Most enterprise AI programmes do not pay back. I work on why, and on what would have to change.

The usual explanations are technical — the wrong model, the wrong data, the wrong vendor. The evidence points somewhere less comfortable. Adoption is broad and shallow, the complementary investment that would let a deployment change anything operationally is rare, and a saving only becomes money when somebody decides something.

That last point is the one I keep returning to. Across the United States, over a six-month window, 95.7% of firms using AI report no AI-driven change in employment of any kind (U.S. Census Bureau, CES-WP-26-25, April 2026, Table 4). Hours were saved somewhere. Nothing downstream of them moved. That is what an unconverted saving looks like at national scale, and no amount of model quality fixes it.

So the question I think is worth asking before an AI programme starts is not which department generates revenue — nearly every firm already answers that the same way, and it has not helped. It is narrower and harder: if this works, what specifically changes, who decides it, and by when? Where there is no answer, the return is zero regardless of the department, the vendor or the model.

I published that argument in a peer-reviewed journal in May 2026, and then built it into something people can run against their own organisation rather than take on faith. Both are on the research page, with permanent identifiers.

Elsewhere

ORCID 0009-0009-2774-0897
GitHub github.com/madinaedigeeva
LinkedIn in/madinayedigeyeva