Invoice recognition in 1C from photo and PDF: cloud or local AI

A supplier invoice can get into 1C in four ways. Each has its own conditions and there is no universal winner. Below are a comparison, a testing checklist and the errors recognition makes most often.

Four ways to enter a document

The main rule: if a document arrives electronically through EDI, there is nothing to recognize. Recognition is for paper, scans and photos, where there is no data structure.

When you need a local neural network

A local setup is chosen in three cases:

1. Data must not leave the company. Invoices contain purchase prices, tax IDs and supplier lists. For some companies this is a trade secret or a security requirement.
2. There is no stable internet or it is restricted by company policy.
3. You need control over filling rules: how items are matched and which accounting policy applies.

If none of these apply, a cloud service is simpler: no computer or model to set up.

What local mode requires

• A program to run the model, for example LM Studio: free, runs on the user's computer.
• A model that can read images (a vision model).
• A computer with enough power. A model can run on a CPU too, but on weak hardware a document takes noticeably longer. Rely on a test on your own computer, not on other people's numbers.
• A 1C extension that sends the image to the model and fills the document.

How to test any solution on your documents

Do not trust demos, test on your data:

1. Take 20–30 real documents of different quality: clean PDFs, scans, phone photos, crumpled ones, multi-page ones.
2. For each, compare: supplier and tax ID, numbers and dates, quantities and prices per line, VAT and the total.
3. Look separately at item matching: how many items matched by themselves and how many needed manual fixes.
4. Measure time: from pressing the button to a filled document.
5. Check what happens to what was not recognized: it should be clearly marked, not a silent error.

The readiness criterion is simple: the document total in 1C equals the paper total, and human review takes less time than manual entry.

Where recognition fails most often

• Similar characters: 0 and O, 1 and I, 5 and S in article numbers and tax IDs.
• Merged or shifted columns in narrow tables.
• Stamps and signatures covering digits.
• Multi-page documents where the header and the total are on different pages.
• Angled photos and poor lighting.
• Units of measure (pcs, pack, kg) that do not match the catalog.

That is why such documents cannot be posted without review. The right scheme: the system fills the document and a person checks and posts it.

How my extension does it

The extension “Local document recognition from photos for 1C” works on a local model through LM Studio: a photo, scan or PDF goes to a model on the user's computer, the document is detected as UPD or TORG-12, the supplier and items are matched, and the document is filled in 30–60 seconds. No internet is needed and data never leaves the computer. It is tested on 1C:Trade Management 11.5 and 1C:Accounting 3.0, and a custom integration is done for other configurations. Details and prices are on the project page and on Infostart. For other 1C work see 1C and accounting systems.