REF / AUTOMATION

Manchester Manufacturer - Invoice OCR and Vendor Management at Scale

Automated invoice processing, vendor management, and predictive maintenance signals for a 1,200-staff Manchester manufacturer - recovered 6 FTE roles.

RoleAutomation Architect
Year2025
Outcome6 FTE roles recovered, 89% touchless AP
DomainAutomation
00
STACK

Tech used.

Claude Opus 4.6Azure Form Recognizern8nMicrosoft Dynamics 365Power BISnowflake

The Problem

A 1,200-staff engineering manufacturer in the Manchester area was processing 14,000+ vendor invoices per month through a 9-person AP team. Each invoice was OCR'd by a legacy system, then manually corrected by humans (the OCR was wrong on roughly 40% of fields). The 3-way match (invoice ↔ PO ↔ goods receipt note) was done in Excel. Discrepancy resolution sat in email threads. Vendor master data was drifting. duplicate vendor records, inconsistent payment terms, banking details updated only when someone noticed an error.

The CFO's two pain points: AP cost was creeping up year on year, and vendor disputes were becoming a recurring P&L line item.

What I Built

A modern AP + vendor management automation that integrates with the existing Dynamics 365 system:

1. Invoice ingestion + AI extraction. Vendors email invoices to a dedicated address. A pipeline using Azure Form Recognizer for layout detection + Claude for semantic extraction pulls every relevant field with high accuracy. Confidence scores on each field; low-confidence flags route to humans.

2. Automated 3-way match. The system compares invoice line items against the matching PO and goods-receipt note. Exact matches (89% of volume after tuning) are auto-approved into Dynamics. Mismatches are routed to AP staff with the specific reason (price variance, quantity short, missing GRN) and a one-click resolution flow.

3. Vendor master data hygiene. A weekly reconciliation flags duplicate vendor records, inconsistent VAT numbers, and bank detail changes that need verification. A monthly AI-generated "vendor health" report goes to procurement: spend concentration, payment-term outliers, late-payment risk.

4. Predictive maintenance signal. Service invoices from equipment vendors are parsed for trend signals: "vendor X has been called out 4 times this quarter on Machine 23 vs once a year historically." Surfaced to the maintenance manager.

3-Way Match Outcome
InputVendor invoice extracted by Form Recognizer + Claude
ClassifierMatch vs PO × goods-receipt note
  • Exact match89%
    → Auto-approve into Dynamics 365touchless
  • Price variance5%
    → AP review with one-click resolutionflagged
  • Quantity short3%
    → AP review + warehouse callbackflagged
  • Missing GRN2%
    → Hold for goods receipt confirmationhold
  • Low extraction confidence1%
    → Human re-keys low-confidence fieldshuman review
89%
Touchless AP
from 12%
6 FTE
Roles recovered
redeployed
−72%
Vendor disputes
£840K
Annualised savings
Invoice Processing Funnel. Monthly Average14,200 invoices → finance ledger
Invoices received14,200
Top of funnel
AI-extracted (high confidence)13,680
96.3% of prior96%
3-way match. auto-approved12,640
92.4% of prior89%
Routed to AP for exception1,560
12.3% of prior
Resolved within 5 days1,480
94.9% of prior

Outcome

In 11 months, the AP team shrank from 9 to 3 (with 6 redeployed into procurement and finance analytics roles, no layoffs). Vendor disputes dropped 72%. Late-payment penalties. a £180K/year P&L line. fell to under £30K. The CFO has used the same automation team to start tackling order-to-cash on the receivables side.