You plan it so the only outage is a failover, not a restore. For a standalone instance that means a side-by-side migration with log shipping or backup and log restore, cut over in a short window. For an availability group it means upgrading a secondary first, then failing over. Microsoft documents that once a secondary runs a newer version, the group can fail over to it but cannot fail back to the older version, so rollback must be planned before the failover, not after.
An AceMQ upgrade plan contains:
An inventory: version, edition, compatibility level, features in use, linked servers, jobs, logins and drivers.
A compatibility test at the target version with Query Store open, so a regressed plan is visible and can be forced back as a stopgap.
A rehearsal on a copy of production data and the real workload.
A cutover runbook with a rollback per step and a named decision point.
A first-week review of waits, plans and errors against the pre-upgrade baseline.
Downtime is the failover window, usually seconds to minutes. It is not zero, and a plan that promises zero has not read the connection strings.