Updated 17 October 2026. Control over a deployment and continuity of service are related but different goals. A system can satisfy a location preference while remaining vulnerable to a dependency failure.
Test the recovery assumptions
Start with the service the organisation needs to preserve. Identify the infrastructure, identity, data and external tools it depends on. A recovery plan that names only the model may overlook the part of the workflow most likely to stop useful work.
An illustrative failover plan might require moving processing to another environment. Before relying on that option, assess whether the data, contracts and applicable rules permit the move. Technical possibility does not establish permission.
Where movement is constrained, consider what degraded service is acceptable. A read-only mode, manual queue or temporary restriction may preserve useful work without pretending the full service remains available. The decision belongs in the operating plan before an incident.
Rehearse recovery with appropriate test data. Record what actually worked, what required manual intervention and what remained unavailable. A successful infrastructure restart is not necessarily a successful business-process recovery.
| Question | Evidence or action |
|---|---|
| Dependency | What the service needs to function |
| Constraint | Where recovery may occur |
| Fallback | Useful degraded operation and restoration |
The practical next step
Keep the sovereignty assessment and continuity exercise in the same conversation. Otherwise one team can design a restriction that another team assumes it may ignore in an emergency.
Keep the resulting decision, its owner and unresolved questions in a short decision record. For claims about agent behaviour, connect the decision to evaluation evidence rather than relying on a successful demonstration alone.
Further primary reading: NIST AI Risk Management Framework. The review map and product recommendations above are my analysis, not a certification or a legal assessment of a particular deployment.
