ℹDemonstration scenario. An example of a possible implementation, not a completed client project of Neural2B.

Recognition and Approval of Incoming Primary Documents

Demonstration scenario: we scan invoices, acts, and waybills, extract details, and check rules. If the model is uncertain, we hand it over to the operator.

Business situation overview

The company receives dozens of invoices, acts of completed work, and waybills from various suppliers. In PDF, scans, or photos. The accountant spends up to 4 hours daily entering amounts manually, verifying the company ID, IBAN account numbers, and searching for which contract to link each document to.

Data flow and architecture diagram

1Sources of documents: accounting mail, uploads via internal web cabinet, or DocuSign webhook

2Multimodal model that recognizes tables and seals

3Structure validator: checks the company ID against the control number and the correctness of IBAN

4Amount verification module: does the amount without VAT + VAT = total to be paid

5Integration: structured JSON goes into the accounting system, where a draft appears

The human role in the loop (Human-in-the-loop)

If the system is less than 95% confident in any field or if the document amount exceeds the limit, the document is highlighted. The accountant must confirm it in a convenient web interface.

Synthetic test data

Example input and system response

User input request:

Scan of invoice No. SF-481 from LLC 'TechnoPostach' for the amount of $1,185.00 (including VAT $198.00) for the supply of cable products.

Generated response / system action:

'Recognized: Supplier: LLC TechnoPostach; company ID: 38492015; Invoice: XX00 0000 0000 0000 0000 00; Amount: $1,185.00; VAT: $198.00; Purpose: payment for cable products for invoice 481. Details matched with the supplier directory. Draft created in the accounting system. Awaiting accountant confirmation.'

Limits of the demo scenario

✕Without the accountant's electronic signature, the system will not withdraw money from the account
✕Damaged or unreadable scans receive a mark 'Requires resubmission of the original'
✕If contractual prices have changed, the system verifies them against the recorded terms of the specifications

What to measure during pilot testing

✓Accuracy of extracting mandatory fields (company ID, amount, number, date) — target: > 98%
✓Time taken for one document (target: < 15 seconds)
✓Share of documents that passed without any manual corrections

Relevant company service

Document Processing Automation

Service overview →

Interested in a similar architecture?

We’ll adapt the logic of this scenario to your internal software, database structure, and operating rules.

Discuss adaptation

Other scenarios

Scenario adaptation: Document Recognition and Approval

Describe the process that’s eating up your team’s time. We’ll suggest what can be automated, what data is needed, and where to start.