Four hundred invoices, one by one
Each one takes between 3 and 8 minutes to open, read, type and file. That's 30 hours a month of an expensive person, doing the most mechanical work in the company.
They arrive by email, by WhatsApp and on paper. Someone opens them one by one, reads the RUC, the date, the amount and the ITBMS, and types them into the system. This reads them and types them in for the team, in seconds.
WhatsApp photo, taken crooked, with the stamp over the amount.
RUC 155712345-2-2021 · Vendor Suministros del Istmo, S.A.
Invoice 00048812 · Date 08/14/2026
Subtotal $1,842.00 · ITBMS 7% $128.94 · Total $1,970.94
14 detail lines read. One flagged for review: the stamp covered a quantity.
Not for lack of judgment or tax complexity. Pure data entry.
Each one takes between 3 and 8 minutes to open, read, type and file. That's 30 hours a month of an expensive person, doing the most mechanical work in the company.
The typo isn't visible the day it happens. It shows up in the reconciliation, three weeks later, when you have to go back and find which of the four hundred it was.
And when the DGI or the auditor asks for the backup for an entry, the search begins: who received it, which email it was in, whether anyone saved it. Sometimes it turns up. Sometimes it doesn't.
And the cost nobody adds up: while you're typing, you're not analyzing. The accountant who should be reviewing margins is copying amounts from a PDF.
It gets the document however it arrives — PDF, photo, scan — understands what it is and pulls out the data that matters. Not just the total: the RUC, the date, the ITBMS and the detail lines. And it records it where it belongs.
Old-school OCR turns the image into text and leaves you there, with a block of letters you still have to read. This understands the document: it knows which number is the total and which is the ITBMS, even if each vendor puts them in a different place.
Old systems required setting up each invoice format, one by one. This one doesn't: it understands a format it has never seen, which is exactly what happens when a new vendor comes in.
If the stamp covers a number or the photo came out blurry, it flags that field and sends it to human review. Made-up data in accounting is worse than an empty field, and we treat it that way.
Estimates for sizing, not a promise. In the assessment we calculate them with your real volume.
All that's left is reviewing the documents the system flagged as doubtful. The rest goes in on its own.
The close stops waiting for someone to finish typing. You close when operations end, not two weeks later.
The $1,490 is recovered with the time freed up in the first quarter. And from there it's profit every month.
There's something that doesn't fit in the table: when the auditor asks for a backup, it shows up in seconds. You only appreciate that once you've been through the opposite.
The hard part here isn't reading the nice-looking document. It's the crooked photo, the stamp over the amount and the vendor who invoices in a format they invented.
You give us between 50 and 100 real documents, including the worst ones: blurry photos, crooked scans, handwritten invoices. Anyone gets the nice ones right; quality is measured with the others.
What has to check out for a document to go through on its own: that the ITBMS matches the subtotal, that the RUC exists, that the invoice isn't already recorded. Here you decide how much risk you accept.
It processes at the same time as your people, and we compare result against result. Only when the accuracy rate convinces you is it released. No accounting entry is bet blindly.
That's it. No options to choose and no fine print: setup is paid once and the monthly fee starts when it begins processing.
One service, two payments.
Full first year: $4,238 · from the second year on: $2,748 a year
Only need documents to be sorted and filed automatically, without extracting the data or recording it? That costs a lot less: Automatic Classification and Filing, $990 and $199 a month.
We're not going to give you a made-up percentage before seeing your documents. What we do guarantee is the mechanism: when it isn't sure, it flags the field and sends it to review instead of guessing. We measure the real rate during the parallel month, with your paperwork, and you decide based on that figure.
It reads them, with less accuracy than printed ones. In practice that means they land in the review inbox more often. If a good part of yours is handwritten, we look at it in the assessment and tell you right there whether it's worth it.
They're processed on the AI provider's servers, outside the country. They're not used to train anyone's models. If your company has a restriction that data can't leave the country, tell us at the start: there's a self-hosted alternative and it's priced differently.
Yes. Vendor invoices are the most common case, but it can be configured for purchase orders, receipts and proofs of payment. Each document type is defined during implementation.
Because the work is in training with your real documents and in the month running in parallel. An assistant that answers badly gets corrected; a wrongly recorded accounting entry gets carried all the way to the tax return. That care costs money, and we'd rather charge for it than do it badly.
A one-hour consultation. We look at your real volume, the quality of what you receive and how much time it takes today. If your volume is low, we tell you and we don't sell you anything.
$1,490 today and $229 a month once it goes live. Processing in 15 business days.
Order nowYou're talking to Génesis, Valtriom's AI