Most of the AI hype won't work for you

Most of the AI projects you read about won't work in your business. It's not because you're doing something wrong. It's because the premises don't match.

The press writes about AI as if every company has the same problems as Microsoft, Google, and eight other American tech firms. You don't. You have a finance team, a sales team, and five systems that don't talk to each other. The AI they write about wasn't designed for that. The AI they're selling you isn't either.

The pattern repeats.

Here's why most AI projects fail — and what actually works.

## Six reasons it doesn't work

1. The strategy that never reached operations.

The consultancy delivers a 30-page report. It sits on the chairman's desk. Nothing changes Monday morning. Three months later it's still the only deliverable.

A strategy that never reaches operations isn't strategy. It's theatre.

2. Pilot purgatory.

You set up a proof-of-concept with one team. It works. There's a celebration. Then it stops. Three years later it's still a pilot. Nobody knows who owns it. Nobody dares say it shouldn't be scaled.

PoCs without an ownership question built in from day one don't become products. They become museum exhibits.

3. ChatGPT tabs as AI strategy.

Twenty-five employees. Each with their own ChatGPT account in the browser. No coordination. No GDPR review. Payroll data pasted from last month. Quotes copied out of Salesforce. Nobody knows what's gone out.

That isn't an AI strategy. It's a data leak that selected itself.

4. The AI champion without a mandate.

You found an enthusiastic employee. She built a flow. It works. She quits. Nobody knows how to maintain it. Nobody dares change it. Six months later you call me.

Champion-only deployments aren't implementations. They're hobby projects with an expiry date.

5. AI on data you don't trust.

The company runs Copilot against SharePoint. Half the sales data sits in Visma. Quote templates live in a shared OneDrive folder nobody has cleaned since 2019. The result: Copilot gives confident answers based on half the picture.

AI on bad data produces better lies, not better answers.

6. The demo that doesn't survive production.

The vendor shows you an impressive demo. It works perfectly. You buy. Three months later you see it works in 70% of cases — enough to sell, not enough to replace a person.

Demos are trailers. Production is the film. They are not the same product.

## What actually works

Forget strategy reports. Forget pilot purgatory. Forget the one AI consultant who's going to "transform the business." Five things actually pay off.

Pick one concrete annoyance. Not a strategy. An annoyance. The one task that takes an hour every Monday. The email template everyone spends too long on. The report nobody reads but everyone produces.

Use what you already pay for. Microsoft 365 Copilot. Bilag scanning in Visma. AI features in Tripletex. You probably already have the AI layer you need. It just isn't switched on.

Build on the data source you trust most. Not all your data. The one that's cleanest. Could be the CRM. Could be the accounting system. Start there. Let the rest wait.

Only deploy AI where you can spot errors fast. AI is good at tasks where you can judge the answer in five seconds. AI is bad at tasks where you have to trust the output blindly. If you can't quality-check it, AI shouldn't be doing the job alone.

One person, one task, one week. Not a whole team. Not six months. Pick one employee with one concrete problem. Let them spend a week solving it with AI. If it works, do the next one. If not, learn and try again.

## The hardest thing to accept

The hardest thing about AI in 2026 isn't learning new tools. It's ignoring 90% of what's being sold to you.

The press isn't going to write about the company that switched on Copilot Monday and saved two hours a week. They write about the next big AI fireworks display. You have to choose for yourself what to believe.

Start with the boring. Let the fireworks burn out.