> Predicting which messages are most likely to be important is a challenging technical problem

I think this is much better solved as a social problem. Use mentions sparingly so they have value and have company conventions about what's expected to be read/skimmed/ignored.

Seems like a case of fitting a solution to a problem.