"It depends", which is probably the most unsatisfactory answer to just about any question. But it is true unfortunately. In the case of fraudulent motor vehicle claims you will need to know something of the shape of the motor insurance distribution and incidence of fraudulent claims. For instance there may be different types of fraudulent claims such as by individuals, by corporations, by organised criminals etc. What you are trying to do with this qualitative analysis is work out whether your population is likely to follow a normal Gaussian distribution, or something different BEFORE you apply mathematical analysis. After determining the likely qualitative characteristics of your population you can apply the guidance on sample size determination, such as the following:

https://www.statisticshowto.datasciencecentral.com/probability-and-statistics/find-sample-size/

I also attach a short paper on motor vehicle fraud which may assist in your qualitative analysis.

Article CASE STUDY: Drive for effectiveness: How an internal audit o...

If the class distribution is imbalanced, it is likely going to affect the performance of the ML model. But there are no hard-and-fast rule that tells whether a sample dataset is imbalanced. But if it is like 1:10, then it is imbalanced.

Thanks Chris Kelly . I will go through the papers you have mentioned. Please share if you have similar paper on Life Insurance also?

Thanks Adam Wong . One question, can we apply SMOTE on data if imbalance is 1:100

Over-sampling is one of the methods to handle imbalanced dataset - it is usually used with an appropriate validation method to avoid creating an overfitting model. Some other methods are explained here: https://www.kdnuggets.com/2017/06/7-techniques-handle-imbalanced-data.html