The defining characteristic of a scientific theory is that it makes falsifiable or testable predictions about things not yet observed. The relevance, and specificity of those predictions determine how (potentially) useful the theory is. A would-be theory which makes no predictions which can be observed is not a useful theory. Predictions which are not sufficiently specific to be tested are similarly not useful. In both cases, the term 'theory' is inapplicable.
In practice a body of descriptions of knowledge is usually only called a theory once it has a minimum empirical basis. That is, it:
is consistent with pre-existing theory to the extent that the pre-existing theory was experimentally verified, though it will often show pre-existing theory to be wrong in an exact sense, and
is supported by many strands of evidence rather than a single foundation, ensuring that it is probably a good approximation, if not totally correct.
Additionally, a theory is generally only taken seriously if it:
is tentative, correctable and dynamic, in allowing for changes to be made as new data is discovered, rather than asserting certainty, and
is the most parsimonious explanation, sparing in proposed entities or explanations, commonly referred to as passing the Occam's razor test.
This is true of such established theories as special and general relativity, quantum mechanics, plate tectonics, evolution, etc. Theories considered scientific meet at least most, but ideally all, of these extra criteria.
Theories do not have to be perfectly accurate to be scientifically useful. The predictions made by Classical mechanics are known to be inaccurate, but they are sufficiently good approximations in most circumstances that they are still very useful and widely used in place of more accurate but mathematically difficult theories.