I am working on a two-class classification model to predict if some lead becomes a sale. At this moment I have all leads try to predict these ones, which are sales. I become good metrics right now like a Recall from 80% on unseen data. The problem I see is in real life, where the model works right now, where I can not see the same Recall.
We have a similar Sales-Funnel like in the image below, but with much more steps:
I am wondering is this the right way to predict from all leads the sale (step 5) directly or it would be better to make a model to predict from all leads the conversion (step 4) and after that a new model to predict from all conversions the sale (step 5).