Gen Re has an excellent article on the role of classification model performance in underwriting. This includes some useful comments on machine learning, plus the gloriously-named 'confusion matrix' with associated metrics for assessing the accuracy of models. There are some broad implications for the future of underwriting which can be drawn, you should check out the whole article, of course, but the thoughts below are my own, based on their article and other sources:
Larger data sets, and especially digital test data sets of sufficient size, are particularly valuable for assessing a model. Companies would do well to both maintain such data sets (over time, with claims performance data attached) and the expertise necessary to assess the value of different classification models. A presumption of the efficiency of current underwriting methodologies is not necessarily well-founded. It is also somewhat self-selecting and may lead to ignoring data not currently collected, which could be of value in developing new methods for making underwriting assessments. Non-traditional data sources - such as credit records, social media profiles, wearable health tech, and so on, could provide new opportunities for assessment models. You won't know until you look, so some investment in collecting better data is necessary R&D expenditure.
Data collected is not always equal to data assessed, also, data required to apply for a product may be reduced, while data-collection may continue after application (and it may be useful and economically rewarding to continue to collect the data) so that better analysis of factors for future modelling may be identified. A practical version of that point might be: we can still have a non-underwritten product (to make application easy) and collect lots of data to feed our machine learning tools to enable better underwriting decisions to be made on the basis of third-party data, for future sales.