The naming problem is the governance problem. Forty years of platforms each renamed the same idea, so a single company can run thousands of non-human identities without anyone agreeing on what to call them or who owns them.
What it's called on each platform
Microsoft Entra: service principal, or managed identity when Azure handles the credentials
AWS: IAM role
Google Cloud and Kubernetes: service account, meaning two different things
Windows Server: gMSA
Where teams get it wrong
Joiner-mover-leaver processes point at people. Almost none of them point at service accounts, so permissions granted in 2021 to fix a problem are still granted today, and nobody has a process that would ask about them. An AI agent doesn't inherit the caution behind those grants. It inherits the permissions.
The Cloud Security Alliance found 65% of organizations saying Zero Trust can't secure their non-human identities. That reads like a gap in Zero Trust. It's a gap in the layer underneath, where the identity should have been defined in the first place.