Three things are needed before the term applies: a model that can be executed, a live link to the data of one specific asset, and a comparison between what the model predicts and what the asset does. Without the data link it is a simulation. Without the comparison it is a dashboard.
Fidelity is a design decision
More detail is not better by default. A twin built to estimate thermal fatigue in a power module needs an accurate thermal network and can ignore the geometry of the enclosure. A twin built for airflow needs the opposite. Deciding what the twin is for, before deciding how detailed it is, is what keeps it maintainable.
What it is actually used for
The divergence between predicted and measured behaviour is the diagnostic signal: a drift that grows steadily points at degradation, a step change points at an event. And a twin lets you ask what happens if a setpoint moves, without moving it on the asset.
In BeAI's products
InverterAI keeps an extended digital twin per inverter, with a Foster or Cauer thermal network fed by that unit's own SCADA history, which is what turns generic component ratings into a remaining useful life for that specific unit.