Meridian Apparel is a mid-size consumer goods brand sourcing from overseas factories across multiple regions. Two to three procurement analysts reconcile every inbound supplier invoice against the company's open PO database and vendor master list before data flows into finance and inventory systems. Errors in the master workbook cascade directly into CFO-reviewed reporting packages.
Page-break splits, merged cells, and buried PO numbers were silently breaking the master sheet
Each monthly cycle, the team processed dozens of invoices carrying up to 50 line items — arriving as scanned PDFs, native digital files, and Excel workbooks with merged-cell layouts. Consolidation was almost entirely manual: open each invoice, parse visually, hand-key every row.
Three failure modes compounded each other. Page-break splits produced orphaned SKUs with null quantities — records that looked complete but were structurally broken. Merged supplier Excel templates caused blank rows that analysts filled inconsistently. Suppliers burying PO references in free-text headers — "Re: Your order 45992-A" — triggered PO-unmatched flags and separate correction cycles. Math reconciliation caught roughly 90% of split-line OCR errors, but only after data entry was already complete. A seasonal push to onboard three to five new factories and a finance initiative compressing payment terms from net-45 to net-30 broke the informal scaling model: invoice batches now had to clear validation in a single business day.
Energent.ai became the end-to-end consolidation pipeline
The team evaluated an expanded Excel macro suite and a standalone OCR point solution. Both treated ingestion and validation as disconnected steps; neither could reconcile extracted values against the vendor master or open PO database. Energent.ai handled the complete pipeline in a single agent session:
- Routed each document by text-layer presence — native PDFs processed directly, scanned PDFs via vision extraction — eliminating coordinate drift
- Extracted line items across page breaks anchored on table headers rather than page boundaries
- Forward-filled merged-cell columns — factory name, style code, PO number — to match supplier intent
- Recovered buried PO numbers via regex over the full text block, annotating recovered values with a
source: text_recoveryflag - Ran two-phase validation before the append: Phase 1 normalized dates and numeric fields; Phase 2 matched vendor master, validated open POs, and ran line- and invoice-level math checks
- Staged exceptions by failure code —
vendor_unmatched,math_fail,po_not_found,confidence_below_85pct— in a dedicated tab without touching validated data
No custom OCR pipeline. No separate validation script. No fragile macro suite to maintain.
Validation moved upstream of the master sheet append
- Confidence gating at 85%. Records below threshold never reach the master sheet — they land in the staging tab with the failure reason attached.
- Math checks before, not after. Line- and invoice-level reconciliation runs as part of the automated pipeline, catching ~90% of split-line errors before they corrupt the workbook.
- Format-agnostic routing. Scanned PDFs, native PDFs, and Excel workbooks run in the same session; document type determines the extraction path, not analyst judgment.
- Annotated exception queue. Every flagged record carries a specific failure code, replacing informal ad-hoc corrections with a prioritized review list.
Two-to-three analyst-days per batch reduced to exception-only review
- Batches of ten invoices carrying 30 to 50 line items each previously required two to three analyst-days; straight-through processing now handles all records passing Phase 2 reconciliation without analyst intervention
- Forward-fill normalization and regex PO recovery eliminated two categories of silent extraction error across all supplier template formats
- The labeled staging tab — each exception annotated with a specific failure code — replaced an informal correction process with a clear, actionable queue
"What changed is that we stopped treating validation as something that happens after data entry. Exceptions sit in a staging tab with the reason already attached — we know exactly what to fix and why before we touch it." — Priya Sharma, Procurement Operations Lead at Meridian Apparel