She drafts the work. You approve it. It happens.

What she can do

Draft purchase orders (one per supplier, split automatically). Propose reorder-point changes with the reasoning shown. Draft promotions for slow days. Flag items for review. Each capability is a switch in your control panel — all OFF by default, enabled by you, per action.

The pipeline

Propose → pending → approved → executed, with rejected and expired as first-class outcomes. High-risk actions can require a password re-entry and a second approver. Every step lands in an append-only audit log, and executed actions carry an undo where reversal is possible.

Why propose-only matters

An AI that acts unsupervised is a liability with a chat window. Maya's power is real precisely because it is gated: the work arrives done, and the authority stays human.

The pipeline, and why each state exists

Proposed, pending, approved, executed — with rejected and expired as first-class outcomes rather than silent disappearances. Expiry matters: a reorder proposal that made sense on Monday should not still be sitting there executable on Friday, because the stock position it was based on has moved.

High-risk actions can require a password re-entry and a second approver, and every step lands in an append-only log. Where an executed action can be reversed, it carries an undo.

Off by default, on by decision

Every capability is a switch, and every switch starts off. Nobody wakes up to discover that AI has been drafting purchase orders because a default said it could — you turn on drafting purchase orders, specifically, when you want that.

This is the claim SeroBooks actually makes about AI, and it is a governance claim rather than a capability one: plenty of accounting products ship AI, and several are good. What we are building is one that proposes and asks permission, with the audit trail and the kill switch to make that meaningful.