Long-form notes on AI governance, model risk, and the audit questions no one's asking yet. Written by someone who sits in the room.
A boundary and a classification model are only as good as the population you run them against. Where to find the AI your organisation does not know it has, including the vendor defaults that connect your data to a model without anyone deciding that.
Most AI registers use one classification field: high, medium, low. This piece proposes five dimensions instead, three that combine into a score and two that can override it outright, so the reasoning behind every tier stays visible.
Most AI registers look complete and are not. This article sets out the three decisions, what counts as an AI system, what a record represents, and what classification is for, that determine whether yours reflects your actual estate.
Control selection is downstream of classification, and classification is a governance output. Where that chain is missing, what looks like a control framework is really just a control list.
"AI security" collapses two different problems into one conversation: attacks on the model itself, and what using AI does to everything you already have. IBM's 2026 data says most incidents are the second kind, and this piece argues you should scope for it first.
A rules engine, a credit scoring model, and a language model share a label and almost nothing else. This piece argues the deterministic/probabilistic split, not impact or autonomy, is what should decide your testing methodology.