Automatic document entry. They go in on their own, in the form you decide.
Orders, transport documents, invoices from foreign suppliers, receipts. They arrive from different channels and in different formats, and someone retypes them by hand into the management software. This piece reads them, puts them into the form you need and enters them, one by one, after a check.
They sort themselves out. The document arrives, is recognised, the data needed is extracted and written where it has to go. Nobody reopens a PDF to copy a quantity.
You decide the form. The fields to extract, the names they must have and the format they come out in are agreed once, and they are yours, not those of the incoming document.
There is a check before the data goes in. Every field carries how sure the system is of having read it. Above the threshold it goes in; below it, it stops and shows it to you. Nothing goes in on a guess.
This is the page for a single piece of the system. The other pieces, and the criterion for choosing which to start from, are on the services page.
What changes in practice
The retyping work disappears from the day. An order that arrives at 9 is in the system at 9, with the lines, quantities and codes already in place, and whoever used to load it by hand finds a document already loaded, to check in ten seconds.
The second change is less visible and matters more: all the data goes in the same way. Same naming, same codes, same units of measure, whichever supplier sent the document and however they wrote it.
Time isn't recovered where the document is read. It is recovered where it doesn't get opened a second time.
You decide the form
During analysis you decide what to extract, field by field, and under what name. If in your management software the item is called quantità ordinata, the system writes the Italian you use even when the incoming document says ordered qty.
The same choice applies to the output format. Where the management software can be accessed from outside, the data is written directly; where it can't, the system prepares the file exactly in the layout that software imports, and the last click stays yours.
A short list, decided together, is extracted, not everything on the document. Extracting everything produces fields to clean up, and the clean-up is the same manual work you wanted to remove.
What the system does, step by step
Every document enters from the same point, whatever channel it arrived through, and every step leaves a log entry saying what happened and when.
| Step | What happens | What you get |
|---|---|---|
| Recognitionwhat document it is | The document is classified by type before being read, and those that arrive already structured are taken as data, not as images. |
Orders, transport documents and receipts each follow their own route, without anyone having to sort them. |
| Extractionfield by field | The agreed fields are extracted, with the name they have in your management software and in the format you need. |
The data arrives already in your internal language, ready to use without renaming anything. |
| Checkhow sure it is | Every field carries a measure of how sure the system is of having read it, and those that decide money have the highest threshold. |
You know which data is certain and which isn't, instead of finding out from a customer. |
| Stopbelow the threshold | The uncertain document doesn't go in: it is shown with the doubtful field highlighted next to the original, ready to correct. |
Ten seconds of checking on a few documents, instead of a week chasing a wrong order. |
| Writinginto the management software | Documents above the threshold are written. Where the management software can't be accessed from outside, the system prepares the file to import. |
The management software stays yours, and nobody has to change it to make the rest work. |
The check, which is the part that makes the rest usable
A model that reads documents always answers, even when it hasn't understood. On an illegible quantity it writes a plausible number, and that number looks just like one read correctly. That is why the check isn't an extra bit of caution: without it, the system moves errors around instead of removing them.
So every field comes out with a confidence measure, and the threshold is decided together, field by field. An uncertain delivery address gets fixed later. An uncertain quantity, taxable amount or order number stops the whole document and sends it to a person.
Stopping is normal, and it is worth knowing before buying anything: in departments that already process documents automatically, the average exception rate measured by Ardent Partners in January 2026 is 18.4%, with an average cost of 9.84 dollars per document and 8.2 days of processing time. Whoever promises zero exceptions is describing a system that guesses.
The useful comparison is with what happens today without the system. Manual transcription gets around 1% of values wrong, and rereading the document by eye doesn't reduce that error: this is measured by a study by Barchard and Pace in Computers in Human Behavior, which compares data entry methods on 195 people.
The point lies in the form of those errors, and the authors' lab says it verbatim: «most of those errors are within the allowable range for the variables». They are plausible numbers in place of the right numbers, so they pass any check done by looking. The automatic check says which ones are uncertain, which is the one thing the eye can't do.
The share of documents stopped is counted, and if it grows you know. It is the signal that documents different from those the system was calibrated on are arriving, and it shows before it becomes a problem.
In Italy the invoice arrives ready, and it's worth knowing
On Italian invoices there is nothing to read, because they already arrive as data. The Agenzia delle Entrate (Italian Revenue Agency) states that since 1 January 2019 all invoices issued for supplies of goods and services between persons resident or established in Italy can only be electronic invoices. In 2025 2.4 billion of them went through the Exchange System, issued by 5.5 million parties.
It means that on the Italian invoice side the data already arrives structured, and reading it with a vision model would add an error where there is none. The system takes them for what they are, and does the real work elsewhere.
The real work is the orders, which arrive as PDFs or written inside an email, the transport documents, which may stay on paper, the receipts gathered at the end of the month, and everything that comes from foreign suppliers and customers, which doesn't go through the Exchange System. That is where retyping happens today.
