Your invoices process themselves Validated field by field

Arconte does not ask you to change how your paperwork arrives. It listens to a dedicated mailbox (the classic invoices@yourcompany), a shared network folder, Drive or SharePoint, a webhook from your supplier portal and even the photo the warehouse supervisor takes with a phone. It reads native PDFs, scanned PDFs, JPG, PNG, multipage TIFF and the XML of the e-invoice, and it opens the compressed attachments that come in the same email.

Before looking at a single field, it does three things. It straightens and cleans the image, because a crooked photo with a shadow explains a good share of reading errors. It splits the file: a forty-page PDF is often a batch of six different invoices, and it detects where one ends and the next one begins. And it discards duplicates by file fingerprint and by document number, so the same invoice that arrived by email and through the supplier portal does not enter the workflow twice.

Only then does it classify: type A, B or C invoice (the Argentine invoice classes), credit note, delivery note, purchase order, contract. It matters because each type has different fields and rules, and a credit note treated as an invoice muddles the supplier's account balance.

Accuracy is not measured per document but field by field, and that detail changes everything. For each data point, Arconte locates where it sits on that specific invoice (not on a template), reads it, normalizes it and validates it.

Normalizing is the boring part and the one that prevents the most errors: the date 03/04 does not mean the same thing for two suppliers, the amount can come with a period as the thousands separator or a comma as the decimal mark, the tax ID (CUIT) with or without hyphens, and the unit of measure abbreviated six different ways. All of that is converted to a single format before it touches your system.

Validation runs on three levels. Format: the CUIT check digit has to add up. Internal arithmetic: the sum of the lines has to equal the subtotal, and the subtotal plus VAT and withholdings has to equal the total. And against your own data: the CUIT against the supplier master, the purchase order number against open orders, the unit price against the agreed price and the invoiced quantity against the delivery note received, which is the classic three-way match.

Each field ends up with three things attached: the value, its own confidence score and the result of its validations.

No serious solution gets 100% right, and the difference between one that works and one that gets abandoned is what it does with the rest. In Arconte, the confidence threshold is set per field: the total and the CUIT are held to a higher bar than the payment terms. If a field falls below it, the whole document is not stopped: only that field goes to the queue and everything else stays loaded.

The review screen shows the doubtful field next to the crop of the original image where it was read. The person confirms or corrects it in one click, without opening the PDF or hunting for the line. The queue is sorted by age and by amount, not by the order in which documents arrived.

Some cases go to review even when the reading is perfect: a supplier that is not in the master data, a difference against the purchase order above the tolerance, a suspected duplicate, an illegible document. It is human-in-the-loop by design, not a patch.

Whatever gets corrected feeds back into the model. The number we look at every week is the percentage that goes straight through without anyone touching it, and that percentage goes up. An honest note: the first month, the queue is uncomfortable.

With clean data, the document enters the workflow you define. Routing rules are built by amount, cost center, supplier and document type: up to a certain amount it goes straight through, above that amount it waits for the right approver, and a purchase without a prior purchase order always requires sign-off. The approver decides from email or from the dashboard, with the document in view.

Then it goes out to the accounting system: header, lines and the link to the original file, so the journal entry is never orphaned from the paper that backs it. Each document carries a unique key (CUIT, type, point of sale and number); if it is reprocessed, it updates instead of duplicating the entry, which is the most expensive mistake in these integrations.

It connects with Tango Gestión, Bejerman, SAP and any system with an API. And when the ERP rejects the load (closed accounting period, inactive supplier, nonexistent account), the document is not lost or blindly retried: it goes back to the queue with the ERP's exact error message, so the person fixes the cause instead of guessing it.

If an AI touches your accounting, you need to be able to reconstruct why it did what it did. Each field keeps its complete trail: the extracted value, the confidence it was read with, the coordinates on the page, the model version that processed it and, if someone corrected it, who and when. The original document stays immutable next to the data.

That pays off three times. For the audit, which asks how a number was arrived at. For your accountant, who needs the supporting document and not a screenshot. And for disputes with the supplier, because you have the document and the reading side by side.

Retention is set in your environment according to your policy: in Argentina, the Civil and Commercial Code requires keeping supporting records for ten years (Section 328), and that period is configured, not improvised.

As for where the data lives: your implementation runs entirely on Salesforce or on Oracle Cloud Infrastructure. The auditable certifications (SOC 2 Type II, ISO 27001, PCI DSS) belong to those platforms, not to Vantegrate. The details on controls, encryption and data residency are in Security and Data Transparency.

Arconte is not an ERP or an accounting system: it does not replace Tango, SAP or your accounting; it hands them clean data. Nor does it assign GL accounts on its own judgment; it applies the coding rules you define, and a new account is your team's call. It does not sign documents digitally, does not file tax returns and does not issue invoices.

With free-form handwriting it is honest: a handwritten note in the margin gets flagged as an exception, not made up. And from a contract it extracts parties, dates, amounts, terms and expiration dates, but the legal reading is still a lawyer's job.

There are three situations where we will tell you it is not worth it, and we say so on the first call. One: low volume. With a handful of invoices a week, the savings do not pay for the implementation and you are better off as you are. Two: if your documents already arrive structured via EDI or XML end to end, the problem is solved and there is no need for AI on top. Three: if nobody on your side is going to work the exceptions queue during the first few weeks. Without that person the workflow stays half-built and the project gets abandoned within two months. It is the only requirement we cannot cover ourselves.