When Novo Nordisk faced a wave of employee ideas for generative AI, the Danish pharmaceutical company chose a route around the usual bottleneck: rather than picking which use cases deserved a build team, it made building cheap enough that 25,000 employees could try their own — at about $10 a month per chatbot.
Novo Nordisk A/S, founded in Denmark in 1923, ran into a problem that most large organizations recognize: talented people had ideas for AI, and the capacity to build and maintain applications sat far from them. Predicting which use cases would pay off was hard. The original idea reframed the whole question.
The original idea
The original idea: make experimentation so cheap that prediction becomes unnecessary. Working with AWS partner Cloud2 Oy, Novo Nordisk built a self-service platform on Amazon Bedrock, DynamoDB and AWS Lambda that lets any employee create and customize a chatbot for a specific task. A proof-of-concept takes about an hour. Moving from proof-of-concept to a minimum viable product takes about a month. At roughly $10 per chatbot per month, the cost of a failed experiment stays trivial — so the organization stopped choosing winners in advance and let usage reveal them.
Verified results
- 25,000 employees now use the platform, with 2,500+ chatbots created.
- The largest single use case trained on 140,000 documents.
- The general-purpose chatbot serves 1,000+ employees and processes 26,000+ prompts each month.
- Tasks that ran a full day now finish in 30 minutes; the innovation cycle compressed from months to days.
From Novo Nordisk's AWS case study. Jens Jepsen, Senior Vice President Data Science Partner, framed the foundation plainly: “Having this foundation through Amazon Bedrock and our platform supports very rapid innovation.”
The friction worth naming
The instructive tension sits in what the platform declines to touch. It serves processes outside the regulated perimeter — a deliberate boundary for a pharmaceutical company under strict oversight. That scoping is the discipline behind the speed: by keeping the self-service model clear of regulated workflows, Novo Nordisk earned the freedom to let thousands of employees build, free of a per-idea compliance review. The boundary is the enabler.
What you can take from this
What you can take from this: the barrier to internal AI adoption is often the cost of being wrong, rather than the cost of being right. Lower the price of a failed experiment far enough, and an organization can replace a selection committee with a market. Novo Nordisk reached 2,500 chatbots because a wrong guess cost about $10 — a figure any mid-sized company can match. We saw a related pattern when JPMorgan put its LLM Suite in front of 200,000 employees as an opt-in. We covered that here.
The universal question: which AI decisions is your organization routing through a committee today, when the cost of letting people try has fallen to the price of a lunch?
Article by SAGA — Success Stories & Real Cases
SAGA curates the stories of real companies that built something with AI and grew into it — the verified before and after, told so the reader can picture doing it too.