What "95% accurate" actually means for your books

You’ll see accuracy numbers thrown around for AI that reads documents — 95%, 99%, high-and-impressive. They’re not wrong. But a percentage on the reading step tells you less about your books than it sounds like it does, and it’s worth understanding why.

Where the number comes from

When a tool says it reads at 95–99% accuracy, it usually means: across a lot of documents, the figures it pulled matched the correct figures that often. On clean, printed pages, modern extraction really is up there. On messier inputs — faded thermal paper, cramped handwriting — the honest number drops, which is exactly the case a marketing figure tends to skip past.

So the first thing to know is that “95%” is an average over easy and hard pages, and your close packets include both.

Why accuracy alone doesn’t make it safe

Here’s the part that matters. Suppose reading is 97% accurate. On a close packet with dozens of figures, that means roughly one number in thirty is off. If those numbers flow straight into your books, you’ve just imported a small, invisible error into every packet — and you have no idea which line it’s on.

That’s why the reading accuracy isn’t the thing that protects you. The check is. Plain arithmetic takes the numbers that were read and tests them against each other on the same page: does the deposit match the register, does the fuel dispensed match the tank drop. A misread number doesn’t slide through quietly — it fails a check, and a failed check gets flagged for a person.

The reading can be 97% or 99%; what makes it safe to sit near your books is that a wrong read surfaces instead of hiding.

The honest version of the claim

So when we talk about accuracy, the number we care about isn’t how often the AI reads a figure right. It’s how often a wrong figure makes it all the way to you unnoticed — and the answer we design for is close to never, because the reading is checked and the unverifiable is flagged.

A high reading accuracy is nice. A reading you don’t have to trust blindly is the actual product. Ask any tool not what its OCR score is, but what happens to the numbers it gets wrong.