The pilot is not the problem. The missing owner for adoption is.

Failure does not look like failure

A pilot that fails does not crash. That would be easy, and it would be dealt with. What actually happens is quieter. The model keeps running in a parallel environment. The dashboard is still reachable, and the number of people who open it settles at two. The shift goes back to the list it trusts.

Nine months later nobody can say when it stopped mattering, because there was no moment when it did. I have seen a pilot celebrated in a steering committee and quietly dead by the following summer, with the same slide still in circulation. Technically nothing had gone wrong. Nothing had gone anywhere at all.

Everybody quotes the number, and reads it wrong

Gartner predicted that at least thirty percent of gen AI projects would be abandoned after proof of concept by the end of 2025, and it listed the reasons: poor data quality, inadequate risk controls, escalating cost, unclear business value. Read that list again. Only one of the four is technical, and underneath it is organizational too.

The number gets quoted as evidence that the technology disappoints. In the plants I have worked in, the technology was rarely the disappointment. The pilot did what it had been asked to do. What was never built was the receiving end.

A pilot has a project manager. Operations has a line.

The handover fails because it has no counterpart. IT hands over to the plant. The plant sees a foreign body that arrives with a training session and no budget. The project manager's job ends at go-live, and the line manager's job never included this in the first place.

So the model becomes everyone's improvement and nobody's obligation. It appears in no operating review, on no shift handover sheet, in nobody's target agreement. An organization does not run on intentions. It runs on what sits in someone's report.

The three questions

Before I commit to an adoption plan I ask three questions, and I have never needed a fourth.

Who is named? A person with a surname, not a function and not a committee. "The digital team" is not an owner.

In whose report does the number appear? If the KPI agreed in week one does not sit in the operating review of the person just named, the answer to the first question was decorative.

What happens in the first week after go-live when the model is wrong? Because it will be. The shift needs to know now who may switch it off, what the fallback is, who gets called and how long an answer may take. A pilot without an escalation path gets switched off once and never switched back on.

What an owner needs in order to own anything

Naming someone and handing them nothing is worse than naming nobody, because it turns a structural problem into a personal failure.

The unglamorous eighty percent needs its own budget line: interfaces, master data quality, work instructions, shift training, on-call cover, model care. None of it demonstrates well. All of it decides.

The owner needs the right to change the process, not only the tool. An AI that recommends, inside a process that does not allow the recommendation to be followed, produces frustration on schedule.

And there has to be a date on which scaling or stopping is decided. Without it the pilot enters the permanent state industry knows so well: not used, not stopped, still funded.

Whoever carries the number in their own report runs the adoption. Everyone else accompanies it.

The objection: we will appoint a change manager

A staff role without P&L responsibility owns nothing. It can support, communicate and train, and none of that survives contact with a production plan.

The best adoption owner I have worked with was a plant manager who was against the project for the first six weeks. He argued about the definition of the number, which was the most useful thing anyone did that quarter. Once the definition was his, the defence was his too. Resistance from the line is not the obstacle to adoption. It is the raw material.

Where the money should go

BCG describes the split as ten, twenty, seventy: roughly ten percent of the effort on algorithms, twenty on technology and data, seventy on people and processes. Every industrial AI budget I have reviewed was closer to the reverse.

That inversion is the entire story of the pilot that dies quietly. The cheapest part of an industrial AI programme is the model. The expensive part is the shift that has to work differently on a Monday morning, and it is the part that is funded last, out of whatever is left.

The pilot proves feasibility. Adoption proves value.

So the owner is named in week one, in the same conversation in which the number is named. The two belong together. A number without an owner is a report. An owner without a number is a volunteer.

Andreas Geiss

Founder, Axiva Industrial. Operator notes from the shop floor and the P&L.