An AI project that does not reach daily use has often not failed. It has stalled.

There is a pilot, it was demonstrated once and it worked, and then everyday work took over. One reason that is easy to recognise: the solution required someone to change how they work without getting anything back in the first week, and nobody had it on their plate when the meeting ended.

Implementation is the least glamorous part of AI work, and the only part that decides whether anything comes of it. What is needed is seldom more technology. It is scoping, a named owner, and patience for the change to survive an ordinary working week.

This page describes how I work with that part, and what I have learned from running my own systems in live use — among them My AI team and Nyheter AI, which both have to work on a Monday and not just in a demonstration.

Five steps from decision to daily use

Scope it to something absurdly small

The first thing built should be smaller than feels meaningful. The reason is not caution — it is that a small first step is cheap to change when it turns out the assumption was wrong. The breadth comes later.

Name a person, not a role

Someone with a name should own the work process afterwards, and that person should be involved already while it is being built. A system handed over to "the business" is in practice handed over to nobody.

Build in the hands of the person who will use it

Short rounds with the actual user, not a requirements specification followed by a delivery. What looks right in a meeting can turn out to be wrong as soon as someone uses it in reality, and you want to know that in week one.

Make sure it wins something in the first week

If the change costs more than it gives at the start, it will not survive. Sometimes that means beginning with a step that is less important but more irritating, in order to bring people along.

Hand it over so that you can change it yourselves

Whoever uses something should be able to change it without calling me. An assignment is finished when you can manage without me — otherwise I have built myself into your business, and that is not a service I want to sell.

Four ways a pilot dies

It was built for a demonstration

A system that is to impress in a meeting is optimised for the neatest case. Reality consists of odd cases, and then it stops working in week two.

Nobody owned it afterwards

The quietest cause of death. Nobody decided to stop — it simply stopped happening, and then it was awkward to bring up.

The checking ate the time saving

So much review was added on top that it was quicker to do things the old way. Then it is right to stop, but it is also a sign that the wrong area of use was chosen.

It solved a problem nobody had

Technically successful, operationally uninteresting. This can be avoided at an early stage, and it is the whole point of doing the mapping before the build.

What the assignments look like

Some assignments are limited, with a clear scope and a fixed quote. Others run with a few hours a week, which can suit better when the aim is to get something truly to settle into everyday work. Builds are quoted according to need, since the scope varies a great deal between different areas of use.

Taking over a pilot that someone else has built works well. The first step is then to find out why it stalled — whether it was the technology, the way of working or the ownership. Sometimes the right answer is to wind it down and start again more narrowly, and if so I would rather say that directly.

If you have not yet decided what to implement, an AI strategy is the right starting point. If you want to know what the technology can actually do before you decide, that is covered on the page about generative AI for businesses.

Back to the AI consulting services

Questions about implementation

How do I take my AI project from pilot to reality?

Five things decide it, in this order: scope the first step to something smaller than feels meaningful, name a person who owns the work process afterwards, build in short rounds together with the person who will use it, make sure the change gives something back in the first week, and hand it over so that you can make changes yourselves without a consultant.

If no owner has been named for the pilot, that is the first thing to solve, before anything more is built. If the pilot was built by someone else, it can be taken over; the first step is then to find out why it stalled.

Why do so many AI pilots get stuck before they reach daily use?

One reason that is easy to recognise: the solution required someone to change how they work without getting anything back in the first week, and nobody had it on their plate when everyday work returned.

The technology is seldom the cause. What is missing is a named owner and a first step small enough.

How long does an AI implementation take?

For a limited area of use in a smaller business, the aim is for the first thing to be in live use within a month.

What tends to drag on is seldom the build, but agreeing who owns the work process afterwards and getting hold of the material that is needed.

Can you take over a pilot that someone else has built?

Yes. The first step is to find out why it stalled — whether it was the technology, the way of working or the ownership.

Sometimes the right answer is to wind it down and start again more narrowly.

What does it cost?

Builds are quoted according to need, since the scope varies a great deal. Some assignments instead run on an ongoing basis with a few hours a week.

You receive a fixed quote or a clear hourly arrangement before the work begins.

Do we need to replace our existing systems?

Normally not. What I build sits alongside what you already have and picks up what is missing. Replacing systems that work is expensive, risky and seldom where the time saving comes from.

Do you have something that has got stuck?

A free half-hour call. Describe what was built and where it stalled, and I will tell you what I think.

Book a first meeting

If what is put into operation should also remove recurring work, there is the page on AI automation, and what is actually worth automating is covered in the article on automating administration.