Agent in a Box

Automated B2B Invoice Reconciliation & Dispute Agent

operations
1
1

Automated B2B Invoice Reconciliation & Dispute Agent

Problem Statement

Mid-market B2B companies, particularly those in logistics, SaaS, and manufacturing, face a "reconciliation debt" that scales linearly with volume. The core problem is the data mismatch between three disparate sources: the accounts receivable (AR) ledger, the bank remittance advice, and the customer’s purchase order (PO).

Currently, finance teams spend 15-20 hours per week manually matching partial payments to line items. When a discrepancy occurs—such as a short-pay due to a damaged shipment or a tax miscalculation—the process grinds to a halt. A human must manually email the customer, request a reason code, cross-reference it with internal shipping logs, and decide whether to issue a credit memo or demand full payment.

This manual intervention leads to high Days Sales Outstanding (DSO), increased "unapplied cash" on the balance sheet, and friction with high-value customers who receive automated "overdue" notices for invoices they believe they’ve already settled. For a startup scaling from 50 to 500 customers, this bottleneck prevents efficient cash flow forecasting and necessitates hiring expensive accounting clerks. This is similar to how an Autonomous Stockout & Inventory Recovery Agent prevents supply chain bottlenecks. The lack of a real-time feedback loop between the bank and the ERP means disputes often sit unresolved for 30+ days, decreasing the likelihood of recovery by 50%.

What the Agent Does/Doesn't Do

Does:

  • Monitors bank feeds and email inboxes for remittance advice.
  • Matches payments to invoices using fuzzy logic for entity names and amounts.
  • Categorizes discrepancies (short-pays, over-pays, missing tax).
  • Drafts and sends professional dispute inquiries to customer AP departments.
  • Updates ERP status to "In Dispute" to pause automated collection dunning.

Doesn't:

  • Authorize large-scale write-offs (over a set threshold, e.g., $500).
  • Move actual money between bank accounts.
  • Handle physical debt collection or legal threats.

Workflow

  1. Ingestion: Agent pulls daily bank transaction exports (CSV/API) and scans the billing@ inbox for PDF remittance advice.
    • Input: Bank API/PDF Remittance. Output: Structured Payment Data.
  2. Entity Matching: Agent uses LLM-based fuzzy matching to link the payer name to a Customer ID in the ERP (e.g., NetSuite/QuickBooks).
    • Input: Payer Name. Output: ERP Customer Record.
  3. Reconciliation Logic: Agent compares the payment amount against open invoices. If it's a 1:1 match, it flags as "Ready to Close." If it's a partial payment, it moves to step 4.
    • Input: Payment Amount + Open Invoices. Output: Match Status (Full/Partial/None).
  4. Discrepancy Analysis: For partial payments, the agent parses the remittance PDF to find "Reason Codes" or notes (e.g., "Damaged goods invoice #123").
    • Input: Remittance Text. Output: Dispute Reason Category.
  5. Internal Verification: Agent queries the warehouse/logistics tool (e.g., ShipStation) to see if a damage report exists for that invoice.
    • Input: Invoice ID. Output: Internal Evidence.
  6. Resolution/Escalation: If evidence matches the dispute, the agent drafts a credit memo request for the manager. If no evidence exists, it sends a clarification email to the customer. Just as an Automated SaaS Churn Intervention Agent manages customer health, this agent protects the financial relationship.

Tool Stack

  • LLM Orchestration: LangChain or CrewAI
  • ERP Integration: Rutter or Unified API (for QuickBooks/Xero/NetSuite access)
  • OCR/Document Parsing: Amazon Textract or Docsumo
  • Email Automation: Front or Zendesk API
  • Database: Supabase (PostgreSQL) for audit logs

Prompt Skeletons

### Prompt 1: Remittance Parser (Extraction)
Act as a B2B Financial Analyst. Extract data from the following OCR text of a Remittance Advice PDF. 
Identify: 
1. Payer Name
2. Total Amount Paid
3. List of Invoice Numbers referenced
4. Amount paid per invoice
5. Any "Reason Codes" or adjustment notes (e.g., "Disc", "Short", "Tax").
Output in JSON format.

Text: {{ocr_text}}

### Prompt 2: Dispute Outreach
You are an Accounts Receivable Specialist. A customer ({{customer_name}}) has paid ${{paid_amount}} for Invoice {{invoice_number}} which totals ${{invoice_total}}.
The difference is ${{difference}}.
The customer noted: "{{reason_code}}".
Internal records show: {{internal_evidence}}.

Draft a polite, professional email to their AP department asking for clarification or providing the proof of delivery required to settle the balance. Do not use aggressive language.

Success Metrics

  • DSO Reduction: Decrease Days Sales Outstanding by >15%.
  • Unapplied Cash Ratio: Reduce the percentage of payments not linked to invoices within 24 hours.
  • Human-Touch Rate: Percentage of invoices resolved without manual intervention (Target: >70%).