By super.AI
Invoice data extraction usually fails in the gap between reading text and putting the right value in the right field. OCR can read every character on a page and still store the subtotal as the total, split a line-item table at a page break, or merge two invoices from one PDF into one record. These failures come from vendor layout variety, poor scans, multi-page tables, and inconsistent formats. A single page-level accuracy number hides all of them. To choose a tool, test it on your own invoices at the field and line-item level. Look for field-level confidence, validation rules, human review of uncertain fields, corrections that carry forward, and export to your system.
super.AI publishes this article and builds document processing software. It describes common failure modes and how to test for them. It does not rank vendors or make accuracy claims about any product.
If you are new to the topic, start with our guides to invoice data capture and extracting data from PDFs. This article picks up where those leave off: where extraction breaks in real workflows, and how to measure it before you commit.
Every vendor lays out invoices its own way, and vendors change their layouts without notice. A template or rule built for last quarter's layout can start returning blanks or wrong values after a redesign. Nothing flags the change unless something checks the output.
Skewed pages, faint print, shadows, and low-resolution phone photos cause character errors. A 5 becomes an S, a decimal point disappears, or a 1 becomes a 7. On an amount field, one wrong character changes what you pay.
Line-item tables often run across pages. Page two may repeat the column headers, drop them, or start with a carried-forward subtotal. A tool that reads each page on its own can duplicate rows, lose rows, or treat a subtotal as a line item. For more on table structure, see our post on automating table extraction from PDFs and scanned images.
Some vendors send one PDF that holds several invoices, or an invoice with remittance slips and terms pages attached. If the tool does not split the file first, fields from two invoices end up in one record.
"Received" stamps, approval initials, handwritten PO numbers, and highlighter marks sit on top of printed text. They can hide a value, or add one that was never on the original invoice.
03/04/2026 is March 4 in the US and 3 April in much of Europe. 1.234,56 and 1,234.56 are the same amount written two ways. The $ sign alone does not tell you whether an amount is in US, Canadian, or Australian dollars. Extraction can read every character correctly and still store the wrong date or amount.
The header total may not equal the sum of the line items. Tax, freight, discounts, and rounding explain some gaps. A missed or duplicated line explains others. If nothing checks the math, a missing line passes straight through.
The hardest failure to spot is text that is read correctly but mapped to the wrong field. The PO number lands in the invoice number field. The ship-to address becomes the vendor address. The due date is stored as the invoice date. Character-level checks will not catch this, because every character is right.
A single accuracy figure for a page or a document tells you little about whether your AP data is right.
Measure these instead, on your own invoices:
Ask every vendor, including us, to show each item on this list working on your own files. In super.AI, low-confidence results go to human review, and corrections carry into the next run. Outputs include CSV, JSON, and Google Sheets. You can read more about data extraction with super.AI and invoice processing with super.AI.
Run this on any tool you are considering, including ours.
OCR reads characters. Invoice data extraction also has to decide which value belongs in which field. Errors in that second step, such as a PO number stored as the invoice number, pass any character-level check.
Measure on your own invoices against an answer key you typed by hand. Score each header field and each line item, and track wrong values the tool did not flag. A single page-level figure is not enough.
The tool needs to split the file into separate invoices before extraction. Include multi-invoice PDFs in your test set and check that each invoice becomes its own record. For general techniques, see our guide to extracting data from PDFs.
It depends on the tool and on how much the mark covers printed text. Include stamped and annotated invoices in your test set and check the affected fields by hand.
For uncertain fields, usually yes. Review is a verification step: a person confirms or corrects the values the software is unsure about. In super.AI, low-confidence results go to human review, and corrections carry into the next run.
This article covers how extraction fails and how to test it. To compare specific products, see our invoice scanning software buyer's guide.
Pick your messiest invoice, upload it, describe the fields you need in plain language, and see each step of the flow super.AI builds. It is free to start, with no credit card and no sales call required. Start for free or book a demo.
The main AP automation benefits, with sourced benchmarks for cost per invoice, cycle time and exceptions, plus how AP automation works and best practices for a rollout.
A buyer's guide to invoice scanning software: seven invoice OCR tools in alphabetical order, who each one fits, and how they compare on trials, pricing, and deployment.
Intelligent character recognition (ICR) reads handwriting on forms and checks. Learn how it works, how it compares with OCR and AI extraction, its limits, and how to test it.
