AI Eval Framework

Most AI evaluations cannot come back "no". That is how you know they are not evaluations.

Last week we worked on prioritisation: which AI use cases are worth doing at all. Value, readiness, risk, a ranking you can defend.

Prioritisation tells you which. It does not tell you whether.

This week we take a single use case all the way down the Business Architecture Spine (stakeholder, value proposition, outcome and KPI, value stream, capability, process, decision, agent, semantic model, data product, critical data elements, records, governance) and we build the AI Evaluation Framework on top of it.

Four quality pillars. Eight evaluation dimensions. Six gates. Then the Data Product Canvas that has to supply it.

The Spine is what makes the judgement call possible. Without it, "is this AI any good?" has nowhere to land. With it, the question becomes answerable: whose outcome, which decision, what evidence, who owns the quality, what closes the gate.

An evaluation that can only ever say yes is a formality. The Spine is what gives it the standing to say no.