How We Automated Document Processing for a Manufacturing Company

The Problem: 300 Invoices a Month, All Processed by Hand

A mid-size manufacturer in Bucharest came to us with a familiar problem. Their accounting team spent roughly 100 hours per month processing supplier invoices. Each invoice arrived in a slightly different format -- some as PDFs, some as scanned images, some as email attachments. Each one needed to be read, validated, and entered into their ERP system.

The process was not just slow. It was error-prone. About 4% of invoices contained manual entry errors: wrong amounts, transposed digits, mismatched supplier codes. Each error took 30-60 minutes to investigate and correct, often involving phone calls to suppliers and internal reconciliation meetings.

Their operations director put it simply: "We have three people spending half their week on work that a computer should handle. And the mistakes cost us more than the salaries."

Understanding the Real Workflow

Before we wrote a single line of code, we spent two days observing how the accounting team actually processed invoices. This is the step most document processing automation projects skip -- and why many of them fail.

What we found was more nuanced than "read PDF, enter data":

1. Invoices arrived through four channels: email, postal mail (scanned by reception), supplier portal downloads, and a shared Google Drive folder.

2. Each invoice required cross-referencing against purchase orders in their ERP system. The accountant was not just entering data -- they were verifying that the invoice matched what was ordered, at the agreed price, in the correct quantity.

3. Exceptions were common. About 15% of invoices had discrepancies: partial shipments, price adjustments, or items not on the original purchase order. These required human judgment.

4. The final step was approval routing. Invoices under EUR 5,000 were auto-approved. Everything above needed a department manager's sign-off.

Understanding this workflow was critical. Automating just the data entry would have solved 60% of the problem. Automating the full workflow -- ingestion, extraction, validation, exception flagging, and routing -- solved 95% of it.

What We Built

Document Ingestion Layer

We created a unified ingestion pipeline that monitors all four input channels:

• Email: A dedicated inbox that automatically downloads attachments
• Scanned mail: A watched folder where reception drops scans
• Supplier portals: Automated downloads on a daily schedule
• Shared drive: Folder monitoring with deduplication

Every document enters the same processing queue regardless of source. The system handles PDFs, scanned images (via OCR), and even photos of invoices taken on phones.

AI Extraction Engine

The core of the system is a document understanding model that extracts structured data from unstructured invoices:

• Supplier name and tax ID
• Invoice number and date
• Line items with descriptions, quantities, and prices
• VAT amounts and totals
• Payment terms and bank details

We trained the extraction model on 500 historical invoices from their top 30 suppliers. The model learns each supplier's invoice format and improves over time as it processes more documents.

Accuracy from day one: 93% of fields extracted correctly without human intervention. After two months of learning, that number climbed to 97%.

Validation Against ERP

Extracted data is automatically cross-referenced against open purchase orders in their ERP system:

• Match found: Invoice is linked to the PO, quantities and prices verified
• Partial match: System flags discrepancies for human review (wrong price, different quantity)
• No match: Invoice routed to exceptions queue for manual investigation

This validation step catches errors that even experienced accountants sometimes miss -- a supplier charging EUR 12.50 when the agreed price was EUR 12.00, for example.

Exception Handling Dashboard

Not everything can be automated, and we never pretended otherwise. The system includes a clean dashboard where the accounting team handles the 15% of invoices that need human judgment:

• Clear display of what the AI extracted vs what the ERP expects
• Side-by-side view of the original document and extracted data
• One-click approval for simple discrepancies
• Notes and routing for complex exceptions

The dashboard turned a 20-minute manual process into a 2-minute review task for exception cases.

Approval Routing

Approved invoices are automatically routed based on the company's existing rules:

• Under EUR 5,000: Auto-approved and entered into ERP
• EUR 5,000 - 20,000: Sent to department manager for approval
• Over EUR 20,000: Requires CFO approval

Approvers receive a notification with a summary and can approve directly from email or the dashboard.

The Results

We measured everything before and after, comparing three months of manual processing against the first three months of automated processing:

Financial impact:

• Staff time saved: ~88 hours/month at EUR 18/hour = EUR 1,584/month
• Error correction savings: ~EUR 400/month
• Faster processing enabled earlier payment discounts: ~EUR 300/month
• Total monthly savings: approximately EUR 2,280

The system paid for itself in under 6 months.

Lessons Learned

1. Observe Before You Automate

If we had built the system based on the brief ("automate invoice processing"), we would have missed the validation step, the exception handling, and the approval routing. Two days of observation saved weeks of rework.

2. Plan for Exceptions, Not Just the Happy Path

The 85% of invoices that process cleanly are easy. The system's real value is how it handles the 15% that do not. A good exception handling interface makes the difference between a system people trust and one they bypass.

3. Start with Your Top Suppliers

We trained the model on invoices from their 30 most frequent suppliers, which covered 80% of monthly volume. Getting those right first meant the system delivered value from week one. Less frequent suppliers were added gradually.

4. Measure Everything

Before-and-after metrics are not just for reporting. They build trust with the team using the system. When the accounting department saw the error rate drop from 4% to 0.4%, they went from skeptical to advocates.

5. Keep Humans in the Loop

The goal was never to replace the accounting team. It was to eliminate the tedious parts of their work so they could focus on analysis, supplier relationships, and financial planning. The three people who used to spend half their week on data entry now spend that time on work that actually needs human judgment.

Is This Right for Your Business?

Document processing automation works well when you have:

• High volume (100+ documents per month)
• Repetitive data entry into an existing system
• Measurable error rates that cost real money
• Staff time you would rather redirect to higher-value work

If you are processing documents manually and wondering whether automation could work for your business, we would love to help you evaluate. We offer a free 30-minute discovery call where we assess your specific workflow and give you an honest ROI estimate. Book a discovery call and let us find the right approach for your automation needs.