95% of AI projects deliver nothing. Here is the pattern.

In 2025 MIT reviewed more than 300 corporate AI initiatives. 95% had no measurable impact on the bottom line. RAND has found that over 80% of AI projects never reach production.

Those are the numbers you are up against. We take them seriously. They are not an argument against starting. They are a list of what to avoid.

The cause is rarely the model

The projects do not fail because the model reads badly. They fail on method. Take an office that sees the same documents every week: leases, invoices, deeds. The same fields must be found and typed in. Here are the five failures we see again and again.

1. Nobody measured first. Nobody wrote down how many hours the invoices took each week, or who sat with them. Without that number nobody can show that anything improved. A project that cannot prove its value is not renewed.
2. Too much was started at once. Several workflows at once. It looked ambitious. The result is things that almost work, which nobody dares use on a real lease.
3. It was built beside the work. A new login, a new system, an extra step. People are busy, so they leave it. What survives sits in the inbox and in the system that is already open.
4. The data processing agreement came last. Something was built. Then the question of where the data sat. The project stalled. That is not the lawyers' fault. The order was wrong.
5. Nobody decided what a human approves. Either the system sent things out on its own, and one day something went wrong. Or everything had to pass several people, and doing it by hand was faster.

What the 5% do

The firms that get something out of this differ in sector and size. They share a method. It is the method we build by.

• We measure first. Hours and kroner in writing, before anything is built.
• One workflow at a time. The most repeated one, not the most interesting one.
• We build where the work is. In the inbox and on top of the systems you already use.
• A human approves. Every field from the deed lands in a queue with a source on every line: page and paragraph. One person says yes.
• The data processing agreement first. It is signed before the first document is sent.
• The system flags what it cannot decide. It does not guess.

A mistake must be cheap to catch

A language model makes mistakes. Not often, if the task is narrow enough. But it does. If it must never be wrong, you will never start. Nobody can meet that requirement.

The right requirement is a different one. A mistake must be cheap to catch. If the draft lands in a queue where a human sees it before it goes out, the mistake is free. If it goes straight to the tenant, it is expensive. The difference is not the model. It is the flow around it.

A system that is right most of the time and says so when it is unsure is worth more than one that is right a little more often and never says anything.

So we always build two things: the automation, and the queue where a human says yes. And we ask it to flag what it is unsure about instead of guessing.

Where Denmark stands

Danish firms are already moving. According to Danmarks Statistik, 42% of Danish companies used AI in 2025, up from 15% in 2023. And in a Wolters Kluwer survey, 65% of Danish SMEs using AI say the savings exceeded their expectations.

So it works. Once the dull parts are done: the measurement, the agreement, the queue.

Documents are the clearest example. But the same method covers the other work that repeats: calls you do not get to, the inbox, quotes, typing between systems, follow-up, invoicing and reminders, reports. Tell us what repeats. Then we find what saves the most.

The next step

The 95% never measured their baseline. It is the first thing we do. You tell us what repeats. We map it for free and write it down within a week: hours, kroner, where to start, and one fixed price — or an honest no. It takes at most 3 hours of your time.

Get my free AI audit