7x throughput,
same team size.
How Novo Nordisk Foundation changed the way its data team operates, and why the same team now ships seven times as much with AI agents writing most of the code.
Each engineer went
from 5 tasks a week
to 37.
The figures are July 2025 against a year later, with the same data engineering team, measured in their own work tracker. The gain comes from a changed division of labor: the engineers specify and review the work, and AI agents write most of the code.
The multiplier came from changing how the team operates. The AI runs inside that operating model.
Novo Nordisk Fonden runs one of the largest grant portfolios in Europe, and behind every grant sits data about applications, decisions, payouts and dates. The knowledge existed, but it was held in places and formats that neither people nor systems could reliably reach.
Two platforms ran in parallel, and the standards that governed the work sat in a wiki, separated from the work itself. The team’s AI agents worked under the same constraint: the one skill built for them existed only to fetch standards from elsewhere, and it could see roughly a quarter of them. As a result, anyone who needed an answer produced their own, which meant the same question returned different answers depending on who asked, and trust in all of them fell.
Handing code production to agents only pays off when the platform, the standards and the definitions are in a form agents can build against.
The platform is now tested, version-controlled and deployed the same way every time, with twenty automated checks gating every change where previously there were none.
It stopped being something anyone could change by hand, which is the property that lets AI agents build on it safely.
The standards moved out of the wiki and into the repository, so the rules now live next to the code they govern and engineers and agents read the same ones.
When the code changes, the documentation changes in the same commit. This is the step that made the agents genuinely useful.
The Foundation now holds 74 architecture decision records, and every key definition has a named owner, a date and a review cadence. A further 120 data contracts make the guarantees machine-checkable.
The definitions function as commitments rather than opinions, for people and agents alike.
Answers now sit in one governed layer with 21 metric views, which anyone can query in plain language and which also feeds five dashboards and a read-only path into Excel.
Once a trusted answer is easier to get than a homemade one, the reason to build your own disappears.
The Foundation's own engineers went on to build DEX-Driven, a spec-driven framework that connects their scrum ceremonies to an agent workflow, and they ship it internally through their own marketplace.
It has seen 38 releases since March, none of them written by us, which is the clearest evidence that the capability outgrew the handover.
The numbers.
The workday looks different as a result. The engineers spend more of their time in conversation, where decisions turn into working software while the discussion is still running. Weak spots that previously stayed hidden for months now surface within minutes, because everything worth knowing is written down where both people and agents read it.
“Our engineers don’t code much anymore. Claude does that for them. They talk a lot more together, which has been a very positive change. What we’ve done here with Heyra has really changed our way of working.”
Building the platform as code means agents can construct it and tests can verify it.
Moving the standards next to the work means people and agents follow the same rules.
Placing the answers in one governed layer means the whole Foundation can use what the data engineering team builds.
The AI sits inside that operating model rather than on top of the old one, which is why the gains have held.
Change how
you operate.
Tell us where you’re stuck. If operational AI can change the outcome we’ll show you how, and if it can’t, we’ll say so.
Talk to us