OCR on Real Invoices: What Accuracy Looks Like Before You Buy

OCR demos always use a clean, well-lit invoice. Real documents arrive as photos taken at an angle, scans with a coffee ring, and PDFs where the table has no lines. Before you commit to automating document entry, it’s worth knowing what accuracy looks like on your own paperwork.

Text recognition is only step one

OCR turns pixels into text. Getting value from it means a second step: working out which text is the invoice number, which is the total and which rows make up the line items. That second step is where layouts that vary from supplier to supplier cause trouble.

Test on a real sample

Gather 50 to 100 documents that reflect the mess you actually receive, including the ugly ones. Decide which fields matter, such as supplier, date, total and tax, and measure how many each approach gets right. That’s a far better guide than a vendor’s headline percentage.

Plan for review, not perfection

Even good systems misread something occasionally. The practical design is a confidence score on each field. High-confidence values flow straight through, and low-confidence ones go to a person with the original image beside the extracted text. Most of the time saving comes from not typing, and a quick check is far cheaper than keying everything.

It isn’t only invoices

The same ideas apply to forms, contracts, ID documents and delivery notes. Beyond OCR, a few related jobs come up often:

• Classifying text, such as sorting support emails by topic or urgency.
• Extracting entities, like names, dates and amounts from free text.
• Image classification, for example flagging damaged goods in uploaded photos.
• Semantic search, so people find what they mean, not only what they typed.

Do you need a custom model?

Often not. Pre-trained models plus light tuning handle many cases. A custom model is worth considering when your documents are unusual or you need to recognise something very specific, and testing on your sample will show whether that’s true.

Other languages

Quality varies by language and script, so check early. Don’t assume a tool that works well in English will do the same on German tax forms or handwritten notes.

If you have a pile of documents or images you’d like turned into usable data, send us a handful of real examples and what you want from them. We’ll test feasibility and tell you what accuracy to expect.

You can read more about our NLP, Search & Computer Vision service on the main site.