Hi Leon,
Perhaps improving forecasting by using non-linear ANNs over linear methods?
In my view, as of today, the most impressive showcases of DL are more in the areas of image recognition, time series, generative models, classification of 1d/2d/3d data and such, than MLPs...
However, if you're looking for business applications specifically, there are some interesting examples on the blog:
On the other hand, if you went for time series, there would for sure be interesting business applications, say, from the stock market, or sales forecasting...