From use cases to an AI factory

An inventory of 38 AI use cases is not a portfolio. It is a list in which vendor-embedded features, exploratory ideas and fundable projects sit side by side, assessed by different people on different criteria. The factory starts when the organisation can tell them apart — and can say, for each, what value it targets, what risk it carries, and what it would take to run it.

The instruments are built one on the limit of the previous one. A five-axis diagnostic with no composite score, where the value is in crossing the axes: fast-evolving AI maintained by vendors without a common steering frame; high-risk systems with opaque supervision. An intake test that separates proofs of interest from projects. A six-dimension synthesis sheet — value, finance, IT solution, social impact, risks, responsibility — with a three-colour verdict and “non-blocking reservations”, so anyone can record a problem without vetoing. A value × feasibility × factorisation grid for programme design, with risk filters.

Evidence is published before headlines. When a collaborator assistant shows 84.6% satisfaction and +45 NPS but under 5% of weekly time saved, both numbers go to the executive committee, the second one first. That is what lets a programme be funded on credible grounds — 21 AI-heavy projects and an acceleration programme in one budget cycle — rather than on enthusiasm that the first audit will puncture.

The factory’s last component is run governance: a common method for predictive models, batch and real-time architecture frameworks validated by the architecture authority, a shared vocabulary, and the agent layer governed as a gateway with autonomy tiers. Without that, the factory produces pilots.