Problem

Many business workflows still begin with unstructured documents: invoices, receipts, statements, forms, and messages. Extracting values is only part of the problem. The harder question is whether a person can review where each value came from and understand what was not found.

Pipeline

The Phase 4 demo focuses on the path from document text to structured, reviewable fields.

Business document -> Extracted text -> Document type -> Structured fields -> Evidence and missing fields

What the examples use

The invoice visual below is synthetic and uses the same public-safe values as the extracted fields shown in the demo. The current backend does not yet perform OCR on scanned images, so this page shows a realistic static visual plus the matching Phase 4 transcript-derived output.

Synthetic invoice visual matching the extracted fields shown on this page
Synthetic public-safe invoice visual. The values match the extracted fields below; this is not a claim that the backend currently performs OCR on scanned images.

Demo 1: Vendor invoice extraction

The invoice demo shows how synthetic vendor invoice text becomes structured fields such as invoice number, vendor name, invoice date, due date, totals, and payment terms.

The point is not that the fixture is complex. The point is that the output is structured enough for a downstream workflow to inspect.

Input transcript invoice.v1 
INVOICE
Invoice No: INV-DEMO-2026-0042
Invoice Date: 15 Jan 2026
Due Date: 14 Feb 2026
From: Northstar Office Supplies LLC
Bill To: Blue Harbor Consulting LLC
Currency: USD
Subtotal: $1250.00
Tax Amount: $100.00
Total Amount: $1350.00
Invoice number INV-DEMO-2026-0042 high Invoice No: INV-DEMO-2026-0042
Vendor Northstar Office Supplies LLC high From: Northstar Office Supplies LLC
Due date 2026-02-14 high Due Date: 14 Feb 2026
Total $1350.00 high Total Amount: $1350.00
Missing fields vendorAddress buyerAddress purchaseOrderNumber lineItems

Demo 2: Expense receipt extraction

The receipt demo applies the same Phase 4 idea to a different business document shape. It extracts merchant, date, category, subtotal, tax, tip, total, currency, and payment method from a synthetic receipt.

This shows that the workflow is not tied to one document template.

Synthetic receipt visual matching the extracted receipt fields shown on this page
Synthetic public-safe receipt visual. The values match the extracted receipt fields below.
Input transcript receipt.v1 
RECEIPT
Merchant: Harbor Cafe Supplies
Receipt No: RCP-DEMO-2026-019
Transaction Date: 22 Jan 2026
Currency: USD
Subtotal: $42.50
Tax Amount: $3.40
Total Amount: $45.90
Payment Method: card
Merchant Harbor Cafe Supplies high Merchant: Harbor Cafe Supplies
Receipt number RCP-DEMO-2026-019 high Receipt No: RCP-DEMO-2026-019
Total $45.90 high Total Amount: $45.90
Payment method card high Payment Method: card
Missing fields transactionTime merchantAddress items

Demo 3: Evidence and missing fields

The evidence viewer shows why structured extraction needs review context. Fields are displayed with values, confidence levels, and supporting snippets. Missing fields are shown explicitly instead of being silently ignored.

That makes the output easier to inspect, explain, and route into the next workflow step.

Synthetic evidence and missing-fields review visual matching the review values shown on this page
Synthetic public-safe review visual. The reviewed value, evidence, confidence, and missing fields match the review panel below.
Reviewed field Total amount: $1350.00 high confidence

Evidence: Total Amount: $1350.00

Needs review Missing fields
  • vendorAddress
  • buyerAddress
  • purchaseOrderNumber
  • lineItems
Invalid or ambiguous No invalid or ambiguous fields in this clean fixture.

The display model still reserves these buckets so future workflows do not hide uncertainty.

What this proves

Phase 4 proves that AI Workflow Lab can produce reviewable structured data from everyday business documents.

  • Document-type-aware field extraction.
  • Public-safe fixture outputs.
  • Confidence and evidence display.
  • Explicit missing-field reporting.
  • A clear boundary between extraction and workflow decisions.

What comes next in Phase 5

Phase 5 should focus on what happens after extraction: human review, validation, exception handling, routing by confidence or missing fields, and audit-friendly workflow state.

Phase 4 does not decide whether a document should be approved, rejected, reimbursed, paid, or posted to accounting. It prepares data that can support those later workflow decisions.

Phase 5 focus
  • Human review screens
  • Validation and exception handling
  • Business-rule routing
  • Workflow state and audit trail

Limitations

  • The current demos use synthetic public-safe fixtures.
  • The public invoice image is a visual reference; extracted fields shown here come from matching synthetic transcripts.
  • The current backend does not perform OCR on scanned image input.
  • The examples are static and do not call production APIs.
  • The extraction outputs are demo artifacts, not a production review queue.
  • The demos do not approve payments, reimbursements, or accounting entries.
  • The demos do not make LLM-based decisions.

Source artifacts

The source demo package is maintained in the platform repository so the website copy, fixture references, QA notes, and claims review remain traceable to the backend platform work.