Essay
Past pilot purgatory: why Nordic enterprise AI stalls, and how to break through
The pattern that traps most Nordic enterprises in their first wave of AI pilots — and the operating-model changes that distinguish the companies that escape it.
By 2026, almost every Nordic enterprise of meaningful size has run AI pilots. Most have run dozens. A smaller number have shipped something to a real customer. A much smaller number have built a second AI product that builds on the first. The gap between pilot and product is what analysts call pilot purgatory, and it is the central problem of enterprise AI right now.
This essay describes why companies get stuck there, and what the few that have escaped have in common. I write from a specific vantage: nearly a decade at the LEGO Group leading enterprise technology behind the mobile, web, and foundational platforms that powered LEGO’s consumer digital experiences — shipping AI-driven content moderation and AI-supported experiences touching tens of millions of consumers — and from twenty years of international digital work across the US, Denmark, and China.
The trap
Pilot purgatory is not a problem of imagination. It is a problem of operating model. Many Nordic enterprises run their first wave of AI in a way that almost guarantees the work cannot scale.
The first pilot is owned by an enthusiastic VP. It runs on a budget held outside the normal product portfolio. It uses tools the platform team has not approved. It works because the team is small and the use case is narrow. It gets shown at the all-hands. Everyone agrees AI is real.
Then the work stalls. The team that built it cannot put it into production because security has not reviewed the vendor, the platform does not support the inference path, the data is in a shape the production data systems cannot use, and the function that would own it day-to-day was not in the room when it was designed. The pilot becomes permanent — neither shipped nor killed, demoed quarterly to visiting executives.
This is not a failure of any individual team. It is the predictable output of treating AI as a side initiative inside a company whose operating model was built for a different kind of work.
What changes in companies that break through
A small number of Nordic enterprises have shipped second and third generations of AI products. They share four traits.
A platform that treats AI as a first-class workload
In the companies that ship, AI is not a guest in the platform. It is a first-class workload alongside web services and data pipelines. The platform team owns model serving, evaluation harnesses, prompt registries, and observability for non-deterministic systems. They do not build foundation models. They build the layer above. This is the shape of the platform work I led inside the LEGO Group across identity, consent, AI-driven content moderation, engagement, and analytics — and it is the work I think every serious Nordic enterprise needs in some form.
A real evaluation function
Companies that ship have someone whose job is to know whether the AI is getting better. Not a vendor’s benchmark. Not a vibes-based product review. A real evaluation function with golden dataset, eval sets, regression tracking, and a process for failure analysis. Evaluation is infrastructure. Without it, every release is a guess.
A governance posture that ships product, not paperwork
The companies stuck in pilot purgatory have one of two governance failures. Either they have no governance — and Legal blocks every launch at the last minute — or they have governance theater, where a committee meets monthly and rubber-stamps work it does not understand. The companies that ship have governance built into the development process: risk classification at design time, human oversight specified in the spec, eval thresholds that gate release. The AI Act is, when used well, a forcing function for this.
Executive sponsorship that survives the first failure
The hardest part. The first real AI product will partially fail. A model will hallucinate in a way that reaches a customer. A vendor will miss an SLA. Some KPI will go the wrong direction. The companies that break through have executive sponsorship that survives this moment — that treats it as a normal feature of shipping AI and not as a reason to retreat to demos.
The role of the advisor
The reason I have ended up advising on this work is that the four traits above are recognizable from the outside, and most of them are fixable. The diagnostic is fast. The roadmap is doable. The execution is the hard part, and that is the company’s work to do.
What I think a good advisor brings is the willingness to say, plainly, which of the four traits a company is missing. Most internal teams know. They have not had the air cover to say it out loud. An outsider can.
If you are stuck in pilot purgatory and recognize the picture, the advisory page explains how I work. The fastest path is a short note on LinkedIn.