Autonomy is a ladder, not a switch

"Should the AI do this on its own" is the wrong question. Every task sits on a rung of a ladder: the agent suggests, the agent drafts and a person reviews, the agent acts and logs it, or the agent runs unattended. The rung is a decision about the task, not a property of the model. Two tasks running on the same model can sit two rungs apart, and should, if what happens when they go wrong is different.

Decision rights

A person places each task on the ladder, and they track two numbers, not one. The ceiling is the highest rung that is safe given what a failure costs. The current rung is where the task actually sits today, based on the evidence you have. The gap between them is the roadmap. A task with a high ceiling and a low current rung is one you are still earning trust on. A task with a low ceiling is one that stays there no matter how good the model gets.

Memory and context

The rung lives on the task record itself, not in a policy doc and not in someone's head. When the task moves to a new owner or a new agent, the rung and the reasoning move with it. If you have to ask someone which rung a task is on, it is not on a rung.

Execution

Each rung costs something different. Suggestions are nearly free. Draft-for-review costs reviewer attention, and that cost scales with volume until reviewers stop reading. Act-with-a-log needs a tested way to reverse the action, not a theoretical one. Unattended needs someone actually watching, with a name against the watching. Picking a rung is picking which of these bills you are going to pay.

Governance

Promotion up the ladder needs two things: a measured error rate on real volume, not a demo, and containment already in place for when it goes wrong. Demotion has to be as ordinary as promotion. If moving a task down a rung feels like an admission of failure, nobody will do it, and the ladder becomes a ratchet that only turns one way.

Observability

Each rung has its own number. Suggestions: how often are they accepted. Drafts: how far does the reviewer edit them. Logged actions: can you reconstruct end to end what happened without asking anyone. Unattended: how fast do you detect a problem. If every AI task in the company sits on the same rung, nobody made a decision. You have a default wearing a policy's clothes.