Source document recognition: how AI enters your paperwork into 1C

Source document recognition is when AI reads a scanned delivery note or VAT invoice, pulls out the supplier, the line items and the amounts on its own, and enters a finished document into 1C, the accounting and ERP platform most Russian companies run on, leaving the accountant only to check and post it. Manual entry drops several times over. But the technology has an honest ceiling that box-product vendors tend to keep quiet about, and there is a fresh reason to deal with it right now — the VAT reform is multiplying the flow of documents.

Below: how much time manual entry eats, why there are more documents in 2026, how AI really differs from old-style scanning, where it stalls and when an off-the-shelf service is not enough.

What is source document recognition?

It is a technology that turns a scan, photo or PDF of a document into a finished record in an accounting system. The input is a source document — a TORG-12 delivery note, a VAT invoice, a UPD, an acceptance act — and the output is a document in 1C with its fields filled in: supplier, tax ID, number, date, line items, amounts, VAT. The person is left to check and post it rather than type it in by hand.

The key word is "understands". Modern AI-based recognition does not merely identify characters in an image, it takes the document apart by meaning: where the header is, where the line item table is, which figure is a number and which is a date. And, most importantly for a business with hundreds of suppliers, it matches the document's line items against the item catalogue in 1C. That matching is the most valuable and the most fragile part, and we will come back to it below.

How much time does a business lose on manual entry?

Dozens of hours a month, and most of it does not go on the header. Estimates from the Russian market put manual entry of a single document at 10 to 30 minutes depending on its complexity. At a volume of 200–500 documents that is 15–40 hours of monotonous work every month for one specialist, peaking before reporting deadlines.

The least obvious part is where the time goes. Most of it is spent not on the document's details but on item matching: every line has to be linked to an entry in the 1C catalogue or added as a new one. Different suppliers call the same thing differently, and a person works through it by hand, line after line — that stage, not typing in the header, is what eats the bulk of the time.

Errors come on top of the lost time. International benchmarks from the Institute of Finance and Management put the average manual entry error rate at 3.6–4%, and according to Ardent Partners processing a single invoice by hand costs around 13 dollars against three for teams with automation. An error in a source document means not only redoing the work but also a risk on VAT and profit tax. There is a separate walk-through of where to start removing this kind of routine at all — automating routine work: where to start.

Why are there more source documents in 2026?

Because of the VAT reform for the simplified tax regime — hundreds of thousands of businesses are handling VAT invoices for the first time. From January 1, 2026 the income threshold at which companies on the simplified regime are exempt from VAT fell from 60 to 20 million rubles. If your 2025 income passed 20 million, you have been a VAT payer since January. Under the law signed on July 4, 2026 the 20 million threshold is frozen for 2026–2029, with a drop to 15 million planned for 2030 and to 10 million for 2031 (figures as of the publication date; the schedule may still change).

What this changes in practice. The base VAT rate has risen from 20 to 22%, and businesses on the simplified regime were given special rates with no right to deduct — 5% on income up to 272.5 million and 7% on income up to 490.5 million rubles. But the tax is only half the story. Once a business becomes a VAT payer, it has to issue VAT invoices with the tax shown separately, keep a purchase ledger and a sales ledger, and file returns. For those who never kept these documents before, the flow of accounting paperwork has grown several times over.

The scale of the reform was described by Finance Minister Anton Siluanov: according to the declaration campaign, keeping the threshold at 20 million exempts more than 360 thousand small and medium-sized businesses from VAT. Those who crossed the threshold got a new obligation and a new wave of documents — at exactly the moment when manual entry turns from an inconvenience into a bottleneck before every reporting period.

How does AI recognition differ from old-style OCR?

Old OCR identifies characters, AI understands the document. Classic optical recognition works in two steps: it finds the text in an image and identifies the letters and digits. It tells you which characters are on the page but not what they mean, and it usually needs its own template for each supplier's form — "the delivery note number is always in the top right corner". Industry sources put the accuracy of classic OCR at 85–92%, and it copes badly with handwriting, non-standard layouts and creased scans.

Modern AI reads a page differently. It works through tables, multi-column layouts and headers and extracts structured fields straight away. The key difference for a business is stated in the Vellum platform's analysis: "Give a language model 100 differently formatted invoices and it will extract the key fields from each one without additional setup or new templates." For a company with hundreds of counterparties, each with its own form, that is the difference between "configure a template for every one of them" and "just feed in the documents".

The price of "understanding" is speed: a neural network takes seconds to parse a document against milliseconds for old OCR. For a flow of source documents that is immaterial.

Want to take manual source document entry off your hands, on your own document flow and your own item catalogue? Custom development at IncubeAi means an integration built around your process: recognition, matching against 1C catalogues and reconciliation in one loop, under contract, with data kept in Russia and support after handover. The team has 300+ automation and AI projects behind it.

How do documents get into 1C without manual entry?

Through the chain "recognised — matched — document created for checking". The person does not disappear from the process but stops being a typist: they control the result instead of moving data by hand. Here is what the path of a single document looks like.

