AI Document Processing / Rebate Statement Automation
I designed and built an end-to-end document processing workflow for a distributor that receives quarterly rebate statements from multiple commercial partners.
Details
Industry:
Overview
Walkthrough Video:
https://www.boomshare.ai/shared/01KX3V82DXR9TWGMJYJ3A6CS7M
I designed and built an end-to-end document processing workflow for a fictional distributor that receives quarterly rebate statements from multiple commercial partners.
Each partner used a different document layout. Some statements contained a single table, others included both current-period rebates and prior-period adjustments, while some arrived as scanned PDFs or PNG images. The original process required Finance to manually read every document and enter each product line into a master spreadsheet.
The automation processes every statement from a Google Drive input folder, identifies the document format, extracts the relevant fields and normalizes them into one consistent data structure.
AI is used only for document understanding and structured extraction. Financial approval is handled separately through deterministic spreadsheet formulas. If even one row or section fails validation, the entire statement is blocked and moved to a Review Queue rather than being partially loaded.
Tools Used / Stack
n8n — workflow orchestration and document routing
Mistral OCR — extraction of structured Markdown from PDFs and scanned images
OpenAI Information Extractor — normalization into partner-specific JSON Schemas
Google Drive — input, approved and review folders
Google Sheets — staging, validation, final tracker and audit logs
JSON Schema — consistent field names, data types and allowed values
HTML Email — end-of-run notification
The workflow was built without custom Code nodes, making its logic easier to inspect and more portable to environments such as Power Automate, SharePoint and Excel Online.
Key Features
Multi-layout document processing
The workflow supports three different partner formats:
Atlas: single-table statements containing current rebate rows
Borealis: statements containing both current-period rebates and prior-period adjustments
Cascade: visually different statements, including scanned files and negative values represented with parentheses
Each layout is extracted through a dedicated JSON Schema but normalized into the same final structure.
Structured AI extraction
The Information Extractor converts OCR output into consistent fields such as:
Statement ID
Partner program
Region
Claim period
Record type
Product code and name
Unit rebate rate
Units and orders
Rebate amount
Printed section totals
Enums and field descriptions constrain the extractor. For example, Atlas can only return current, while Borealis can return either current or adjustment.
Deterministic financial validation
The workflow does not ask AI whether a statement “looks correct.”
Every extracted row is written to a staging area and validated using spreadsheet formulas:
Units × Rebate Rateis recalculatedThe calculated amount is compared with the printed row amount
Extracted row amounts are summed
Section totals are reconciled against the totals printed in the statement
Borealis current and adjustment sections are validated separately
If any row or section exceeds the configured tolerance, the entire statement is flagged.
Statement-level approval
Statements are processed conservatively:
Approved: all rows are loaded into the Rebate Tracker
Flagged: no rows are loaded, and the original file is moved to the Review Queue
Unknown layout: the workflow does not guess the mapping and sends the file for manual review
Duplicate: previously processed Statement IDs are skipped
Auditability
The solution includes:
Rebate Tracker for approved rows
Staging Rows for temporary extracted data
Validation for calculations, deltas and final status
Review Queue for flagged statements and reasons
Run Log for processing history
Processed Registry for duplicate prevention
Approved and Review Queue folders for the original source documents
Completion email sent after the entire batch has finished processing
Outcome
The final workflow was tested on a batch of 11 rebate statements containing clean documents, scanned files, adjustments, mathematical inconsistencies and an unsupported partner layout.
The run produced:
11 statements processed
8 statements approved
3 statements flagged
46 product rows loaded into the Rebate Tracker
0 known invalid statements loaded
The workflow correctly identified:
a statement whose printed total did not reconcile with its product rows;
a scanned statement containing an incorrect row-level rebate amount;
an unsupported partner format that required manual review.
The project demonstrates how AI-assisted document extraction can be combined with deterministic controls to create a finance workflow that is efficient, explainable and auditable.