The same thing changes name with the trade
The mechanism is identical; the document that gets retyped isn't. It is worth looking at your own case, because that is where you see how much it costs today.
| Sector | The document that gets retyped | Where you see it |
|---|---|---|
| Food and agricultureand export | The order from a foreign distributor written inside an email in English, with product codes that aren't yours, outside the Exchange System. |
|
| Restaurantsand bars | Suppliers' transport documents, which arrive every morning on paper and have to be checked against what actually came into the warehouse. |
The service of a dining room where the phone rings unanswered |
| Hospitalityaccommodation and events | A season's expense notes and receipts, gathered at the end of the month in an envelope and rewritten one by one before passing them to whoever keeps the accounts. |
What this piece doesn't do
It doesn't replace the management software or impose a new one. If yours can't be accessed from outside, the system stops one step earlier and delivers the file to import.
It doesn't do accounting. It prepares the data and delivers it in order, which shortens the monthly collection, and the bookkeeping stays with whoever signs it. Monthly reporting works along the same line, putting that data into a readable form once a month. The reasoning about the human constraint is on the page about the principles we build with.
It doesn't decide what to do with the document. The requests that arrive along with the documents, and that need a reply, are a different problem: that is handled by Inbox AI, the first reply to every request.
Questions and answers
Which documents can it read?
Orders as PDFs or written in the body of an email, transport documents even when photographed, receipts and expense notes, invoices from foreign suppliers, price lists and order confirmations. In general any recurring document for which there is a reasonable sample to start from.
Italian electronic invoices are taken as data and not as images, because since 1 January 2019 they go through the Exchange System and arrive already in a structured format.
How can I trust what it writes into the management software?
Because every extracted field carries how sure the system is of having read it, and you decide the threshold field by field. Above the threshold the data goes in; below it the document stops and is shown to you with the doubtful field highlighted next to the original.
The fields that decide money, such as quantity, taxable amount and order number, have the highest threshold of all. A wrong figure written into management software is worse than a missing one, because nobody rereads it.
Does the data come out in our management software's format?
Yes, and it is a choice made during analysis. The fields to extract, the names they must have and the output layout are agreed once and remain yours, even when those items are called something else or are in another language on the incoming document.
Where the management software can be accessed from outside, the data is written directly. Where it can't, the system prepares the file in the layout that software imports.
Is it for accounting? Does it replace the accountant?
No. The system prepares the data in an orderly form and delivers it ready, which shortens the collection work done by hand every month today.
Bookkeeping entries, the tax treatment of a document and every choice with consequences for the tax return stay with the accountant, and we give no guidance on that front.
How do you measure whether it's working?
With the share of documents that go through the route without anyone touching anything, agreed as the indicator before starting. You measure the time spent today on a batch of real documents, set the threshold the work must pass not to be considered a failure, and compare afterwards.
The second number is the share of documents stopped by the check, which must stay stable. If it grows, documents different from those the system was calibrated on are arriving.
Notes on sources
- The electronic invoicing obligation between parties resident or established in Italy since 1 January 2019 is reported by the Agenzia delle Entrate. The 2.4 billion invoices processed and the 5.5 million issuing parties are in the Agency's own document I risultati 2025, pages 3 and 25.
- The 18.4% exception rate comes from Ardent Partners, January 2026, published by the firm that runs the research. The survey states neither sample size nor geographical area and measures accounts payable departments, largely big and American: we say so because whoever reads this page isn't that sample.
- The error rate of around 1% in manual transcription, and the fact that rereading by eye doesn't reduce it, come from Barchard and Pace, Preventing human error: The impact of data entry methods on data accuracy and statistical results, Computers in Human Behavior 27(5), 2011, pages 1834-1839, summarised by the authors' lab at the University of Nevada. It is a controlled experiment on 195 people transcribing research data, not employees loading business documents: we cite it for the form of the error, which stays within the plausible range and therefore can't be seen.
- We don't publish a percentage of time saved. The most cited figure on this subject comes from a 2016 survey of data professionals, and it measures a different job and a different activity from retyping business documents. The time is counted on your own batch of real documents, before any quote.
- This page doesn't report results obtained for a client, because this piece hasn't yet been delivered to a client. The tests cited are functional checks carried out in testing.
The other pieces in this group
Hours spent doing by hand what repeatsFifteen minutes, with your case in front of us.
How many documents a month come into the business and get retyped by hand somewhere? If the answer is that nobody knows, that is already the most useful information, and it can be counted in thirty days. In fifteen minutes on the phone we look at it together and tell you where it makes sense to start, even if we don't end up working together.
You get Mattia Esposito, who then builds the system: there's no salesperson in between. If you'd rather measure on your own before talking, the Diagnostico (in Italian) is twenty questions and five minutes.