Intelligent Document Processing (IDP)

Intelligent document processing (IDP) is a set of technologies that turns unstructured documents (invoices, delivery notes, purchase orders, contracts, scanned forms) into structured data that a system can read, validate and process automatically. Unlike traditional scanning, IDP doesn't just "take a picture" of the document: it understands it, identifying what type of document it is, where each piece of data sits and whether that information is consistent.

IDP combines several layers: optical character recognition (OCR) to read the text, document classification to know what kind of document it is dealing with, data extraction to pull out the key fields and document validation to confirm that the data complies with business rules. For financial and administrative documents, IDP is part of the engine that powers the automation in Arconte.

What sets modern IDP apart from the template-based readers of the past is its use of machine learning and AI: you no longer need to manually define the exact position of each field. The model learns to recognize an "invoice number" or a tax ID (the CUIT, in Argentina) even when the supplier's format changes, which makes the solution scalable to hundreds of different formats without reconfiguring anything each time.

Why IDP matters in a real operation

In any midsize or large company, a huge share of critical information lives in documents: the supplier invoice that has to be paid, the delivery note that confirms a delivery, the contract that sets the terms, the withholding certificate that affects tax reconciliation. Historically, a person read each document and typed the data by hand into the ERP or accounting system. That work is slow, error-prone and becomes a bottleneck as volume grows. Intelligent document processing tackles exactly that problem: it automates reading, interpretation and data entry, leaving people only the exceptions that truly require human judgment.

How it works, step by step

A typical IDP workflow goes through five linked stages:

1. Capture and ingestion: the document enters the system from an email, a scanner, a shared folder, an API or a photo taken with a phone. At this stage the image is normalized (straightened, contrast-enhanced, pages separated).
2. Classification: the system determines what type of document it is (a type A invoice, which in Argentina is issued to VAT-registered buyers; a delivery note; a credit note; a purchase order) using document classification. This defines which fields to look for.
3. Extraction: the relevant fields (supplier, tax ID, number, date, amount, VAT rate, line items) are identified and read through OCR on the text and data extraction models that understand semantics, not just position.
4. Validation: the extracted data is checked against business rules and external sources (does the tax ID exist? does the total match the sum of the line items? is the purchase order approved?). This is the document validation stage, which often includes a three-way match between the invoice, the purchase order and the goods receipt.
5. Integration and exceptions: documents that pass every validation are loaded into the target system on their own; those that fail are routed to a person for review, an approach known as human-in-the-loop.

The leap that generative AI brought

For a long time, IDP depended on models trained specifically for each document type and field. With the arrival of generative AI and large language models (LLMs), a new capability appeared: you can ask the model "extract the total amount and the due date from this invoice" in natural language, without training anything by hand, and it works even with documents it has never seen before. That drastically lowered the cost of getting an IDP project off the ground. The flip side is that these models can hallucinate (make up a piece of data that looks plausible), so the validation layer and human oversight remain essential for financial documents, where a one-digit error has real consequences.

A concrete example in Argentina

Think of a distributor in Buenos Aires that receives 1,500 supplier invoices a month, each in a different format. Without IDP, two people in accounts payable spend the whole day entering documents and reconciling. With intelligent document processing, the system reads every invoice that arrives by email and recognizes the tax ID (CUIT), the point-of-sale code, the invoice number, the net amount, the itemized VAT and the line items; it validates that the CUIT appears in the tax authority's registry, that the total adds up and that an approved purchase order exists; and it automatically loads the ones that meet every check. People go from typing to reviewing only the 10% that has some inconsistency, and a month-end close that used to take days speeds up significantly.

Common mistakes when implementing IDP

• Expecting 100% automation from day one. There is always a percentage of exceptions; the realistic goal is to automate most documents and handle the rest well.
• Skipping validation. Extracting data without checking it against business rules just replicates errors faster. Validation is what makes IDP reliable.
• Confusing IDP with RPA. RPA automates clicks and repetitive tasks on screens; IDP understands the content of a document. They are often used together, but they are not the same thing.
• Not measuring. Without metrics for the correct extraction rate and the percentage of exceptions, you can't tell whether the solution is getting better or worse.

IDP vs. traditional OCR

The most frequent confusion is assuming that IDP and OCR are synonyms. OCR is one piece within IDP, not its equivalent. This table sums up the differences:

In practice, IDP uses OCR for the initial reading and adds the layers of intelligence on top (classification, semantic extraction, validation) that turn a pile of text into actionable information. That is why, when a company says it wants to "digitize documents," what it almost always needs is IDP, not just a scanner with OCR.