The idea of a 'straight and level' 'fallback' autopilot function is a good one, except that as you will both doubtless know, once you start coding for this you will find there will probably be situations where that is not the best solution.
The code is basically a series of conditional statements: "if x is true then do y" so any situation that can be foreseen can be written into the code. In situations where that 'straight and level fallback' is not the best solution, other options could be in place, including reverting to the current logic (handing control to the pilot).
A lay pseudocode example of an obvious stall avoidance scenario:
"If altitude>35,000, if airspeed unreliable, if thrust idle, if TCAS not blaring, and if AOA >10% set pointNoseDownFast=true! otherwise do nothing"
Add some better thought-out version of the above logic to the flight computer and without affecting anything else you could avoid this accident.
My issue is more with the piecemeal approach to this automation, where pilots are not just pilots, they're computer operators interfacing with the software that actually does the flying. Either pilots should be rigorously trained in the software (including how it was written), and that software should be responsible for handling all possible scenarios (loss of airspeed being quite a common one) while providing intuitive feedback to the "computer operator", or it should be removed apart from the simplest functionality (more along Boeing lines).