I have no idea whether the current telematics algorithms are any good at assessing risk.

However, with a large enough corpus of training data including years of telematics on drivers with and without insurance claims it certainly seems like it should be possible to build models that can make much more accurate risk assessments than they can with the data they currently collect such as age, zipcode, and past accidents.

Put another way, I would be surprised if there was zero correlation between driving behavior and insurance claims.