The answer is not so straight forward. From a mathematical perspective, if the importance is derived from tree based models, then yes, because the sum of importances adds up to 100 as explained by @Mattravel.

However, random forests tend to give more importance to features with higher cardinality, and hence binary features, like those coming from OHE will inherently show lower importance.

So, while it is true that we can add importance, to truly assess the importance of a categorical variable we might want to use additional methods, like using a different encoding, or a different feature selection process that can take up categorical variables as inputs.

For a list of feature selection process that support categorical variables check out feature-engine's documentation.