AI adoption as an operating-model programme

A consumer goods group arrived with what most organisations had in 2025: fourteen AI pilots, a platform subscription, a steering committee, and no verified operating result. The pilots reported to IT; the decisions the models were meant to change — ordering, allocation, customer service — reported elsewhere. Eighteen months of activity had produced demonstrations.

The group’s question was whether to double the AI budget or halve it. The answer was neither: it was to change who owned the work.

Screening the pilot portfolio against one test — which operating decision does this change, and who owns that decision? — was quietly brutal. Of 42 candidate use cases (fourteen pilots and twenty-eight proposals), 31 could not name the decision they would change. They were killed on paper, before further spend.

The eleven survivors clustered where the operating value actually sat: demand forecasting feeding weekly ordering decisions, and service automation resolving contact classes that never needed a queue.

Ownership moved first: AI became an operating-committee agenda item with a value ledger, and each use case got a named business owner — the person whose decision it changed. Each was priced with a value case and a month-0 baseline; platform and data work were costed inside the use cases that needed them, never as a programme of their own.

The eleven use cases were then drawn in three phases under the standard operating model — three or four at a time, released into production only when the previous set had passed verification. HolisticAutomation built the pipelines and integrations; Markham held the architecture and the oversight. At month 10, all eleven were in production with verified per-case value.

Every use case carries its own baseline and attribution line; month-10 readings verified under MKM-F-003.