DocuFlow AI

We started following Croatia's Fiscalisation 2.0 while nobody yet knew quite how it would work or why it was being introduced. Digging deeper, we noticed what the announcements left out: mandatory e-invoicing does not mean the end of paper. A small business still receives PDFs, photographed receipts, scans and cash slips — only now alongside e-invoices. Two worlds, and no tool for both.

• E-invoices arrive automatically through a certified access point (ePoslovanje, doku.hr). The data is already structured, so no AI processing is spent on it.
• Everything else — PDF, scan, photo — is read by AI (Claude), which extracts amounts, due dates, IBAN and tax ID into structured, reviewable data.
• An action engine that suggests the next step — a payment reminder, an accounting task — and only sends it once a person approves it, with a real email behind every approved reminder
• Per-call AI cost tracking built in from day one, so the product knows its own margins before its first customer does
• A security pass most teams skip entirely: an audit of every database function that runs with elevated privileges, closing grants nobody remembered were open

• A real e-invoice was sent through the access point's test environment, fetched and shown in the app — without a single AI call. "E-invoices at no AI processing cost" is measured, not assumed.
• AI extraction was measured on real photographed and scanned Croatian receipts and payment slips. Uncertain IBAN and tax-ID digits are settled by checksums (mod-97, mod 11-10), not guesswork.
• In production, the whole product in Croatian

The AI call that reads a document is the easy half. The part that earns trust is everything around it — the review step nothing skips, the cost you can account for, the access you can prove is locked down. And a clear boundary: DocuFlow deliberately does not fiscalise receipt of invoices, because that is the access point's legal role. It points the user through that step rather than pretending to take it over.