Does it follow from this that while you could train a large machine learning program to do a task then separating it from a monolith into separable parts would

Does it follow from this that while you could train a large machine learning program to do a task then separating it from a monolith into separable parts would be highly unlikely to be successful? For example,you could train an algorithm that does object identification and then train an algorithm that does object boundary identification considering only the red value in a given photo from the original large program. However if you start with the components of an object identifier you might not be able to reach SOA or at least those two methods would look very different.