The three technology layers of IDR: OCR, contextual understanding and validation with cross-checking
How does Scan & Recognise (IDR/OCR) work?

The technology behind automatic PDF recognition: OCR, AI and the self-learning system for invoices and purchase orders.

The Intelligent Document Recogniser (IDR) is eConnect's proprietary technology that converts PDF documents, images and receipts into validated electronic documents. This is the engine behind Scan & Recognise. The system processes 100% of all submitted documents and continuously improves through use.

eConnect positions this processing as Recognition as a Service (RAAS): not a customer-operated scan-and-recognise tool with a reject bin, but a service in which eConnect performs the recognition, validation and quality control. The customer receives a structured document for further processing in their own ERP or accounting package.

In addition to invoices, the IDR can also recognise and convert purchase orders. This functionality uses the same three technology layers as invoice recognition. Order recognition is currently available as a beta and is actively being developed.

Three layers, one result. The IDR combines OCR text recognition, LLM interpretation and proprietary document interpretation into an accurate, self-learning system that correctly recognises more than 98% of all core fields.

Three technology layers

The IDR combines three layers that work together to recognise and process documents:

OCR layer: The first step is pure text recognition. This layer reads the text on the document, regardless of language (over 200 languages are supported) or format.

LLM layer: A Large Language Model helps interpret the recognised text. The model understands the context of the information and assists in assigning text to the correct invoice elements.

Document interpretation: This is the core of the IDR and entirely eConnect's own technology. This layer determines what the document means, which fields are relevant, how the data should be structured and whether everything is correct. The intelligence embedded here, including the self-learning system, validation rules and feedback loop, is eConnect's differentiating capability.

The OCR and LLM are commodity services. The real value lies in the interpretation and validation layer that eConnect builds and maintains.

The diagram below shows how a PDF invoice is processed through the three layers of the IDR into a validated e-invoice.

What is recognised?

The IDR automatically recognises all relevant invoice elements:

  • Supplier and buyer (name, address, company registration number, VAT number)
  • Invoice number and invoice date
  • Amounts (subtotal, VAT, total)
  • Payment details (IBAN, payment reference)
  • Order numbers and references (with Professional subscription)
  • Invoice lines with description, quantity, unit price and line amount (line recognition)
Line recognition

Since July 2024, the IDR also offers line recognition: splitting PDF invoice lines into individual transaction lines. Per line, description, price, quantity, line amount and reference fields are recognised.

By default, 80% of invoice lines are recognised automatically. The remaining 20% falls back to standard processing, using the subtotal per VAT rate. The "No Lines No Pay" guarantee applies: you only pay for lines that are actually successfully recognised.

Optionally, you can activate line validation. The eConnect validation team then reviews the lines that were not automatically recognised, giving you a 100% guarantee on invoice recognition at line level.

The self-learning system

The IDR improves as more invoices are processed. This works through a feedback loop:

  • With every processed invoice, the system compares the recognised values with previously stored patterns from the same supplier.
  • If the IDR is not confident enough, the invoice goes to the Quality Control team for manual verification.
  • Every correction by the QC team is fed back to the system as a "hint", training the AI for future recognition.
  • Customer feedback is also processed: reported errors are analysed, incorporated into test procedures and structurally prevented.

The result is a system that continuously improves. An error that occurs once doesn't persist; it is resolved and the system learns from it.

IBAN blacklist and phantom invoice detection

During the validation phase, IDR compares the IBAN field on the invoice against a blacklist of known fraudulent IBANs. If there is a match, the invoice is not automatically processed but flagged for manual review. The goal is phantom invoice detection: invoices that appear correct at first glance but are intended to redirect payments to a fraudulent account number.

Status May 2026: the exact current implementation — which source feeds the blacklist, how flag resolution works and in which subscription tier this validation is active by default — is currently under review. For specific deployment at a client: check with sales or support.

Accuracy and processing guarantee

The IDR scores high on accuracy:

  • On core fields (invoice number, date, amounts): > 98% accurate, supported by periodic audits
  • SLA guarantee: 99% of documents processed within 8 working hours

The architecture prevents structural errors by validating each recognised value against another source (document-internal or external). When in doubt, the document is routed to the QC team, providing an additional safety layer.

Processing times

The pure processing time for automatically processed documents averages 20 to 30 seconds. The total turnaround time including queue wait is approximately 3.5 minutes on average. With the priority option (available with higher subscriptions), documents are given precedence.

Structured payment references

The IDR also recognises country-specific payment references:

  • Belgian OGM (structured communication): the format +++XXX/XXXX/XXXXX+++ with built-in checksum validation
  • Norwegian KID number: structured payment reference for Norwegian invoices
  • Swiss QR code: payment details from the QR code on Swiss invoices

These recognition features are part of the Professional version of the IDR.

Frequently asked questions
How accurate is the IDR's automatic recognition?

On core fields (invoice number, date and amounts), the IDR scores above 98% accuracy, supported by periodic audits. This percentage applies to automatic processing; 100% of submitted documents are processed because uncertain cases are handled by the Quality Control team. Both figures together — 100% processed and more than 98% correct — form the IDR guarantee. When in doubt, the document is forwarded to the Quality Control team for manual verification.

What if an invoice cannot be automatically recognised?

If the IDR is not confident enough about the recognition, the conversion task goes to eConnect's Quality Control team. They review the document manually and correct any errors. Every correction is fed back to the system as training data, so that similar invoices are automatically processed in the future.

Does recognition improve as more invoices are processed?

Yes, the IDR is a self-learning system. With every processed invoice, patterns are built up per supplier. Corrections by the QC team and customer feedback are fed back to the system as hints. The more invoices from a supplier that have been processed, the better and faster the recognition becomes.


Want to experience how the IDR works? Submit a test invoice and view the result in your Inbox.

Submit your first invoice