How do you go from AI pilot to AI in production?

Why do so many AI pilots stop before production?

I meet more and more companies that have already done the work. Someone internally has got something of their own running with AI support. It solves a real problem. It works when they show it.

And still it is not used day to day.

They got 80 percent of the way on their own. Then it stood still for six months.

Because the last 20 percent never looks like the first 80. The first 80 is the fun part: the idea, the test, the feeling that this could really become something. The last 20 is something else entirely:

• The connection to the system where the information actually lives.
• Who owns it when the champion is on holiday.
• What happens the day the answer is wrong in front of a customer.
• How colleagues fit it into their week so it feels like relief, not an extra task.

That is where most AI initiatives stop. Rarely on the technology. Almost always on ownership.

What separates an AI pilot from an AI service in production?

A pilot shows that something can work. A service has to work every day, for everyone, even when nobody is watching.

Which three questions decide whether a pilot can go into production?

When someone shows me something they got running themselves, I ask three questions:

• Who answers when it breaks on a Tuesday morning?
• Which decision in the business is this meant to affect?
• What happens to it if the person behind it changes jobs?

If we can answer those three, the last mile is usually a few weeks of work. Not because it is hard, but because it takes a different job than the one that got you here.

Checklist: what needs to be in place before go-live?

Process and ownership

• A named owner in the business, not a steering group.
• A clear decision or task the service is meant to affect.

Integration and data

• The service reads from the systems where the information actually lives, not from a copy.
• Permissions are set: the service sees only what it needs.

Quality and error handling

• You know what an error costs, and where a human reviews before anything reaches a customer.
• There is a routine for what happens when an answer is wrong.

Regulation

• The service is classified under the EU AI Act. Most AI support in sales, quotes and customer service is not high-risk, but the transparency requirements have applied since 2 August 2026 (Article 50), for example that someone chatting with an AI service must be told so.
• Data flows are mapped under GDPR. Where processing may involve high risk, a data protection impact assessment is carried out.

Operations

• Monitoring: you can see whether the service responds, how often and how well.
• Cost control: you know what it costs per month and per case.
• Someone is responsible for updating the service when models and regulations change.

Adoption

• The people helped shape the way of working.
• Training in how the AI support is used day to day and where its limits are.
• Change management until the new way of working sticks. We work with Prosci and the ADKAR model.

Measurement

• A baseline measured before start, so the impact can be shown in numbers.

How do you design a pilot that makes it to production?

By treating the pilot as the first step of production, not as an experiment on the side. That is the core of our method Prove & expand:

• Real users. The pilot is run by the people who will use it, in their normal work.
• Real cases. Data comes from the systems the service will connect to in operation.
• An owner from day one. Whoever owns the process owns the pilot.
• Regulation as a design requirement. Classification and data flows are done in the pilot, not afterwards.
• Measure from the start. The baseline is measured before anything changes.

Then the decision after the pilot is simple. You have numbers, an owner and a service that already works day to day. The next step is to expand, process by process. Read more in the guide What is AI transformation?

Who should own the last mile?

Whoever owns the process. Not IT, and not whoever built the pilot.

It takes a little longer the first time. But when the owner has been involved all the way, the know-how stays in the company, and the next process goes considerably faster. See how we work in our service ladder and solutions we have built.

Sources: Regulation (EU) 2024/1689 (AI Act), EUR-Lex · Prosci, the ADKAR model