Why the first step is classification
You could try to extract fields immediately. But then someone — or something — would first need to know which model belongs to which document. That is exactly the information classification creates, and it creates it in a form the machine can process further.
Documents emerge from a stack
What arrives at intake is rarely a document — it is a scan stack. Classification recognizes the boundaries and writes each subdocument with its own pages out as a separate file. Everything after that works on a cleanly delimited single document.
The category is the routing key
Recognition models are configured per document type. Without a category, there is no deterministic path to the right model — with it, model selection becomes a lookup the machine can perform without asking.
The cheap step comes before the expensive one
Only the first pages are classified. The costly field extraction then runs exactly once, with exactly one model, on exactly the matching subdocument — instead of trying “all models against every document.”
Confidence comes free with it
Every category comes with a confidence score. This makes it possible to decide before analysis whether a document may continue automatically — an uncertain assignment never creates wrong metadata in the target system, but instead a review case.