The largest LP in the world, Norway's $2 trillion oil fund, recently told the world how they use AI. A phone ringing at 3 am in Oslo. A five-person team with one hour to answer Goldman. An accountant in love with a footnote. The real story runs a decade longer than the one they tell. Plus what 727 real estate players told us in our AI survey, and what we're seeing in the classroom.
Sometime last year, Goldman Sachs called Oslo.
Ferrari’s largest shareholder wanted out: 30 billion kroner of stock, more than three weeks of normal trading volume, offered quietly to a handful of the world’s biggest investors.
Goldman needed three answers within an hour:
Are you in, how much, and at what price?
The five-person team fielding that call manages ~$200 billion in European equities. They get about 200 of these calls a year, and over time, their answers have added billions to the fund’s excess returns.
Who is selling?
Why are they selling?
Does the market see this coming?
Will the trade trigger index buying?
What’s a fair price?
All of it, in under an hour, with the data scattered across databases, filings, and the open web.
How five people answer that call is the story of the world’s largest fund’s AI transformation, and it doesn’t start where they say it starts.
“It all started about two years ago,” their AI lead told the audience at a seminar this year, “when Nikolai had Sam Altman from OpenAI and Dario Amodei from Anthropic on his podcast.”
The CEO walked away from those interviews with one number in his head, 20%, and handed down a mandate to make the fund that much more efficient.
That’s the origin story Norges Bank Investment Management tells. It’s a good one. It’s also off by about nine years, and the nine years are the part worth stealing.
NBIM manages Norway’s oil money: roughly $2.2 trillion at year-end 2025, with stakes in about 1.5% of every listed company on earth. This year they put on a public seminar showing exactly how they use AI. Ten use cases, presented by the people who built them. We watched the whole thing. Told in order, it's a different story than the one on stage.
2015: “You are going to look very stupid”
The story actually starts with a database nobody wanted to clean.
Beginning in 2015, NBIM made three moves that had nothing to do with AI.
They fired their back-office vendor and brought settlement, accounting, and valuation in-house, because “we wanted to own the knowledge.”
They moved everything to public cloud, because they’d hit what they called a data ceiling and “wanted a data horizon.”
They discovered the old databases they’d dragged into the cloud couldn’t scale, which meant rebuilding them from scratch.
Data cleaning?
In practice, AI is more about data cleaning than use cases.
Nobody volunteers for that. The CEO knows it, and his exchange with the executive who ran the migration is the most honest moment in the seminar:
CEO: “Is it fun to clean data?”
AI lead: “It’s no fun at all.”
CEO: “Does anybody thank you for cleaning your data?”
AI lead: “No.”
CEO: “How do you get people to clean data?”
AI lead: “You basically tell them that 31st of January we are going to turn the old data off... if you sit there the day after and have no data, you are going to look very stupid.”
Late nights, rewritten code across the whole firm, and out the other side: one clean warehouse for everything.
Notice what’s absent from this chapter. AI. They insourced for expertise, moved to cloud for scale, rebuilt the databases because the old ones broke. The presenter calls these moves “the foundation of our success within AI,” but that framing came later. When the AI moment arrived, they discovered they’d built the prerequisite stack years earlier, by accident, for other reasons.
This fits squarely with an observation from our recent study of AI in CRE. We surveyed 727 real estate professionals, skewed toward senior decision-makers at LPs and GPs, and found that data readiness was a common barrier. Those who have organized data are able to leverage AI much more extensively.
Early winners from our study: a few big platforms that saw this coming and have been organizing data for years, and small, nimble firms with less data and fewer compliance complications.
2024: “You have to be on them like a wasp”
So when the podcast lands, it lands on prepared ground, and the CEO starts pushing. His AI lead describes him as “a Duracell battery that never goes out of energy... he has pushed everyone.”
Claude for every employee.
Twenty volunteer ambassadors, one use case each with a two-month training program from Anthropic.
Firm-wide training: Seven unenjoyable but mandatory sessions.
“Do people like mandatory? They hate mandatory... Can it be voluntary? No. Because the people who don’t want to do it are the people who need it the most. It has to be mandatory and you have to be on them like a wasp.”
By 2025, in their words, whenever anything happened at NBIM, it had AI in it. Leader summits, tech days, everything.
2025: “We weren’t that inefficient before we started”
Thousands of experiments were now running bottom-up. But the mandate was 20%, so leadership went hunting for the big one. They interviewed the chiefs, the global heads, the CEO himself. They ran workshops. And then came the admission that makes the whole seminar worth your time:
“We found 171 new projects. We did not find the one great AI use case.”
Followed by the line that should probably be shared in every corporate strategy meeting:
“The good thing is that we weren’t that inefficient before we started this. The bad thing is that we had to do all of these smaller projects to actually become more efficient.”
The best-resourced AI adopter in institutional investing, with every advantage money and mandate can buy, found no silver bullet. They found 171 boring projects and did them anyway.
The search killed some sacred cows too. Scrum, the ceremony-heavy development ritual the whole corporate world imported from the 1990s, didn’t survive contact. The new unit is two developers and one business person, “autonomous but also empowered to take all the decisions they need.”
2026: “It’s never lazy”
Which brings us to the ten use cases, and to the people in them.
