There's probably a danger-maximum somewhere before 100% replacement, though. Some mix of highly-predictable self-driving cars and poorly-predictable human-driven cars will cause the most accidentsThat's an interesting point. Since the technology will be improving at the same time adoption is increasing, I don't know if we'll be able to observe that maximum, but I think you're right. One thing to consider is that once there are a significant number of self-driving cars on the road, they'll start communicating in a much faster and richer way than human drivers are able to communicate. If self-driving cars are able to achieve human collision rates with human drivers, they will be able to achieve much, much lower collision rate with each other. That effect might lower the percentage at which safety starts to improve.
Another variable will be the built environment. Over the years we have developed road systems that are highly enriched with features to assist human drivers, and of course we will begin to do the same for self-driving cars.
Nobody is talking about those improvements right now because they aren't relevant to getting the first cars on the road. The engineers working on self-driving cars have to tackle the hardest version of the problem that will ever exist first, and only then can they start making it easier.
Confession: for years I predicted to my friends that the first autonomous passenger vehicles to coexist with untrained members of the public would be in a theme park or a city center closed to other vehicles, with sensors and other assistance built into the environment, because I thought we'd have to start with an easier version of the problem.