One example of the necessity of balloon launches is in convection. Knowing only surface convergence, temperature and moisture is important, but does not tell the whole story. If the 850 mb temperatures are too warm, it won't matter how good the surface observations look, you won't get convection. How much energy the storm can tap into depends on the temperatures all the way up to the tropopause (and higher if you want to know how high the cloud tops will be). Knowing what type of storm to expect depends on knowing the winds in at least the bottom 6 km of the atmosphere. Having vertical information provided by a radiosonde is the difference in only being able to say "the air is warm and juicy" versus being able to say "expect isolated supercells today with large hail and straight-line wind damage" or "despite the high dew points, the capping inversion and associated CIN will suppress all convection today -- no storms" given the exact same surface observations.

And yes, models can give us the upper air structure to base these kind of statements on, but remember the radiosonde launches are where a lot of the upper-air data comes from that allows the models to work well comes from.