Scans in, verified data out
The agent reads invoices, acts, waybills and design documentation. It extracts fields, shows where each one came from, checks them against your rules and exports to your system. A person reviews anything uncertain.
From email, a shared folder, e-document exchange or a web upload. PDF, DOCX, photos and scans.
Scans go through OCR. Document structure — tables, requisites, signature marks — is preserved.
AI finds the fields you need and records the location on the page and a confidence score for each. A field without a source is rejected.
Tax IDs against your directory, totals against line items, dates against the contract. Rules are written in code, not left to the model’s judgment.
Low-confidence fields and broken rules go to an employee, who confirms, corrects or enters the value manually.
To 1C, CRM, e-document exchange or a file in the required schema. XML and other strict formats are built by code, not by the model, and validated before sending.
One task, three different projects. Pick the one that looks like you and press “I want this” — your request arrives tagged, and we send an estimate in 1–2 days.
One document stream, say incoming invoices, that eats hours of an accountant’s week.
What we build
Invoices, acts, waybills and contracts flow in from email and e-document exchange, and several people check them.
What we build
Documents hold personal data and trade secrets, exports go into strict industry formats, and security has its own requirements.
What we build
Pilot from RUB 300,000. Beyond that we price by task and hours.
Before we start, we take a sample of your real documents, including the worst ones, and see what gets recognized. A field read with low confidence never moves on without a person. If a scan is unreadable, we’ll say so plainly: it needs a rescan.
It can, which is why we never rely on the model alone. Every field is tied to a spot in the document, totals and requisites are checked by rules, and a person reviews anything in doubt. In the XML project for construction documentation, the model only suggests fields; the XML itself is assembled by code against the XSD schema.
OCR gives you text. You need fields: who the supplier is, the total including VAT, which contract the invoice belongs to. AI extracts them from the recognized text, then rules check the result.
Depends on the setup. In the cloud version, on the project server. If documents contain personal data or trade secrets, we deploy everything on your servers and connect the model so data never leaves the perimeter. The architecture is designed for 152-FZ.
A working version for one document type is usually ready in about a month. Complex document sets take longer: stage one of the construction XML project was scoped at 45 days. Cost is estimated by task and hours: we send a staged estimate within 1–2 days after the brief.
Name the process that eats time. You get a pilot range, not a 40-slide deck.