A leading finance company was drowning in 40,000 invoices a month across 200+ supplier formats. We replaced manual keying with an LLM-in-the-loop ingestion pipeline — and lifted throughput by 95% without adding a single headcount.
Invoices came in as clean PDFs, scanned images, email bodies, supplier-portal downloads and EDI feeds — with GST layouts that shifted per state and per vendor. Six clerks spent their days retyping the same fields into the ERP, and the queue only ever grew.
A time-and-motion study of a typical week showed clerks spending 84% of their shift on data entry and correction, with format-specific handling driving the tail. Scanned invoices were the worst — OCR quality made every field a manual re-check.
PainFormat varianceManual keyingDuplicate posting risk
We collapsed six manual handoffs into a single automated flow with a human review lane for anything the model wasn't sure about. The pipeline learns from every correction.
Email drop, SFTP, portal scrapers, and mobile capture all land in one queue with source metadata preserved.
An LLM pre-labels every field against your GST + PO taxonomy. Per-field confidence attached.
Only low-confidence fields hit the reviewer queue. Everything else is posted straight through.
Two-way sync with the ERP, with every change traceable back to the model version and reviewer.
The same team clears an invoice queue 20× larger without overtime, and the model's straight-through rate keeps climbing as it learns from every reviewer correction.
"The team stopped being data entry clerks. They became the exception handlers — and the pipeline handles everything else. Volume is now a growth question, not an ops question."
Every artifact we shipped is theirs — the pipeline, the taxonomy, the model, and the tooling. No lock-in.