Insights

Document extraction for SMEs: invoices to data

An Australian accounts payable team spends $12 to $15 processing a single supplier invoice. That covers opening the email, reading the PDF, keying numbers into the accounting package, matching against a purchase order, and flagging exceptions. At best-in-class automated operations, the cost drops to $2.78 (Ardent Partners, 2025).

For a WA business processing 200 invoices a month, the difference comes to roughly $2,000 per month in recovered labour. Over a year, that more than covers the integration build.

Most businesses already own the data they need. It sits in PDF attachments and scanned forms. Document extraction reads those unstructured files, pulls out the fields that matter, and drops them somewhere structured: a row in Xero, a line in a job costing sheet.

How the pipeline works

A working extraction pipeline for invoices has four stages.

Capture. Invoices arrive by email, upload, or forwarding address. Some tools accept hard copy via scanner. Capture normalises everything into an image or PDF the extraction engine can read.

Extract. Older systems use template-based OCR: you define which box on the page holds the invoice number for each supplier. That breaks when a supplier redesigns their layout or sends a handwritten note. Newer approaches use LLMs with vision capabilities. A 2026 study from Purdue researchers (arXiv:2511.05547) found that multimodal LLMs reading invoice images directly outperformed traditional OCR on variable layouts, especially with handwritten annotations or poor scan quality. A hybrid approach, OCR for structured regions plus LLM for the rest, was the strongest performer.

Validate. Extraction is never 100%. Validation checks extracted values against known data: does the ABN match a registered business? Does the line total equal the sum of its parts? Does this supplier normally send invoices in this format? Human review catches the rest. A good pipeline flags only the exceptions.

Push. Validated data lands in the accounting system via API or webhook. Xero and MYOB both support invoice creation through their APIs, though the paths differ in maturity and rate limits.

Where it fails

Three failure points show up consistently in production.

Multi-page PDFs with line items that span columns. Accuracy drops when layouts have tables inside tables or when the document mixes English and another language on the same page.

Handwritten additions. A supplier scrawling a revised amount in pen over a printed invoice defeats most template-based systems. Vision-capable LLMs handle these better, but at higher cost per page.

GST treatment ambiguity. Australian invoices frequently mix GST-inclusive and GST-exclusive line items, or apply different tax codes to different rows. Getting the GST wrong on extraction means getting the BAS wrong.

What it costs

A focused implementation for a mid-size business (200 to 500 invoices per month) takes six to ten weeks. Connect the invoice inbox, train and validate extraction on real supplier invoices, wire the approval rules, integrate with the accounting package, and run in parallel for two weeks before cutover. Agenty IS (July 2026) benchmarks this at $8,000 to $15,000 depending on supplier format count and accounting platform.

Ongoing costs are API usage for the extraction engine ($50 to $200 per month at this volume) plus human review time at around 10 to 15 minutes per day once the pipeline runs stably.

At 200 invoices per month and $12 per manual process, the payback period is measured in months. Whether your invoices are consistent enough to learn, and whether your team will actually use the review queue instead of printing PDFs anyway, are the questions that matter.

Sources