Document classification, automatic categorization, intelligent sorting... all refer to a key capability in a professional world overwhelmed by documents. Whether it's payslips, supplier invoices, contracts, or identity documents, the need to efficiently organize and sort information has become critical across many industries.

Banking, insurance, healthcare, logistics, human resources, and even the public sector all face a massive influx of heterogeneous and often sensitive documents that must be processed quickly and accurately.But manual document handling quickly reaches its limits: it’s slow, error-prone, and labor-intensive.

This is where Intelligent Document Processing (IDP) technologies come into play.Artificial intelligence, machine learning, and optical character recognition (OCR) now make it possible to automate the analysis and classification of large volumes of documents.

What is Document Classification?

It refers to the ability to automatically recognize the type of a document (e.g., “contract,” “passport,” “invoice,” etc.) without human intervention, in order to route it to the appropriate workflow or database.

This step is crucial in any document automation process, as it directly impacts the subsequent phases such as data extraction, validation, and archiving.

In this article, we illustrate this capability through a real-world use case: the automatic classification of multilingual identity documents (national ID cards, passports, driver’s licenses, residence permits) using the Koncile solution.

Document Classification for Identity Documents

Koncile is an intelligent OCR solution specialized in extracting accurate data from complex documents such as contracts, payslips, financial statements, or transport documents.

‍Our goal is simple: to provide a fast, reliable system with no manual configuration required.

In this example, we tested our automatic classification engine on a set of identity documents (national ID cards, passports, driver’s licenses, residence permits) from multiple countries.

Test Protocol

We used a set of 18 documents (driver’s licenses, residence permits, national ID cards, passports), including:

• French identity documents
• British documents
• And Italian documents, which were deliberately not targeted in our test in order to verify automatic exclusion.

Test Dataset Composition

Test 1: Classification by Document Type

Objective: Evaluate the performance of our system without any contextual guidance.
Here are the steps of this first test:

1. Create a folder for identity documents

‍

2- Add the available extraction templates

In this first test, we do not provide any description of the type of information to retrieve from this folder.We simply add the document templates to be extracted, using the existing extraction models already created in the application.

3- Import the different types of documents without classification

In this phase, we performed a raw import of all 18 identity documents, without applying any sorting rules or specific instructions.
The goal is to observe the default behavior of Koncile’s auto-classification engine when faced with a collection of diverse document types.

We aim to assess whether the engine is capable of:

• Identifying relevant documents
• Determining their type (ID card, passport, driver’s license...)
• Automatically classifying them without human input

Test 1 Results

Despite the absence of explicit instructions, our engine was able to identify the document types (ID card, passport, driver’s license, etc.) and apply automatic classification based on their visual and textual features.

Each document was accurately categorized, despite the diversity of languages used.

Soit un taux de réussite de 100%.

Test 2: Classification by Country

Objective: Test the automatic differentiation between French and British identity documents, also without any contextual indication.

1- Creation of Two Identity Document Folders
For this new test, we are increasing the level of complexity by creating two separate folders: one containing French identity documents and the other containing documents exclusively from the United Kingdom. As in the previous test, classification will rely solely on the folder name and model name, with no additional description provided.

2- Adding the Available Extraction Models
As in the previous step, we add the various extraction models to both folders.

3. Import of various document types without predefined classification

Test 2 Results

With 15 documents correctly classified out of 18, the engine achieved a success rate of 83.33% in this scenario.
Driver’s licenses and residence permits were perfectly recognized (4/4 and 3/3 respectively).
The observed errors involved two national ID cards (one British and one Italian) that were incorrectly placed in the French folder, as well as one Italian passport, also mistakenly classified as French.

These errors are understandable, particularly in the case of the Italian documents (which account for 2 out of the 3 misclassifications), since no dedicated category for Italy had been created at this stage, and no explicit instruction was given to exclude foreign documents.
The engine therefore classified these items based on linguistic or visual similarity, in the absence of precise guidance.

Test 3: Adding a Description to the Folder

Objective: Observe the impact of a descriptive prompt on classification accuracy.

Prompt for the folder "Foreign Identity Documents: "Extract only identity documents issued by official foreign authorities, excluding any documents issued by France. The language used (French, English, or others) does not matter, all documents originating from countries other than France must be included: passports, driver's licenses, identity cards, etc. French documents (i.e., issued by French institutions) must always be ignored, even if written in English or any other language."

Prompt for the folder "French Identity Documents: "Extract only identity documents issued by official French authorities. All documents must have been issued by French institutions: passports, identity cards, driver’s licenses, residence permits, etc. Documents issued by foreign countries must always be ignored, even if they are written in French."

Test 3 Results

This third test of automatic identity document classification shows a clear improvement compared to the previous one.

During the second test, the success rate was 83.33%, with several misclassifications for national ID cards.

In this new iteration, all 18 documents were correctly classified, resulting in a 100% success rate.

This significant improvement can be attributed mainly to the addition of a contextualized prompt, which enabled the tool to interpret the documents more accurately.
By providing a clear framework and explicit instructions, the model’s performance improved considerably, especially on cases that had previously been prone to errors.

Solutions to Improve Reliability

In cases that are even more complex than the one presented here, such as documents lacking explicit labels, with few distinctive visual features, or highly technical content — several optimization strategies can be explored to further enhance classification accuracy.

• Create additional categories/folders to improve classification granularity: By adding more specific folders, the engine can better group documents by origin, reducing misclassifications caused by visual or linguistic similarities with other European documents, for example.
• Use document-specific prompts: Tailored descriptions can be defined for each document type (e.g., ID card, passport, driver's license), enabling the system to apply more precise criteria such as issuing authority, language, or format.
• Leverage the visual structure of the document: Considering the layout of key elements (photo, fields, signatures) can help distinguish between visually similar documents.
• Implement a fallback or human verification mechanism: In low-confidence cases, routing documents for manual review helps ensure quality while also refining the models through human feedback.

Automatic document classification is no longer just a technical feature, it is a strategic asset for any organization dealing with large volumes of heterogeneous documents.

As demonstrated by our tests, our solution delivers high accuracy even without prior configuration, thanks to its advanced visual, textual, and contextual analysis capabilities.

By integrating descriptive prompts or refining sorting categories, even higher performance levels can be achieved, making document management smoother, more reliable, and significantly less dependent on human intervention.