When self driving cars will be consistently more reliable than 99% of human drivers, that will be enough. That is not a lot of nines.
How could you even calculate this? By running massive experiments and seeing if fewer people die over a given period of time driving versus self-driving? Do you really think we can run such an experiment?
It is easier to do it if you consider 1% worst drivers, either older people with long reaction times, distracted drivers, drunk drivers etc. If you have a car that is obeying all traffic rules, it never goes above the speed limit and it has sensors that makes it brake before rear-ending other cars you are already better than the 1% limit. It is a simplistic approach, just giving an example of how 99% is not that hard to reach.