Of course it can. As a matter of fact, it is a typical use case for ML. Two years ago, I worked on a project for a bank (one of the top 10 biggest banks in the world) and they had a ML pipeline that was able to detect outliers within 50 miliseconds.
A typical way of detecting fraud is looking at anomalies:
- are there big sums of money moved around?
- are there many small transactions from linked accounts to some common place?
- were these accounts sleeping but then suddenly woke up and started to move money?
And things like that. A Machine Learning algorithm can detect such suspicious activity and flag transactions for futher analysis