Every AI accuracy problem we opened this week was a governance problem in costume.
Ironic subject line on file: “Model Grounded In Undefined Data Is Now Confidently Wrong, With Citations”
Retrieval-augmented generation was supposed to make the model accurate by feeding it your documents. It turns out the accuracy lives in the documents — in whether they are current, owned, and consistent — which is a data-quality program no one scoped before the demo. The companion story is lineage: required for the audit, present in the slide, partial in the warehouse, and absent the moment the number leaves it. The thread is that AI does not retire governance debt; it draws interest on it, then asks why the statements don't reconcile.
Retrieval-augmented generation is a data-quality project nobody scoped.
The retriever inherits every undefined term in the corpus. The model just reads it aloud.
Lineage is mandatory for the audit and partial in practice.
A graph that stops at the warehouse door explains everything except where the number came from.
Before the model can be trusted, the data has to be — and trust is a decision someone signs, not a feature you enable.