CyberFreezeDevAccounts payable
Purchase invoices for input tax credit, matched to orders and vendor records.
Read more about Accounts payable
Purchase invoices for GST input tax credit, matched to purchase orders and vendor master data.
- Home
- AI glossary
- Intelligent document processing (IDP)
What is intelligent document processing (IDP)?
Documents in, checked data out: classify, extract and validate, with a person for doubtful cases.
The plain definition
IDP takes documents in any form, identifies them, extracts and checks fields, sends doubtful ones to a person and writes results to your systems.
Read more about The plain definition
Intelligent document processing (IDP) is a pipeline, not a single tool. It takes documents in whatever form they arrive (email attachments, scans, phone photos, PDFs), works out what each one is, pulls out the fields you need, checks them against rules and your own records, sends doubtful ones to a person and writes the results into your systems.
Modern IDP uses vision models and large language models for extraction, so it copes with layouts it has never seen, instead of needing a template per vendor.
How it works
-
01 Capture
Collect files from email, WhatsApp, a shared folder or an upload form.
Read more about Capture
Collect files from email, WhatsApp, a shared folder or an upload form.
-
02 Classify
Decide what each file is, from purchase invoice to KYC document.
Read more about Classify
Decide the type: purchase invoice, credit note, bank statement, KYC document, delivery note.
-
03 Extract
Pull fields like vendor, GSTIN, invoice number, line items and tax amounts.
Read more about Extract
OCR the text if needed, then pull fields such as vendor name, GSTIN, invoice number, date, line items, taxable value, CGST, SGST and IGST.
-
04 Validate
Check totals, GSTIN format and dates, and catch invoices already entered.
Read more about Validate
Check that totals add up, GSTIN format is valid, dates are sensible, and the invoice is not already entered.
-
05 Review and export
Doubtful items go to a reviewer; approved data goes to Tally, an ERP or a sheet.
Read more about Review and export
Low-confidence or failed items go to a reviewer; approved data goes to Tally, an ERP or a sheet.
Where it matters in business
Onboarding and KYC
Insurance, lending and property forms with IDs and proofs attached.
Read more about Onboarding and KYC
Insurance, lending and real estate forms with IDs and proofs attached.
Operations paperwork
Delivery notes, lab reports, claim forms and contracts that are retyped today.
Read more about Operations paperwork
Delivery notes, lab reports, claim forms and contracts that today get retyped.
A practical example
Our OCR pipeline for invoices, IDs and forms routes, extracts and checks each upload, and shows doubtful values to a reviewer before anything moves on.
Read more about A practical example
One of our anonymous engagements was an OCR pipeline for invoices, IDs and forms with human review built in. The pattern is the useful part: a classifier routes each upload; an extractor fills a fixed schema; rule checks catch arithmetic and format errors; a reviewer screen shows the document beside the extracted fields, with doubtful values highlighted; nothing reaches the downstream system until it passes or a person approves it. See OCR and document intelligence.
Common mistakes
No validation layer
Extraction is sometimes wrong. Arithmetic, format and duplicate checks catch most of it.
Read more about No validation layer
Extraction alone will sometimes be wrong. Arithmetic, format and duplicate checks catch most of it cheaply.
Skipping the reviewer screen
Human review makes the system trustworthy from day one and turns corrections into test data.
Read more about Skipping the reviewer screen
Human-in-the-loop review is what makes the system trustworthy from day one, and the corrections become test data.
Measuring the wrong thing
Track field accuracy and documents passing with no edits, not documents processed.
Read more about Measuring the wrong thing
Track field-level accuracy and documents that pass with no edits, not just "documents processed". See model evaluation.
Questions
How is IDP different from OCR?
OCR converts an image to text. IDP adds classification, field extraction, validation, review and export, so the output is structured, checked data rather than a block of text.
How is IDP different from RPA?
Robotic process automation repeats fixed clicks and keystrokes in software. IDP handles the reading and understanding of documents. They are often combined: IDP reads the invoice, RPA or an API enters it.
Does IDP need a template for each vendor?
Older systems did. AI-based extraction generally does not, though giving the model a few examples of difficult vendors still improves results.
Last updated