Organisations should decide which parts of a growing AI system they need to control to avoid lock-in.
Proposed mechanism: Select control points in workflows and infrastructure.
Conditions: Ability to change critical components.
Limit: Technical control does not itself establish enforceable rights.