AI Drafts, Humans Decide: Designing the Human-in-the-Loop Boundary

'Human in the loop' is one of the most-cited and least-specified phrases in AI governance. Most organizations that claim to have it cannot answer a simple follow-up question: for which decisions, exactly, is a human required, and what happens if that human defers to the AI's output by default.

The PMAIS approach: draft versus decide

PMAIS's principle is deliberately narrow: AI may draft, summarize, accelerate, or recommend. Humans decide, and every AI-generated artefact carries a named human reviewer of record. This draws an explicit line between what the system may do unsupervised and what requires a human decision before it has effect, and assigns the reviewer role to a specific named individual — not to 'the team' in the abstract.

Designing the boundary without slowing delivery

A well-designed oversight boundary is faster than the alternative, because ambiguous oversight produces slower delivery: work stalls in informal escalation, or ships without real review and creates rework later. The design principles: scope review to what carries actual risk; make the reviewer's decision binary and time-boxed; and route review to the person with the shortest path to a confident answer.

Why this has to be an explicit artefact

An unwritten commitment degrades under delivery pressure. A written, gate-owned Human Oversight Plan with a named reviewer per artefact type is much harder to quietly erode, because skipping it becomes a visible, attributable decision.