Gartner first published the 85% figure in 2017. Nine years later, it has not moved. That alone should tell you the problem is not in the choice of framework, the size of the model, or the GPU budget.
Across a decade of rescue engagements, the pattern is remarkably consistent. The project was scoped against a metric no business owner had agreed to. The data lived in a system the data team did not own. The model was handed off to an operations team that had never seen it. And the success criterion was ‘we have a model in production’ rather than ‘a decision is now better’.
Each of those is a strategic failure dressed up as a technical one. Fixing them does not require better algorithms. It requires three things almost every failed project skipped: a named business owner accountable for the outcome, a written definition of what ‘better’ means in monetary or operational terms, and an operating model for who runs the system after launch.
Where we have seen the 85% number reverse, and we have, repeatedly, it has not been because of new tooling. It has been because someone with authority drew a line around a single decision, defined what improving it was worth, and insisted the project be measured on that and nothing else.
The next decade of enterprise AI will not be won by whoever has the best models. It will be won by whoever has the discipline to attach those models to decisions that matter, and to keep them attached when the original team has moved on.
