"Managed Kubernetes" covers everything from a control plane you still have to operate to a team that carries the pager for your workloads. Before signing anything, these are the questions that separate the two.
Who performs upgrades, and how often?
Kubernetes moves quickly. Ask who schedules minor version upgrades, how far behind the provider allows a cluster to fall, what the rollback plan is, and who tests workload compatibility beforehand.
Where does responsibility stop?
Control plane only, or nodes too? Ingress? Certificates? Storage classes? The failure modes that hurt most tend to sit exactly on the boundary between "platform" and "application" — so agree where that line is while everyone is calm.
What happens at 3 a.m.?
Ask for severity definitions, response expectations and who is actually on call — the same engineers who built the cluster, or a first line reading a script. Ask what happened during the last three incidents.
How is recovery tested?
Cluster state, persistent volumes and databases recover differently. Ask when a full restore was last performed, how long it took, and whether the result was written down.
How do we leave?
Standard Kubernetes, standard storage interfaces and infrastructure as code mean an exit is an engineering task. Custom control planes and undocumented glue mean it is a negotiation. Decide which you are buying.