I’ve watched something like a dozen to fifteen AI pilots up close at operating companies across the region in the last eighteen months. Rough number, not a dataset. These were real pilots with real budgets, most of them in the hundreds of thousands of dollars, running over three to nine months. I was involved in parts of most of them.
Maybe two got far enough for the model itself to become the thing being evaluated. The rest stalled earlier, upstream of any real model question.
About six stalled on data access. The pilot team couldn’t get at the information the model needed because the data lived in a system owned by another business unit, or across laptops and SaaS tools nobody could get data out of cleanly. The pilot was nominally running and actually waiting three months for a permission that never arrived. One ended when the sponsor rotated out and the replacement didn’t prioritize the access request.
Another three or four stalled because nobody outside the pilot team cared enough to change a workflow or use the output. The pilot could be a technical success and still go nowhere because the decisions it was meant to inform were being made by people who had never been involved in setting it up.
Two or so stalled on what I’d call the missing step. IT or a digital transformation office built the pilot. The operators who would’ve used it were brought in at the end as reviewers. By then the thing fit a workflow nobody ran.
The remainder stalled for the usual reasons: vendor acquisition midway through, or a budget cut before the first stakeholder review.
What stands out in the number is that model quality was rarely the issue, rather than the stall rate itself, which is high in my experience almost everywhere. In nearly every case, the tech was good enough. The pilot exposed operational dysfunction that had been tolerable before and that the AI project forced into view.