BEFORE AI TAKES OVER: CAN WE PROVE IT STOPS WHEN WE TELL IT TO?

A stop command ends a process, not what the process set in motion. Before it stops, an AI agent can issue a credential, queue an instruction, schedule a job or delegate work. This paper proposes a conceptual framework for judging whether that retired authority can still produce a protected business effect. Business-impact analysis defines the effect first, and reconciled evidence and negative tests then challenge the routes. A surviving route is FAIL. Unresolved state, attribution or timing is INCONCLUSIVE. Adequate evidence over a justified window supports a bounded PASS. The cutoff is the accepted stop demand, so later containment does not erase a late accepted effect. Independent authorization needs its own assessment. All three outcomes enter a scenario-conditioned loss interval, shown with three constructed payment cases and an OpenAI sandbox incident (scope errors only). The contribution is a decision framework. No implementation, detection gain or reviewer validation is claimed.