Authority is the interface.

The interesting question about AI in a company was never what it could do. It’s what it’s allowed to do, who knows about it, and who can undo it.

Most products answer by hiding the question. Autonomy appears as a setting, the actions appear as results, and the boundary between them lives in someone's mental model of a settings page they visited once.

We think the boundary is the product.

Four verbs, in order

Watch any serious deployment of automated work and you will find the same four verbs underneath:

Observe. The system assembles context and reports. Nothing is at risk yet, so nothing needs permission.

Draft. The system prepares work and recommendations. Still reversible, still quiet.

Act. The system executes authorized internal work. This is where trust is spent, so this is where visibility has to live: every action inspectable, every outcome logged, and the record of what changed kept whole.

Escalate. The system holds uncertain or consequential work and returns it to a person. Not as a failure state. As the designed route for anything that deserves judgment.

A product that respects the order widens what it may do the way you would widen a new hire's remit: inside limits you can see, until you decide the limits can move.

Why this is an interface problem

Because authority that is not visible cannot be trusted, and trust that cannot be inspected cannot grow. The screen a founder looks at should make the answer to "what is the system doing on my behalf?" available at all times, in plain language, with an undo.

Anything less asks people to extend faith instead of confidence. Companies do not scale on faith.