AI action authority

AI action authority answers a concrete question: what may this system do now, for this property, with this evidence? It is not the same as model capability. A model may be able to draft a notice or select a vendor, while policy allows it only to summarize the situation. Authority should be defined by action, dollar or risk limit, required inputs, property scope, actor role, and the conditions that force human review.

Useful authority has levels. The system may observe and summarize, prepare a reversible draft, execute a bounded routine step, or be prohibited from acting. A human-in-the-loop checkpoint belongs before actions whose consequences, ambiguity, or governing rules exceed the approved level. The checkpoint must show the proposed action and evidence; a generic “approve” button without context is not meaningful oversight.

Suppose a tenant reports a leaking faucet. Policy may let AI acknowledge the request, collect photos, and offer appointment windows. It may dispatch a previously approved vendor only when the issue matches a routine category, the estimate stays below the owner’s cap, access consent is recorded, and no safety or habitability ambiguity appears. Missing consent, an active leak, or a higher estimate changes the action from execute to escalate.

Authority fails when it is expressed as a broad product setting such as “autopilot on.” It also fails when stale property data, an expired vendor credential, or a changed policy is ignored. Evaluate authority at action time, record the policy version and facts used, and issue a receipt for any external call. Human review is a control boundary, not a transfer of blame for a proposal the reviewer could not inspect.