RPA Intelligent Document Processing

Intelligent Document Processing (IDP) is a technology that reads, extracts, classifies, and validates data from documents — including invoices, contracts, identity cards, medical forms, and bank statements — automatically. IDP combines OCR, AI, and machine learning to handle documents that vary in format, layout, and quality, which basic RPA text extraction cannot reliably manage alone.

The Problem IDP Solves

A company receives invoices from 500 different vendors. Each vendor has a unique invoice layout. One puts the total at the top right. Another puts it at the bottom left. A third writes it in a table inside the document. Basic RPA looks for a fixed position or keyword — it breaks the moment the layout changes. IDP learns what an "invoice total" looks like across all formats and finds it reliably regardless of position.

Diagram: IDP vs Basic Text Extraction

 BASIC RPA TEXT EXTRACTION:
 ─────────────────────────────────────────────────────────
 Rule: "Total amount is always in cell B12 of the PDF"
 Works for: Invoices from Vendor A only
 Fails for: Any other vendor with a different layout

 INTELLIGENT DOCUMENT PROCESSING:
 ─────────────────────────────────────────────────────────
 AI model trained on thousands of invoice examples
 Understands: "Total Amount" = the final payable figure,
              regardless of where it appears on the page
 Works for: Invoices from any vendor, any layout

IDP Pipeline: The Five Stages

Stage 1: Digitise

Convert the document into a format the AI can analyse. For native PDFs, the text is already extractable. For scanned documents or image-based PDFs, OCR runs first to convert the image into text while preserving the document's spatial structure (where each word appears on the page).

Stage 2: Classify

Determine what type of document this is. A single submission might contain a cover letter, an invoice, a purchase order, and a delivery note — all in one PDF. The classifier identifies each document type so the correct extraction model is applied to each section.

 Multi-document PDF received:
 Page 1 → Classified as: Cover Letter (skip)
 Page 2 → Classified as: Invoice (extract financial fields)
 Page 3 → Classified as: Purchase Order (extract PO fields)
 Page 4 → Classified as: Delivery Note (extract shipment fields)

Stage 3: Extract

The extraction model reads the classified document and pulls out specific data fields. Extraction models are trained on examples of the document type and learn which part of the page corresponds to which field.

 INVOICE EXTRACTION RESULTS:
 ┌──────────────────────────────────────────────────┐
 │ Field            │ Extracted Value │ Confidence  │
 ├──────────────────┼─────────────────┼─────────────┤
 │ Vendor Name      │ Acme Ltd        │ 99%         │
 │ Invoice Number   │ INV-2024-0451   │ 98%         │
 │ Invoice Date     │ 15/03/2024      │ 97%         │
 │ Due Date         │ 14/04/2024      │ 95%         │
 │ Line Item 1      │ Office Chairs   │ 94%         │
 │ Line Item 1 Qty  │ 10              │ 99%         │
 │ Line Item 1 Price│ $150.00         │ 98%         │
 │ Tax Amount       │ $270.00         │ 97%         │
 │ Total Amount     │ $1,770.00       │ 99%         │
 └──────────────────────────────────────────────────┘

Stage 4: Validate

Validation checks extracted values against business rules and cross-references data within the document and against external sources.

• Does the sum of line items equal the stated total?
• Is the vendor in the approved vendor master list?
• Is the invoice number a duplicate of one already processed?
• Is the currency valid for this vendor's country?
• Are any extracted fields below the confidence threshold?

Fields with low confidence scores or failed validation go to a human for review rather than being processed automatically. This is the human-in-the-loop checkpoint that prevents costly errors.

Stage 5: Export

Once validated, the structured data is exported to the destination system — the ERP, a database, a spreadsheet, or a downstream RPA workflow that uses the data to complete a business transaction.

Confidence Scores and Human-in-the-Loop

Every field extracted by an IDP model comes with a confidence score — a percentage indicating how certain the model is about the extracted value. Your workflow defines a threshold: fields above it are accepted automatically; fields below it are sent to a human validator.

 CONFIDENCE THRESHOLDS:
 ─────────────────────────────────────────────────────────
 Confidence ≥ 95%   → Auto-accept, no human review needed
 Confidence 70–94%  → Flag for human confirmation
 Confidence < 70%   → Reject extraction, manual entry required

 Human Validation Station (Action Center in UiPath):
 Bot pauses the transaction and creates a task.
 A human opens the task, sees the document side by side
 with the extracted values, corrects any errors, and submits.
 Bot resumes processing with the verified data.

UiPath Document Understanding Framework

UiPath provides Document Understanding as a built-in framework for IDP. It includes:

• Taxonomy Manager: Define what fields you want to extract (field names and data types)
• Digitise Activity: Convert the document to processable format using OCR
• Classify Document Scope: Route different document types to the correct model
• Data Extraction Scope: Run the extraction model against the classified document
• Validation Station: Web interface where humans review and correct low-confidence extractions
• AI Center: Platform for training custom ML models on your specific document types

Pre-Built vs Custom Models

Real-World IDP Use Case: Mortgage Application Processing

 APPLICANT SUBMITS: 8 different documents
 ├── Loan application form
 ├── 3 months of bank statements
 ├── Payslip (last 3 months)
 ├── Identity document (passport)
 └── Property valuation report

 IDP PIPELINE:
 ─────────────────────────────────────────────────────────
 Digitise all 8 documents (OCR for scanned ones)
         │
         ▼
 Classify each page into its document type
         │
         ▼
 Extract fields from each document:
   Loan form:       Applicant name, requested amount, term
   Bank statements: Monthly income, average balance, debits
   Payslip:         Employer, net salary, deductions
   Passport:        Full name, DOB, nationality, expiry date
   Valuation:       Property address, assessed value
         │
         ▼
 Validate: Names match across all documents?
           Income sufficient for loan amount?
           Passport not expired?
           Property value ≥ loan amount?
         │
         ▼
 Low-confidence fields → Human review queue
 All high-confidence + validated → Auto-proceed
         │
         ▼
 RPA bot creates application record in loan management system
 with all extracted and validated data

IDP Metrics to Track

• Straight-Through Processing Rate (STP): Percentage of documents processed without any human intervention
• Extraction Accuracy: Percentage of fields extracted correctly vs ground truth
• Throughput: Number of documents processed per hour
• Human Review Rate: Percentage of documents requiring human validation
• Model Improvement Rate: How accuracy improves as the model learns from human corrections

Summary

Intelligent Document Processing transforms the manual extraction of data from documents into an automated, AI-driven pipeline. The five stages — Digitise, Classify, Extract, Validate, Export — handle documents of any format and layout. Confidence scores determine when human review is needed. Pre-built models handle common document types immediately. Custom models handle industry-specific forms after training. IDP combined with RPA enables the automation of document-heavy processes in finance, insurance, healthcare, and legal — processes that were previously impossible to automate with basic RPA alone.