Procurement Reconciliation Agent

Agent Description:

The Procurement Reconciliation Agent is an AI-powered automation system designed to validate and reconcile purchasing transactions by cross-referencing data from purchase orders, invoices, and goods receipts. It performs structured extraction from these three document sources and flags any mismatches in quantity, product ID, or amounts for human review. This enables finance and procurement teams to streamline payment approvals while reducing errors and fraud.

Purpose and Components
  • Purpose: To automate the procurement reconciliation process between invoices, purchase orders, and goods receipts, flagging any inconsistencies for human review.
  • Components:
    • An invoice extractor to retrieve invoice records.
    • A purchase order extractor to pull structured PO data.
    • A goods receipt extractor to validate what was actually received.
    • A downstream matching engine (optional) to compare the datasets and surface discrepancies.
Supported Capabilities
  • Extracting structured data from invoice, PO, and receipt documents (PDFs).
  • Preserving relational links (for example, PO_ID, ProductID).
  • Structuring outputs for downstream reconciliation or analytics.
  • Flagging mismatches for manual validation (if extended with a comparison engine).
  • Enabling accurate payment processing by validating billing, ordering, and delivery alignment.
LLM Used
  • Google Vertex
    Note: To learn more about the LLM and to modify its behavior, refer to the Configuring LLM settings section.

Sub-Agents

1. Invoice Extractor Agent

  • Role:Data Extractor.
  • Scope:Extracts relevant invoice fields from PDF documents for reconciliation.
  • Description: This sub-agent processes invoice documents to retrieve:

    • InvoiceID, PO_ID, ProductID, QuantityBilled, TotalAmount, and InvoiceDate.

    • It maintains structural relationships (for example, which invoice belongs to which PO) and outputs data in clean JSON or tabular format.

    • No matching or computation is performed — only structured extraction.

  • LLM Used: Google Vertex (inherits from parent).
  • Tool Used: PDF Reader Tool

2. Purchase Order Extractor Agent

  • Role:Data Extractor.
  • Scope:Parses purchase order PDFs and organizes relevant information for each linked invoice.
  • Description: This sub-agent extracts from PO documents the fields:

    • InvoiceID, PO_ID (foreign key), ProductID, QuantityBilled, TotalAmount, and InvoiceDate.

    • Ensures all data maintains original field structure and data types.

    • Outputs in a structured, machine-readable JSON or tabular format.

  • LLM Used: Google Vertex (inherits from parent).
  • Tool Used: PDF Reader Tool

3. Goods Receipt Extractor Agent

  • Role:Data Extractor.
  • Scope:Pulls delivery confirmation data from goods receipt PDFs.
  • Description: This sub-agent extracts the following from the GoodsReceipts table:

    • GR_ID, PO_ID, ProductID, QuantityReceived, and ReceiptDate.

    • It ensures relational integrity between PO_ID and receipt data.

    • Output is structured and clean, without performing any matching.

  • LLM Used: Google Vertex (inherits from parent).
  • Tool Used: PDF Reader Tool
Tools Used:
  • PDF Reader Tool: Used by all three sub-agents to extract structured data from invoice, purchase order, and goods receipt PDF documents.
Note: For details on modifying the Tools, refer Tools Library section.
Agent Workflow Behavior Summary
  • The Invoice Extractor Agent processes PDF invoices and extracts structured fields such as InvoiceID, PO_ID, and QuantityBilled.
  • The Purchase Order Extractor Agent retrieves the corresponding PO information, maintaining links between purchase orders and invoiced items.
  • The Goods Receipt Extractor Agent extracts delivery data to verify actual quantities received against what was ordered and billed.
  • Each agent outputs clean JSON or tabular data for downstream use in reconciliation or exception reporting.
  • (Optional extension) A reconciliation engine can compare all three datasets to detect discrepancies in quantity, amount, or missing references, then flag records for human validation.
Sample Questions:
  • Are there any mismatches between purchase orders, goods receipts, and invoices?
  • Which transactions have quantity mismatches?
  • List all POs where the invoice total doesn’t match the expected amount.