So, what about using exponential smoothing or 95% peak estimation to get a better "current" spending rate? There is also the possibility to do some data mining on the actual usage (assuming you have collect that data and are willing to use it in aggregate to improve service) — i.e. customers who spend like you do and have a balance like yours usually run out of money after x days.

Another option I just thought of: try and estimate usage several different ways, and then check the variance — if its low, then issue an alert. I wouldn't be surprised if there is some research on how to do this, so it should be cook-up-able relatively easily if its a feature you want to add.