Intelligent document processing (IDP) is software that reads documents such as invoices, claim forms, and identity papers, works out what each one is, extracts the fields a business needs, checks them against rules and other systems, and passes uncertain cases to a person. The output is structured, validated data ready for an ERP, a claims system, or a case file.
IDP is a pipeline, not a single model. Optical character recognition (OCR) turns pixels into text. IDP adds what OCR lacks: knowing that this page is an invoice and not a delivery note, that "Total" in the bottom-right corner is the amount due, and that the VAT number must match the supplier record before anything gets posted.
The shift in recent years is in the understanding step. Models trained on text and page layout together read forms better than text-only methods,1 and some newer models skip separate OCR entirely.2 Large multimodal models now handle layouts that would once have needed a template per supplier.
The IDP pipeline, step by step
A supplier invoice arrives as a PDF attached to an email. Each step below either adds structure or catches an error before it reaches the ledger.
OCR and layout
Converts the page to text with positions, including handwriting and tables. Cloud services such as Amazon Textract extract text, forms, and tables from scans,5 and open-source engines like Tesseract cover plain OCR.4
Classification
Decides what each document, or each page in a bundle, is: invoice, credit note, claim form, ID card, bank statement. Azure's documentation gives the example of a loan application package that mixes forms, pay slips, and bank statements.8
Extraction
Pulls named fields into a schema: supplier, date, line items, totals, policy number. Google describes Document AI as turning unstructured document data into structured fields suitable for a database.6
Validation and review
Checks totals, formats, and lookups against master data, then routes low-confidence results to a person. AWS built a service specifically for human review of low-confidence predictions from Textract.7
What extraction output looks like
The extraction step returns fields with a confidence score. Validation then decides what happens to each document.
{
"doc_type": "invoice",
"supplier_vat": {"value": "PL5260001234", "confidence": 0.99},
"invoice_no": {"value": "FV/2026/11/0412", "confidence": 0.98},
"issue_date": {"value": "2026-11-14", "confidence": 0.97},
"po_number": {"value": "PO-77310", "confidence": 0.62},
"net": {"value": 4120.00, "currency": "PLN"},
"vat": {"value": 947.60},
"gross": {"value": 5067.60},
"checks": {
"net_plus_vat_equals_gross": true,
"supplier_in_master_data": true,
"po_found_in_erp": false
},
"route": "human_review",
"reason": "PO number low confidence and not found in ERP"
}Illustrative example. The supplier, numbers, and confidence values are invented. Thresholds and routing rules are set per company.
IDP, OCR, RPA, and AI agents
These tools are often sold together, and they solve different parts of the same back-office process.
| Term | What it does | Input it handles | Where it stops |
|---|---|---|---|
| OCR | Turns an image of text into machine-readable text | Scans, photos, image PDFs | Does not know what any field means |
| IDP | Classifies, extracts, and validates fields from documents | Semi-structured and unstructured documents | Hands clean data to another system; does not run the wider process |
| RPA | Repeats scripted clicks and keystrokes across applications | Structured data in stable screens | Breaks on unfamiliar documents or changed screens |
| AI agent | Pursues a goal across systems, choosing its own steps | Documents, messages, and API data | Within its permissions; escalates exceptions to people |
In practice the combination is common: IDP reads the document, and an AI agent or an RPA bot acts on the result. An agent can also call an IDP service as one of its tools, for example when a customer uploads a damage photo and a repair invoice during a claim chat.
Back-office documents IDP handles
IDP pays off where documents arrive in volume, vary in layout, and feed a decision with clear rules.
Supplier invoices
A retail chain matches invoices to purchase orders and goods receipts, posts clean ones, and queues mismatches for accounts payable.
extract → three-way match → post or reviewClaims documents
An insurer classifies claim forms, repair estimates, and medical notes, extracts amounts and dates, and opens the claim with the file already indexed.
classify → extract → open claimKYC documents
A lender reads ID cards and proof-of-address documents, checks expiry dates and name matches, and flags inconsistencies for compliance review.
read ID → cross-check → approve or flagRefund and compensation claims
An airline reads boarding passes, receipts, and bank details attached to refund requests and pre-fills the case for an agent.
extract → validate booking → pre-fill caseContract and porting forms
An operator extracts customer and number details from signed forms and checks them against the order before activation.
extract → match order → activateSupplier onboarding
Registration certificates, bank letters, and tax forms are read and checked before a new supplier is created in the ERP.
classify → verify → create supplierWhere IDP fails, and the usual controls
Extraction accuracy on a test set is only the start. The failures that cost money come from confident errors and from documents nobody planned for.
Confident wrong values
ControlA misread digit in a bank account number can pass a format check. Validate against master data and require a second source for payment details.
Unseen document types
ControlA new supplier layout or a foreign ID card lowers accuracy without warning. Monitor confidence by document type and route unknown classes to a person.
Review queues that grow
ControlIf thresholds are too strict, people re-key everything. Tune thresholds per field, and track the share of documents that pass with no human touch.
Personal data in documents
ControlIDs, claims, and bank letters are full of personal data. Keep processing and storage in approved regions and set retention for both images and extracted fields.
Fraudulent documents
ControlExtraction does not prove a document is genuine. Add authenticity checks for high-risk flows such as KYC and large payments, with a human in the loop for exceptions.
Common questions
What is the difference between OCR and IDP?
OCR converts an image of text into text. IDP uses OCR as one step, then classifies the document, extracts named fields, and validates them against rules and other systems before passing the data on.
Is IDP the same as RPA?
No. RPA repeats scripted actions in applications and needs structured input. IDP turns unstructured documents into structured data. They are often combined, with IDP reading the document and RPA or an AI agent entering the result.
Do large language models replace IDP?
They change how the extraction step is built, because a multimodal model can read many layouts without templates. The rest of IDP still applies: classification, validation against business data, confidence routing, audit trails, and human review.
Which documents are best to start with?
Start with one high-volume type that has clear validation rules, such as supplier invoices matched to purchase orders. The rules give you an objective way to measure accuracy from the first week.
References
- Xu, Y. et al. LayoutLM: Pre-training of Text and Layout for Document Image Understanding. KDD 2020.
- Kim, G. et al. OCR-free Document Understanding Transformer. ECCV 2022.
- Jaume, G., Ekenel, H. K., and Thiran, J.-P. FUNSD: A Dataset for Form Understanding in Noisy Scanned Documents. ICDAR Workshops 2019.
- Tesseract OCR. Tesseract Open Source OCR Engine. GitHub repository.
- Amazon Web Services. What is Amazon Textract?. Amazon Textract Developer Guide.
- Google Cloud. Document AI overview. Google Cloud documentation.
- Amazon Web Services. Using Amazon Augmented AI for Human Review. Amazon SageMaker Developer Guide.
- Microsoft. What is Azure Document Intelligence?. Microsoft Learn.