1. The document arrives through any channel: a scan, a phone photo, a PDF, an electronic document from EDI, the legally binding electronic document exchange used between companies in Russia.
2. The system determines the document type and extracts the fields — supplier, number, date, line items, amounts, VAT.
3. It matches the counterparty, the contract and every line item against the 1C catalogues — the "smartest" step.
4. It creates a finished document (a goods receipt, a sale, a VAT invoice) with the status "to be checked".
5. The accountant checks the questionable lines and posts the document.

The market benchmark at which automation starts to pay off is roughly 50 documents a month per accountant. Below that volume manual entry is cheaper than automation; above it, the routine starts eating a noticeable share of the working day, and every further document costs more than it should.

Is it true that AI recognises "up to 98%"?

Per character, yes; per whole document, no — and an honest conversation matters more here than advertising. Recognition vendors claim accuracy of "up to 98%" or "from 95%", but those figures are almost always counted per individual character. One wrong character in a long item name and the field already has to be corrected, even though "per character" accuracy stayed high.

The honest benchmark is set by an analysis on Habr from a practitioner who built such systems: 83–85% of fields recognised correctly is already a good result, and 100% without a human is unattainable in principle. The reason is mundane: "Scanning glitches, sending files through certain messengers and poorly printed text all reduce quality" (the author's own wording). Re-shooting a delivery note on a phone in bad light hits accuracy harder than any algorithm.

The practical conclusion for an owner: do not expect a magic button that replaces the accountant 100%. The working model is AI taking over the retyping and the matching while a person checks whatever the system is unsure about. That is still several times faster than manual entry, but honest about expectations. It is exactly at this honest threshold that projects break down where "full automation with no people" was promised.

Where does AI recognition stall?

On handwriting, stamps, non-standard forms and — most of all — on item matching. Recognising fields is in itself a solved problem; the difficulty lies in what comes after it. Here are the typical places where box-product recognition stumbles.

• Item matching. The system will not automatically link "C2H5OH 0.2 L" to "Ethyl alcohol 0.2" or "ARKHYZ water 25L" to "Artesian water 25 L" — your catalogue needs rules and training.
• Handwritten text, stamps and seals are recognised poorly.
• Bad scans and photos re-sent through messengers reduce accuracy regardless of the model.
• Reconciliation against orders and contracts is separate logic on top of recognition; OCR on its own does not do it.
• Non-standard and specialised documents require additional training, and that is separate money and time.

All these points share one common denominator: recognising a field is easy, but fitting it into your accounting logic is a job for a specific business. Errors at this stage are especially dangerous because they slip into the books unnoticed. A similar story with manual data was covered in the article on when it is time for a business to leave Excel.

An off-the-shelf box or an integration built around your process?

A box recognises fields; an integration covers your accounting logic. The market has ready services — 1C:Source Document Recognition, Entera, Smart Engines and others; they charge per page or by subscription and suit standard document flows well. There is one problem: not everyone has a standard flow.

Our case with a wholesale and manufacturing company was built exactly on that line: a two-way integration with 1C where items, stock balances, shipments and payments sync by themselves, and order processing sped up from hours to minutes. Source document recognition in such a loop is not a separate service but part of the general flow: the document is recognised, reconciled against the order and posted in a single system. How different systems are tied into a loop like this was covered in the article on integrating 1C, CRM, the bank and Telegram.

Where do you start automating source document entry?

With an honest count of the volume and an assessment of exactly where the hours are lost. Automation does not pay off everywhere, so the first step is numbers, not buying a service.

1. Count the volume: how many source documents a month pass through accounting and how much time entering them takes.
2. Check the payback threshold: the market benchmark is roughly 50 documents a month per accountant; below that, automation may not pay for itself.
3. Assess the complexity: standard forms and a standard item catalogue mean an off-the-shelf service will do; your own catalogue, reconciliations and non-standard documents mean an integration is needed.
4. Build in a checking stage: not "100% without people", but recognition plus a person checking the questionable lines.
5. Connect recognition to the rest of your accounting, so the document does not live apart from orders, stock and payments.

The VAT reform has made this calculation relevant for those who never thought about it before: the volume has grown and the number of hands has not. If you want to run the numbers and work out what is more profitable in your case — an off-the-shelf service or an integration — tell us about your task. We will go through your document flow, count the savings in hours and money and propose a solution for your process: under contract, with data kept in Russia and support after handover.

Sources

• Federal Tax Service of Russia, the country's tax authority — tax changes from 2026 (the VAT threshold under the simplified regime, the 22% rate)
• RIA Novosti — Putin signs the law freezing the VAT threshold under the simplified regime at 20 million rubles (04.07.2026)
• Kommersant — Siluanov on keeping the 20 million threshold: more than 360 thousand SMEs exempt from VAT
• BUH.1S — how to automate source document entry into 1C
• Habr — document recognition systems: what to know before you start (the honest 83–85% field accuracy)
• Vellum — document data extraction: language models versus OCR
• DocuClipper — the cost of processing an invoice and the manual entry error rate (Ardent Partners and IOFM benchmarks)