AI Invoice Processing Automation

We built an unattended, end-to-end invoice processing robot on a self-hosted n8n platform. It takes invoice entry off the team entirely and writes straight into the accounting system they already use.

From scanned batch to posted entry

1. Intake. Batches are collected from the client's SharePoint library through the Microsoft Graph API, per company and per period.
2. AI splitting and classification. A custom-trained Google Document AI splitter cuts the multi-page scan into individual documents and separates real invoices from receipts, returns and everything else.
3. AI extraction, trained on Bulgarian documents. Supplier, UIC and VAT number, invoice number and date, net, VAT and total, and every line item with quantity and unit price, handwritten documents included. F1 score 0.95.
4. Nothing guessed. A document that cannot be read with confidence is left for a person, not posted on a best guess.
5. Business rules. Advance, final and advance-deduction invoices are skipped; prior-year invoices are held back; supplier plus invoice number must be unique, so a duplicate never posts.
6. Validation and enrichment. Unknown suppliers are looked up in the Bulgarian Commercial Register and created with their default accounts. Invoices already marked as paid are detected and skipped.
7. Account mapping the accountants own. An Excel workbook of description keywords, transaction types, IBANs and debit and credit accounts decides how each line posts. The accounting rules stay in the accountants' hands, not in code.
8. Direct posting. Our own Microinvest connector writes validated documents straight into each company's database over MSSQL, through a VPN tunnel the robot opens before every run.
9. Report and exceptions. Every row is logged as it is processed and compiled into an Excel report, filed to SharePoint and emailed. Anything not posted is marked as a system or business exception with a plain-language reason.

The question that is only ever asked once

When a supplier name does not match, written in Latin script on the invoice and in Cyrillic in the accounting system, the accountant gets a one-click confirmation form. The answer is stored, so the same question is never asked twice. The exception list shrinks every month on its own.

Built to run unattended

• Scheduled daily or triggered on demand through an authenticated link.
• Safe to repeat after a failure: no duplicate entries in the report or in the books.
• Monitored through a hypercare period after go-live.

Technology

The accountant opens the accounting system and the data is already there: split, read, mapped to the right client company and posted.

• 97% of documents processed without correction over six months in production.
• 2,731 line items across 60 invoices posted in a single production run, every row successful.
• About 576 hours, 72 working days, of data entry removed a year. On the client's own ROI model the robot pays for itself inside the first year and returns more than 300% by year three.
• The team moved from typing invoices to reviewing a short exception report.
• Reusable components. The Microinvest connector and the Commercial Register lookup built here now power the firm's bank reconciliation robot too.

The risk most teams miss

New suppliers and new invoice layouts arrive every month. Without an exception loop, extraction accuracy drifts quietly and errors creep back into the books. Here every answered exception is stored, so the system gets more complete the longer it runs.

Part of a three-robot programme for the same firm: bank statements download themselves, invoices post themselves, and payments reconcile themselves.