This sounds like the beginnings of the McNamara fallacy: that there is some abstract "data" where all the relevant variables can be explored to decide the best course of action. The reason it's a fallacy is that data is itself only a reflection of conscientious decisions about what to measure and how to measure, and these are invariably limited to what we already appreciate and understand, not accounting for the things we don't know how to measure and don't know we should be measuring.
To be more concrete, let's say we follow your suggestion and focus on crashes. Maybe it's found that automatic breaking reduces the number of crashes so it's mandated. But maybe automatic breaking also increases the likelihood that a crash is deadly in some sub-population like pedestrians, construction workers, and motorcyclists. So overall we see crashes decline because a large number of minor crashes are averted but we also see deadly crashes increase because pedestrians, construction workers, and motorcyclists are killed more frequently.
So data isn't a free lunch. Every decision about what to measure and what to optimize involves some prior contextual understanding. A misunderstanding of the context means the data will misrepresent some aspect of the problem and you'll overoptimize yourself into a local maximum.