99% Accurate. Still Wrong on One Invoice in Five.

Your vendor quotes field accuracy. That is the wrong number, and the arithmetic takes thirty seconds to check.

🔢 Run the maths on a whole document

Your vendor says their OCR is 99% accurate. Sounds great. Almost perfect.

Now look at a normal invoice. Supplier name. Invoice number. Date. Due date. Tax ID. Line items. Quantities. Prices. Tax rate. Total. Easily 20 fields.

If each field is 99% right, the chance that all 20 are right is about 82%.

Roughly one invoice in five has at least one mistake.

On 1,000 invoices a month, that is about 180 invoices someone has to catch and fix. 😬

Here is the same sum at other accuracy levels, so you can find your own row:

Per field	10 fields	20 fields	30 fields
95%	60% clean	36% clean	21% clean
98%	82% clean	67% clean	55% clean
99%	90% clean	82% clean	74% clean
99.5%	95% clean	90% clean	86% clean

Check it yourself. Accuracy to the power of the number of fields. Thirty seconds in a spreadsheet.

📌 And the errors are not spread evenly

This is the part that surprises people.

Those 180 invoices are not scattered randomly across your suppliers. They cluster. The phone photograph someone emailed in. The supplier whose layout changed last quarter. The scan that went through at a slight angle. The one credit note that looks almost like an invoice.

In practice, a small number of document types generate most of your rework. Which is good news, because it means the problem has an address. Most teams have simply never looked at which suppliers cost them the most checking time.

🧱 Why field extraction hits a wall

Classic OCR reads a document the way a scanner reads a barcode. Find the box. Copy the text. Move on.

But documents are not just boxes.

• A credit note looks almost like an invoice. Same supplier, same layout, same fields. One minus sign decides whether you are paying or being paid.
• A table continues on page two. Where field extraction sees two unrelated chunks of text, a person sees one table split across a page break.
• A total at the bottom should match the lines above it. Classic OCR has no opinion about that. It copies both and never notices they disagree.
• A new supplier means a new layout. Which means new configuration. Which means somebody maintains a growing pile of templates forever, and nobody budgeted for that job.

Think of a new assistant who copies numbers perfectly but never once asks "wait, does this add up?" Fast, and quietly risky.

🤖 What understanding actually means

The difference is not better scanning. It is that the system knows what it is looking at.

• It recognises the document type. Invoice, credit note, delivery note, dunning letter. Before extracting anything, it knows which one this is.
• It keeps structure intact. A table stays a table. Quantity stays attached to its description and its price. Sections, headings and page breaks survive rather than collapsing into a wall of text.
• It checks internal consistency. Do the line items add up to the total? Is the due date after the invoice date? Does the tax rate match the tax amount? These are the questions a person asks automatically and a template never asks at all.
• It needs no configuration for a supplier it has never seen. Because there was no template to configure.

❓ What to ask any vendor, including us

Nine questions. The first four matter most.

1. What is your document level straight through rate, not field accuracy? The share of documents completed with no human touch at all.
2. Do you output confidence per field, and can low confidence be routed to a person automatically?
3. How long does it take to onboard a supplier you have never seen? Template systems answer in days. It should be close to zero.
4. Do you validate against our systems, or only against format? A perfectly extracted invoice number that does not exist in our ERP is still wrong.
5. Does a table survive as a table, including across a page break?
6. Can you tell an invoice from a credit note without being told?
7. What happens to a document that fails validation? Does it hold, or guess?
8. Where is the processing done, and can it run inside our own environment?
9. Can we see, per document, what was extracted and how confident the system was?

If a vendor cannot answer the first question, the accuracy figure they quoted means very little.

📊 The number that actually matters

Stop asking for field accuracy. Ask for straight through rate: the percentage of documents that complete with nobody touching them.

One of those numbers flatters the vendor. The other one is your business case.

And measure your current rate before you start. Most organisations never have, which makes any "after" figure impossible to argue with, in either direction.

🗃 What we offer

Meju-Dox handles any document, in any format, with its structure intact. Tables stay tables. Layout survives. What goes in as a document comes out as something you can actually work with.

MeJuScan turns any scan or PDF into structured, searchable data. Document OCR as a service, and free while in beta.

⚖ What we will not claim

No document system is right every time, and anyone telling you otherwise is describing a clean demo set.

What changes is where the human effort goes. Instead of checking every document because any of them might be wrong, your team sees only the ones the system flagged, and the flag is usually right. Smaller job, better job, and a measurable one.

It also will not fix a document that is genuinely illegible. A photograph taken at an angle in bad light is hard for a person too.

📄 Try it on your own documents

Not on a sample. The best test is your messiest real document. The phone photo. The fax. The supplier whose layout nobody understands.

Send us one and see what comes back.

👉 Request MeJuScan beta access