Of all the figures we researched for this series, the one about intelligent document processing is the most interesting. Because, unlike WhatsApp's "98% open rate", which is flat-out made up with no source, the famous "99% accuracy" of AI invoice readers is… true. More or less. Sometimes. With fine print.
And that fine print is exactly the difference between a successful implementation and a project that fails three months in. That's why, in this fourth post of the series on the real numbers behind enterprise AI, we're not going to debunk the 99%. We're going to do something more useful: explain when it's true, when it isn't, and how not to be fooled by the average.
The 99% is real… for one part of the invoice
Let's start by acknowledging what is true. The leading intelligent document processing vendors, what is called IDP (Intelligent Document Processing), really do reach very high accuracy. ABBYY, one of the global leaders, reports 99.5% accuracy on structured invoices, across 1,500 million invoices processed per year (IDP tools comparison, 2026).
But here comes the question almost nobody asks: 99% accuracy on what, exactly?
An invoice is not a uniform block of information. It has two very different parts:
The header fields: the invoice number, the date, the supplier's tax ID (CUIT), the total. There are few of them, they are almost always in the same place, and they have a predictable format.
The line items: each billed item, with its description, quantity, unit price and subtotal. There can be 3 or there can be 300, in formats that vary from one supplier to another.
The "99% accuracy" that gets advertised almost always refers to the header fields under clean conditions. And there, yes, it is achievable. The problem is that when you go down to the line items (which are usually what really matters for entering the invoice into your system), accuracy drops: the same comparison treats the header as solved and places the difference between tools in the line items, the tables and the poor-quality scans (IDP tools comparison, 2026).
The difference between 99% and 90% doesn't sound dramatic until you translate it into volume: on 10,000 invoices, it's the difference between manually reviewing 100 or reviewing 1,000. That is real, and it's money.
The other fine print: "out of the box" vs. trained
There is a second trap in the 99%, and it has to do with when it is measured.
When a vendor like Nanonets states the accuracy of its system freshly installed, without having trained it on your company's specific documents (what is called "out of the box"), the figure is around 93% to 96%, and it improves only after training the model on your suppliers, your formats and your particular cases (Nanonets / comparative analysis, 2026).
This matters a lot for your expectations. If someone sells you "99% from day one", you will be disappointed. The truth is that the system starts out very well (93-96% in Nanonets' case) and improves after a few weeks of use and tuning. Knowing this up front is the difference between a project that is managed well and one that generates frustration.
That's why on our site you won't see a "99%" or an accuracy percentage of our own. You'll see what Gartner publishes: machine learning can take accuracy to the high end of the 90% range on machine-readable documents, not on images (Gartner). The accuracy that counts for your company is the one measured on your own documents. Less eye-catching in a headline. Infinitely more credible when someone on your team audits it.
Where the case is compelling: savings
If accuracy needs asterisks, on savings the evidence is among the strongest in the whole series. And here we do have independent backing, not just from vendors.
Processing an invoice costs 9.40 dollars on average, and the most efficient teams bring it down to 2.78 dollars (Ardent Partners, via Tungsten Automation). The key data point: according to Ardent Partners, an independent industry research firm, the top 20% of accounts payable teams process each invoice at a cost 78% lower than the rest (Ardent Partners, 2024).
The fact that the data comes from an industry analyst and not from a software vendor is what makes it valuable. Ardent Partners doesn't sell invoice readers; it measures the market. When they say 78%, it carries different weight. A caveat, though: it compares the most efficient teams with the rest, not an automated process with a manual one.
On time, the case of Rossum with its customer Fugro (a company that processes 300,000 invoices per year across four global sites) is illustrative: processing time per invoice went from about 2 minutes to around 35 seconds, a 70% reduction (Rossum, Fugro case). It is a case from a named customer, with real, verifiable volume, so we present it for what it is: a good example, not a universal guarantee.
Productivity: the best teams process 82% faster
The last number worth looking at is productivity. According to the same Ardent Partners analysis, the most efficient accounts payable teams process each invoice 82% faster than the rest and have 59% fewer exceptions (Ardent Partners, 2024).
What we did not find is an independent measurement of how many more invoices a person processes with IDP: the "4 to 8 times" that circulates comes from vendors, with no published methodology, which is why we don't use it. The real result depends on the complexity of your invoices and the level of exceptions you tolerate.
And in case you need to size the opportunity: McKinsey estimates that in banking alone, generative AI could add between 200,000 and 340,000 million dollars of value per year if all of its use cases were implemented (McKinsey, 2023). Document processing is not an administrative detail: it is one of the places where AI generates concrete, measurable value.
What you will read on the Arconte site
After all this research, these are the figures we chose for Arconte, our intelligent document processing product:
"High 90% range on machine-readable documents, according to Gartner": a category reference, not a promise. The accuracy in your case is measured with your own documents.
"78% lower cost per invoice at the most efficient teams": Ardent Partners, 2024, an independent firm. It compares the top 20% with the rest.
"From 2 minutes to under 1 minute per invoice": the Rossum / Fugro case, 300,000 invoices a year.
"Every field validated, and anything doubtful goes to human review": no figure, because this is how Arconte works, not a benchmark.
What you won't read is a bare "99% accuracy," with no context, as if your invoice with 80 detail lines would be read perfectly on day one. We tell you the whole truth: AI reads documents extraordinarily well, improves with use, and saves you a fortune in time and cost. But the header and the line items are not the same thing, and day one is not day thirty.
That transparency doesn't make us look less capable. It makes us look like the vendor that already thought through the problems other vendors prefer you discover on your own, after signing.
The full series: The real numbers of enterprise AI
The AI industry lies to you with numbers: we investigated which ones are real
Sellium: Your chatbot doesn't resolve 80% of inquiries. We lie to you less
Revio: The "98% open rate" on WhatsApp is made up. Here is the real number
Metrix: AI won't answer everything for your team. It will answer well what matters
Arconte: Reading invoices with AI. Where 99% is true and where it's marketing (you are here)
Trazzo: Last mile with AI. How much fuel, time and calls you really save
Every figure in this series is backed by its source, with date and methodology when available. If you find a figure you think we should review, write to us: we take this seriously.
![Reading Invoices with AI: Where 99% Is True, Where It's Marketing [and How Much You Really Save]](https://info.vantegrate.com/pulse/images/ia-facturas-procesamiento-documentos-idp-precision-marketing-empresarial.png.webp)














