The report The GenAI Divide, produced by MIT from 52 executive interviews, 153 surveys and an analysis of 300 public deployments, reaches a conclusion worth reading slowly: 95% of generative AI pilots had no measurable impact on profit and loss.

More than 80% of organisations have trialled tools of this kind and close to 40% say they have deployed them. Adoption is massive. Transformation is not.

The wrong question

When a pilot ends, the question almost always asked is «did it work?». It is a comfortable question because the answer is usually yes: the demo impresses, the team is pleased, the vendor gets a case study.

The question that matters is different: «who will stop doing what they were doing?». If nobody changes task, there has been no transformation. There has been one more layer on top of the same process, and that layer costs money.

Why they stall

The report identifies the main cause as what it calls the learning gap: generic tools work very well for an individual because they are flexible, and stall inside a company because they neither learn from the workflow nor adapt to it.

That matches what we see. A pilot runs under ideal conditions: a clean case, motivated people, none of the exceptions that take up 30% of the real work. Going to production is precisely the step into the exceptions, and that is where the project meets the actual organisation.

Three conditions before starting

We have learned not to authorise a trial unless these three come answered:

  • Which specific process is replaced or changed. With a name, an owner and a volume. Not «improve customer service».
  • What will stop being done if it works. If the answer is «nothing, it adds on top», the project has already failed and has not started yet.
  • Who will maintain it a year from now. Pilots are held up by enthusiasm. Processes are held up by someone with the job in writing.

The cost of the endless trial

There is a recurring pattern: organisations that have been in pilot phase for three years. They change tool, change vendor, repeat the demo with fresh data. Every iteration looks like progress and none reaches production.

The cost is not the budget spent. It is internal credibility. After two pilots with no consequences, when a project that could genuinely change something finally arrives, the organisation no longer believes it.

What sets the 5% apart

According to the study, only 5% of integrated systems generated significant value. What they have in common is neither the model nor the budget: it is that they sit inside a process, not beside it.

Sitting inside means that if the system goes down, somebody notices within an hour because the work stops. It sounds like risk, and it is precisely the proof that it has been integrated.

A pilot that inconveniences nobody is not a pilot. It is an expensive demo.