1,200 invoices a month. 45 seconds each. 94% accuracy.

The challenge

Managing 800+ properties means dealing with a constant stream of invoices — maintenance contractors, utility providers, cleaning services, landscaping companies. Each of the 300+ vendors sends invoices in their own format: some are structured PDFs, others are scanned paper invoices, some are emails with attached images, a few are still handwritten.

The company's accounts payable team was manually keying invoice data — vendor name, invoice number, amount, line items, due date, property reference — into their accounting system. Each invoice took an average of 12 minutes to process. At 1,200 invoices per month, that was 240 hours of manual data entry, and the error rate was running at 8%. Errors meant payment delays, duplicate payments, and reconciliation disputes with vendors.

The company had evaluated invoice processing software, but template-based tools required configuring a separate template for each vendor's format. With 300+ vendors — and new ones added regularly — maintaining those templates was itself a full-time job.

The solution

Gilligan Tech deployed a Document Intelligence pipeline using GPT-4o Vision as the core extraction engine. Unlike template-based OCR tools, GPT-4o understands document structure semantically — it can read any invoice format without configuration, extracting the same fields regardless of layout, font, language, or whether the document is typed or scanned.

Every incoming invoice (via email attachment or file upload) is automatically passed to GPT-4o Vision with a structured output schema. The model returns a JSON object containing all required fields — vendor, invoice number, date, line items, totals, property reference, payment terms. A validation layer then checks the extracted data against business rules: known vendor names, expected amount ranges, required fields.

High-confidence extractions are automatically posted to the accounting system. Low-confidence extractions and anomalies (unexpected vendor, unusually high amount) are routed to a human review queue with the AI's extracted data pre-filled, so the reviewer only needs to confirm or correct rather than type from scratch.

Implementation

1. Intake pipeline: An email monitor and file drop folder were set up as invoice intake points. New attachments trigger automatic processing — no manual upload or routing required from staff.
2. Pre-processing: Scanned PDFs and image attachments are rendered to high-resolution images. Multi-page invoices are split and each page processed individually before reassembly.
3. GPT-4o extraction: Each invoice image is sent to GPT-4o Vision with a structured output schema specifying required fields and data types. The model returns a validated JSON object in under 10 seconds.
4. Validation layer: Extracted data is checked against a vendor master list, expected amount ranges per property, and required field completeness. A confidence score is computed for each extraction.
5. Routing: High-confidence extractions post directly to the accounting system. Low-confidence items enter a review queue with pre-filled data. Human reviewers confirm, correct, or reject. All outcomes feed back to improve thresholds.

Results

Technology