relational databases they used day to day and found the data model feeding statistics and report text into LLMs was too complex and too rigid — dozens of narrow tables that looked nothing like the Python dictionaries the team worked in, schema migrations blocking every application change, and query-time joins degrading performance. Moving to a document model let a single driver call retrieve the snippet, its vector and its metadata together. Teams building regulated document automation routinely underestimate how much the retrieval data model, not the prompt, governs iteration speed.

The measured outcome

Waheed Jowiya, Digitalization Strategy Director at Novo Nordisk, gave Anthropic the cleanest statement of the writing-side effect.

Claude has helped us cut writing times on CSRs by 90% so we can get documentation directly into human hands for review and approval

— Waheed Jowiya, Digitalization Strategy Director, Novo Nordisk, Anthropic case study

The second-order effects are arguably larger than the headline. Review-to-approval cycles shortened as reviewers grew satisfied with the clinical interpretation and output quality. The programme spread beyond CSRs: device verification protocols now take 95% fewer resources, and a complete study booklet and patient guide — materials that previously required months of back-and-forth with external agencies — can be produced in under a minute. Skov's framing of the milestone, given to MongoDB, is about industry position as much as cycle time.

With NovoScribe, we are the first in the industry to generate a complete Clinical Study Report in minutes rather than weeks.

— Louise Lind Skov, Head of Content Digitalisation, Novo Nordisk, MongoDB case study

Claude Code as a prototyping layer, not just a coding tool

The part of this deployment most often skipped in summaries is organisational. Novo Nordisk used Claude Code to build and extend NovoScribe with a development team of eleven people — and, more interestingly, to let non-engineers participate in product development. Jowiya holds a PhD in molecular biology rather than a programming background; with Claude Code he generates a proof of concept in natural language instead of translating an idea into tickets for a development team.

It allows our scientific domain experts, people like myself and my director Louise, to have a much bigger impact on how things are developed

— Waheed Jowiya, Digitalization Strategy Director, Novo Nordisk, Anthropic case study

Anthropic reports that the ability to spin up a prototype in hours rather than weeks let the team hold its size while expanding scope, avoiding headcount growth and the overhead of traditional agile ceremonies. Jowiya's summary: Claude Code lets us do experiments easily, make decisions fast, and iterate quickly to get products out. For anyone evaluating which model to put where, note that the deployment uses two different model surfaces for two different jobs — generation inside the product, and agentic coding to build the product.

Timeline of the public record

That April 2026 announcement is worth reading alongside the NovoScribe story rather than as a replacement for it. CEO Mike Doustdar framed the rationale in the company's own release: Integrating AI in our everyday work gives us the ability to analyse datasets at a scale that was previously impossible, identify patterns we could not see, and test hypotheses faster than ever. Sam Altman's accompanying line was that the collaboration will help Novo Nordisk accelerate scientific discovery, run smarter global operations, and redefine the future of patient care. Novo Nordisk states the partnership is structured with strict data protection, governance and human oversight.

What the public record does not establish

Honest documentation of a vendor-published deployment means naming the gaps, and there are several here.

No quality metrics. Nowhere in the public record is there an error rate, a defect-per-document figure, a hallucination rate, or a comparison of AI-drafted versus human-drafted sections at review. Anthropic states that NovoScribe output has received positive feedback from regulators — but no agency, filing or decision is named, and no regulator has publicly commented. That claim should be read as a vendor characterisation, not a verified regulatory outcome.

The headline comparison is not apples to apples. Twelve weeks describes an end-to-end multidisciplinary cycle; ten minutes describes generation. Novo Nordisk itself says the LLMs take minutes and the rest of the time is spent in QA. The real end-to-end compression is therefore smaller than 12-weeks-to-10-minutes implies, and the company has not published it.

No financial disclosure. There is no published cost saving, licence cost, infrastructure cost or ROI calculation. The widely repeated figure that each day sooner to market can add around $15 million in revenue is a company estimate of the value of speed, not a measured saving attributable to this system.

Coverage is a forecast, not a confirmed result. The expectation of over 90% CSR coverage was stated prospectively. We found no public confirmation that the target was met.

Workforce effects are unaddressed. Anthropic notes that documentation once requiring entire departments now requires single users, and that the dev team avoided scaling headcount. Neither source discusses what happened to the technical authors and external agencies previously doing that work.

The two AI programmes are not publicly connected. The OpenAI press release does not mention NovoScribe, Claude or MongoDB. How the group-wide partnership interacts with an existing multi-model platform is not disclosed.

What an integration team should take from this

Pick the artefact, not the department. NovoScribe did not automate 'writing'. It automated a specific, high-volume, template-governed document with an existing QA function and a measurable baseline of 2.3 reports per writer per year. That baseline is what made the outcome legible.

Design for verification before fluency. The lineage requirement — every generated assertion traceable to an approved snippet and a specific statistic — is what let a 100-page regulatory document pass human review quickly. In most regulated document workflows, review time dominates authoring time, so provenance is the lever.

Abstract the model early. LangChain in front of Bedrock meant Novo Nordisk could trial different embedding and generation models per report type without application changes. In 2023 that looked like engineering hygiene. After April 2026 it looks like strategy.

Fix the data model before the prompt. The team's migration from relational tables to a document store is the least glamorous and most transferable lesson in the case study: iteration speed on RAG systems is usually gated by retrieval plumbing, not by model capability.

Let domain experts build. An eleven-person team expanded from CSRs to device protocols to patient materials partly because the molecular biologists could prototype directly. If you are scoping an equivalent programme, the shortest path is usually to put the agentic coding tool in the hands of the people who know what 'correct' looks like.