AI does not fail on model quality. It fails on ownership.
Why AI initiatives that report to IT underperform those that report to the operating committee.
AI does not fail on model quality. It fails on ownership.
We tracked 63 AI initiatives across the transformation cohort from approval to month 18. The single strongest predictor of value delivery was not vendor choice, data readiness or model quality. It was the reporting line. Initiatives owned by the operating committee and priced per use case delivered a median 68% of their value case; initiatives owned by IT delivered 23%. The technology was frequently identical. Reporting line and pricing discipline travel together almost everywhere in this data, so we cannot say how much of the gap belongs to ownership and how much belongs to the approval discipline that tends to arrive with it.
The mechanism is the same one that governs every transformation in our data: AI creates value only when the operating model changes around it. A forecast has to change an ordering decision. An automated answer has to close a case without a queue. Those are operating decisions, and IT does not own them. This paper sets out the ownership test, the per-use-case pricing discipline, and the four failure patterns that account for most of the write-offs in the sample.
Median value delivery at month 18 for AI initiatives owned by the operating committee. Initiatives owned by IT delivered 23%.
Use cases in production in the strongest programme in the sample, each priced and verified separately.
Higher write-off rate for initiatives approved as platforms rather than as priced use cases.
Share of sampled initiatives that could not state, at approval, which operating decision the model would change.
Move AI ownership to the operating committee. If the initiative cannot name the operating decision it changes, it is not ready to be funded.
Approve use cases, not platforms. Each use case carries its own value case, baseline and kill test.
Verify per use case at a fixed cadence, using the same standard you would apply to any other transformation outcome.
Treat data and platform work as enabling spend inside a priced use case, never as a value case of its own.
Markham Institute, AI adoption as an operating-model problem, MKM-R-2026-005, v1.0 (March 2026). Citation permitted with attribution.