Unpacking Novo Nordisk’s AI Magic: A Closer Look (and a Grain of Salt)

Harry Glorikian | March 27, 2025

My recent LinkedIn post on Novo Nordisk’s use of generative AI clearly struck a chord. With so much interest, I felt compelled to dig a little deeper to see what I could uncover about their approach. Keep in mind, this is just my informed sleuthing based on publicly available insights – so please take this breakdown with a grain of salt.

Novo Nordisk didn’t simply adopt AI – they built a thoughtfully integrated solution called NovoScribe. Here’s my understanding of how it might work and some ideas on alternative tools that could similarly do the job.

How NovoScribe (Probably) Works:

1. Picking the Right AI Model: NovoScribe leverages Anthropic Claude 3.5 Sonnet, deployed through Amazon Web Services (AWS). Earlier AI models like OpenAI GPT-4 or Meta’s LLaMa initially fell short of pharma’s demanding accuracy standards, though recent advancements mean these models could now perform much better in similar tasks.

2. Retrieval-Augmented Generation (RAG): At its core, NovoScribe seems to use a RAG approach powered by MongoDB’s Atlas Vector Search. RAG helps the AI avoid generating incorrect information by pulling from a database of pre-approved clinical definitions, language, and real trial data. Today, other robust tools like Activeloop’s Deep Lake could also provide strong alternative capabilities for RAG implementations.

3. Integration with LangChain: LangChain likely plays a critical role by connecting Claude with the database and the internal human review processes. LangChain simplifies experimentation and lets companies easily swap in new AI models, including popular ones like OpenAI’s GPT, offering flexibility as new and improved AI tools emerge.

4. Human Oversight Remains Crucial: Novo Nordisk reportedly reduced their team from 50+ to just 3 experts reviewing AI-generated drafts in real-time, ensuring compliance and accuracy. The system likely shows reviewers exactly what data or existing text informed each AI-generated passage, continually improving future outputs based on their feedback.

5. Built-In Security and Compliance: Given pharma’s intense regulatory scrutiny, Novo Nordisk would have prioritized secure cloud infrastructure and stringent data privacy measures from the start. Any enterprise deploying AI for sensitive documentation would need similarly rigorous security measures.

From Weeks to Minutes – The Results:

Novo Nordisk has achieved great results:

Drafting clinical reports: reduced from about 15 weeks to under 10 minutes.

Personnel required: from over 50 to just 3 reviewers.

Annual cost: reportedly less than the salary of a single medical writer.

Crucially, they didn’t reduce headcount; instead, they strategically redeployed their skilled professionals towards innovation and complex problem-solving tasks. This aligns perfectly with the “Driving Intelligence to Zero” concept I’ve discussed – drastically cutting routine cognitive costs to unleash higher-value human creativity and innovation.

Strategic Lessons Beyond Pharma:

Novo Nordisk’s approach is broadly applicable. Industries like finance, insurance, and legal, dealing heavily with regulated documentation, can adopt a similar strategy, using AI, RAG techniques, and careful human oversight. And with constantly improving AI tools, it’s becoming easier than ever to customize these strategies.

Again, this is just my interpretation of their impressive work based on available information. Alternative tools could easily substitute some of those mentioned here – this space is evolving rapidly (Daily to be honest).

I hope this breakdown helps spark some actionable ideas.

Here’s to navigating – and leading – in this exciting new AI-driven reality.

One final word – don’t wait – jump in – play with these systems – push the envelope.