If a tech works 80% of the time, then I know that I need to be vigilant and I will review the output. The entire team structure is aware of this. There will be processes to offset this 20%.
The problem is that when the AI becomes > 95% accurate (if at all) then humans will become complacent and the checks and balances will be ineffective.
Maybe people here are used to good code bases, so it doesn't make sense that 80% is good enough there, but I've seen some bad code bases (that still made money) that would be much easier to work on by not reinventing the wheel and not following patterns that are decades old and no one does any more.
My list so far is:
* Runs locally on local data and does not connect to the internet in any way (to avoid most security issues)
* Generated by users for their own personal use (so it isn't some outside force inflicting bad, broken software on them)
* Produces output in standard, human-readable formats that can be spot-checked by users (to avoid the cases where the AI fakes the entire program & just produces random answers)