The head of risk asks the AI model to check a proposed AI use against the written policy and to draft the committee's decision note. The model receives the policy and the proposal as written, and the risk register with its client-name column masked. The evidence section is the head of risk's own search of Kovac's refusal record: a project manager not cleared for contract values, for instance, refused that column. The note says what that record is today: refusals only, not signed.
Fields you mark sensitive reach the AI model masked with placeholders; fields you don't mark reach it as they are. The same value always gets the same placeholder, so masked data can still be linked to a person and remains personal data.
On records loaded into Kovac, someone not cleared for a column you have marked who asks for it is refused, and the refusal is recorded.
Clearance decides what a person may read. It never unmasks a field for the AI model: a marked field reaches the model masked on every call, whoever asks.