Using advanced AI to tidy email is like putting a McKinsey consultant in the mailroom

The AI market has grown up, and the bills from the tech giants are starting to land. TellusR CEO Morten Krogh-Moe argues we should stop talking about AI bubbles and start talking about smarter architecture.

Two years ago, artificial intelligence was defined by boundless enthusiasm, wild optimism — and subsidised prices from the American tech giants. Now, as summer is upon us, AI has truly arrived in Norwegian organisations.

When The Economist recently debated whether AI investments actually justify their valuations — or in plain terms, whether the AI wave is a bubble — Morten Krogh-Moe felt something essential was missing from the analysis.

“It’s a good article, but they discuss AI almost exclusively from a financial perspective,” says Krogh-Moe. “They miss that the question is largely also about technology architecture, and organisations’ ability to establish strategic flexibility in their AI architecture. When companies build that kind of architecture, the big AI vendors lose some of their power to define the terms.”

What happens when this architecture is missing?

“Two things happen: dependency increases, and prices skyrocket. The trend of ‘tokenmaxxing’ — burning as much compute as possible without a coherent strategy — is hitting hard now that agentic solutions are being scaled and costs are running away.”

The TellusR CEO is preparing for summer holiday, but AI never takes time off. And it certainly doesn’t have a resting heart rate. Especially not now, as the AI economy reveals its first reality check. The phase Krogh-Moe calls tokenmaxxing — where the goal seemingly was to burn as much compute as possible without thinking about the bill — has reached a pain threshold.

Just ask Microsoft. It wasn’t long ago they had to tell their developers to stop using Claude Code. The reason? Costs going through the roof. And they’re not alone. American Uber burned through its entire 2026 AI budget in the first four months of the year. Here at home, Kode24 reports that major players like NAV now pay three to four times more for their hundreds of Copilot users.

“It’s an incredibly unfortunate situation for a CFO to get that bill in their lap,” Morten observes dryly.

McKinsey in the mailroom

At the bottom of the billing shock, according to Krogh-Moe, lies an immature approach to the technology: an inherent belief that we must always use the very largest American language models for absolutely every task. Because why not use the best?

Well, according to Morten Krogh-Moe, it’s far from certain that using the most powerful language models available is actually best.

Why is it expensive?

“We need to look at what the big tech companies are trying to achieve. OpenAI, Microsoft and Anthropic are all trying to build platforms where customers get locked into a single ecosystem. Much like Facebook did with social media a decade ago. Platform theory says that when switching costs get high enough, the vendor gains almost all power to dictate the terms and prices of the relationship.”

Already, there are municipal departments quietly nervous about how locked in they are to technology vendors they’ve “always” used — who now suddenly also deliver AI. When technology is woven more or less uncritically throughout an organisation, two things happen at once: dependency increases. And prices skyrocket.

This trend will intensify enormously as the market moves increasingly towards autonomous AI agents. Swarms of AI instances working together to solve intricate tasks for Norwegian organisations sounds good at first glance. But that’s only at first glance.

“In complex, agentic workflows, a single call can pull in schemas for over 40 different tools, even when only two are actually used. You might burn 80,000 tokens per task completely unnecessarily,” says Krogh-Moe.

Is this where the bill grows?

“Yes, it is. And the majority of use cases in Norwegian business actually don’t require a massive frontier model to solve the tasks at hand. It’s very rare that Norwegian organisations need a model fundamentally built to solve advanced maths puzzles!”

He points to a fresh example to illustrate the point.

“We recently tested this for NVE,” says Krogh-Moe. “We set up a smaller, open Mistral model on a local server and benchmarked it against Microsoft’s enormous Azure OpenAI model in the cloud. The result is very illustrative. The giant model is best on the heaviest tasks, but the small basement model handles the daily housekeeping just as well — at a fraction of the price. This points to a direction for AI architecture going forward.”

Norwegian organisations are building on quicksand

Rising costs are one thing. Something else — and far more serious — is the ripple effects of technological dependency. Lock-in to foreign vendors means an organisation becomes extremely vulnerable to geopolitical swings.

Hardly had the US asked Anthropic to pull the plug on its most powerful model Fable 5 (twin to Mythos 5), before news arrived in late June that the Trump administration sent a similar message to OpenAI ahead of the launch of GPT-5.6.

Through Morten Krogh-Moe’s eyes, this is far more than a financial question. It’s about our digital sovereignty. It’s about great-power politics.

“The current US administration is not shy about dictating terms to European democracies. Recently, the director of the new KI Norge, Hans Christian Holte, expressed deep concern that American giants like Anthropic deliberately limit European access to their newest models. If our entire digital foundation is built exclusively on software from American tech giants, we’re building on quicksand. It can be pulled out from under us at any moment.”

It doesn’t exactly look like we have any choice but to think differently?

“No, we have no alternative if we want to preserve control. And fortunately, politicians are starting to wake up. The European Commission is now looking at stricter legal requirements for which cloud providers authorities can use — which could also push Norway away from American cloud giants.”

The tech giants’ models aren’t built for us

Another aspect worth noting comes from the French AI company Mistral. Behind the challenger Le Chat, they point out quite simply: general AI isn’t built for you. It isn’t built for your company.

“And that’s absolutely right,” says Morten Krogh-Moe. “American models are built on American culture and American business principles. If you build your business intelligence blindly on such a model, you import American culture straight into your organisation. I’m not saying you can’t use these models, but we must be conscious of this aspect.”

He points out that local, good alternatives exist if we dare to use them.

“I recently had a meeting with the National Library. They’re training their own Norwegian models. For many of the daily tasks in Norwegian organisations, such a Norwegian model will be just as good — and much safer for the public sector.”

Krogh-Moe stresses that Europe is far from hopeless, but it requires action, awareness and an independent architecture. The solution he and the rest of TellusR have worked on for many years is called AI Communications Hub.

“If we insert such an independent layer between the model vendors and the line-of-business systems, we give organisations the power to swap out the AI engine under the hood exactly when they want. Then the giants can’t build the dominant position. TellusR gives its customers control and strategic flexibility.”