Document Classification

Document classification is the task of assigning each incoming file a category or type (for example: invoice, delivery note, purchase order, contract, credit note) so the system knows what to do with it. It is the first step in any document workflow: before extracting data or validating a receipt, you need to know what kind of document it is.

In its modern form, classification no longer depends on the file name or on manual folders; it reads the document's actual content (text, structure, stamps, fields) using AI models and OCR. That lets it tell an invoice from a delivery note even when they arrive as a scanned PDF or a phone photo. It is one of the pieces that make up intelligent document processing, and within the Vantegrate stack it is part of what Arconte automates in documents and finance.

Done well, classification turns a chaotic inbox of files into an orderly flow where each document goes to the process it belongs to, without a person having to open it and decide by hand.

How it works in practice

The typical flow starts when a document comes in through some channel: an email to an accounts payable mailbox, an upload to a portal, a scan or a photo sent over WhatsApp. The system first digitizes it with OCR to get the text, and then a model analyzes signals such as keywords ("invoice", "CAE", "delivery note"), the layout structure (where the totals fall, whether there is a table of line items), the presence of typical fields (tax ID, document number, VAT status) and even the overall format. With those signals it assigns a category and a confidence level. If confidence is high, the document moves on by itself; if it is low or ambiguous, it is routed to a person for review, a pattern known as human-in-the-loop.

Why it matters for a business

A midsize company in Argentina can receive hundreds of documents a day from suppliers, logistics and customers. If someone has to open every PDF to decide whether it is a Type A invoice, a delivery note or a credit note, a bottleneck is guaranteed: delays in paying suppliers, data-entry errors and misplaced documents. Automatic classification is what makes everything that comes after scalable (extraction, validation, reconciliation). Without classifying well, you cannot automate the rest of the document workflow.

A concrete example: a distributor receives supplier invoices, carrier delivery notes and credit notes all mixed together in a single mailbox. The classifier separates the three types instantly, sends the invoices to the accounts payable flow for reconciliation and three-way match, the delivery notes to goods receipt control, and the credit notes to balance adjustments. What used to take an administrative clerk all morning is done in seconds.

Classification vs data extraction

People often confuse classifying with extracting, but they are two distinct, sequential steps. Classifying answers "what type of document is it"; extracting answers "what does it say inside".

The relationship is one of dependency: good classification tells the data extraction engine which template or which fields to look for, because you do not extract the same thing from an invoice as from a contract. If the document is misclassified, extraction looks for fields that do not exist and the workflow breaks.

Common mistakes when implementing it

• Relying on the file name or the source folder instead of the content: an "invoice_final_v2.pdf" could be anything.
• Not defining a confidence threshold or a human review path, which leads to silent misclassification.
• Training the classifier only on "clean" documents, so it fails on skewed scans, low-quality photos or documents in unusual formats.
• Treating very similar categories (Type A vs Type B invoice, delivery note vs priced delivery note) as if they were one, losing the detail that downstream processes need.
• Forgetting that categories evolve: new document types appear, or documents from a supplier nobody anticipated, and the model needs to keep learning.

Where it fits in the full workflow

Classification rarely lives on its own. It is usually the trigger for a pipeline that continues with extraction, document validation and, in finance, reconciliation against the purchase order. Thinking of it as the gatekeeper of the workflow helps: its job is not to resolve the document but to send it through the right door so each specialized process can do its part.