Invoice processing automation

Finance teams manually retype information from incoming invoices into spreadsheets and accounting systems.

An automated pipeline that reads invoice PDFs, extracts the key fields, checks whether the extraction is confident enough, routes uncertain invoices to a person for review, and writes approved records into a spreadsheet.

Demonstrates a production-shaped document workflow with confidence gating and human review. Built and tested against sample invoice PDFs, not live production invoices.

• Workflow design
• Architecture
• Prompt engineering
• Evaluation & testing

1. 01A new invoice PDF arrives in a watched folder and enters the queue.
2. 02Document extraction pulls the key fields: vendor, invoice number, dates, line items, totals.
3. 03Each extracted field carries a confidence score. Anything below the configured threshold stops the automatic path.
4. 04Confident invoices are written to the destination spreadsheet as structured rows. Uncertain ones are routed to a person with the extracted values shown next to the source document.
5. 05Line item descriptions are generated where the source document is terse, so the spreadsheet is readable to someone who never sees the PDF.

Orchestrated in UiPath with Document Understanding handling classification and field extraction.

Confidence gating sits between extraction and output rather than after it, so a low-confidence value never reaches the spreadsheet unreviewed.

A language model step generates line item descriptions from extracted fields; it does not produce numeric values.

Results are written to Google Sheets with a fixed column mapping, and every run logs what it processed and what it deferred.

• Tested against sample invoice PDFs, not a live accounts payable inbox. Real invoice variety is wider and messier.
• Nothing writes to an accounting system. The output is a spreadsheet, which is one step short of production.
• The confidence threshold is a configured number, not a value tuned against a labeled dataset of real invoices.
• Scanned or photographed invoices with poor image quality degrade extraction, and the workflow defers rather than recovers.