Challenges & Solutions
Built an AI agent that matches carrier invoices to rate confirmations, consignment notes, PODs, and ERP shipment records.
Challenge 1: Manual invoice matching
Qevrixa's billing team keyed more than 9,000 carrier invoices a month into the ERP by hand. Each invoice took 12 to 15 minutes to match against its rate card, consignment note, and POD, and the month-end backlog delayed both carrier payments and customer rebilling.
Solution: Through our AI agent development services, we built an intake agent on LangGraph that collects invoices from EDI, email and carrier portals. Azure AI Document Intelligence extracts structured fields from PDFs and scans, and GPT-4o on Azure OpenAI normalizes carrier-specific line items, subcontractor invoices without standard layouts, and tax lines into one standard invoice schema.
Challenge 2: Invoices in every format
Invoices arrived as EDI files, emailed PDFs, scanned images and carrier portal exports from more than 200 carriers, from national linehaul operators to regional subcontractors. Line items, charge codes, Australian Business Numbers (ABNs) and Goods and Services Tax (GST) amounts sat in different places on every carrier's tax invoice.
Solution: Our team connected the agent to the legacy ERP without changing its core. A read-only replica synced over ODBC gives the agent current shipment, rate, and carrier data. Approved invoices go back only through the ERP’s own batch import files, which the ERP validates the same way it validates manual entries.
Challenge 3: Undetected overbilling
Accessorial charges such as waiting time, unloading fees, tail-lift and fuel surcharges often did not match the agreed rate. With staff checking under time pressure, overbilling went undetected, and disputes surfaced weeks after payment.
Solution: Drawing on our experience as a Logistics Application Development Company, we built a rule-based rate-check engine between the agent and the ERP. It compares every charge with the contracted rate, the weekly fuel surcharge table, and the accessorial terms in each carrier agreement. Any invoice with a variance above Qevrixa’s tolerance is held for review, so the model never makes the final payment decision.
Challenge 4: A legacy ERP with no API
The ERP had no REST or SOAP API, only scheduled flat-file imports and read access through ODBC. IT would not allow an AI system to write to accounts payable tables, and auditors needed a clear record of who approved every payment.
Solution: Our AI engineers worked as an extension of Qevrixa’s billing team and built a React review portal where staff see each flagged invoice next to its rate confirmation, consignment note, and POD, approve or dispute it, and send an auto-drafted dispute email to the carrier. Every decision is logged with the user, timestamp, and reason for audits.