The instinct when giving software the ability to act is to put an approval in front of every action. It feels careful. It reliably produces the opposite of care.
An approval wall in front of routine work generates a stream of requests that are almost always fine. People learn the pattern within days and start approving without reading, which is a rational response to a queue of items that have never once needed intervention. The wall is then providing no protection and a documented record of consent that nobody gave attention to.
The default we run
Inside its assigned purview, with access it has already been granted, an agent acts. Access granted is treated as authorisation to use it, which removes the second guessing that makes a permission model incoherent. Reading, drafting, analysis, recommending and modelling were never gated at all, because none of them changes anything outside the system.
What replaces the wall is a record. Every automatically executed action writes to a standing audit trail, which does not prevent a bad action and does make it visible, attributable and reversible where reversal is possible. The record is what makes the permissive default defensible.
Where the walls do belong
Six kinds of decision stop and wait, and they are chosen by consequence rather than by frequency. Spending above a threshold, client-outbound email needing discernment, legal and contractual positions, employment actions, anything irreversible, and any conclusion that belongs to a qualified professional.
Six is a number a person can hold in their head. A model with sixty exceptions is not a governance model; it is a list nobody has read, and the actions it fails to stop will be the ones it never contemplated.