Context Is Your Only Moat

Most AI projects fail before the first prompt is written. Not because the technology doesn't work. Because nobody asked what the AI actually needs to know.

The 95% problem

Pål Erik Waagbø opens with a number that should stop anyone in their tracks: 95% of companies that implemented AI in a recent study saw no measurable difference on their P&L.

Not a small difference. Not a disappointing ROI. Nothing.

His explanation is blunt: companies are buying capability without building the structure to use it. They're bolting AI on top of unstructured, semi-structured, or just plain messy data and expecting the machine to sort it out.

"AI is an amplifier," he told me. "It amplifies whatever context you give it."

If you give it noise, you get more noise. At scale.

The money that's already there

Before we even get to AI, Pål describes a version of this problem that's been around for decades.

A restaurant chain. A steady customer. Seven months since his last visit, and his last satisfaction score was low. The CDP can tell you he stopped coming. But if you connect the operations data, you might find out that two chefs were off sick the last time he visited. That the food was slow. That the experience was bad for reasons that had nothing to do with him.

That's money hiding in the system. Not missing data. Data that hasn't been connected.

"You will already have this data. What you haven't done is connect it."

Context engineering, as Pål defines it, is the work of connecting those dots. Finding the commercial pain, identifying what data the AI actually needs to surface it, and stripping out everything else.

Meet Kevin

The most memorable idea from our conversation came early.

Imagine you've paid for a spot on stage at the industry's biggest conference. The room is full of buyers. Then, minutes before you go on, the organizer changes the format. You're not presenting. An audience member will present on your behalf.

You've got Kevin.

Kevin is Pål's metaphor for AI search. Perplexity, ChatGPT, Gemini. They are presenting your brand to potential buyers right now. And most companies, he says, haven't written Kevin a brief.

The brief is your website. It's your schema markup. It's the clarity of how you explain what you do, for whom, and why it matters. And right now, small companies that have done this work well are showing up as the authority on topics they have no right to own by size.

"That should terrify you. And make you excited."

The layer nobody owns

One of the more useful distinctions Pål draws is between what a marketing team does with data and what IT does with data. He's not dismissive of IT. But his point is that the commercial instinct, the hunting for revenue, the pattern-matching across customer journeys, that belongs on the marketing side.

He describes himself as existing below the departments. Seeing all the wires. Connecting them.

"I like to lean into owning the system," he said. "Not just checking that it works."

That's the context engineering gap. Most companies have one team that owns the pipes and another team that knows what should flow through them. The gap between those two teams is where AI investment goes to die.

The Jevons paradox and the closing window

Pål introduces two frameworks that I haven't been able to stop thinking about since we spoke.

The first is the arbitrage window. Right now, understanding what's happening in AI before the market fully wakes up gives you a measurable advantage. That window is closing. Twelve months ago, Pål estimated 12 to 18 months. Six months ago, he said 3 to 6. The window compresses every time he's asked.

The second is the Jevons paradox. In the 1860s, coal became cheaper. So everyone used more of it. The total bill climbed. The same thing is happening with AI. As the cost per token falls, usage expands, and the scarce resource shifts. What becomes scarce isn't access to AI. It's judgment about what to build with it.

"The scarcity always moves somewhere."

Two roles survive this, in Pål's framing. The Context Engineer, who builds the infrastructure layer. And the Tastemaker, the strategic leader with the judgment to decide what gets built and why.

Most people reading this are probably Tastemakers. And if you don't start paying attention to the infrastructure layer, Pål's prediction for the next two years is essentially the great flattening. Being okay at cognitive work, producing reasonable output, becomes harder to distinguish from everyone else who has AI help doing the same.

What's unpromptable

The question Pål leaves with is a sharp one.

He's tried, multiple times, to get Claude to write like him. He says it has no chance. His writing style, his analytical instincts, the way he pulls patterns across industries, that is unpromptable.

And he thinks that's the question every professional needs to sit with: what do you have that can't just be typed into a box?

Not as a thought experiment. As career strategy.

The bottom line

The reason most AI investment fails isn't the AI. It's the context. The data isn't connected, the brief hasn't been written, and the machine is being asked to do work without knowing what value it's supposed to deliver.

Pål's framing for the opportunity is one I think gets overlooked: the companies moving now, before the window closes, don't need to be big. They need to be precise.

The best line from our conversation: "Kevin will speak your truth." The question is whether you've told him what it is.

Find Pål on LinkedIn at linkedin.com/in/palerikwaagbo and his writing at palerikwaagbo.com. If this conversation made you think differently about what your AI needs to know, I'd love to connect on LinkedIn at linkedin.com/in/oddmorten.