Human-in-the-loop AI for government means consequential determinations are made by accountable officials, with AI preparing the decision — assembling the file, checking policy, drafting the determination with citations — and engineered gates ensuring nothing consequential executes without authorisation. MAIA implements the loop as architecture: tiered policy boundaries, approval queues with full evidence attached, and a sealed record of who approved what, when.
The gate is only as good as its evidence
A human-in-the-loop control that presents an official with a yes/no button and no file is compliance theatre. The gate must deliver everything the decision needs: the signals, the policy basis, the draft, the blast radius, and the alternatives considered. Officials move faster with complete files — the loop, done properly, accelerates government rather than slowing it.
Tiered autonomy, written as policy
Not everything needs a signature. Routine, reversible actions execute inside pre-authorised policy boundaries; elevated actions queue for review; critical actions require explicit sign-off. The tiering itself is written policy an administrator can read and change — so the autonomy boundary is a decision the institution owns, not a property of a model.