Great question Victor.  While you are correct that you could accomplish this with one very large model for all these channels via context, this make the model very monolithic in a few ways.  It means that the model has the same operational model for all channels, which often is not the case (ie, for certain channels the rate of learning is much faster than others and you want the parameters to be different for those channels).  Some channels may have different data needs, such as predictors.  Lastly, mixing very different channel insights together has some overarching influence on the overall model.  This introduces 'noice' into the model.  As such, what has been implemented in the NBA Designer strategy framework here forms a best practice of how to segregate models.  Our data science team has put in a lot of time to think through the best practices and apply them in this framework.  

I hope this helps clarify things for you.