Autonomy is not a property of an AI system. It is a decision made about that system by the organization that deploys it.
The Space of Autonomy is the bounded region — inputs, magnitudes, contexts, populations — inside which the system is permitted to act on its own. Everything outside that space requires a human in the loop. Everything inside it does not.
Naming the Space of Autonomy is one of the two or three most consequential governance decisions a leader will make about an AI system.
Why it must be explicit
Left implicit, the Space of Autonomy tends to expand. Every successful run inside the boundary is treated as evidence that the boundary could be wider. Every failure is treated as an exception. Over time, the system ends up making decisions no one ever authorized it to make — not because anyone chose that outcome, but because no one ever wrote the boundary down.
Made explicit, the Space of Autonomy becomes a contract between the organization and the system. Inside these limits, we accept what you do. Outside them, we require a human.
The four questions
For any AI system with autonomy, the leadership team should be able to answer, on one page:
- On what? Which decisions may the system make without human intervention?
- Within what magnitudes? What are the limits — dollar amount, population size, exposure — beyond which autonomy stops?
- For whom? Which populations, customers, or cases are excluded from autonomous handling because the stakes are too high or the model is too uncertain?
- Under what conditions of the world? In what circumstances (regulatory change, model drift, incident aftermath) is autonomy suspended pending review?
If the leadership team cannot answer these, the system does not have a Space of Autonomy — it has room to act.
How it interacts with the framework
The Space of Autonomy sits inside the Second Map for the workflow. It is enforced through the Challenge Path (which reveals when the boundary has been drawn wrong) and audited through the Judgment Trail (which records both the decisions inside the space and the human calls made outside it).
The failing test
An AI system whose Space of Autonomy has never been reduced is almost certainly one whose Space of Autonomy has never been examined. Working governance produces contractions as well as expansions. If the boundary only ever grows, the organization is not learning — it is drifting.