Rédiger un PREreview

Proposed research protocol. No empirical results. Not peer reviewed. Organizations need outside people and outside AI agents to work inside their systems, while sovereign AI policy and data protection law govern how that access may be granted. The question is not whether external capability exists but whether it can be delegated without transferring control. This protocol specifies a comparative evaluation at the level of a single engagement: can outside practitioners, assisted by AI agents, deliver work a client judges useful while the client retains demonstrable and testable control over data access, consequential decisions, revocation and exit? A bounded set of workflows is compared under a combined control package against a control condition that must meet a stated security adequacy floor, so no arm is deliberately insecure. Tooling and practitioner preparation are separated by a two by two factorial design, with a staged fallback, so an improvement cannot be credited to the wrong cause. Co-primary outcomes are authority integrity (unauthorized consequential attempts blocked before effect, independently corroborated) and client acceptance on a pre-agreed rubric. Secondary outcomes cover revocation latency, leakage under seeded probes, recovery to a verified clean state, reviewer sensitivity and specificity, full cost per accepted deliverable, and completion of exit. Failures and adverse events are publishable outputs with the same standing as successes. No results from the proposed study are reported. Version 1.0. Author-approved preprint. This proposed evaluation does not claim empirical findings, ethics approval, institutional endorsement or completed implementation arrangements. Generative AI assistance and competing interests are disclosed in the manuscript.